test: cover terminal limactl resolver errors

.claude/skills/ask-internal/SKILL.md

@@ -1,152 +0,0 @@
----
-name: ask-internal
-description: Answer questions about BrowserOS internal stuff (setup, features, architecture, design decisions) by reading the private internal-docs submodule and the codebase. Use for "how do I X", "where is Y", "what is the deal with Z", or any question that mixes ops/setup knowledge with code knowledge. Can execute steps with per-command confirmation.
-allowed-tools: Bash, Read, Grep, Glob, Edit, Write
----
-
-# Ask Internal
-
-Answer team-internal questions by reading `.internal-docs/` and the codebase, synthesizing a direct answer with file:line citations, and optionally running surfaced commands with confirmation.
-
-**Announce at start:** "I'm using the ask-internal skill to answer this from internal-docs and the codebase."
-
-## When to use
-
-- "How do I reset my dogfood profile?"
-- "What's the deal with the OpenClaw VM startup?"
-- "Where do we configure release signing?"
-- Any question whose answer lives in setup runbooks, feature notes, architecture docs, or the code that produced them.
-
-## Hard rules — never do these
-
-- NEVER execute a state-mutating command without per-command `y` confirmation from the user.
-- NEVER edit BrowserOS code in response to an ask-internal question. The skill answers; it does not modify code. Use `/document-internal` for writes.
-- NEVER guess. If grep finds nothing useful in docs or code, say so plainly.
-- NEVER run this skill if `.internal-docs/` is missing. Stop with the init command.
-- NEVER cite a file or line number you have not actually read.
-
-## Voice rules
-
-Apply the same voice rules as `document-internal` to the synthesized answer:
-
-- Lead with the point.
-- Concrete nouns. Name files, functions, commands.
-- Short sentences. Active voice. No em dashes.
-- Banned words: delve, crucial, robust, comprehensive, nuanced, multifaceted, furthermore, moreover, additionally, pivotal, landscape, tapestry, underscore, foster, showcase, intricate, vibrant, fundamental, significant, leverage, utilize.
-- No filler intros.
-
-## Workflow
-
-### Step 0: Pre-flight
-
-```bash
-if git submodule status .internal-docs 2>/dev/null | grep -q '^-'; then
-  echo "internal-docs submodule not initialized. Run: git submodule update --init .internal-docs"
-  exit 0
-fi
-[ -d .internal-docs ] && [ -n "$(ls -A .internal-docs 2>/dev/null)" ] || {
-  echo ".internal-docs/ missing or empty. Submodule not configured?"
-  exit 0
-}
-```
-
-### Step 1: Parse the question
-
-Pull the keywords from the user's question. Drop stop words. Identify intent:
-
-- **Setup-question** ("how do I", "how to", "where do I configure"): bias the search toward `setup/`.
-- **Feature-question** ("what is X", "why does X work this way"): bias toward `features/` and `architecture/`.
-- **Free-form** ("anything about Y"): search all categories.
-
-### Step 2: Multi-source search
-
-Run grep in parallel across two sources.
-
-**Internal docs:**
-
-```bash
-grep -rni --include='*.md' '<keyword>' .internal-docs/
-```
-
-Search each keyword separately. Collect top hits by relevance (more keyword matches = higher).
-
-**Codebase (skip vendored Chromium and `node_modules`):**
-
-```bash
-grep -rni --include='*.ts' --include='*.tsx' --include='*.js' --include='*.json' --include='*.sh' \
-     --exclude-dir=node_modules --exclude-dir=chromium --exclude-dir=.grove \
-     '<keyword>' packages/ scripts/ .config/ .github/
-```
-
-Read the top 3-5 doc hits and top 3-5 code hits. Do not skim — read the relevant section fully so citations are accurate.
-
-### Step 3: Synthesize answer
-
-Structure the response:
-
-1. **Direct answer.** First sentence answers the question. No preamble.
-2. **Steps if applicable.** Numbered list with exact commands.
-3. **Citations.** Every factual claim references `path/to/file.md:42` or `path/to/code.ts:117`. Run the voice self-check before printing.
-
-If multiple docs cover the topic at different layers (e.g., a setup runbook and a feature note both mention dogfood profiles), reconcile them in the answer rather than dumping both.
-
-### Step 4: Offer execution (only if commands surfaced)
-
-If Step 3 produced executable commands the user could run, ask:
-
-> Run these for you? (y / n / dry-run)
-
-- **y:** Execute one at a time. For any command that mutates state (writes a file, modifies config, kills a process, deletes anything), ask "run this? <command>" before each. Read-only commands (`ls`, `cat`, `git status`) run without per-command confirmation but still print before running.
-- **n:** Skip. Done.
-- **dry-run:** Print the full sequence as a `bash` block. Do not execute.
-
-### Step 5: Doc-not-found path
-
-If Step 2 returned nothing useful (no doc hits AND no clear code answer):
-
-1. Tell the user: "No doc covers this. Tangentially relevant files: <list>."
-2. Ask: "Draft a new doc and open a PR to internal-docs?"
-3. On yes: invoke the full `/document-internal` flow (four sharp questions, draft, voice check, PR), forced to `setup/` doc type, with the code-grep findings handed in as initial context.
-
-### Step 6: Completion status
-
-Report one of:
-
-- **DONE** — answer delivered, citations verified.
-- **DONE_WITH_CONCERNS** — answered, but flag uncertainty (e.g., docs and code disagreed; user should reconcile).
-- **BLOCKED** — submodule missing or other pre-flight failure.
-- **NEEDS_CONTEXT** — question too vague to search effectively. Ask one clarifying question.
-
-## Citation discipline
-
-Every "X is at Y" claim in the answer must point to a file:line that the skill actually read. Do not approximate. If you didn't read it, don't cite it.
-
-If a doc says one thing and the code says another, surface the conflict explicitly:
-
-> The setup runbook (`setup/dogfood-profile.md:23`) says to delete `~/.cache/browseros/dogfood`, but the actual code path in `packages/cli/src/cleanup.ts:47` removes `~/.local/share/browseros/dogfood`. The doc looks stale. Recommend updating it.
-
-## Common Mistakes
-
-**Skimming and then citing**
-- **Problem:** Citation points to a line that doesn't actually contain the claim.
-- **Fix:** Read the section fully before citing. If you didn't read line 117, don't cite line 117.
-
-**Executing without per-command confirmation for mutations**
-- **Problem:** User says "y" to "run all", skill blasts through `rm -rf`-style commands.
-- **Fix:** "y" means "run this sequence with per-mutation confirmations". Per-command y is required for writes.
-
-**Searching only docs, not code**
-- **Problem:** Doc says X but code does Y; answer is wrong.
-- **Fix:** Always grep both sources in Step 2.
-
-## Red Flags
-
-**Never:**
-- Cite a file:line you haven't read.
-- Run mutations without per-command confirmation.
-- Modify BrowserOS code from this skill (use `/document-internal` for writes).
-
-**Always:**
-- Pre-flight check before any search.
-- Reconcile doc vs code conflicts in the answer, don't hide them.
-- Plain "no doc covers this" when grep is empty — never invent.

.claude/skills/document-internal/SKILL.md

@@ -1,208 +0,0 @@
----
-name: document-internal
-description: Draft a 1-page internal doc (feature, architecture, or design) for the private browseros-ai/internal-docs repo. Use when wrapping up a feature on a branch, after the PR is open or about to be opened. Skill drafts from the diff, asks four sharp questions, enforces voice rules, and opens a PR to internal-docs.
-allowed-tools: Bash, Read, Write, Edit, Grep, Glob
----
-
-# Document Internal
-
-Draft a 1-page internal doc (feature note, architecture note, or design spec) from the current branch's diff and open a PR to `browseros-ai/internal-docs`.
-
-**Announce at start:** "I'm using the document-internal skill to draft a doc for internal-docs."
-
-## When to use
-
-After finishing implementation on a feature branch, when the work is doc-worthy (a major feature, a new subsystem, a setup runbook for something internal, or a design decision that future engineers need to know).
-
-## Hard rules — never do these
-
-- NEVER `git add -A` or `git add .` inside the tmp clone of internal-docs. Always specific paths.
-- NEVER write outside the tmp clone (no spillover into the OSS repo's working tree).
-- NEVER fabricate filler content for empty template sections. Empty stays empty.
-- NEVER touch the OSS repo's `.gitmodules` or submodule pointer — the sync workflow handles that.
-- NEVER run this skill if `.internal-docs/` is missing. Stop with the init command.
-- NEVER push to `internal-docs/main` directly. Always a feature branch + PR.
-
-## Voice rules — enforced by Step 4
-
-The skill MUST follow these and refuse to draft otherwise. After generation, scan for violations and regenerate offending sentences (max 3 attempts).
-
-- Lead with the point. First sentence answers "what is this?"
-- Concrete nouns. Name files, functions, commands. Not "the system" or "the component".
-- Short sentences. Average <20 words. No deeply nested clauses.
-- Active voice. "X does Y" not "Y is done by X".
-- No em dashes. Use commas, periods, or rephrase.
-- Banned words: delve, crucial, robust, comprehensive, nuanced, multifaceted, furthermore, moreover, additionally, pivotal, landscape, tapestry, underscore, foster, showcase, intricate, vibrant, fundamental, significant, leverage, utilize.
-- "110 IQ" target. Write for a smart engineer who has not seen this code yet.
-- No filler intros ("This document describes..."). Start with the substance.
-- Empty sections stay empty. Do not write "N/A" or fabricate content.
-
-## Workflow
-
-### Step 0: Pre-flight
-
-Bail with a clear message on any failure.
-
-```bash
-# Submodule must be initialized
-if git submodule status .internal-docs 2>/dev/null | grep -q '^-'; then
-  echo "internal-docs submodule not initialized. Run: git submodule update --init .internal-docs"
-  exit 0
-fi
-[ -d .internal-docs ] || { echo ".internal-docs/ missing. Submodule not configured?"; exit 0; }
-
-# Must be on a feature branch
-BRANCH=$(git branch --show-current)
-if [ "$BRANCH" = "main" ] || [ "$BRANCH" = "dev" ]; then
-  echo "On $BRANCH. Run from a feature branch."
-  exit 0
-fi
-
-# Determine base branch (default: dev for this repo, fall back to main).
-# Suppress rev-parse's SHA output on stdout so it doesn't get captured into BASE.
-BASE=$(git rev-parse --verify origin/dev >/dev/null 2>&1 && echo dev || echo main)
-
-# Gather context
-git log "$BASE..HEAD" --oneline
-git diff "$BASE...HEAD" --stat
-gh pr view --json body -q .body 2>/dev/null  # may be empty if no PR yet
-```
-
-### Step 1: Identify the doc
-
-Ask the user for three things in one prompt:
-
-1. **Doc type:** `feature` (default for `feat/*` branches), `architecture`, or `design`
-2. **Slug:** kebab-case, short (e.g., `cowork-mcp`, `auto-skill-suggest`)
-3. **Owner:** GitHub handle (default = `git config user.name` or current `gh api user --jq .login`)
-
-### Step 2: Decision brief — four sharp questions
-
-Ask one question at a time. Each answer constrains the next. These force compression before drafting.
-
-1. "In one sentence: what can someone now DO that they could not before?"
-2. "What is the one design decision a future engineer needs to know?"
-3. "Which 3-5 files are the heart of this change?" (suggest candidates from the diff)
-4. "Any sharp edges or gotchas? (or 'none')"
-
-Skip any question that is N/A for the doc type. Architecture notes don't need question 1; design specs don't need question 4.
-
-### Step 3: Draft from the template
-
-Read the matching template from `.internal-docs/_templates/`:
-
-- `feature` → `feature-note.md`
-- `architecture` → `architecture-note.md`
-- `design` → `design-spec.md`
-
-If `.internal-docs/_templates/` does not exist (first run, before seeding), fall back to the seeds bundled with this skill at `.claude/skills/document-internal/seeds/_templates/`.
-
-Generate the 1-pager from the template, the four answers, and the diff context.
-
-### Step 4: Voice self-check
-
-Scan the draft for violations:
-
-- Em dash present (`—`).
-- Any banned word from the list.
-- Average sentence length > 20 words.
-- Body line count > 60 (feature notes only — architecture/design have no cap).
-
-If any violation found, regenerate the offending sentences in place. Max 3 attempts. If still failing after 3 attempts, stop and report which rules are violated.
-
-If the body is over 60 lines for a feature note, ask: "This is N lines, target is 60. Trim, or promote to `architecture/` (no length cap)?"
-
-### Step 5: Show + iterate
-
-Print the full draft. Ask:
-
-> Edit needed? Paste any changes, or say "looks good".
-
-Apply user edits with the Edit tool. Re-run Step 4. Loop until the user approves.
-
-### Step 6: Open PR to internal-docs
-
-Use a tmp clone. Never the user's `.internal-docs` checkout — keeps the user's submodule clean.
-
-```bash
-TMP=$(mktemp -d)
-trap 'rm -rf "$TMP"' EXIT  # cleans up even if any step below fails
-git clone -b main git@github.com:browseros-ai/internal-docs.git "$TMP"
-cd "$TMP"
-git checkout -b "docs/<slug>"
-
-# Write the doc
-mkdir -p "<type>"  # features, architecture, designs, or setup
-cat > "<type>/$(date -u +%Y-%m)-<slug>.md" <<'DOC'
-<draft content>
-DOC
-
-# Update the root README index — insert one line under the matching section
-# Use Edit tool to add: "- [<title>](<type>/YYYY-MM-<slug>.md) — <one-line description>"
-
-git add "<type>/$(date -u +%Y-%m)-<slug>.md" README.md
-git commit -m "docs(<type>): <slug>"
-git push -u origin "docs/<slug>"
-
-PR_URL=$(gh pr create -R browseros-ai/internal-docs --base main \
-  --head "docs/<slug>" \
-  --title "docs(<type>): <slug>" \
-  --body "$(cat <<'BODY'
-## Summary
-<one-line of what this doc covers>
-
-## Source
-- BrowserOS branch: <branch>
-- Related PR: <#NNN if any>
-BODY
-)")
-
-cd -
-echo "PR opened: $PR_URL"
-# trap above cleans up $TMP on EXIT
-```
-
-If the slug contains characters that won't shell-escape cleanly, sanitize before substitution.
-
-### Step 7: Completion status
-
-Report one of:
-
-- **DONE** — file written, branch pushed, PR opened. Print PR URL.
-- **DONE_WITH_CONCERNS** — same as DONE but list concerns (e.g., voice check needed multiple regens, user skipped a question).
-- **BLOCKED** — submodule missing, auth fail, or template missing. State exactly what's needed.
-
-## Doc type defaults
-
-| Branch pattern | Default doc type | Default location |
-|----------------|------------------|------------------|
-| `feat/*`       | feature          | `features/`      |
-| `arch/*` or refactor branches with >10 files in `packages/` | architecture | `architecture/` |
-| `rfc/*` or `design/*` | design          | `designs/`       |
-| Otherwise      | ask              | ask              |
-
-## Common Mistakes
-
-**Drafting before asking the four questions**
-- **Problem:** Output is generic filler that says nothing concrete.
-- **Fix:** Always ask Step 2 first, even if the diff "looks obvious".
-
-**Touching `.internal-docs/` directly**
-- **Problem:** User's submodule HEAD moves, parent repo shows dirty state.
-- **Fix:** Always use the tmp clone in Step 6.
-
-**Skipping voice check on user edits**
-- **Problem:** User pastes prose with em dashes or filler; ships as-is.
-- **Fix:** Re-run Step 4 after every user edit.
-
-## Red Flags
-
-**Never:**
-- Push to `internal-docs/main`. Always branch + PR.
-- Modify the OSS repo's `.gitmodules` or submodule pointer.
-- Fabricate content for empty template sections.
-
-**Always:**
-- Pre-flight check before doing any work.
-- One-pager rule for feature notes (60-line body cap).
-- File:line citations when referencing code.

.claude/skills/document-internal/seeds/README.md

@@ -1,51 +0,0 @@
-# BrowserOS Internal Docs
-
-Private team docs for `browseros-ai`. Mounted as a submodule into the public OSS repo at `.internal-docs/`.
-
-If you are reading this from a public clone of BrowserOS without team access — this submodule is for the BrowserOS internal team. Nothing here is required to build or use BrowserOS.
-
-## How to find what you need
-
-- Setup task ("how do I X locally") → look in [`setup/`](setup/)
-- Recently shipped feature → look in [`features/`](features/)
-- Cross-cutting subsystem → look in [`architecture/`](architecture/)
-- A design decision or RFC → look in [`designs/`](designs/)
-
-Or run `/ask-internal "<your question>"` from any BrowserOS checkout. The skill greps these docs and the codebase, then synthesizes an answer with citations.
-
-## How to add a doc
-
-Run `/document-internal` from a feature branch. The skill drafts a 1-pager from your branch's diff, asks four sharp questions, enforces voice rules, and opens a PR back to this repo.
-
-## Index
-
-### Setup
-<!-- one line per setup runbook: -->
-<!-- - [Dev environment](setup/dev-environment.md): first-time machine setup -->
-
-### Features
-<!-- one line per shipped feature, newest first: -->
-<!-- - [Cowork MCP](features/2026-04-cowork-mcp.md): bring outside MCPs into the BrowserOS agent -->
-
-### Architecture
-<!-- one line per cross-cutting subsystem: -->
-<!-- - [Chrome fork overview](architecture/chrome-fork-overview.md): what we patched and why -->
-
-### Designs
-<!-- one line per design spec, newest first: -->
-<!-- - [Internal docs submodule](designs/2026-04-30-internal-docs-submodule.md): this system -->
-
-## Templates
-
-When `/document-internal` runs, it reads from [`_templates/`](_templates/). Edit the templates here when the team's preferred shape changes.
-
-## Voice
-
-Docs in this repo follow these rules. The `/document-internal` skill enforces them; humans editing by hand should match.
-
-- Lead with the point.
-- Concrete nouns. Name files, functions, commands.
-- Short sentences, active voice, no em dashes.
-- No filler words: delve, crucial, robust, comprehensive, nuanced, multifaceted, leverage, utilize, etc.
-- Empty sections stay empty. Do not write "N/A" or fake content.
-- Feature notes target one screen, body 60 lines max.

.claude/skills/document-internal/seeds/_templates/architecture-note.md

@@ -1,31 +0,0 @@
----
-title: <subsystem name>
-owner: <github handle>
-status: current | deprecated
-date: YYYY-MM-DD
-related-features: [feature-slug-1, feature-slug-2]
----
-
-# <subsystem name>
-
-## What this subsystem does
-<1-2 paragraphs. The top-level responsibility. Boundaries.>
-
-## Architecture
-<Diagram (ASCII or mermaid) plus prose. Components and how they talk.>
-
-## Constraints
-<Hard rules the design enforces. "X must never call Y" type statements.>
-
-## Decisions made
-<Numbered list of non-obvious decisions and the reason for each.>
-
-## Key files
-- `path/to/file.ts` — role
-- `path/to/dir/` — what lives here
-
-## How to evolve this
-<Where to add things. Which tests to expect to update. What NOT to touch.>
-
-## Open questions
-<What is still being figured out. Empty if none.>

.claude/skills/document-internal/seeds/_templates/design-spec.md

@@ -1,34 +0,0 @@
----
-title: <design name>
-owner: <github handle>
-status: proposed | accepted | rejected | superseded
-date: YYYY-MM-DD
-supersedes: <design-slug or none>
----
-
-# <design name>
-
-## Goal
-<2-4 sentences. What this design is trying to accomplish.>
-
-## Context
-<1-2 paragraphs. The current state, what is failing, why this needs to change.>
-
-## Selected Approach
-<The chosen design at a high level. Architecture, components, data flow.>
-
-## Alternatives Considered
-### 1. <name>
-<2-3 sentences on what this would look like, then pro/con and why rejected (or deferred).>
-
-### 2. <name>
-<Same shape.>
-
-## Out of Scope
-<What this design does NOT cover. Defer references.>
-
-## Rollout
-<Numbered steps from "nothing exists" to "fully shipped".>
-
-## Open Questions
-<Resolved during design? Empty. Unresolved? List with owner.>

.claude/skills/document-internal/seeds/_templates/feature-note.md

@@ -1,29 +0,0 @@
----
-title: <feature name>
-owner: <github handle>
-status: shipped | wip | deprecated
-date: YYYY-MM-DD
-prs: ["#NNN"]
-tags: [agent, browser, mcp]
----
-
-# <feature name>
-
-## What it does
-<2-3 sentences. What can someone now do that they could not before. Lead with user-facing impact, not implementation.>
-
-## Why we built it
-<1-2 sentences. Motivation. What pain it removed or what unlocked.>
-
-## How it works
-<3-6 sentences. The flow at a high level. Name the key files.>
-
-## Key files
-- `path/to/file.ts` — what it does
-- `path/to/other.ts` — what it does
-
-## How to run / test it locally
-<bullet list of commands. Empty section if N/A — do not fake.>
-
-## Gotchas
-<known sharp edges. "If you see X, that's why." Empty if N/A.>

.github/workflows/build-agent.yml

@@ -0,0 +1,157 @@
+name: build-agent
+
+on:
+  workflow_dispatch:
+    inputs:
+      agent:
+        description: "Agent name from bundle.json"
+        required: true
+        type: string
+        default: openclaw
+      publish:
+        description: "Upload to R2 and merge manifest slice"
+        required: false
+        default: false
+        type: boolean
+  pull_request:
+    paths:
+      - "packages/browseros-agent/packages/build-tools/**"
+      - ".github/workflows/build-agent.yml"
+
+env:
+  BUN_VERSION: "1.3.6"
+  PKG_DIR: packages/browseros-agent/packages/build-tools
+
+permissions:
+  contents: read
+
+jobs:
+  check:
+    runs-on: ubuntu-24.04
+    steps:
+      - uses: actions/checkout@v4
+      - uses: oven-sh/setup-bun@v2
+        with:
+          bun-version: ${{ env.BUN_VERSION }}
+      - working-directory: packages/browseros-agent
+        run: bun install --frozen-lockfile
+      - working-directory: packages/browseros-agent
+        run: bun run --filter @browseros/build-tools typecheck
+      - working-directory: packages/browseros-agent
+        run: bun run --filter @browseros/build-tools test
+
+  build:
+    needs: check
+    strategy:
+      fail-fast: false
+      matrix:
+        include:
+          - arch: arm64
+            runner: ubuntu-24.04-arm
+    runs-on: ${{ matrix.runner }}
+    steps:
+      - uses: actions/checkout@v4
+      - uses: oven-sh/setup-bun@v2
+        with:
+          bun-version: ${{ env.BUN_VERSION }}
+      - name: Install podman
+        run: |
+          sudo apt-get update
+          sudo apt-get install -y podman
+      - working-directory: packages/browseros-agent
+        run: bun install --frozen-lockfile
+      - name: Build tarball
+        working-directory: ${{ env.PKG_DIR }}
+        env:
+          AGENT: ${{ inputs.agent || 'openclaw' }}
+          OUT: ${{ github.workspace }}/dist/images
+        run: bun run build:tarball -- --agent "$AGENT" --arch "${{ matrix.arch }}" --output-dir "$OUT"
+      - uses: actions/upload-artifact@v4
+        with:
+          name: tarball-${{ inputs.agent || 'openclaw' }}-${{ matrix.arch }}
+          path: dist/images/
+          retention-days: 7
+
+  smoke:
+    needs: build
+    runs-on: ubuntu-24.04-arm
+    steps:
+      - uses: actions/checkout@v4
+      - uses: oven-sh/setup-bun@v2
+        with:
+          bun-version: ${{ env.BUN_VERSION }}
+      - uses: actions/download-artifact@v4
+        with:
+          name: tarball-${{ inputs.agent || 'openclaw' }}-arm64
+          path: dist/images
+      - name: Install podman
+        run: |
+          sudo apt-get update
+          sudo apt-get install -y podman
+      - working-directory: packages/browseros-agent
+        run: bun install --frozen-lockfile
+      - name: Smoke test tarball
+        working-directory: ${{ env.PKG_DIR }}
+        env:
+          AGENT: ${{ inputs.agent || 'openclaw' }}
+        run: |
+          set -euo pipefail
+          tarball="$(find "$GITHUB_WORKSPACE/dist/images" -name "${AGENT}-*-arm64.tar.gz" -print -quit)"
+          if [ -z "$tarball" ]; then
+            echo "missing arm64 tarball artifact for ${AGENT}" >&2
+            exit 1
+          fi
+          bun run smoke:tarball -- --agent "$AGENT" --arch arm64 --tarball "$tarball"
+
+  publish:
+    needs: [build, smoke]
+    if: ${{ github.event_name == 'workflow_dispatch' && inputs.publish == true }}
+    runs-on: ubuntu-24.04
+    environment: release
+    concurrency:
+      group: r2-manifest-publish
+      cancel-in-progress: false
+    steps:
+      - uses: actions/checkout@v4
+      - uses: oven-sh/setup-bun@v2
+        with:
+          bun-version: ${{ env.BUN_VERSION }}
+      - uses: actions/download-artifact@v4
+        with:
+          pattern: tarball-*
+          path: dist/images
+          merge-multiple: true
+      - working-directory: packages/browseros-agent
+        run: bun install --frozen-lockfile
+      - name: Upload tarballs to R2
+        working-directory: ${{ env.PKG_DIR }}
+        env:
+          R2_ACCOUNT_ID: ${{ secrets.R2_ACCOUNT_ID }}
+          R2_ACCESS_KEY_ID: ${{ secrets.R2_ACCESS_KEY_ID }}
+          R2_SECRET_ACCESS_KEY: ${{ secrets.R2_SECRET_ACCESS_KEY }}
+          R2_BUCKET: ${{ secrets.R2_BUCKET }}
+        run: |
+          set -euo pipefail
+          for file in "$GITHUB_WORKSPACE"/dist/images/*.tar.gz; do
+            base="$(basename "$file")"
+            bun run upload -- --file "$file" --key "vm/images/$base" --content-type "application/gzip" --sidecar-sha
+          done
+      - name: Merge agent slice into manifest
+        working-directory: ${{ env.PKG_DIR }}
+        env:
+          AGENT: ${{ inputs.agent || 'openclaw' }}
+          R2_ACCOUNT_ID: ${{ secrets.R2_ACCOUNT_ID }}
+          R2_ACCESS_KEY_ID: ${{ secrets.R2_ACCESS_KEY_ID }}
+          R2_SECRET_ACCESS_KEY: ${{ secrets.R2_SECRET_ACCESS_KEY }}
+          R2_BUCKET: ${{ secrets.R2_BUCKET }}
+        run: |
+          set -euo pipefail
+          mkdir -p dist/images
+          cp -R "$GITHUB_WORKSPACE"/dist/images/* dist/images/
+          bun run download -- --key vm/manifest.json --out dist/baseline-manifest.json
+          bun run emit-manifest -- \
+            --slice "agents:${AGENT}" \
+            --dist-dir dist \
+            --merge-from dist/baseline-manifest.json \
+            --out dist/manifest.json
+          bun run upload -- --file dist/manifest.json --key vm/manifest.json --content-type "application/json"

.github/workflows/eval-weekly.yml

@@ -14,7 +14,7 @@ on:
       config:
         description: 'Eval config file (relative to apps/eval/)'
         required: false
-        default: 'configs/legacy/browseros-agent-weekly.json'
+        default: 'configs/browseros-agent-weekly.json'
 
 permissions:
   contents: read
@@ -42,12 +42,10 @@ jobs:
 
       - name: Install dependencies
         working-directory: packages/browseros-agent
-        run: bun install --ignore-scripts
+        run: bun install --ignore-scripts && bun run build:agent-sdk
 
       - name: Install Python eval dependencies
-        # agisdk pinned so silent upstream releases can't shift task definitions
-        # or grader behavior. Bump intentionally with a documented re-baseline.
-        run: pip install agisdk==0.3.5 requests
+        run: pip install agisdk requests
 
       - name: Clone WebArena-Infinity
         run: git clone --depth 1 https://github.com/web-arena-x/webarena-infinity.git /tmp/webarena-infinity
@@ -62,27 +60,33 @@ jobs:
           curl -sL -o /tmp/nopecha.zip https://github.com/NopeCHALLC/nopecha-extension/releases/latest/download/chromium_automation.zip
           unzip -qo /tmp/nopecha.zip -d extensions/nopecha
 
-      - name: Run eval and publish to R2
+      - name: Run eval
         working-directory: packages/browseros-agent/apps/eval
         env:
           FIREWORKS_API_KEY: ${{ secrets.FIREWORKS_API_KEY }}
           OPENROUTER_API_KEY: ${{ secrets.OPENROUTER_API_KEY }}
           CLAUDE_CODE_OAUTH_TOKEN: ${{ secrets.CLAUDE_CODE_OAUTH_TOKEN }}
           NOPECHA_API_KEY: ${{ secrets.NOPECHA_API_KEY }}
+          BROWSEROS_BINARY: /usr/bin/browseros
+          WEBARENA_INFINITY_DIR: /tmp/webarena-infinity
+          EVAL_CONFIG: ${{ github.event.inputs.config || 'configs/browseros-agent-weekly.json' }}
+        run: |
+          echo "Running eval with config: $EVAL_CONFIG"
+          xvfb-run --auto-servernum --server-args="-screen 0 1440x900x24" bun run src/index.ts -c "$EVAL_CONFIG"
+
+      - name: Upload runs to R2
+        if: success()
+        working-directory: packages/browseros-agent/apps/eval
+        env:
           EVAL_R2_ACCOUNT_ID: ${{ secrets.EVAL_R2_ACCOUNT_ID }}
           EVAL_R2_ACCESS_KEY_ID: ${{ secrets.EVAL_R2_ACCESS_KEY_ID }}
           EVAL_R2_SECRET_ACCESS_KEY: ${{ secrets.EVAL_R2_SECRET_ACCESS_KEY }}
           EVAL_R2_BUCKET: ${{ secrets.EVAL_R2_BUCKET }}
           EVAL_R2_CDN_BASE_URL: ${{ secrets.EVAL_R2_CDN_BASE_URL }}
-          BROWSEROS_BINARY: /usr/bin/browseros
-          WEBARENA_INFINITY_DIR: /tmp/webarena-infinity
-          # OpenClaw container runtime is macOS-only; opt the Linux runner
-          # into the no-op stub so the server can boot and the eval can run.
-          BROWSEROS_SKIP_OPENCLAW: '1'
-          EVAL_CONFIG: ${{ github.event.inputs.config || 'configs/legacy/browseros-agent-weekly.json' }}
+          EVAL_CONFIG: ${{ github.event.inputs.config || 'configs/browseros-agent-weekly.json' }}
         run: |
-          echo "Running eval with config: $EVAL_CONFIG"
-          xvfb-run --auto-servernum --server-args="-screen 0 1440x900x24" bun run src/index.ts suite --config "$EVAL_CONFIG" --publish r2
+          CONFIG_NAME=$(basename "$EVAL_CONFIG" .json)
+          bun scripts/upload-run.ts "results/$CONFIG_NAME"
 
       - name: Generate trend report
         if: success()
@@ -103,11 +107,3 @@ jobs:
         with:
           name: eval-report-${{ github.run_id }}
           path: /tmp/eval-report.html
-
-      - name: Upload server stderr logs (for post-mortem on startup failures)
-        if: always()
-        uses: actions/upload-artifact@v4
-        with:
-          name: browseros-server-logs-${{ github.run_id }}
-          path: /tmp/browseros-server-logs/
-          if-no-files-found: ignore

.github/workflows/release-agent-sdk.yml

@@ -1,11 +1,168 @@
-name: Release BrowserOS Agent SDK (disabled)
+name: Release BrowserOS Agent SDK
 
 on:
   workflow_dispatch:
 
+concurrency:
+  group: release-agent-sdk
+  cancel-in-progress: false
+
 jobs:
-  disabled:
-    if: ${{ false }}
+  publish:
+    if: github.ref == 'refs/heads/main'
     runs-on: ubuntu-latest
+    permissions:
+      contents: write
+      pull-requests: write
+    defaults:
+      run:
+        working-directory: packages/browseros-agent/packages/agent-sdk
+
     steps:
-      - run: echo "Agent SDK publishing is disabled."
+      - uses: actions/checkout@v6
+        with:
+          fetch-depth: 0
+
+      - uses: oven-sh/setup-bun@v2
+
+      - uses: actions/setup-node@v6
+        with:
+          node-version: "20"
+          registry-url: "https://registry.npmjs.org"
+
+      - name: Install dependencies
+        run: bun ci
+        working-directory: packages/browseros-agent
+
+      - name: Build
+        run: bun run build
+
+      - name: Test
+        run: bun test
+
+      - name: Get version
+        id: version
+        run: |
+          echo "version=$(node -p "require('./package.json').version")" >> "$GITHUB_OUTPUT"
+          echo "release_sha=$(git rev-parse HEAD)" >> "$GITHUB_OUTPUT"
+
+      - name: Generate release notes
+        env:
+          GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+        run: |
+          SDK_PATH="packages/browseros-agent/packages/agent-sdk"
+          CURRENT_TAG="agent-sdk-v${{ steps.version.outputs.version }}"
+          # Find the previous tag, excluding the current version's tag
+          # (which may already exist from a prior failed run)
+          PREV_TAG=$(git tag -l "agent-sdk-v*" --sort=-v:refname | grep -v "^${CURRENT_TAG}$" | head -n 1)
+
+          if [ -z "$PREV_TAG" ]; then
+            echo "Initial release" > /tmp/release-notes.md
+          else
+            # Get commits scoped to the SDK directory
+            COMMITS=$(git log "$PREV_TAG"..HEAD --pretty=format:"%H" -- "$SDK_PATH")
+
+            if [ -z "$COMMITS" ]; then
+              echo "No notable changes." > /tmp/release-notes.md
+            else
+              echo "## What's Changed" > /tmp/release-notes.md
+              echo "" >> /tmp/release-notes.md
+
+              # For each commit, find the associated PR and format with author
+              CONTRIBUTORS=""
+              while IFS= read -r SHA; do
+                # Get commit subject and author
+                SUBJECT=$(git log -1 --pretty=format:"%s" "$SHA")
+                AUTHOR=$(git log -1 --pretty=format:"%an" "$SHA")
+                GITHUB_USER=$(gh api "/repos/${{ github.repository }}/commits/${SHA}" --jq '.author.login // empty' 2>/dev/null)
+
+                # Find associated PR number
+                PR_NUM=$(gh api "/repos/${{ github.repository }}/commits/${SHA}/pulls" --jq '.[0].number // empty' 2>/dev/null)
+
+                # Format line: skip PR number if already in the commit subject
+                # (squash merges include "(#123)" in the subject automatically)
+                if [ -n "$PR_NUM" ] && ! echo "$SUBJECT" | grep -qF "(#${PR_NUM})"; then
+                  echo "- ${SUBJECT} (#${PR_NUM})" >> /tmp/release-notes.md
+                else
+                  echo "- ${SUBJECT}" >> /tmp/release-notes.md
+                fi
+              done <<< "$COMMITS"
+            fi
+          fi
+        working-directory: ${{ github.workspace }}
+
+      - name: Publish
+        run: npm publish --access public
+        env:
+          NODE_AUTH_TOKEN: ${{ secrets.NPM_TOKEN }}
+
+      - name: Create GitHub release
+        env:
+          GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+        run: |
+          TAG="agent-sdk-v${{ steps.version.outputs.version }}"
+          RELEASE_SHA="${{ steps.version.outputs.release_sha }}"
+          TITLE="BrowserOS Agent SDK - v${{ steps.version.outputs.version }}"
+
+          # Create or reuse tag (idempotent for re-runs)
+          if git rev-parse "$TAG" >/dev/null 2>&1; then
+            echo "Tag $TAG already exists, skipping tag creation"
+          else
+            git tag "$TAG" "$RELEASE_SHA"
+          fi
+
+          # Push tag (skip if already on remote)
+          if git ls-remote --tags origin "$TAG" | grep -q "$TAG"; then
+            echo "Tag $TAG already on remote, skipping push"
+          else
+            git push origin "$TAG"
+          fi
+
+          # Create or update release
+          if gh release view "$TAG" >/dev/null 2>&1; then
+            echo "Release $TAG already exists, updating"
+            gh release edit "$TAG" --title "$TITLE" --notes-file /tmp/release-notes.md
+          else
+            gh release create "$TAG" --title "$TITLE" --notes-file /tmp/release-notes.md
+          fi
+        working-directory: ${{ github.workspace }}
+
+      - name: Update CHANGELOG.md via PR
+        env:
+          GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+        run: |
+          VERSION="${{ steps.version.outputs.version }}"
+          DATE=$(date -u +"%Y-%m-%d")
+          BRANCH="docs/agent-sdk-changelog-v${VERSION}"
+          CHANGELOG="packages/browseros-agent/packages/agent-sdk/CHANGELOG.md"
+
+          # Return to main before branching
+          git checkout main
+
+          # Use head/tail to safely insert without sed quoting issues
+          {
+            head -n 1 "$CHANGELOG"
+            echo ""
+            echo "## v${VERSION} (${DATE})"
+            echo ""
+            cat /tmp/release-notes.md
+            echo ""
+            tail -n +2 "$CHANGELOG"
+          } > /tmp/new-changelog.md
+          mv /tmp/new-changelog.md "$CHANGELOG"
+
+          git config user.name "github-actions[bot]"
+          git config user.email "github-actions[bot]@users.noreply.github.com"
+          git checkout -b "$BRANCH"
+          git add "$CHANGELOG"
+          git commit -m "docs: update agent-sdk changelog for v${VERSION}"
+          git push origin "$BRANCH"
+
+          gh pr create \
+            --title "docs: update agent-sdk changelog for v${VERSION}" \
+            --body "Auto-generated changelog update for BrowserOS Agent SDK v${VERSION}." \
+            --base main \
+            --head "$BRANCH"
+
+          gh pr merge "$BRANCH" --squash --auto || true
+        working-directory: ${{ github.workspace }}

.github/workflows/sync-internal-docs.yml

@@ -1,53 +0,0 @@
-name: Sync internal-docs submodule
-
-on:
-  schedule:
-    - cron: '0 */4 * * *'
-  workflow_dispatch:
-
-jobs:
-  sync:
-    name: Bump internal-docs submodule pointer on dev
-    runs-on: ubuntu-latest
-    steps:
-      - name: Rewrite SSH submodule URL to HTTPS-with-token
-        env:
-          TOKEN: ${{ secrets.INTERNAL_DOCS_SYNC_TOKEN }}
-        run: |
-          git config --global "url.https://x-access-token:${TOKEN}@github.com/.insteadOf" "git@github.com:"
-
-      - uses: actions/checkout@v4
-        with:
-          token: ${{ secrets.INTERNAL_DOCS_SYNC_TOKEN }}
-          submodules: true
-          ref: dev
-          fetch-depth: 50
-
-      - name: Bump submodule pointer if internal-docs has new commits
-        env:
-          GH_TOKEN: ${{ secrets.INTERNAL_DOCS_SYNC_TOKEN }}
-        run: |
-          set -e
-
-          # Skip if submodule not yet configured (handoff window before someone adds it)
-          if ! git config --file .gitmodules --get-regexp '^submodule\..internal-docs\.path$' >/dev/null 2>&1; then
-            echo "internal-docs submodule not yet configured in .gitmodules. Skipping."
-            exit 0
-          fi
-
-          git submodule update --remote --merge .internal-docs
-
-          if git diff --quiet .internal-docs; then
-            echo "No internal-docs changes to sync."
-            exit 0
-          fi
-
-          git config user.name  "browseros-bot"
-          git config user.email "bot@browseros.ai"
-          git add .internal-docs
-          git commit -m "chore: sync internal-docs submodule"
-
-          # Rebase onto latest dev to absorb any commits that landed during the run,
-          # then push. set -e takes care of failing the run on rebase conflict.
-          git pull --rebase origin dev
-          git push origin dev

.github/workflows/test.yml

@@ -58,20 +58,28 @@ jobs:
             command: (cd apps/server && bun run test:lib)
             junit_path: test-results/server-lib.xml
             needs_browser: false
+          - suite: server-sdk
+            command: (cd apps/server && bun run test:sdk)
+            junit_path: test-results/server-sdk.xml
+            needs_browser: true
           - suite: server-root
             command: (cd apps/server && bun run test:root)
             junit_path: test-results/server-root.xml
             needs_browser: false
           - suite: agent
-            command: (cd apps/agent && bun run test)
+            command: bun run test:agent
             junit_path: test-results/agent.xml
             needs_browser: false
           - suite: eval
-            command: (cd apps/eval && bun run test)
+            command: bun run test:eval
             junit_path: test-results/eval.xml
             needs_browser: false
+          - suite: agent-sdk
+            command: bun run test:agent-sdk
+            junit_path: test-results/agent-sdk.xml
+            needs_browser: false
           - suite: build
-            command: bun run ./scripts/run-bun-test.ts ./scripts/build
+            command: bun run test:build
             junit_path: test-results/build.xml
             needs_browser: false
 

README.md

@@ -188,21 +188,6 @@ We'd love your help making BrowserOS better! See our [Contributing Guide](CONTRI
 - [ungoogled-chromium](https://github.com/ungoogled-software/ungoogled-chromium) — BrowserOS uses some patches for enhanced privacy. Thanks to everyone behind this project!
 - [The Chromium Project](https://www.chromium.org/) — at the core of BrowserOS, making it possible to exist in the first place.
 
-## Citation
-
-If you use BrowserOS in your research or project, please cite:
-
-```bibtex
-@software{browseros2025,
-  author = {Nithin Sonti and Nikhil Sonti and {BrowserOS-team}},
-  title = {BrowserOS: The open-source Agentic browser},
-  url = {https://github.com/browseros-ai/BrowserOS},
-  year = {2025},
-  publisher = {GitHub},
-  license = {AGPL-3.0},
-}
-```
-
 ## License
 
 BrowserOS is open source under the [AGPL-3.0 license](LICENSE).

packages/browseros-agent/README.md

@@ -1,6 +1,6 @@
 # BrowserOS Agent
 
-The agent platform powering [BrowserOS](https://github.com/browseros-ai/BrowserOS) — contains the MCP server, agent UI, CLI, and evaluation framework.
+The agent platform powering [BrowserOS](https://github.com/browseros-ai/BrowserOS) — contains the MCP server, agent UI, CLI, evaluation framework, and SDK.
 
 ## Monorepo Structure
 
@@ -12,6 +12,7 @@ apps/
   eval/            # Evaluation framework for benchmarking agents
 
 packages/
+  agent-sdk/       # Node.js SDK (@browseros-ai/agent-sdk)
   cdp-protocol/    # Type-safe Chrome DevTools Protocol bindings
   shared/          # Shared constants (ports, timeouts, limits)
 ```
@@ -22,6 +23,7 @@ packages/
 | `apps/agent` | Agent UI — Chrome extension for the chat interface |
 | `apps/cli` | Go CLI — control BrowserOS from the terminal or AI coding agents |
 | `apps/eval` | Benchmark framework — WebVoyager, Mind2Web evaluation |
+| `packages/agent-sdk` | Node.js SDK for browser automation with natural language |
 | `packages/cdp-protocol` | Auto-generated CDP type bindings used by the server |
 | `packages/shared` | Shared constants used across packages |
 
@@ -79,15 +81,14 @@ cp apps/server/.env.example apps/server/.env.development
 cp apps/agent/.env.example apps/agent/.env.development
 cp apps/server/.env.production.example apps/server/.env.production
 
-# Install deps and generate agent code
+# Install deps, generate agent code, and sync the VM cache
 bun run dev:setup
 
 # Start the full dev environment
 bun run dev:watch
 ```
 
-`dev:watch` starts the server immediately. OpenClaw VM/image prewarm runs from
-the server startup path and pulls the configured GHCR image on demand.
+`dev:watch` exits when the VM cache manifest is missing, but setup stays in `dev:setup`.
 
 ### Environment Variables
 
@@ -157,14 +158,9 @@ bun run build:server          # Build production server resource artifacts and u
 bun run build:agent           # Build agent extension
 
 # Test
-bun run test                  # Run all tests
-bun run test:all              # Run all tests
-bun run test:main             # Run key server tools and integration tests
-
-# App-specific test groups (from packages/browseros-agent)
-cd apps/server && bun run test:tools
-cd apps/server && bun run test:cdp
-cd apps/server && bun run test:integration
+bun run test                  # Run standard tests
+bun run test:cdp              # Run CDP-based tests
+bun run test:integration      # Run integration tests
 
 # Quality
 bun run lint                  # Check with Biome

packages/browseros-agent/apps/agent/components/chat/ChatProviderSelector.helpers.ts

@@ -1,50 +0,0 @@
-import type { Provider } from './chatComponentTypes'
-
-export interface ProviderOptionGroup {
-  key: 'llm' | 'acp'
-  label: string
-  options: Provider[]
-}
-
-export function groupProviderOptions(
-  providers: Provider[],
-): ProviderOptionGroup[] {
-  const llm = providers.filter((provider) => provider.kind !== 'acp')
-  const acp = providers.filter((provider) => provider.kind === 'acp')
-
-  return [
-    ...(llm.length
-      ? [{ key: 'llm' as const, label: 'AI Providers', options: llm }]
-      : []),
-    ...(acp.length
-      ? [{ key: 'acp' as const, label: 'Agents', options: acp }]
-      : []),
-  ]
-}
-
-export function getProviderSearchValue(
-  provider: Provider,
-  groupLabel: string,
-): string {
-  return [
-    provider.id,
-    provider.name,
-    provider.type,
-    groupLabel,
-    provider.adapterName,
-    provider.modelLabel,
-  ]
-    .filter(Boolean)
-    .join(' ')
-}
-
-export function getProviderSubtitle(provider: Provider): string | undefined {
-  if (provider.kind !== 'acp') return undefined
-  return [
-    provider.adapterName,
-    provider.modelLabel,
-    provider.modelControl === 'best-effort' ? 'best effort' : undefined,
-  ]
-    .filter(Boolean)
-    .join(' · ')
-}

packages/browseros-agent/apps/agent/components/chat/ChatProviderSelector.test.tsx

@@ -1,72 +0,0 @@
-import { describe, expect, it } from 'bun:test'
-import {
-  getProviderSearchValue,
-  getProviderSubtitle,
-  groupProviderOptions,
-} from './ChatProviderSelector.helpers'
-import type { Provider } from './chatComponentTypes'
-
-const options: Provider[] = [
-  { kind: 'llm', id: 'browseros', name: 'BrowserOS', type: 'browseros' },
-  {
-    kind: 'llm',
-    id: 'anthropic-sonnet',
-    name: 'Anthropic Sonnet',
-    type: 'anthropic',
-  },
-  {
-    kind: 'acp',
-    id: 'agent-claude-review',
-    name: 'Review Bot',
-    type: 'acp',
-    adapterName: 'Claude Code',
-    modelLabel: 'Haiku',
-    modelControl: 'best-effort',
-  },
-  {
-    kind: 'acp',
-    id: 'agent-codex-browser',
-    name: 'Browser Driver',
-    type: 'acp',
-    adapterName: 'Codex',
-    modelLabel: 'GPT-5.5',
-    modelControl: 'runtime-supported',
-  },
-]
-
-describe('groupProviderOptions', () => {
-  it('groups normal providers separately from created agents', () => {
-    expect(groupProviderOptions(options)).toEqual([
-      {
-        key: 'llm',
-        label: 'AI Providers',
-        options: [options[0], options[1]],
-      },
-      {
-        key: 'acp',
-        label: 'Agents',
-        options: [options[2], options[3]],
-      },
-    ])
-  })
-})
-
-describe('getProviderSearchValue', () => {
-  it('matches created-agent group labels and item labels', () => {
-    expect(getProviderSearchValue(options[2], 'Agents')).toContain('Agents')
-    expect(getProviderSearchValue(options[2], 'Agents')).toContain('Review Bot')
-    expect(getProviderSearchValue(options[2], 'Agents')).toContain(
-      'Claude Code',
-    )
-  })
-})
-
-describe('getProviderSubtitle', () => {
-  it('describes created-agent runtime context without model-target copy', () => {
-    expect(getProviderSubtitle(options[2])).toBe(
-      'Claude Code · Haiku · best effort',
-    )
-    expect(getProviderSubtitle(options[3])).toBe('Codex · GPT-5.5')
-    expect(getProviderSubtitle(options[0])).toBeUndefined()
-  })
-})

packages/browseros-agent/apps/agent/components/chat/ChatProviderSelector.tsx

@@ -1,4 +1,4 @@
-import { Bot, Check, Plus } from 'lucide-react'
+import { Check, Plus } from 'lucide-react'
 import type { FC, PropsWithChildren } from 'react'
 import { useState } from 'react'
 import {
@@ -17,11 +17,6 @@ import {
 import { BrowserOSIcon, ProviderIcon } from '@/lib/llm-providers/providerIcons'
 import type { ProviderType } from '@/lib/llm-providers/types'
 import { cn } from '@/lib/utils'
-import {
-  getProviderSearchValue,
-  getProviderSubtitle,
-  groupProviderOptions,
-} from './ChatProviderSelector.helpers'
 import type { Provider } from './chatComponentTypes'
 
 interface ChatProviderSelectorProps {
@@ -34,58 +29,54 @@ export const ChatProviderSelector: FC<
   PropsWithChildren<ChatProviderSelectorProps>
 > = ({ children, providers, selectedProvider, onSelectProvider }) => {
   const [open, setOpen] = useState(false)
-  const groups = groupProviderOptions(providers)
 
   return (
     <Popover open={open} onOpenChange={setOpen}>
       <PopoverTrigger asChild>{children}</PopoverTrigger>
-      <PopoverContent side="bottom" align="start" className="w-64 p-0">
+      <PopoverContent side="bottom" align="start" className="w-48 p-0">
         <Command>
-          <CommandInput
-            placeholder="Search providers or agents..."
-            className="h-9"
-          />
+          <CommandInput placeholder="Search providers..." className="h-9" />
           <CommandList>
+            <div className="my-2 px-2 font-semibold text-muted-foreground text-xs uppercase tracking-wide">
+              AI Provider
+            </div>
             <CommandEmpty>No provider found</CommandEmpty>
-            {groups.map((group) => (
-              <CommandGroup key={group.key} heading={group.label}>
-                {group.options.map((provider) => {
-                  const isSelected = selectedProvider.id === provider.id
-                  const subtitle = getProviderSubtitle(provider)
-                  return (
-                    <CommandItem
-                      key={provider.id}
-                      value={getProviderSearchValue(provider, group.label)}
-                      onSelect={() => {
-                        onSelectProvider(provider)
-                        setOpen(false)
-                      }}
-                      className={cn(
-                        'flex w-full items-center gap-3 rounded-md p-2 transition-colors',
-                        isSelected && 'bg-[var(--accent-orange)]/10',
+            <CommandGroup>
+              {providers.map((provider) => {
+                const isSelected = selectedProvider.id === provider.id
+                return (
+                  <CommandItem
+                    key={provider.id}
+                    value={`${provider.id} ${provider.name}`}
+                    onSelect={() => {
+                      onSelectProvider(provider)
+                      setOpen(false)
+                    }}
+                    className={cn(
+                      'flex w-full items-center gap-3 rounded-md p-2 transition-colors',
+                      isSelected && 'bg-[var(--accent-orange)]/10',
+                    )}
+                  >
+                    <span className="text-muted-foreground">
+                      {provider.type === 'browseros' ? (
+                        <BrowserOSIcon size={18} />
+                      ) : (
+                        <ProviderIcon
+                          type={provider.type as ProviderType}
+                          size={18}
+                        />
                       )}
-                    >
-                      <span className="text-muted-foreground">
-                        <ProviderOptionIcon provider={provider} />
-                      </span>
-                      <span className="min-w-0 flex-1 text-left">
-                        <span className="block truncate text-sm">
-                          {provider.name}
-                        </span>
-                        {subtitle && (
-                          <span className="block truncate text-muted-foreground text-xs">
-                            {subtitle}
-                          </span>
-                        )}
-                      </span>
-                      {isSelected && (
-                        <Check className="h-3.5 w-3.5 text-[var(--accent-orange)]" />
-                      )}
-                    </CommandItem>
-                  )
-                })}
-              </CommandGroup>
-            ))}
+                    </span>
+                    <span className="flex-1 text-left text-sm">
+                      {provider.name}
+                    </span>
+                    {isSelected && (
+                      <Check className="h-3.5 w-3.5 text-[var(--accent-orange)]" />
+                    )}
+                  </CommandItem>
+                )
+              })}
+            </CommandGroup>
             <div className="border-border border-t p-1">
               <button
                 type="button"
@@ -105,9 +96,3 @@ export const ChatProviderSelector: FC<
     </Popover>
   )
 }
-
-function ProviderOptionIcon({ provider }: { provider: Provider }) {
-  if (provider.kind === 'acp') return <Bot size={18} />
-  if (provider.type === 'browseros') return <BrowserOSIcon size={18} />
-  return <ProviderIcon type={provider.type as ProviderType} size={18} />
-}

packages/browseros-agent/apps/agent/components/chat/chatComponentTypes.ts

@@ -1,14 +1,7 @@
 import type { ProviderType } from '@/lib/llm-providers/types'
 
-export type ChatProviderType = ProviderType | 'acp'
-
 export interface Provider {
   id: string
   name: string
-  type: ChatProviderType
-  kind: 'llm' | 'acp'
-  agentId?: string
-  adapterName?: string
-  modelLabel?: string
-  modelControl?: 'runtime-supported' | 'best-effort'
+  type: ProviderType
 }

packages/browseros-agent/apps/agent/entrypoints/app/agent-command/AgentCard.tsx

@@ -0,0 +1,136 @@
+import { Bot, Loader2, Wrench } from 'lucide-react'
+import type { FC } from 'react'
+import type { AgentCardData } from '@/lib/agent-conversations/types'
+import { cn } from '@/lib/utils'
+
+interface AgentCardProps {
+  agent: AgentCardData
+  onClick: () => void
+  active?: boolean
+}
+
+function formatTimestamp(timestamp?: number): string {
+  if (!timestamp) return 'No activity yet'
+  const diff = Date.now() - timestamp
+  const minutes = Math.floor(diff / 60000)
+  if (minutes < 1) return 'just now'
+  if (minutes < 60) return `${minutes}m ago`
+  const hours = Math.floor(minutes / 60)
+  if (hours < 24) return `${hours}h ago`
+  return `${Math.floor(hours / 24)}d ago`
+}
+
+function getStatusLabel(status: AgentCardData['status']): string {
+  if (status === 'working') return 'Working'
+  if (status === 'error') return 'Error'
+  return 'Ready'
+}
+
+function getStatusTone(status: AgentCardData['status']): string {
+  if (status === 'working') return 'bg-amber-500'
+  if (status === 'error') return 'bg-destructive'
+  return 'bg-emerald-500'
+}
+
+function formatCost(usd: number): string {
+  if (usd < 0.005) return `$${usd.toFixed(4)}`
+  return `$${usd.toFixed(2)}`
+}
+
+export const AgentCardExpanded: FC<AgentCardProps> = ({
+  agent,
+  onClick,
+  active,
+}) => (
+  <button
+    type="button"
+    onClick={onClick}
+    className={cn(
+      'group flex min-h-32 w-full min-w-0 flex-col rounded-2xl border p-4 text-left shadow-sm transition-all duration-200',
+      active
+        ? 'border-border/80 bg-card shadow-md ring-1 ring-[var(--accent-orange)]/20'
+        : 'border-border/60 bg-card/85 hover:border-border hover:bg-card hover:shadow-md',
+    )}
+  >
+    <div className="flex items-start justify-between gap-3">
+      <div className="flex min-w-0 items-center gap-3">
+        <div
+          className={cn(
+            'flex size-10 shrink-0 items-center justify-center rounded-xl',
+            active
+              ? 'bg-[var(--accent-orange)]/10 text-[var(--accent-orange)]'
+              : 'bg-muted text-muted-foreground',
+          )}
+        >
+          <Bot className="size-5" />
+        </div>
+        <div className="min-w-0">
+          <div className="truncate font-semibold text-sm">{agent.name}</div>
+          <div className="truncate text-muted-foreground text-xs">
+            {agent.model ?? 'OpenClaw agent'}
+          </div>
+        </div>
+      </div>
+      <div className="flex items-center gap-2 rounded-full border border-border/60 bg-background/70 px-2.5 py-1 text-[11px] text-muted-foreground">
+        <span
+          className={cn('size-2 rounded-full', getStatusTone(agent.status))}
+        />
+        <span>{getStatusLabel(agent.status)}</span>
+      </div>
+    </div>
+
+    <div className="mt-4 flex-1">
+      <p className="line-clamp-2 text-foreground/90 text-sm">
+        {agent.lastMessage ??
+          'Start a conversation to see recent work and summaries.'}
+      </p>
+    </div>
+
+    <div className="mt-4 space-y-1.5 text-muted-foreground text-xs">
+      <div className="flex items-center justify-between gap-3">
+        <span>{formatTimestamp(agent.lastMessageTimestamp)}</span>
+        {agent.costUsd ? (
+          <span className="tabular-nums opacity-70">
+            {formatCost(agent.costUsd)}
+          </span>
+        ) : null}
+      </div>
+      {agent.status === 'working' && agent.currentTool ? (
+        <div className="flex items-center gap-1.5 text-[var(--accent-orange)]/70">
+          <Loader2 className="size-3 shrink-0 animate-spin" />
+          <span className="truncate">{agent.currentTool}</span>
+        </div>
+      ) : agent.activitySummary ? (
+        <div className="flex items-center gap-1.5 text-muted-foreground/60">
+          <Wrench className="size-3 shrink-0" />
+          <span className="truncate">{agent.activitySummary}</span>
+        </div>
+      ) : null}
+    </div>
+  </button>
+)
+
+export const AgentCardCompact: FC<AgentCardProps> = ({
+  agent,
+  onClick,
+  active,
+}) => (
+  <button
+    type="button"
+    onClick={onClick}
+    className={cn(
+      'inline-flex items-center gap-2 rounded-full border px-3 py-2 text-sm transition-colors',
+      active
+        ? 'border-border bg-card shadow-sm ring-1 ring-[var(--accent-orange)]/20'
+        : 'border-border/60 bg-card/85 text-foreground hover:border-border hover:bg-card',
+    )}
+  >
+    <span
+      className={cn(
+        'size-2 rounded-full',
+        active ? 'bg-[var(--accent-orange)]' : getStatusTone(agent.status),
+      )}
+    />
+    <span className="truncate">{agent.name}</span>
+  </button>
+)

packages/browseros-agent/apps/agent/entrypoints/app/agent-command/AgentCardDock.tsx

@@ -1,71 +1,70 @@
 import { Plus } from 'lucide-react'
 import type { FC } from 'react'
-import type {
-  HarnessAdapterDescriptor,
-  HarnessAdapterHealth,
-  HarnessAgent,
-  HarnessAgentAdapter,
-} from '@/entrypoints/app/agents/agent-harness-types'
+import type { AgentCardData } from '@/lib/agent-conversations/types'
 import { cn } from '@/lib/utils'
-import { HomeAgentCard } from './HomeAgentCard'
+import { AgentCardCompact, AgentCardExpanded } from './AgentCard'
 
 interface AgentCardDockProps {
-  agents: HarnessAgent[]
-  adapters: HarnessAdapterDescriptor[]
+  agents: AgentCardData[]
   activeAgentId?: string
   onSelectAgent: (agentId: string) => void
   onCreateAgent?: () => void
+  compact?: boolean
 }
 
-function CreateAgentButton({ onCreateAgent }: { onCreateAgent: () => void }) {
+function CreateAgentButton({
+  compact,
+  onCreateAgent,
+}: {
+  compact?: boolean
+  onCreateAgent: () => void
+}) {
   return (
     <button
       type="button"
       onClick={onCreateAgent}
       className={cn(
-        'flex min-h-32 shrink-0 items-center justify-center gap-2 rounded-2xl border border-dashed px-5 py-4 text-muted-foreground transition-colors',
-        'hover:border-[var(--accent-orange)] hover:text-[var(--accent-orange)]',
+        'flex shrink-0 items-center justify-center gap-2 border border-dashed text-muted-foreground transition-colors hover:border-[var(--accent-orange)] hover:text-[var(--accent-orange)]',
+        compact
+          ? 'rounded-full px-3 py-2 text-sm'
+          : 'min-h-32 rounded-2xl px-5 py-4',
       )}
     >
-      <Plus className="size-5" />
-      <span>Create agent</span>
+      <Plus className={compact ? 'size-3.5' : 'size-5'} />
+      <span>{compact ? 'New' : 'Create agent'}</span>
     </button>
   )
 }
 
-/**
- * 3-column grid of HomeAgentCards plus a trailing "Create agent"
- * tile. The previous `compact` mode (rendered a horizontal pill rail)
- * had no callers and was dropped along with the legacy AgentCard.
- */
 export const AgentCardDock: FC<AgentCardDockProps> = ({
   agents,
-  adapters,
   activeAgentId,
   onSelectAgent,
   onCreateAgent,
+  compact,
 }) => {
   if (agents.length === 0 && !onCreateAgent) return null
 
-  const adapterHealth = new Map<HarnessAgentAdapter, HarnessAdapterHealth>()
-  for (const descriptor of adapters) {
-    if (descriptor.health) adapterHealth.set(descriptor.id, descriptor.health)
-  }
+  const Card = compact ? AgentCardCompact : AgentCardExpanded
 
   return (
-    <div className="grid gap-4 md:grid-cols-3">
+    <div
+      className={cn(
+        compact
+          ? 'flex items-center gap-2 overflow-x-auto pb-1'
+          : 'grid gap-4 md:grid-cols-3',
+      )}
+    >
       {agents.map((agent) => (
-        <HomeAgentCard
-          key={agent.id}
+        <Card
+          key={agent.agentId}
           agent={agent}
-          adapter={agent.adapter}
-          adapterHealth={adapterHealth.get(agent.adapter) ?? null}
-          active={agent.id === activeAgentId}
-          onClick={() => onSelectAgent(agent.id)}
+          active={agent.agentId === activeAgentId}
+          onClick={() => onSelectAgent(agent.agentId)}
         />
       ))}
       {onCreateAgent ? (
-        <CreateAgentButton onCreateAgent={onCreateAgent} />
+        <CreateAgentButton compact={compact} onCreateAgent={onCreateAgent} />
       ) : null}
     </div>
   )

packages/browseros-agent/apps/agent/entrypoints/app/agent-command/AgentCommandConversation.tsx

@@ -1,13 +1,7 @@
 import { ArrowLeft, Bot, Home } from 'lucide-react'
-import { type FC, useEffect, useMemo, useRef } from 'react'
+import { type FC, useEffect, useMemo, useRef, useState } from 'react'
 import { Navigate, useNavigate, useParams, useSearchParams } from 'react-router'
 import { Button } from '@/components/ui/button'
-import {
-  cancelHarnessTurn,
-  useEnqueueHarnessMessage,
-  useHarnessAgents,
-  useRemoveHarnessQueuedMessage,
-} from '@/entrypoints/app/agents/useAgents'
 import {
   type AgentEntry,
   getModelDisplayName,
@@ -21,9 +15,10 @@ import {
   filterTurnsPersistedInHistory,
   flattenHistoryPages,
 } from './claw-chat-types'
-import { QueuePanel } from './QueuePanel'
 import { useAgentConversation } from './useAgentConversation'
+import { useClawChatHistory } from './useClawChatHistory'
 import { useHarnessChatHistory } from './useHarnessChatHistory'
+import { useOutboundQueue } from './useOutboundQueue'
 
 function StatusBadge({ status }: { status: string }) {
   return (
@@ -181,10 +176,19 @@ function getAgentEntryMeta(agent: AgentEntry | undefined): string {
   return getModelDisplayName(agent?.model) ?? 'OpenClaw agent'
 }
 
+function getConversationStatusCopy(status: string | undefined): string {
+  if (status === 'running') return 'Ready'
+  if (status === 'starting') return 'Connecting'
+  if (status === 'error') return 'Attention'
+  if (status === 'stopped') return 'Offline'
+  return 'Setup'
+}
+
 function AgentConversationController({
   agentId,
   initialMessage,
   onInitialMessageConsumed,
+  status,
   agents,
   agentPathPrefix,
   createAgentPath,
@@ -192,6 +196,7 @@ function AgentConversationController({
   agentId: string
   initialMessage: string | null
   onInitialMessageConsumed: () => void
+  status: ReturnType<typeof useAgentCommandData>['status']
   agents: AgentEntry[]
   agentPathPrefix: string
   createAgentPath: string
@@ -199,67 +204,121 @@ function AgentConversationController({
   const navigate = useNavigate()
   const initialMessageSentRef = useRef<string | null>(null)
   const onInitialMessageConsumedRef = useRef(onInitialMessageConsumed)
+  const [streamSessionKey, setStreamSessionKey] = useState<string | null>(null)
   const agent = agents.find((entry) => entry.agentId === agentId)
   const agentName = agent?.name || agentId || 'Agent'
-  // Routing is now harness-only. Every OpenClaw agent has a harness
-  // record post the gateway → harness backfill, so the chat panel
-  // always talks to /agents/<id>/chat. The legacy ClawChat surface
-  // was deleted with the /claw/agents/:id/chat server route.
-  const harnessHistoryQuery = useHarnessChatHistory(agentId, Boolean(agent))
+  const isAgentHarnessAgent = agent?.source === 'agent-harness'
+  const clawHistoryQuery = useClawChatHistory({
+    agentId,
+    sessionKey: streamSessionKey,
+    enabled: Boolean(agent) && !isAgentHarnessAgent,
+  })
+  const harnessHistoryQuery = useHarnessChatHistory(
+    agentId,
+    Boolean(agent) && isAgentHarnessAgent,
+  )
 
   const historyMessages = useMemo(
     () =>
       flattenHistoryPages(
-        harnessHistoryQuery.data ? [harnessHistoryQuery.data] : [],
+        isAgentHarnessAgent
+          ? harnessHistoryQuery.data
+            ? [harnessHistoryQuery.data]
+            : []
+          : (clawHistoryQuery.data?.pages ?? []),
       ),
-    [harnessHistoryQuery.data],
+    [
+      clawHistoryQuery.data?.pages,
+      harnessHistoryQuery.data,
+      isAgentHarnessAgent,
+    ],
   )
   const chatHistory = useMemo(
     () => buildChatHistoryFromClawMessages(historyMessages),
     [historyMessages],
   )
-
-  // Listing query feeds queue + active-turn state for this agent. We
-  // already poll it every 5s for the rail; reusing the same cache
-  // keeps cross-tab queue state in sync without a second poll.
-  const { harnessAgents } = useHarnessAgents()
-  const harnessAgent = harnessAgents.find((entry) => entry.id === agentId)
-  const queue = harnessAgent?.queue ?? []
-  const activeTurnId = harnessAgent?.activeTurnId ?? null
+  const resolvedSessionKey =
+    streamSessionKey ??
+    (isAgentHarnessAgent
+      ? null
+      : (clawHistoryQuery.data?.pages?.[0]?.sessionKey ?? null))
 
   const { turns, streaming, send } = useAgentConversation(agentId, {
-    runtime: 'agent-harness',
-    sessionKey: null,
+    runtime: isAgentHarnessAgent ? 'agent-harness' : 'openclaw',
+    sessionKey: resolvedSessionKey,
     history: chatHistory,
-    activeTurnId,
     onComplete: () => {
-      void harnessHistoryQuery.refetch()
+      if (isAgentHarnessAgent) {
+        void harnessHistoryQuery.refetch()
+      }
+    },
+    onSessionKeyChange: (sessionKey) => {
+      setStreamSessionKey(sessionKey)
     },
-    onSessionKeyChange: () => {},
   })
-  const enqueueMessage = useEnqueueHarnessMessage()
-  const removeQueuedMessage = useRemoveHarnessQueuedMessage()
-
-  const handleStop = () => {
-    void cancelHarnessTurn(agentId, {
-      turnId: activeTurnId ?? undefined,
-      reason: 'user pressed stop',
-    })
-  }
   const visibleTurns = useMemo(
-    () => filterTurnsPersistedInHistory(turns, historyMessages),
-    [historyMessages, turns],
+    () =>
+      isAgentHarnessAgent
+        ? filterTurnsPersistedInHistory(turns, historyMessages)
+        : turns,
+    [historyMessages, isAgentHarnessAgent, turns],
   )
+  const outboundQueue = useOutboundQueue({
+    agentId,
+    sessionKey: resolvedSessionKey,
+    enabled: Boolean(agent) && !isAgentHarnessAgent,
+  })
   onInitialMessageConsumedRef.current = onInitialMessageConsumed
 
-  const disabled = !agent
+  // Refetch history whenever a server-dispatched queue item completes.
+  // The server worker streams the queued turn into OpenClaw directly, so
+  // the client never observes the live tokens — we only see the new
+  // assistant turn once the JSONL is updated. Watching the queue for
+  // any 'sending' item dropping out is the cleanest "turn finalized"
+  // signal we have without exposing per-turn SSE.
+  const previousSendingIdsRef = useRef<Set<string>>(new Set())
+  useEffect(() => {
+    if (isAgentHarnessAgent) return
+    const currentSending = new Set(
+      outboundQueue.queue
+        .filter((item) => item.status === 'sending')
+        .map((item) => item.id),
+    )
+    const dropped = [...previousSendingIdsRef.current].filter(
+      (id) => !currentSending.has(id),
+    )
+    previousSendingIdsRef.current = currentSending
+    if (dropped.length > 0) {
+      void clawHistoryQuery.refetch()
+    }
+  }, [clawHistoryQuery, isAgentHarnessAgent, outboundQueue.queue])
+
+  const disabled =
+    !agent || (!isAgentHarnessAgent && status?.status !== 'running')
+  // Two-part gate: cover both "still fetching" AND "just got enabled but
+  // hasn't started fetching yet". When `enabled` flips true (baseUrl
+  // resolves), there's a render frame where React Query reports
+  // isLoading=false but hasn't run the queryFn yet — `isFetched` is still
+  // false. Without this we render EmptyState during that one frame.
+  const isInitialLoading =
+    !isAgentHarnessAgent &&
+    (clawHistoryQuery.isLoading ||
+      (!clawHistoryQuery.isFetched && !clawHistoryQuery.isError))
+
   const historyReady =
-    harnessHistoryQuery.isFetched || harnessHistoryQuery.isError
+    (isAgentHarnessAgent &&
+      (harnessHistoryQuery.isFetched || harnessHistoryQuery.isError)) ||
+    (!isAgentHarnessAgent &&
+      (clawHistoryQuery.isFetched || clawHistoryQuery.isError))
   const initialMessageKey = initialMessage
     ? `${agentId}:${initialMessage}`
     : null
-  const error = harnessHistoryQuery.error ?? null
+  const error = isAgentHarnessAgent
+    ? (harnessHistoryQuery.error ?? null)
+    : (clawHistoryQuery.error ?? null)
 
+  const enqueueRef = useRef(outboundQueue.enqueue)
+  enqueueRef.current = outboundQueue.enqueue
   const sendRef = useRef(send)
   sendRef.current = send
 
@@ -281,8 +340,18 @@ function AgentConversationController({
 
     initialMessageSentRef.current = initialMessageKey
     onInitialMessageConsumedRef.current()
-    void sendRef.current({ text: query })
-  }, [disabled, historyReady, initialMessage, initialMessageKey])
+    if (isAgentHarnessAgent) {
+      void sendRef.current({ text: query })
+    } else {
+      enqueueRef.current({ text: query })
+    }
+  }, [
+    disabled,
+    historyReady,
+    initialMessage,
+    initialMessageKey,
+    isAgentHarnessAgent,
+  ])
 
   const handleSelectAgent = (entry: AgentEntry) => {
     navigate(`${agentPathPrefix}/${entry.agentId}`)
@@ -295,26 +364,32 @@ function AgentConversationController({
         historyMessages={historyMessages}
         turns={visibleTurns}
         streaming={streaming}
-        isInitialLoading={harnessHistoryQuery.isLoading}
+        isInitialLoading={
+          isAgentHarnessAgent ? harnessHistoryQuery.isLoading : isInitialLoading
+        }
         error={error}
-        hasNextPage={false}
-        isFetchingNextPage={false}
-        onFetchNextPage={() => {}}
+        hasNextPage={
+          isAgentHarnessAgent ? false : Boolean(clawHistoryQuery.hasNextPage)
+        }
+        isFetchingNextPage={
+          isAgentHarnessAgent ? false : clawHistoryQuery.isFetchingNextPage
+        }
+        onFetchNextPage={() => {
+          if (!isAgentHarnessAgent) {
+            void clawHistoryQuery.fetchNextPage()
+          }
+        }}
         onRetry={() => {
-          void harnessHistoryQuery.refetch()
+          if (isAgentHarnessAgent) {
+            void harnessHistoryQuery.refetch()
+          } else {
+            void clawHistoryQuery.refetch()
+          }
         }}
       />
 
       <div className="border-border/50 border-t bg-background/88 px-4 py-3 backdrop-blur-md">
-        <div className="mx-auto max-w-3xl space-y-3">
-          {queue.length > 0 ? (
-            <QueuePanel
-              queue={queue}
-              onRemove={(messageId) =>
-                removeQueuedMessage.mutate({ agentId, messageId })
-              }
-            />
-          ) : null}
+        <div className="mx-auto max-w-3xl">
           <ConversationInput
             variant="conversation"
             agents={agents}
@@ -329,30 +404,31 @@ function AgentConversationController({
                 name: a.name,
                 dataUrl: a.dataUrl,
               }))
-              // When the agent already has an in-flight turn, route
-              // the new message into the durable queue instead of
-              // starting a parallel turn. Drains automatically as
-              // soon as the active turn ends.
-              if (streaming || activeTurnId) {
-                enqueueMessage.mutate({
-                  agentId,
-                  message: input.text,
+              if (isAgentHarnessAgent) {
+                void send({ text: input.text, attachments, attachmentPreviews })
+              } else {
+                outboundQueue.enqueue({
+                  text: input.text,
                   attachments,
+                  attachmentPreviews,
+                  history: chatHistory,
                 })
-                return
               }
-              void send({ text: input.text, attachments, attachmentPreviews })
             }}
             onCreateAgent={() => navigate(createAgentPath)}
-            onStop={handleStop}
             streaming={streaming}
             disabled={disabled}
-            status="running"
-            attachmentsEnabled={true}
-            placeholder={
-              streaming
-                ? `Type to queue another message for ${agentName}...`
-                : `Message ${agentName}...`
+            status={isAgentHarnessAgent ? 'running' : status?.status}
+            attachmentsEnabled={!isAgentHarnessAgent}
+            placeholder={`Message ${agentName}...`}
+            outboundQueue={
+              isAgentHarnessAgent ? undefined : outboundQueue.queue
+            }
+            onCancelQueued={
+              isAgentHarnessAgent ? undefined : outboundQueue.cancel
+            }
+            onRetryQueued={
+              isAgentHarnessAgent ? undefined : outboundQueue.retry
             }
           />
         </div>
@@ -377,7 +453,7 @@ export const AgentCommandConversation: FC<AgentCommandConversationProps> = ({
   const { agentId } = useParams<{ agentId: string }>()
   const [searchParams, setSearchParams] = useSearchParams()
   const navigate = useNavigate()
-  const { agents } = useAgentCommandData()
+  const { status, agents } = useAgentCommandData()
   const shouldRedirectHome = !agentId
   const resolvedAgentId = agentId ?? ''
   const agent = agents.find((entry) => entry.agentId === resolvedAgentId)
@@ -395,11 +471,10 @@ export const AgentCommandConversation: FC<AgentCommandConversationProps> = ({
     navigate(`${agentPathPrefix}/${entry.agentId}`)
   }
 
-  // Every visible agent runs through the harness now, so per-agent
-  // runtime status doesn't gate chat the way OpenClaw's legacy
-  // gateway lifecycle did. Show "Ready" once the agent record is
-  // resolved from the rail, "Setup" otherwise.
-  const statusCopy = agent ? 'Ready' : 'Setup'
+  const statusCopy =
+    agent?.source === 'agent-harness'
+      ? 'Ready'
+      : getConversationStatusCopy(status?.status)
 
   return (
     <div className="absolute inset-0 overflow-hidden bg-background md:pl-[theme(spacing.14)]">
@@ -425,6 +500,7 @@ export const AgentCommandConversation: FC<AgentCommandConversationProps> = ({
           key={resolvedAgentId}
           agentId={resolvedAgentId}
           agents={agents}
+          status={status}
           initialMessage={initialMessage}
           onInitialMessageConsumed={() =>
             setSearchParams({}, { replace: true })

packages/browseros-agent/apps/agent/entrypoints/app/agent-command/AgentCommandHome.tsx

@@ -1,25 +1,18 @@
 import { Plus } from 'lucide-react'
-import { type FC, useEffect, useMemo, useState } from 'react'
+import { type FC, useEffect, useState } from 'react'
 import { useNavigate } from 'react-router'
 import { Button } from '@/components/ui/button'
 import { Card, CardContent } from '@/components/ui/card'
 import { Separator } from '@/components/ui/separator'
-import type {
-  HarnessAdapterDescriptor,
-  HarnessAgent,
-} from '@/entrypoints/app/agents/agent-harness-types'
-import {
-  useAgentAdapters,
-  useHarnessAgents,
-} from '@/entrypoints/app/agents/useAgents'
 import type { AgentEntry } from '@/entrypoints/app/agents/useOpenClaw'
 import { ImportDataHint } from '@/entrypoints/newtab/index/ImportDataHint'
 import { SignInHint } from '@/entrypoints/newtab/index/SignInHint'
 import { useActiveHint } from '@/entrypoints/newtab/index/useActiveHint'
+import type { AgentCardData } from '@/lib/agent-conversations/types'
 import { AgentCardDock } from './AgentCardDock'
 import { useAgentCommandData } from './agent-command-layout'
 import { ConversationInput } from './ConversationInput'
-import { orderHomeAgents } from './home-agent-card.helpers'
+import { buildAgentCardData } from './useAgentCardData'
 
 function EmptyAgentsState({ onOpenAgents }: { onOpenAgents: () => void }) {
   return (
@@ -45,13 +38,11 @@ function EmptyAgentsState({ onOpenAgents }: { onOpenAgents: () => void }) {
 function RecentThreads({
   activeAgentId,
   agents,
-  adapters,
   onOpenAgents,
   onSelectAgent,
 }: {
   activeAgentId?: string | null
-  agents: HarnessAgent[]
-  adapters: HarnessAdapterDescriptor[]
+  agents: AgentCardData[]
   onOpenAgents: () => void
   onSelectAgent: (agentId: string) => void
 }) {
@@ -77,7 +68,6 @@ function RecentThreads({
       </div>
       <AgentCardDock
         agents={agents}
-        adapters={adapters}
         activeAgentId={activeAgentId ?? undefined}
         onSelectAgent={onSelectAgent}
         onCreateAgent={onOpenAgents}
@@ -89,32 +79,25 @@ function RecentThreads({
 export const AgentCommandHome: FC = () => {
   const navigate = useNavigate()
   const activeHint = useActiveHint()
-  // The conversation input still consumes the merged AgentEntry list
-  // from the layout context (handles legacy /claw/agents entries that
-  // haven't yet been backfilled into the harness store). The Recent
-  // Agents grid below reads the richer harness payload directly.
-  const { agents: legacyAgents, status } = useAgentCommandData()
-  const { harnessAgents } = useHarnessAgents()
-  const { adapters } = useAgentAdapters()
+  const { agents, status } = useAgentCommandData()
   const [selectedAgentId, setSelectedAgentId] = useState<string | null>(null)
-
-  const orderedAgents = useMemo(
-    () => orderHomeAgents(harnessAgents),
-    [harnessAgents],
-  )
+  const cardData = buildAgentCardData(agents, status?.status, undefined)
 
   useEffect(() => {
-    if (legacyAgents.length === 0) {
-      if (selectedAgentId) setSelectedAgentId(null)
+    if (agents.length === 0) {
+      if (selectedAgentId) {
+        setSelectedAgentId(null)
+      }
       return
     }
+
     if (
       !selectedAgentId ||
-      !legacyAgents.some((agent) => agent.agentId === selectedAgentId)
+      !agents.some((agent) => agent.agentId === selectedAgentId)
     ) {
-      setSelectedAgentId(legacyAgents[0].agentId)
+      setSelectedAgentId(agents[0].agentId)
     }
-  }, [legacyAgents, selectedAgentId])
+  }, [agents, selectedAgentId])
 
   const handleSend = (input: { text: string }) => {
     if (!selectedAgentId) return
@@ -127,7 +110,7 @@ export const AgentCommandHome: FC = () => {
     setSelectedAgentId(agent.agentId)
   }
 
-  const selectedAgent = legacyAgents.find(
+  const selectedAgent = agents.find(
     (agent) => agent.agentId === selectedAgentId,
   )
   const selectedAgentReady = selectedAgent
@@ -135,15 +118,13 @@ export const AgentCommandHome: FC = () => {
     : false
   const selectedAgentStatus =
     selectedAgent?.source === 'agent-harness' ? 'running' : status?.status
-  const selectedAgentName =
-    selectedAgent?.name ?? orderedAgents[0]?.name ?? 'your agent'
-
-  const hasAgents = legacyAgents.length > 0
+  const selectedCard =
+    cardData.find((agent) => agent.agentId === selectedAgentId) ?? cardData[0]
 
   return (
     <div className="min-h-full px-4 py-6">
       <div className="mx-auto flex w-full max-w-5xl flex-col gap-8">
-        {hasAgents ? (
+        {cardData.length > 0 ? (
           <>
             <div className="flex flex-col items-center gap-5 pt-[max(10vh,24px)] text-center">
               <div className="space-y-3">
@@ -159,7 +140,7 @@ export const AgentCommandHome: FC = () => {
               <div className="w-full max-w-3xl">
                 <ConversationInput
                   variant="home"
-                  agents={legacyAgents}
+                  agents={agents}
                   selectedAgentId={selectedAgentId}
                   onSelectAgent={handleSelectAgent}
                   onSend={handleSend}
@@ -170,7 +151,7 @@ export const AgentCommandHome: FC = () => {
                   attachmentsEnabled={false}
                   placeholder={
                     selectedAgentReady
-                      ? `Ask ${selectedAgentName} to handle a task...`
+                      ? `Ask ${selectedCard?.name ?? 'your agent'} to handle a task...`
                       : 'Agent runtime is not running...'
                   }
                 />
@@ -181,8 +162,7 @@ export const AgentCommandHome: FC = () => {
 
             <RecentThreads
               activeAgentId={selectedAgentId}
-              agents={orderedAgents}
-              adapters={adapters}
+              agents={cardData}
               onOpenAgents={() => navigate('/agents')}
               onSelectAgent={(agentId) => navigate(`/home/agents/${agentId}`)}
             />

packages/browseros-agent/apps/agent/entrypoints/app/agent-command/ConversationInput.tsx

@@ -1,4 +1,5 @@
 import {
+  AlertTriangle,
   ArrowRight,
   Bot,
   ChevronDown,
@@ -8,6 +9,7 @@ import {
   Loader2,
   Mic,
   Paperclip,
+  RefreshCw,
   Square,
   X,
 } from 'lucide-react'
@@ -36,6 +38,7 @@ import { cn } from '@/lib/utils'
 import { useVoiceInput } from '@/lib/voice/useVoiceInput'
 import { useWorkspace } from '@/lib/workspace/use-workspace'
 import { AgentSelector } from './AgentSelector'
+import type { OutboundMessage } from './useOutboundQueue'
 
 export interface ConversationInputSendInput {
   text: string
@@ -54,40 +57,34 @@ interface ConversationInputProps {
   placeholder?: string
   attachmentsEnabled?: boolean
   variant?: 'home' | 'conversation'
-  /**
-   * When set, a Stop button surfaces to the left of the voice mic
-   * while `streaming === true`. Click cancels the active turn
-   * server-side via the chat-cancel endpoint. Absent → no Stop
-   * button (legacy behaviour for the home composer).
-   */
-  onStop?: () => void
+  // Outbound queue: when present, the composer renders the queue strip
+  // above the textarea and lets the user keep sending while a previous
+  // turn is in flight. Optional so non-conversation variants (the home
+  // page) can opt out — the queue only makes sense in the conversation
+  // page where each enqueued message will eventually be delivered to the
+  // active agent.
+  outboundQueue?: OutboundMessage[]
+  onCancelQueued?: (id: string) => void
+  onRetryQueued?: (id: string) => void
 }
 
 function InputActionButton({
   disabled,
   onClick,
   streaming,
-  hasContent,
 }: {
   disabled: boolean
   onClick: () => void
   streaming: boolean
-  hasContent: boolean
 }) {
-  // Show the spinner while streaming only when there's nothing to
-  // send — once the user types something, the icon flips back to the
-  // paper-plane so it reads as "queue this message" instead of
-  // "still working".
-  const showSpinner = streaming && !hasContent
   return (
     <Button
       onClick={onClick}
       size="icon"
       disabled={disabled}
-      title={streaming && hasContent ? 'Queue message' : undefined}
       className="h-10 w-10 flex-shrink-0 rounded-xl bg-primary text-primary-foreground hover:bg-primary/90"
     >
-      {showSpinner ? (
+      {streaming ? (
         <Loader2 className="h-5 w-5 animate-spin" />
       ) : (
         <ArrowRight className="h-5 w-5" />
@@ -96,22 +93,6 @@ function InputActionButton({
   )
 }
 
-function StopButton({ onStop }: { onStop: () => void }) {
-  return (
-    <Button
-      type="button"
-      size="icon"
-      variant="ghost"
-      onClick={onStop}
-      title="Stop current turn — queued messages will start next."
-      aria-label="Stop current turn"
-      className="h-8 w-8 flex-shrink-0 rounded-lg bg-destructive/10 text-destructive transition-colors hover:bg-destructive/15 hover:text-destructive"
-    >
-      <Square className="h-3.5 w-3.5 fill-current" />
-    </Button>
-  )
-}
-
 function VoiceButton({
   isRecording,
   isTranscribing,
@@ -330,7 +311,9 @@ export const ConversationInput: FC<ConversationInputProps> = ({
   placeholder,
   attachmentsEnabled = true,
   variant = 'conversation',
-  onStop,
+  outboundQueue,
+  onCancelQueued,
+  onRetryQueued,
 }) => {
   const [input, setInput] = useState('')
   const [selectedTabs, setSelectedTabs] = useState<chrome.tabs.Tab[]>([])
@@ -411,17 +394,15 @@ export const ConversationInput: FC<ConversationInputProps> = ({
   }
 
   const hasContent = input.trim().length > 0 || attachments.length > 0
-  // Queue-aware composers (the conversation panel passes `onStop`)
-  // accept input while streaming — the parent decides whether the
-  // submission opens a new turn or enqueues onto the active one.
-  // Surfaces without a Stop hook (home) keep the legacy behaviour
-  // and block input until the current turn finishes.
-  const queueAware = Boolean(onStop)
+  const queueEnabled = outboundQueue !== undefined
 
   const handleSend = () => {
     const text = input.trim()
+    // The outbound queue accepts new messages while streaming; legacy
+    // direct-send callers (e.g., the home composer) keep the original
+    // streaming-blocks-send semantic.
     if (disabled || isStaging) return
-    if (streaming && !queueAware) return
+    if (!queueEnabled && streaming) return
     if (!text && attachments.length === 0) return
     onSend({ text, attachments })
     setInput('')
@@ -513,6 +494,13 @@ export const ConversationInput: FC<ConversationInputProps> = ({
             error={attachmentError}
           />
         ) : null}
+        {queueEnabled && outboundQueue && outboundQueue.length > 0 ? (
+          <OutboundQueueStrip
+            messages={outboundQueue}
+            onCancel={onCancelQueued}
+            onRetry={onRetryQueued}
+          />
+        ) : null}
         <div
           className={cn(
             'flex gap-3',
@@ -551,7 +539,6 @@ export const ConversationInput: FC<ConversationInputProps> = ({
               )}
             />
           </div>
-          {streaming && onStop ? <StopButton onStop={onStop} /> : null}
           <VoiceButton
             isRecording={voice.isRecording}
             isTranscribing={voice.isTranscribing}
@@ -569,13 +556,15 @@ export const ConversationInput: FC<ConversationInputProps> = ({
               !!disabled ||
               voice.isRecording ||
               voice.isTranscribing ||
-              (streaming && !queueAware)
+              // Only block on `streaming` for the legacy direct-send path
+              // (no queue). With the queue active the press always
+              // succeeds — it just enqueues instead of dispatching.
+              (!queueEnabled && streaming)
             }
             onClick={handleSend}
             // Spinner stays the user-facing "agent is busy" hint; with the
             // queue active we still spin while a turn is in flight.
             streaming={streaming}
-            hasContent={hasContent}
           />
         </div>
         {voice.error ? (
@@ -606,6 +595,117 @@ export const ConversationInput: FC<ConversationInputProps> = ({
   )
 }
 
+function OutboundQueueStrip({
+  messages,
+  onCancel,
+  onRetry,
+}: {
+  messages: OutboundMessage[]
+  onCancel?: (id: string) => void
+  onRetry?: (id: string) => void
+}) {
+  return (
+    <div className="border-border/40 border-b px-4 pt-3 pb-2">
+      <ul className="flex flex-col gap-1">
+        {messages.map((message) => (
+          <OutboundQueueItem
+            key={message.id}
+            message={message}
+            onCancel={onCancel}
+            onRetry={onRetry}
+          />
+        ))}
+      </ul>
+    </div>
+  )
+}
+
+function OutboundQueueItem({
+  message,
+  onCancel,
+  onRetry,
+}: {
+  message: OutboundMessage
+  onCancel?: (id: string) => void
+  onRetry?: (id: string) => void
+}) {
+  const preview = message.text.trim() || '(attachments only)'
+  return (
+    <li className="flex items-center gap-2 rounded-md px-2 py-1 text-xs">
+      <OutboundQueueStatusIcon status={message.status} />
+      <span className="min-w-0 flex-1 truncate text-muted-foreground">
+        {preview}
+      </span>
+      {message.attachmentPreviews.length > 0 ? (
+        <span className="inline-flex items-center gap-1 text-muted-foreground/70">
+          <Paperclip className="size-3" />
+          <span className="tabular-nums">
+            {message.attachmentPreviews.length}
+          </span>
+        </span>
+      ) : null}
+      {message.status === 'queued' && onCancel ? (
+        <button
+          type="button"
+          onClick={() => onCancel(message.id)}
+          className="ml-1 inline-flex size-5 items-center justify-center rounded-full text-muted-foreground hover:bg-accent hover:text-foreground"
+          aria-label="Cancel queued message"
+          title="Cancel"
+        >
+          <X className="size-3" />
+        </button>
+      ) : null}
+      {message.status === 'failed' ? (
+        <span className="ml-1 inline-flex items-center gap-2 text-destructive">
+          <span className="max-w-[160px] truncate" title={message.error}>
+            {message.error ?? 'Failed'}
+          </span>
+          {onRetry ? (
+            <button
+              type="button"
+              onClick={() => onRetry(message.id)}
+              className="inline-flex size-5 items-center justify-center rounded-full hover:bg-accent hover:text-foreground"
+              aria-label="Retry failed message"
+              title="Retry"
+            >
+              <RefreshCw className="size-3" />
+            </button>
+          ) : null}
+          {onCancel ? (
+            <button
+              type="button"
+              onClick={() => onCancel(message.id)}
+              className="inline-flex size-5 items-center justify-center rounded-full hover:bg-accent hover:text-foreground"
+              aria-label="Discard failed message"
+              title="Discard"
+            >
+              <X className="size-3" />
+            </button>
+          ) : null}
+        </span>
+      ) : null}
+    </li>
+  )
+}
+
+function OutboundQueueStatusIcon({
+  status,
+}: {
+  status: OutboundMessage['status']
+}) {
+  if (status === 'sending') {
+    return (
+      <Loader2 className="size-3.5 shrink-0 animate-spin text-muted-foreground" />
+    )
+  }
+  if (status === 'failed') {
+    return <AlertTriangle className="size-3.5 shrink-0 text-destructive" />
+  }
+  return (
+    <span className="inline-block size-2 shrink-0 rounded-full bg-muted-foreground/40" />
+  )
+}
+
 function AttachmentStrip({
   attachments,
   onRemove,

packages/browseros-agent/apps/agent/entrypoints/app/agent-command/HomeAgentCard.tsx

@@ -1,243 +0,0 @@
-import { Quote, TriangleAlert } from 'lucide-react'
-import type { FC } from 'react'
-import { Badge } from '@/components/ui/badge'
-import {
-  HoverCard,
-  HoverCardContent,
-  HoverCardTrigger,
-} from '@/components/ui/hover-card'
-import { adapterLabel } from '@/entrypoints/app/agents/AdapterIcon'
-import { formatRelativeTime } from '@/entrypoints/app/agents/agent-display.helpers'
-import type {
-  HarnessAdapterHealth,
-  HarnessAgent,
-  HarnessAgentAdapter,
-} from '@/entrypoints/app/agents/agent-harness-types'
-import { AgentTile } from '@/entrypoints/app/agents/agent-row/AgentTile'
-import {
-  firstNonBlankLine,
-  truncate,
-} from '@/entrypoints/app/agents/agent-row/agent-row.helpers'
-import type { AgentLiveness } from '@/entrypoints/app/agents/LivenessDot'
-import { cn } from '@/lib/utils'
-
-interface HomeAgentCardProps {
-  agent: HarnessAgent
-  adapter: HarnessAgentAdapter | 'unknown'
-  /** Per-adapter health snapshot, shared across cards rendering the
-   *  same adapter. `null` when the /adapters response hasn't surfaced
-   *  health yet (we treat that as healthy until proven otherwise). */
-  adapterHealth: HarnessAdapterHealth | null
-  /** Highlights the card with an accent ring; tells the user which
-   *  agent the conversation input is bound to. */
-  active?: boolean
-  onClick: () => void
-}
-
-const PREVIEW_CHARS = 100
-
-/**
- * Grid-shaped card for the /home Recent agents section. Composition
- * mirrors the rail's `AgentRowCard` but the layout is a vertical
- * column sized for a 1/3-width tile rather than a full-width row.
- *
- * Reuses `<AgentTile>`, `<LivenessDot>`, `livenessDetail`,
- * `formatRelativeTime`, `firstNonBlankLine`, `truncate`, and the
- * inline `Unavailable` chip pattern so the visual language is
- * continuous between rail and grid.
- */
-export const HomeAgentCard: FC<HomeAgentCardProps> = ({
-  agent,
-  adapter,
-  adapterHealth,
-  active,
-  onClick,
-}) => {
-  const status = agent.status ?? 'unknown'
-  const lastUsedAt = agent.lastUsedAt ?? null
-  const isWorking = status === 'working'
-  const isAsleep = status === 'asleep'
-  const isError = status === 'error'
-  const hasActiveTurn = Boolean(agent.activeTurnId)
-
-  return (
-    <button
-      type="button"
-      onClick={onClick}
-      className={cn(
-        'group flex min-h-32 w-full min-w-0 flex-col rounded-2xl border bg-card p-4 text-left shadow-sm transition-colors',
-        active && 'ring-1 ring-[var(--accent-orange)]/30',
-        isWorking
-          ? 'border-[var(--accent-orange)]/40'
-          : isError
-            ? 'border-destructive/30'
-            : 'border-border/60 hover:border-[var(--accent-orange)]/30',
-      )}
-    >
-      <div className="flex items-start gap-3">
-        <AgentTile adapter={adapter} status={status} lastUsedAt={lastUsedAt} />
-        <div className="min-w-0 flex-1">
-          <div className="flex items-center gap-1.5">
-            <span className="truncate font-semibold text-sm">
-              {displayName(agent)}
-            </span>
-            {isWorking && (
-              <Badge
-                variant="secondary"
-                className="ml-auto bg-amber-50 text-amber-900 hover:bg-amber-50"
-              >
-                Working
-              </Badge>
-            )}
-          </div>
-          <SummaryLine
-            adapter={adapter}
-            modelId={agent.modelId ?? null}
-            reasoningEffort={agent.reasoningEffort ?? null}
-            adapterHealth={adapterHealth}
-          />
-        </div>
-      </div>
-
-      <LastMessage message={agent.lastUserMessage ?? null} />
-
-      <div className="mt-3 flex items-center justify-between gap-2 text-muted-foreground text-xs">
-        <span>{statusFootnote(status, lastUsedAt)}</span>
-        {hasActiveTurn ? (
-          <ResumeChip />
-        ) : isAsleep ? (
-          <Badge variant="outline" className="text-muted-foreground">
-            Asleep
-          </Badge>
-        ) : isError ? (
-          <ErrorChip lastError={agent.lastError ?? null} />
-        ) : null}
-      </div>
-    </button>
-  )
-}
-
-const SummaryLine: FC<{
-  adapter: HarnessAgentAdapter | 'unknown'
-  modelId: string | null
-  reasoningEffort: string | null
-  adapterHealth: HarnessAdapterHealth | null
-}> = ({ adapter, modelId, reasoningEffort, adapterHealth }) => {
-  const parts = [adapterLabel(adapter)]
-  if (modelId) parts.push(modelId)
-  if (reasoningEffort) parts.push(reasoningEffort)
-  const unhealthy = adapterHealth?.healthy === false
-  return (
-    <div
-      className={cn(
-        'mt-0.5 flex items-center gap-1.5 text-muted-foreground text-xs',
-        unhealthy && 'text-muted-foreground/70',
-      )}
-    >
-      <span className="truncate">{parts.join(' · ')}</span>
-      {unhealthy && (
-        <HoverCard openDelay={200}>
-          <HoverCardTrigger asChild>
-            <Badge
-              variant="outline"
-              className="h-5 cursor-default gap-1 border-amber-500/40 bg-amber-50 px-1.5 text-amber-900 hover:bg-amber-50"
-            >
-              <TriangleAlert className="size-2.5" />
-              <span className="font-normal">Unavailable</span>
-            </Badge>
-          </HoverCardTrigger>
-          <HoverCardContent side="right" className="w-72 text-sm">
-            <div className="font-medium">
-              {adapterLabel(adapter)} CLI not available
-            </div>
-            <div className="mt-1 text-muted-foreground text-xs">
-              {adapterHealth?.reason ??
-                'Adapter binary missing on $PATH. Install it from the adapter docs to use this agent.'}
-            </div>
-          </HoverCardContent>
-        </HoverCard>
-      )}
-    </div>
-  )
-}
-
-const LastMessage: FC<{ message: string | null }> = ({ message }) => {
-  if (!message) {
-    return (
-      <p className="mt-3 flex-1 text-muted-foreground/70 text-xs italic">
-        No messages yet — start a chat
-      </p>
-    )
-  }
-  return (
-    <p className="mt-3 line-clamp-2 flex flex-1 items-start gap-1.5 text-foreground/85 text-sm italic leading-snug">
-      <Quote
-        className="mt-1 size-3 shrink-0 text-muted-foreground/60"
-        aria-hidden
-      />
-      <span className="line-clamp-2">
-        {truncate(firstNonBlankLine(message), PREVIEW_CHARS)}
-      </span>
-    </p>
-  )
-}
-
-const ResumeChip: FC = () => (
-  <span className="inline-flex items-center gap-1.5 rounded-full bg-[var(--accent-orange)] px-2.5 py-0.5 font-medium text-[11px] text-white shadow-sm">
-    <span className="relative flex size-1.5">
-      <span className="absolute inline-flex h-full w-full animate-ping rounded-full bg-white/70 opacity-75" />
-      <span className="relative inline-flex size-1.5 rounded-full bg-white" />
-    </span>
-    Resume
-  </span>
-)
-
-const ErrorChip: FC<{ lastError: string | null }> = ({ lastError }) => {
-  if (!lastError) {
-    return <Badge variant="destructive">Attention</Badge>
-  }
-  return (
-    <HoverCard openDelay={200}>
-      <HoverCardTrigger asChild>
-        <Badge variant="destructive" className="cursor-default">
-          Attention
-        </Badge>
-      </HoverCardTrigger>
-      <HoverCardContent
-        side="left"
-        className="max-w-xs whitespace-pre-wrap font-mono text-xs"
-      >
-        {lastError}
-      </HoverCardContent>
-    </HoverCard>
-  )
-}
-
-/**
- * Footer left side: relative time on every state EXCEPT working,
- * which shows `now` (the dot is already pulsing — restating it as
- * "Working" would duplicate the pill in the title row).
- */
-function statusFootnote(
-  status: AgentLiveness,
-  lastUsedAt: number | null,
-): string {
-  if (status === 'working') return 'now'
-  return formatRelativeTime(lastUsedAt)
-}
-
-const UUID_PATTERN =
-  /^[0-9a-f]{8}-[0-9a-f]{4}-[0-9a-f]{4}-[0-9a-f]{4}-[0-9a-f]{12}$/i
-const OC_UUID_PATTERN =
-  /^oc-[0-9a-f]{8}-[0-9a-f]{4}-[0-9a-f]{4}-[0-9a-f]{4}-[0-9a-f]{12}$/i
-
-function displayName(agent: HarnessAgent): string {
-  const name = agent.name?.trim()
-  const id = agent.id
-  if (!name || name === id) {
-    if (OC_UUID_PATTERN.test(id)) return id.slice(0, 11)
-    if (UUID_PATTERN.test(id)) return id.slice(0, 8)
-    return id
-  }
-  return name
-}

packages/browseros-agent/apps/agent/entrypoints/app/agent-command/QueuePanel.tsx

@@ -1,94 +0,0 @@
-import { ListPlus, X } from 'lucide-react'
-import type { FC } from 'react'
-import {
-  Queue,
-  QueueItem,
-  QueueItemAction,
-  QueueItemActions,
-  QueueItemAttachment,
-  QueueItemContent,
-  QueueItemFile,
-  QueueItemImage,
-  QueueList,
-  QueueSection,
-  QueueSectionContent,
-  QueueSectionLabel,
-  QueueSectionTrigger,
-} from '@/components/ai-elements/queue'
-import type {
-  HarnessQueuedMessage,
-  HarnessQueuedMessageAttachment,
-} from '@/entrypoints/app/agents/agent-harness-types'
-import { firstNonBlankLine } from '@/entrypoints/app/agents/agent-row/agent-row.helpers'
-
-interface QueuePanelProps {
-  queue: HarnessQueuedMessage[]
-  onRemove: (messageId: string) => void
-}
-
-/**
- * Renders the agent's pending message queue using the shared AI
- * Elements `Queue` primitives. Caller is expected to gate render on
- * `queue.length > 0` — when empty, this returns null so the panel
- * disappears cleanly between turns.
- */
-export const QueuePanel: FC<QueuePanelProps> = ({ queue, onRemove }) => {
-  if (queue.length === 0) return null
-  return (
-    <Queue>
-      <QueueSection>
-        <QueueSectionTrigger>
-          <QueueSectionLabel
-            count={queue.length}
-            label={queue.length === 1 ? 'queued message' : 'queued messages'}
-            icon={<ListPlus className="size-3.5" />}
-          />
-        </QueueSectionTrigger>
-        <QueueSectionContent>
-          <QueueList>
-            {queue.map((entry) => (
-              <QueueItem key={entry.id}>
-                <div className="flex items-center gap-2">
-                  <QueueItemContent>
-                    {firstNonBlankLine(entry.message)}
-                  </QueueItemContent>
-                  <QueueItemActions>
-                    <QueueItemAction
-                      aria-label="Remove from queue"
-                      onClick={() => onRemove(entry.id)}
-                    >
-                      <X className="size-3" />
-                    </QueueItemAction>
-                  </QueueItemActions>
-                </div>
-                {entry.attachments && entry.attachments.length > 0 ? (
-                  <QueueItemAttachment>
-                    {entry.attachments.map((attachment, idx) =>
-                      renderAttachment(entry.id, attachment, idx),
-                    )}
-                  </QueueItemAttachment>
-                ) : null}
-              </QueueItem>
-            ))}
-          </QueueList>
-        </QueueSectionContent>
-      </QueueSection>
-    </Queue>
-  )
-}
-
-function renderAttachment(
-  messageId: string,
-  attachment: HarnessQueuedMessageAttachment,
-  idx: number,
-) {
-  if (attachment.mediaType.startsWith('image/')) {
-    const src = `data:${attachment.mediaType};base64,${attachment.data}`
-    return <QueueItemImage key={`${messageId}-${idx}`} src={src} />
-  }
-  return (
-    <QueueItemFile key={`${messageId}-${idx}`}>
-      {attachment.mediaType}
-    </QueueItemFile>
-  )
-}

packages/browseros-agent/apps/agent/entrypoints/app/agent-command/agent-command-layout.tsx

@@ -26,15 +26,7 @@ export const AgentCommandLayout: FC = () => {
   const { agents: harnessAgents, loading: harnessAgentsLoading } =
     useHarnessAgents()
   const visibleOpenClawAgents = openClawEnabled ? openClawAgents : []
-  // Dual-created OpenClaw agents appear in both `/claw/agents` (gateway
-  // record) and `/agents` (harness record) under the same id. Prefer the
-  // harness entry so the chat panel can route through the harness path
-  // and the rail doesn't show duplicates.
-  const harnessAgentIds = new Set(harnessAgents.map((entry) => entry.agentId))
-  const dedupedOpenClawAgents = visibleOpenClawAgents.filter(
-    (entry) => !harnessAgentIds.has(entry.agentId),
-  )
-  const agents = [...dedupedOpenClawAgents, ...harnessAgents]
+  const agents = [...visibleOpenClawAgents, ...harnessAgents]
 
   return (
     <Outlet

packages/browseros-agent/apps/agent/entrypoints/app/agent-command/claw-chat-types.ts

@@ -23,9 +23,9 @@ export interface BrowserOSChatHistoryToolCall {
   toolName: string
   label: string
   subject?: string
-  status: 'pending' | 'running' | 'completed' | 'failed'
-  input?: unknown
-  output?: unknown
+  status: 'completed' | 'failed'
+  input?: Record<string, unknown>
+  output?: string
   error?: string
   durationMs?: number
 }

packages/browseros-agent/apps/agent/entrypoints/app/agent-command/harness-history-mapper.ts

@@ -1,71 +0,0 @@
-import { buildToolLabel } from '../../../lib/tool-labels'
-import type { HarnessAgentHistoryPage } from '../agents/agent-harness-types'
-import type {
-  AgentHistoryPageResponse,
-  BrowserOSChatHistoryItem,
-  BrowserOSChatHistoryToolCall,
-} from './claw-chat-types'
-
-export function mapHarnessHistoryPage(
-  page: HarnessAgentHistoryPage,
-): AgentHistoryPageResponse {
-  const items: BrowserOSChatHistoryItem[] = page.items.map((item, index) => {
-    const toolCalls = item.toolCalls?.map(
-      (tool): BrowserOSChatHistoryToolCall => {
-        const input = asRecord(tool.input)
-        const { label, subject } = buildToolLabel(tool.toolName, input)
-        return {
-          toolName: tool.toolName,
-          label,
-          status: tool.status,
-          ...(tool.toolCallId ? { toolCallId: tool.toolCallId } : {}),
-          ...(subject ? { subject } : {}),
-          ...(tool.input !== undefined ? { input: tool.input } : {}),
-          ...(tool.output !== undefined ? { output: tool.output } : {}),
-          ...(tool.error ? { error: tool.error } : {}),
-          ...(tool.durationMs != null ? { durationMs: tool.durationMs } : {}),
-        }
-      },
-    )
-
-    return {
-      id: item.id,
-      role: item.role,
-      text: item.text,
-      timestamp: item.createdAt,
-      messageSeq: index + 1,
-      sessionKey: 'main',
-      source: 'user-chat',
-      ...(item.reasoning ? { reasoning: item.reasoning } : {}),
-      ...(toolCalls && toolCalls.length > 0 ? { toolCalls } : {}),
-    }
-  })
-  const updatedAt =
-    page.items.length > 0
-      ? Math.max(...page.items.map((item) => item.createdAt))
-      : Date.now()
-
-  return {
-    agentId: page.agentId,
-    sessionKey: 'main',
-    session: {
-      key: 'main',
-      updatedAt,
-      sessionId: 'main',
-      agentId: page.agentId,
-      kind: 'agent-harness',
-      source: 'user-chat',
-    },
-    items,
-    page: {
-      hasMore: false,
-      limit: items.length,
-    },
-  }
-}
-
-function asRecord(value: unknown): Record<string, unknown> | undefined {
-  return value && typeof value === 'object' && !Array.isArray(value)
-    ? (value as Record<string, unknown>)
-    : undefined
-}

packages/browseros-agent/apps/agent/entrypoints/app/agent-command/home-agent-card.helpers.test.ts

@@ -1,69 +0,0 @@
-import { describe, expect, it } from 'bun:test'
-import type { HarnessAgent } from '@/entrypoints/app/agents/agent-harness-types'
-import { orderHomeAgents } from './home-agent-card.helpers'
-
-function agent(overrides: Partial<HarnessAgent>): HarnessAgent {
-  return {
-    id: overrides.id ?? 'agent-x',
-    name: overrides.name ?? overrides.id ?? 'agent-x',
-    adapter: overrides.adapter ?? 'codex',
-    permissionMode: 'approve-all',
-    sessionKey: `agent:${overrides.id ?? 'agent-x'}:main`,
-    createdAt: 1000,
-    updatedAt: 1000,
-    ...overrides,
-  }
-}
-
-describe('orderHomeAgents', () => {
-  it('places active-turn agents before everyone else', () => {
-    const sorted = orderHomeAgents([
-      agent({ id: 'a', lastUsedAt: 5000 }),
-      agent({ id: 'b', lastUsedAt: 9000, activeTurnId: 'turn-1' }),
-      agent({ id: 'c', lastUsedAt: 7000 }),
-    ])
-    expect(sorted.map((a) => a.id)).toEqual(['b', 'c', 'a'])
-  })
-
-  it('orders non-active agents by lastUsedAt desc', () => {
-    const sorted = orderHomeAgents([
-      agent({ id: 'old', lastUsedAt: 1000 }),
-      agent({ id: 'new', lastUsedAt: 9000 }),
-      agent({ id: 'mid', lastUsedAt: 5000 }),
-    ])
-    expect(sorted.map((a) => a.id)).toEqual(['new', 'mid', 'old'])
-  })
-
-  it('puts the gateway `main` seed agent above other never-used agents', () => {
-    const sorted = orderHomeAgents([
-      agent({ id: 'oc-aaaaaa', lastUsedAt: null }),
-      agent({ id: 'main', lastUsedAt: null }),
-      agent({ id: 'oc-bbbbbb', lastUsedAt: null }),
-    ])
-    expect(sorted.map((a) => a.id)).toEqual(['main', 'oc-aaaaaa', 'oc-bbbbbb'])
-  })
-
-  it('sends never-used agents to the bottom even when `main` is among them', () => {
-    const sorted = orderHomeAgents([
-      agent({ id: 'main', lastUsedAt: null }),
-      agent({ id: 'used', lastUsedAt: 5000 }),
-    ])
-    expect(sorted.map((a) => a.id)).toEqual(['used', 'main'])
-  })
-
-  it('does NOT sort by pinned — pinned agents are treated like any other', () => {
-    const sorted = orderHomeAgents([
-      agent({ id: 'unpinned-recent', lastUsedAt: 9000, pinned: false }),
-      agent({ id: 'pinned-old', lastUsedAt: 1000, pinned: true }),
-    ])
-    expect(sorted.map((a) => a.id)).toEqual(['unpinned-recent', 'pinned-old'])
-  })
-
-  it('falls back to id-stable ordering when lastUsedAt ties', () => {
-    const sorted = orderHomeAgents([
-      agent({ id: 'b', lastUsedAt: 5000 }),
-      agent({ id: 'a', lastUsedAt: 5000 }),
-    ])
-    expect(sorted.map((a) => a.id)).toEqual(['a', 'b'])
-  })
-})

packages/browseros-agent/apps/agent/entrypoints/app/agent-command/home-agent-card.helpers.ts

@@ -1,42 +0,0 @@
-import type { HarnessAgent } from '@/entrypoints/app/agents/agent-harness-types'
-
-/**
- * Order for the /home Recent agents grid.
- *
- * 1. Active turn first — agents mid-turn float to the top so the
- *    Resume affordance is the first thing the user sees on /home.
- * 2. The protected gateway-side `main` agent stays pinned-to-top in
- *    the never-used group on a fresh install (mirrors the rail).
- * 3. Recency (`lastUsedAt` desc).
- * 4. `id` tiebreaker for stability so the grid doesn't reshuffle on
- *    every 5-second poll.
- *
- * Pin is NOT a sort key. The home grid is action-oriented and trusts
- * recency + active-turn to surface the right agent; pinning is an
- * organisation tool that lives on the rail at /agents.
- */
-export function orderHomeAgents(agents: HarnessAgent[]): HarnessAgent[] {
-  return [...agents].sort((a, b) => {
-    const aActive = a.activeTurnId != null
-    const bActive = b.activeTurnId != null
-    if (aActive !== bActive) return aActive ? -1 : 1
-
-    // Recency wins outright. Never-used agents (`lastUsedAt == null`)
-    // both fall to the same `-Infinity` bucket and the seed/id rules
-    // below decide their order — but a used agent always beats any
-    // never-used agent regardless of id.
-    const aValue = a.lastUsedAt ?? Number.NEGATIVE_INFINITY
-    const bValue = b.lastUsedAt ?? Number.NEGATIVE_INFINITY
-    if (aValue !== bValue) return bValue - aValue
-
-    // Inside the never-used (or exact-tie) group: pin the gateway
-    // `main` seed to the top of the group on a fresh install, then
-    // fall back to id-stable order so the grid doesn't reshuffle on
-    // every poll.
-    const aSeed = a.id === 'main' && a.lastUsedAt == null
-    const bSeed = b.id === 'main' && b.lastUsedAt == null
-    if (aSeed !== bSeed) return aSeed ? -1 : 1
-
-    return a.id.localeCompare(b.id)
-  })
-}

packages/browseros-agent/apps/agent/entrypoints/app/agent-command/useAgentCardData.ts

@@ -0,0 +1,53 @@
+import {
+  type AgentEntry,
+  getModelDisplayName,
+  type OpenClawStatus,
+} from '@/entrypoints/app/agents/useOpenClaw'
+import type { AgentCardData } from '@/lib/agent-conversations/types'
+import type { AgentOverview } from './useAgentDashboard'
+
+function resolveAgentStatus(
+  gatewayStatus: OpenClawStatus['status'] | undefined,
+  liveStatus: AgentOverview['status'] | undefined,
+): AgentCardData['status'] {
+  // Gateway-level errors take precedence
+  if (gatewayStatus === 'error') return 'error'
+  if (gatewayStatus === 'starting') return 'working'
+
+  // Per-agent live status from the WS observer
+  if (liveStatus === 'working') return 'working'
+  if (liveStatus === 'error') return 'error'
+
+  return 'idle'
+}
+
+/**
+ * Build agent card display data by merging the raw agent entries from
+ * the gateway with enriched overview data from the dashboard API.
+ *
+ * Pure function — no hooks, no IndexedDB, no async.
+ */
+export function buildAgentCardData(
+  agents: AgentEntry[],
+  status: OpenClawStatus['status'] | undefined,
+  dashboard: AgentOverview[] | undefined,
+): AgentCardData[] {
+  return agents.map((agent) => {
+    const overview = dashboard?.find((d) => d.agentId === agent.agentId)
+
+    return {
+      agentId: agent.agentId,
+      name: agent.name,
+      model: getModelDisplayName(agent.model),
+      status:
+        agent.source === 'agent-harness'
+          ? 'idle'
+          : resolveAgentStatus(status, overview?.status),
+      lastMessage: overview?.latestMessage?.slice(0, 200) ?? undefined,
+      lastMessageTimestamp: overview?.latestMessageAt ?? undefined,
+      activitySummary: overview?.activitySummary ?? undefined,
+      currentTool: overview?.currentTool ?? undefined,
+      costUsd: overview?.totalCostUsd ?? undefined,
+    }
+  })
+}

packages/browseros-agent/apps/agent/entrypoints/app/agent-command/useAgentConversation.ts

@@ -1,12 +1,13 @@
 import { useEffect, useRef, useState } from 'react'
 import {
   type AgentHarnessStreamEvent,
-  attachToHarnessTurn,
-  cancelHarnessTurn,
   chatWithHarnessAgent,
-  fetchActiveHarnessTurn,
 } from '@/entrypoints/app/agents/useAgents'
-import type { OpenClawChatHistoryMessage } from '@/entrypoints/app/agents/useOpenClaw'
+import {
+  chatWithAgent,
+  type OpenClawChatHistoryMessage,
+  type OpenClawStreamEvent,
+} from '@/entrypoints/app/agents/useOpenClaw'
 import type {
   AgentConversationTurn,
   AssistantPart,
@@ -28,23 +29,11 @@ export interface SendInput {
 }
 
 interface UseAgentConversationOptions {
-  // The hook always speaks to the harness chat path now; the OpenClaw
-  // legacy /claw/agents/:id/chat surface was removed in Step 12. The
-  // option remains for forward-compatibility.
-  runtime?: 'agent-harness'
+  runtime?: 'openclaw' | 'agent-harness'
   sessionKey?: string | null
   history?: OpenClawChatHistoryMessage[]
   onComplete?: () => void
   onSessionKeyChange?: (sessionKey: string) => void
-  /**
-   * Server-side active turn id, surfaced via the listing query. When
-   * this changes from null/<id> to a different non-null id while we
-   * aren't already streaming (e.g. the server just popped a queued
-   * message and started a new turn), the hook reattaches via
-   * /chat/active so the chat panel picks up the live stream without
-   * waiting for a remount.
-   */
-  activeTurnId?: string | null
 }
 
 export function useAgentConversation(
@@ -60,11 +49,6 @@ export function useAgentConversation(
   const streamAbortRef = useRef<AbortController | null>(null)
   const onCompleteRef = useRef(options.onComplete)
   const onSessionKeyChangeRef = useRef(options.onSessionKeyChange)
-  // Per-turn resume bookkeeping. `turnId` is captured from the response
-  // header; `lastSeq` advances with every SSE event so a reconnect can
-  // resume via Last-Event-ID.
-  const turnIdRef = useRef<string | null>(null)
-  const lastSeqRef = useRef<number | null>(null)
 
   useEffect(() => {
     sessionKeyRef.current = options.sessionKey ?? ''
@@ -88,12 +72,6 @@ export function useAgentConversation(
     }
   }, [])
 
-  // Indirection for the resume effect below: lets it call the latest
-  // event handler without re-subscribing on every render.
-  const processEventRef = useRef<(event: AgentHarnessStreamEvent) => void>(
-    () => {},
-  )
-
   const updateCurrentTurnParts = (
     updater: (parts: AssistantPart[]) => AssistantPart[],
   ) => {
@@ -104,6 +82,85 @@ export function useAgentConversation(
     })
   }
 
+  const processStreamEvent = (event: OpenClawStreamEvent) => {
+    switch (event.type) {
+      case 'text-delta': {
+        appendTextDelta((event.data.text as string) ?? '')
+        break
+      }
+
+      case 'thinking': {
+        appendThinkingDelta((event.data.text as string) ?? '')
+        break
+      }
+
+      case 'tool-start': {
+        const rawName = (event.data.toolName as string) ?? 'unknown'
+        const args = event.data.args as Record<string, unknown> | undefined
+        const { label, subject } = buildToolLabel(rawName, args)
+        const tool = {
+          id: (event.data.toolCallId as string) ?? crypto.randomUUID(),
+          name: rawName,
+          label,
+          subject,
+          status: 'running' as const,
+        }
+        updateCurrentTurnParts((parts) => {
+          const last = parts[parts.length - 1]
+          if (last?.kind === 'tool-batch') {
+            return [
+              ...parts.slice(0, -1),
+              { ...last, tools: [...last.tools, tool] },
+            ]
+          }
+          return [...parts, { kind: 'tool-batch', tools: [tool] }]
+        })
+        break
+      }
+
+      case 'tool-end': {
+        const toolId = event.data.toolCallId as string
+        const toolStatus: 'completed' | 'error' =
+          (event.data.status as string) === 'error' ? 'error' : 'completed'
+        const durationMs = event.data.durationMs as number | undefined
+        updateCurrentTurnParts((parts) => {
+          for (let i = parts.length - 1; i >= 0; i--) {
+            const part = parts[i]
+            if (
+              part.kind === 'tool-batch' &&
+              part.tools.some((t) => t.id === toolId)
+            ) {
+              const updatedTools = part.tools.map((t) =>
+                t.id === toolId ? { ...t, status: toolStatus, durationMs } : t,
+              )
+              return [
+                ...parts.slice(0, i),
+                { ...part, tools: updatedTools },
+                ...parts.slice(i + 1),
+              ]
+            }
+          }
+          return parts
+        })
+        break
+      }
+
+      case 'done': {
+        markCurrentTurnDone()
+        break
+      }
+
+      case 'error': {
+        const msg =
+          (event.data.message as string) ??
+          (event.data.error as string) ??
+          'Unknown error'
+        appendErrorText(msg)
+        break
+      }
+    }
+  }
+
   const appendTextDelta = (delta: string) => {
     textAccRef.current += delta
     const text = textAccRef.current
@@ -218,105 +275,6 @@ export function useAgentConversation(
         break
     }
   }
-  processEventRef.current = processAgentHarnessStreamEvent
-
-  const activeTurnIdDep = options.activeTurnId ?? null
-
-  // On mount, on agent change, and whenever the listing reports a
-  // *new* active turn id, check whether the server has an in-flight
-  // turn for this agent and reattach to it. This catches three
-  // cases at once: the chat resilience flow (tab close/reopen),
-  // navigation between agents, AND queue drain (the server starts a
-  // new turn from a queued message → activeTurnId flips → attach).
-  useEffect(() => {
-    let cancelled = false
-    const abortController = new AbortController()
-    // Reference the dep inside the body so biome's exhaustive-deps
-    // rule sees it consumed; the value is just an "any non-null
-    // active turn id" trigger — the actual id we attach to comes
-    // from the fresh fetchActiveHarnessTurn call below.
-    void activeTurnIdDep
-
-    const attemptResume = async () => {
-      // Track whether *we* started a stream in this run. When the
-      // early-return paths fire (no active turn, or a `send()` /
-      // earlier resume already owns `streamAbortRef`), the finally
-      // block must NOT touch streaming/turnIdRef/lastSeqRef —
-      // otherwise we clobber the in-flight stream's state and the
-      // Stop button drops out mid-turn while events keep arriving.
-      let weStartedStream = false
-      try {
-        const active = await fetchActiveHarnessTurn(agentId)
-        if (cancelled || !active || active.status !== 'running') return
-        if (streamAbortRef.current) return // someone else already owns the stream
-
-        // Stage a placeholder turn so the streamed events have a row
-        // to render into. The server now persists the kicking-off
-        // prompt on the active turn, so we render it as the user
-        // bubble immediately — no empty-bubble flicker when a queued
-        // message starts running.
-        setTurns((prev) => [
-          ...prev,
-          {
-            id: crypto.randomUUID(),
-            userText: active.prompt ?? '',
-            parts: [],
-            done: false,
-            timestamp: active.startedAt,
-          },
-        ])
-        textAccRef.current = ''
-        thinkAccRef.current = ''
-        turnIdRef.current = active.turnId
-        lastSeqRef.current = null
-        streamAbortRef.current = abortController
-        setStreaming(true)
-        weStartedStream = true
-
-        const response = await attachToHarnessTurn(agentId, {
-          turnId: active.turnId,
-          signal: abortController.signal,
-        })
-        if (!response.ok) return
-        await consumeSSEStream<AgentHarnessStreamEvent>(
-          response,
-          (event, meta) => {
-            if (typeof meta.seq === 'number') lastSeqRef.current = meta.seq
-            processEventRef.current(event)
-          },
-          abortController.signal,
-        )
-      } catch {
-        // Resume is best-effort; transient errors fall back to the
-        // user starting a new turn manually.
-      } finally {
-        // Always release `streamAbortRef` if we owned it — even when
-        // the effect was cancelled mid-stream (a listing poll
-        // captured the next queue-drain turn id, for example). If we
-        // don't, the next effect run hits `if (streamAbortRef.current)
-        // return` against our now-aborted controller and never
-        // reattaches, leaving `streaming === true` with no live stream.
-        if (weStartedStream && streamAbortRef.current === abortController) {
-          streamAbortRef.current = null
-        }
-        // The other state (streaming flag, turn id, lastSeq) is the
-        // *current run's* lifecycle: only reset it on a clean exit.
-        // When `cancelled` is true the next run will set these
-        // itself, so resetting here would only cause a brief flicker.
-        if (!cancelled && weStartedStream) {
-          turnIdRef.current = null
-          lastSeqRef.current = null
-          setStreaming(false)
-        }
-      }
-    }
-
-    void attemptResume()
-    return () => {
-      cancelled = true
-      abortController.abort()
-    }
-  }, [agentId, activeTurnIdDep])
 
   const send = async (input: string | SendInput) => {
     const normalized: SendInput =
@@ -346,25 +304,17 @@ export function useAgentConversation(
     streamAbortRef.current = abortController
 
     try {
-      let response = await chatWithHarnessAgent(
-        agentId,
-        trimmed,
-        abortController.signal,
-        attachments,
-      )
-      // 409 means the server already has an active turn for this
-      // agent (e.g. a previous tab kicked one off and we're a fresh
-      // mount that missed the resume window). Attach to it instead of
-      // double-sending.
-      if (response.status === 409) {
-        const body = (await response.json()) as { turnId?: string }
-        if (body.turnId) {
-          response = await attachToHarnessTurn(agentId, {
-            turnId: body.turnId,
-            signal: abortController.signal,
-          })
-        }
-      }
+      const response =
+        options.runtime === 'agent-harness'
+          ? await chatWithHarnessAgent(agentId, trimmed, abortController.signal)
+          : await chatWithAgent(
+              agentId,
+              trimmed,
+              sessionKeyRef.current || undefined,
+              historyRef.current,
+              abortController.signal,
+              attachments,
+            )
       const responseSessionKey =
         response.headers.get('X-Session-Key') ??
         response.headers.get('X-Session-Id')
@@ -372,11 +322,6 @@ export function useAgentConversation(
         sessionKeyRef.current = responseSessionKey
         onSessionKeyChangeRef.current?.(responseSessionKey)
       }
-      const responseTurnId = response.headers.get('X-Turn-Id')
-      if (responseTurnId) {
-        turnIdRef.current = responseTurnId
-        lastSeqRef.current = null
-      }
       if (!response.ok) {
         const err = await response.text()
         updateCurrentTurnParts((parts) => [
@@ -385,14 +330,19 @@ export function useAgentConversation(
         ])
         return
       }
-      await consumeSSEStream<AgentHarnessStreamEvent>(
-        response,
-        (event, meta) => {
-          if (typeof meta.seq === 'number') lastSeqRef.current = meta.seq
-          processAgentHarnessStreamEvent(event)
-        },
-        abortController.signal,
-      )
+      if (options.runtime === 'agent-harness') {
+        await consumeSSEStream<AgentHarnessStreamEvent>(
+          response,
+          processAgentHarnessStreamEvent,
+          abortController.signal,
+        )
+      } else {
+        await consumeSSEStream<OpenClawStreamEvent>(
+          response,
+          processStreamEvent,
+          abortController.signal,
+        )
+      }
     } catch (err) {
       if (abortController.signal.aborted) return
       const msg = err instanceof Error ? err.message : String(err)
@@ -404,35 +354,14 @@ export function useAgentConversation(
       if (streamAbortRef.current === abortController) {
         streamAbortRef.current = null
       }
-      turnIdRef.current = null
-      lastSeqRef.current = null
       onCompleteRef.current?.()
       setStreaming(false)
     }
   }
 
-  /**
-   * Stop button. The fetch abort only detaches *this* SSE subscriber
-   * now — the underlying turn would otherwise keep running on the
-   * server. So we explicitly cancel via the new endpoint, then unwind
-   * the local stream.
-   */
-  const stop = async () => {
-    const turnId = turnIdRef.current ?? undefined
+  const resetConversation = () => {
     streamAbortRef.current?.abort()
     streamAbortRef.current = null
-    try {
-      await cancelHarnessTurn(agentId, {
-        turnId,
-        reason: 'user pressed stop',
-      })
-    } catch {
-      // Best-effort — UI already aborted.
-    }
-  }
-
-  const resetConversation = () => {
-    void stop()
     setTurns([])
     setStreaming(false)
   }
@@ -442,7 +371,6 @@ export function useAgentConversation(
     streaming,
     sessionKey: sessionKeyRef.current,
     send,
-    stop,
     resetConversation,
   }
 }

packages/browseros-agent/apps/agent/entrypoints/app/agent-command/useAgentDashboard.ts

@@ -0,0 +1,95 @@
+import { useQuery, useQueryClient } from '@tanstack/react-query'
+import { useEffect } from 'react'
+import { useAgentServerUrl } from '@/lib/browseros/useBrowserOSProviders'
+
+export interface AgentOverview {
+  agentId: string
+  status: 'working' | 'idle' | 'error' | 'unknown'
+  latestMessage: string | null
+  latestMessageAt: number | null
+  activitySummary: string | null
+  currentTool: string | null
+  totalCostUsd: number
+  sessionCount: number
+}
+
+export interface DashboardResponse {
+  agents: AgentOverview[]
+  summary: {
+    totalAgents: number
+    totalCostUsd: number
+  }
+}
+
+interface StatusEvent {
+  agentId: string
+  status: AgentOverview['status']
+  currentTool: string | null
+  error: string | null
+  timestamp: number
+}
+
+const DASHBOARD_QUERY_KEY = ['claw', 'dashboard']
+
+export function useAgentDashboard(enabled: boolean) {
+  const { baseUrl, isLoading: urlLoading } = useAgentServerUrl()
+  const queryClient = useQueryClient()
+  const ready = enabled && Boolean(baseUrl) && !urlLoading
+
+  // Initial data load + periodic refresh as fallback
+  const query = useQuery<DashboardResponse>({
+    queryKey: [...DASHBOARD_QUERY_KEY, baseUrl],
+    queryFn: async () => {
+      const url = new URL('/claw/dashboard', baseUrl as string)
+      const response = await fetch(url.toString())
+      if (!response.ok) throw new Error('Failed to fetch dashboard')
+      return response.json()
+    },
+    enabled: ready,
+  })
+
+  // SSE subscription for real-time status patches
+  useEffect(() => {
+    if (!ready || !baseUrl) return
+
+    const streamUrl = new URL('/claw/dashboard/stream', baseUrl)
+    const eventSource = new EventSource(streamUrl.toString())
+
+    eventSource.addEventListener('snapshot', (event) => {
+      try {
+        const dashboard = JSON.parse(event.data) as DashboardResponse
+        queryClient.setQueryData([...DASHBOARD_QUERY_KEY, baseUrl], dashboard)
+      } catch {}
+    })
+
+    eventSource.addEventListener('status', (event) => {
+      try {
+        const status = JSON.parse(event.data) as StatusEvent
+        queryClient.setQueryData<DashboardResponse>(
+          [...DASHBOARD_QUERY_KEY, baseUrl],
+          (prev) => {
+            if (!prev) return prev
+            return {
+              ...prev,
+              agents: prev.agents.map((agent) =>
+                agent.agentId === status.agentId
+                  ? {
+                      ...agent,
+                      status: status.status,
+                      currentTool: status.currentTool,
+                    }
+                  : agent,
+              ),
+            }
+          },
+        )
+      } catch {}
+    })
+
+    return () => {
+      eventSource.close()
+    }
+  }, [ready, baseUrl, queryClient])
+
+  return query
+}

packages/browseros-agent/apps/agent/entrypoints/app/agent-command/useClawChatHistory.ts

@@ -0,0 +1,71 @@
+import { useInfiniteQuery } from '@tanstack/react-query'
+import { useAgentServerUrl } from '@/lib/browseros/useBrowserOSProviders'
+import type { AgentHistoryPageResponse } from './claw-chat-types'
+
+const HISTORY_QUERY_KEY = 'claw-agent-history'
+
+async function fetchClawJson<T>(url: string): Promise<T> {
+  const response = await fetch(url)
+
+  if (!response.ok) {
+    let message = `Request failed with status ${response.status}`
+    try {
+      const body = (await response.json()) as { error?: string }
+      if (body.error) message = body.error
+    } catch {}
+    throw new Error(message)
+  }
+
+  return response.json() as Promise<T>
+}
+
+function buildClawUrl(baseUrl: string, path: string): URL {
+  return new URL(`/claw${path}`, baseUrl)
+}
+
+export function useClawChatHistory({
+  agentId,
+  sessionKey,
+  enabled = true,
+  limit = 50,
+}: {
+  agentId: string
+  // null lets the server resolve the most recent user-chat session for the
+  // agent — avoids an extra /session round-trip and the race that came with it.
+  sessionKey: string | null
+  enabled?: boolean
+  limit?: number
+}) {
+  const {
+    baseUrl,
+    isLoading: urlLoading,
+    error: urlError,
+  } = useAgentServerUrl()
+
+  const query = useInfiniteQuery<AgentHistoryPageResponse, Error>({
+    queryKey: [HISTORY_QUERY_KEY, baseUrl, agentId, sessionKey],
+    initialPageParam: undefined as string | undefined,
+    queryFn: async ({ pageParam }) => {
+      const url = buildClawUrl(baseUrl as string, `/agents/${agentId}/history`)
+      url.searchParams.set('limit', String(limit))
+
+      if (sessionKey) {
+        url.searchParams.set('sessionKey', sessionKey)
+      }
+      if (typeof pageParam === 'string' && pageParam) {
+        url.searchParams.set('cursor', pageParam)
+      }
+
+      return fetchClawJson<AgentHistoryPageResponse>(url.toString())
+    },
+    getNextPageParam: (lastPage) =>
+      lastPage.page.hasMore ? lastPage.page.cursor : undefined,
+    enabled: enabled && Boolean(baseUrl) && !urlLoading && Boolean(agentId),
+  })
+
+  return {
+    ...query,
+    error: query.error ?? urlError,
+    isLoading: query.isLoading || urlLoading,
+  }
+}

packages/browseros-agent/apps/agent/entrypoints/app/agent-command/useHarnessChatHistory.test.ts

@@ -1,55 +0,0 @@
-import { describe, expect, it } from 'bun:test'
-import { mapHarnessHistoryPage } from './harness-history-mapper'
-
-describe('mapHarnessHistoryPage', () => {
-  it('maps rich harness history into chat history items', () => {
-    const page = mapHarnessHistoryPage({
-      agentId: 'agent-1',
-      sessionId: 'main',
-      items: [
-        {
-          id: 'agent:agent-1:main:1',
-          agentId: 'agent-1',
-          sessionId: 'main',
-          role: 'assistant',
-          text: 'Done.',
-          createdAt: 1000,
-          reasoning: { text: 'checking state' },
-          toolCalls: [
-            {
-              toolCallId: 'tool-1',
-              toolName: 'read_file',
-              status: 'completed',
-              input: { path: 'src/index.ts' },
-              output: 'file contents',
-            },
-          ],
-        },
-      ],
-    })
-
-    expect(page.items).toEqual([
-      {
-        id: 'agent:agent-1:main:1',
-        role: 'assistant',
-        text: 'Done.',
-        timestamp: 1000,
-        messageSeq: 1,
-        sessionKey: 'main',
-        source: 'user-chat',
-        reasoning: { text: 'checking state' },
-        toolCalls: [
-          {
-            toolCallId: 'tool-1',
-            toolName: 'read_file',
-            label: 'Read file',
-            subject: 'index.ts',
-            status: 'completed',
-            input: { path: 'src/index.ts' },
-            output: 'file contents',
-          },
-        ],
-      },
-    ])
-  })
-})

packages/browseros-agent/apps/agent/entrypoints/app/agent-command/useHarnessChatHistory.ts

@@ -1,8 +1,11 @@
 import { useQuery } from '@tanstack/react-query'
+import type { HarnessAgentHistoryPage } from '@/entrypoints/app/agents/agent-harness-types'
 import { fetchHarnessAgentHistory } from '@/entrypoints/app/agents/useAgents'
 import { useAgentServerUrl } from '@/lib/browseros/useBrowserOSProviders'
-import type { AgentHistoryPageResponse } from './claw-chat-types'
-import { mapHarnessHistoryPage } from './harness-history-mapper'
+import type {
+  AgentHistoryPageResponse,
+  BrowserOSChatHistoryItem,
+} from './claw-chat-types'
 
 const HISTORY_QUERY_KEY = 'harness-agent-history'
 
@@ -27,3 +30,39 @@ export function useHarnessChatHistory(agentId: string, enabled = true) {
     isLoading: query.isLoading || urlLoading,
   }
 }
+
+function mapHarnessHistoryPage(
+  page: HarnessAgentHistoryPage,
+): AgentHistoryPageResponse {
+  const items: BrowserOSChatHistoryItem[] = page.items.map((item, index) => ({
+    id: item.id,
+    role: item.role,
+    text: item.text,
+    timestamp: item.createdAt,
+    messageSeq: index + 1,
+    sessionKey: 'main',
+    source: 'user-chat',
+  }))
+  const updatedAt =
+    page.items.length > 0
+      ? Math.max(...page.items.map((item) => item.createdAt))
+      : Date.now()
+
+  return {
+    agentId: page.agentId,
+    sessionKey: 'main',
+    session: {
+      key: 'main',
+      updatedAt,
+      sessionId: 'main',
+      agentId: page.agentId,
+      kind: 'agent-harness',
+      source: 'user-chat',
+    },
+    items,
+    page: {
+      hasMore: false,
+      limit: items.length,
+    },
+  }
+}

packages/browseros-agent/apps/agent/entrypoints/app/agent-command/useOutboundQueue.ts

@@ -0,0 +1,271 @@
+import { useCallback, useEffect, useRef, useState } from 'react'
+import type { OpenClawChatHistoryMessage } from '@/entrypoints/app/agents/useOpenClaw'
+import type { UserAttachmentPreview } from '@/lib/agent-conversations/types'
+import type { ServerAttachmentPayload } from '@/lib/attachments'
+import { useAgentServerUrl } from '@/lib/browseros/useBrowserOSProviders'
+
+export type OutboundMessageStatus = 'queued' | 'sending' | 'failed'
+
+export interface OutboundMessage {
+  id: string
+  text: string
+  attachments: ServerAttachmentPayload[]
+  attachmentPreviews: UserAttachmentPreview[]
+  status: OutboundMessageStatus
+  error?: string
+  createdAt: number
+}
+
+export interface OutboundQueueEnqueueInput {
+  text: string
+  attachments?: ServerAttachmentPayload[]
+  attachmentPreviews?: UserAttachmentPreview[]
+  history?: OpenClawChatHistoryMessage[]
+}
+
+export interface OutboundQueueApi {
+  queue: OutboundMessage[]
+  enqueue(input: OutboundQueueEnqueueInput): void
+  cancel(id: string): void
+  retry(id: string): void
+}
+
+interface UseOutboundQueueOptions {
+  agentId: string | null | undefined
+  sessionKey?: string | null
+  enabled?: boolean
+}
+
+interface ServerQueuedItem {
+  id: string
+  status: 'queued' | 'dispatching' | 'failed'
+  message: string
+  attachmentsPreview: Array<{
+    kind: 'image' | 'file'
+    mediaType: string
+    name?: string
+  }>
+  error?: string
+  createdAt: number
+}
+
+function makeId(): string {
+  if (typeof crypto !== 'undefined' && crypto.randomUUID) {
+    return crypto.randomUUID()
+  }
+  return `${Date.now().toString(36)}-${Math.random().toString(36).slice(2, 10)}`
+}
+
+/**
+ * Server-backed outbound message queue. The browser is purely a
+ * projection of server state — closing the tab is safe because the queue
+ * keeps draining server-side via the OutboundQueueService.
+ *
+ * Single id-keyed list: the client generates the queue id and hands it
+ * to the server in the POST body, so the optimistic row and the SSE
+ * snapshot reconcile on the same key from frame zero — there is no
+ * window in which the message renders twice.
+ */
+export function useOutboundQueue(
+  options: UseOutboundQueueOptions,
+): OutboundQueueApi {
+  const { agentId, enabled = true, sessionKey } = options
+  const { baseUrl } = useAgentServerUrl()
+  const sessionKeyRef = useRef<string | null | undefined>(sessionKey)
+  sessionKeyRef.current = sessionKey
+
+  const [items, setItems] = useState<OutboundMessage[]>([])
+  // Track which ids the server has confirmed seeing in any SSE snapshot.
+  // We use this to know whether a missing-from-snapshot id is "drained
+  // by the server" (drop it) or "still in flight client-side" (keep
+  // showing the optimistic row).
+  const everSeenByServerRef = useRef<Set<string>>(new Set())
+  // Local-only attachment previews, keyed by queue id. Data URLs never
+  // leave the browser — the SSE feed only carries metadata, so we hold
+  // them here so the chip strip keeps rendering after server takeover.
+  const previewMapRef = useRef<Map<string, UserAttachmentPreview[]>>(new Map())
+
+  useEffect(() => {
+    if (!enabled || !baseUrl || !agentId) {
+      setItems([])
+      everSeenByServerRef.current = new Set()
+      previewMapRef.current = new Map()
+      return
+    }
+    let cancelled = false
+    const url = `${baseUrl}/claw/agents/${encodeURIComponent(agentId)}/queue/stream`
+    const source = new EventSource(url)
+    source.onmessage = (event) => {
+      if (cancelled) return
+      try {
+        const parsed = JSON.parse(event.data) as { items: ServerQueuedItem[] }
+        const snapshotIds = new Set(parsed.items.map((item) => item.id))
+        for (const id of snapshotIds) everSeenByServerRef.current.add(id)
+
+        setItems((prev) => {
+          const next: OutboundMessage[] = parsed.items.map((item) => ({
+            id: item.id,
+            text: item.message,
+            attachments: [],
+            attachmentPreviews: previewMapRef.current.get(item.id) ?? [],
+            status: serverStatusToClient(item.status),
+            error: item.error,
+            createdAt: item.createdAt,
+          }))
+          // Carry forward any optimistic / failed entries the server
+          // doesn't know about yet (POST in flight) or has finished
+          // dispatching but the client wants to keep visible (failed).
+          const carried = prev.filter((local) => {
+            if (snapshotIds.has(local.id)) return false
+            if (everSeenByServerRef.current.has(local.id)) {
+              // Server saw it before and it's gone now — drained.
+              previewMapRef.current.delete(local.id)
+              return false
+            }
+            return local.status !== 'failed' || Boolean(local.error)
+          })
+          return [...carried, ...next]
+        })
+      } catch {
+        // Malformed event — ignore; next snapshot will recover.
+      }
+    }
+    source.onerror = () => {
+      // Auto-reconnects; nothing to do here.
+    }
+    return () => {
+      cancelled = true
+      source.close()
+    }
+  }, [baseUrl, agentId, enabled])
+
+  const enqueue = useCallback(
+    (input: OutboundQueueEnqueueInput) => {
+      if (!enabled || !baseUrl || !agentId) return
+      const trimmed = input.text.trim()
+      const attachments = input.attachments ?? []
+      if (!trimmed && attachments.length === 0) return
+
+      const id = makeId()
+      const previews = input.attachmentPreviews ?? []
+      previewMapRef.current.set(id, previews)
+      setItems((prev) => [
+        ...prev,
+        {
+          id,
+          text: trimmed,
+          attachments,
+          attachmentPreviews: previews,
+          status: 'queued',
+          createdAt: Date.now(),
+        },
+      ])
+
+      void (async () => {
+        try {
+          const response = await fetch(
+            `${baseUrl}/claw/agents/${encodeURIComponent(agentId)}/queue`,
+            {
+              method: 'POST',
+              headers: { 'Content-Type': 'application/json' },
+              body: JSON.stringify({
+                id,
+                message: trimmed,
+                attachments: attachments.length > 0 ? attachments : undefined,
+                sessionKey: sessionKeyRef.current ?? undefined,
+                history: input.history,
+              }),
+            },
+          )
+          if (!response.ok) {
+            const text = await response.text().catch(() => '')
+            previewMapRef.current.delete(id)
+            setItems((prev) =>
+              prev.map((item) =>
+                item.id === id
+                  ? {
+                      ...item,
+                      status: 'failed',
+                      error:
+                        text || `Failed to enqueue (status ${response.status})`,
+                    }
+                  : item,
+              ),
+            )
+          }
+        } catch (err) {
+          // Only mark as failed if the SSE snapshot hasn't already
+          // taken ownership of the entry (i.e. the request actually
+          // reached the server).
+          if (everSeenByServerRef.current.has(id)) return
+          previewMapRef.current.delete(id)
+          setItems((prev) =>
+            prev.map((item) =>
+              item.id === id
+                ? {
+                    ...item,
+                    status: 'failed',
+                    error:
+                      err instanceof Error
+                        ? err.message
+                        : 'Failed to enqueue message',
+                  }
+                : item,
+            ),
+          )
+        }
+      })()
+    },
+    [baseUrl, agentId, enabled],
+  )
+
+  const cancel = useCallback(
+    (id: string) => {
+      // If the server has never seen this id, just drop it locally.
+      if (!everSeenByServerRef.current.has(id)) {
+        previewMapRef.current.delete(id)
+        setItems((prev) => prev.filter((item) => item.id !== id))
+        return
+      }
+      if (!enabled || !baseUrl || !agentId) return
+      void fetch(
+        `${baseUrl}/claw/agents/${encodeURIComponent(agentId)}/queue/${encodeURIComponent(id)}`,
+        { method: 'DELETE' },
+      ).catch(() => {})
+    },
+    [baseUrl, agentId, enabled],
+  )
+
+  const retry = useCallback(
+    (id: string) => {
+      if (!everSeenByServerRef.current.has(id)) {
+        // Optimistic-only entry, never made it to the server. Reset
+        // status so the user can press Send again.
+        setItems((prev) =>
+          prev.map((item) =>
+            item.id === id
+              ? { ...item, status: 'queued', error: undefined }
+              : item,
+          ),
+        )
+        return
+      }
+      if (!enabled || !baseUrl || !agentId) return
+      void fetch(
+        `${baseUrl}/claw/agents/${encodeURIComponent(agentId)}/queue/${encodeURIComponent(id)}/retry`,
+        { method: 'POST' },
+      ).catch(() => {})
+    },
+    [baseUrl, agentId, enabled],
+  )
+
+  return { queue: items, enqueue, cancel, retry }
+}
+
+function serverStatusToClient(
+  status: ServerQueuedItem['status'],
+): OutboundMessageStatus {
+  if (status === 'dispatching') return 'sending'
+  if (status === 'failed') return 'failed'
+  return 'queued'
+}

packages/browseros-agent/apps/agent/entrypoints/app/agents/AdapterIcon.tsx

@@ -1,42 +0,0 @@
-import { Bot, Cpu, Sparkles } from 'lucide-react'
-import type { FC } from 'react'
-import type { HarnessAgentAdapter } from './agent-harness-types'
-
-/**
- * Single icon component for any adapter the agent rail can render.
- * Falls back to a generic bot when the adapter is unknown so future
- * adapters land without a code change at the call site.
- */
-interface AdapterIconProps {
-  adapter: HarnessAgentAdapter | 'unknown'
-  className?: string
-}
-
-export const AdapterIcon: FC<AdapterIconProps> = ({ adapter, className }) => {
-  switch (adapter) {
-    case 'claude':
-      // Claude Code — text-based agent, sparkles to evoke the "AI assistant" feel.
-      return <Sparkles className={className} aria-label="Claude Code" />
-    case 'codex':
-      // Codex — code-leaning, CPU mark.
-      return <Cpu className={className} aria-label="Codex" />
-    case 'openclaw':
-      // OpenClaw — bot/automation framing.
-      return <Bot className={className} aria-label="OpenClaw" />
-    default:
-      return <Bot className={className} aria-label="Agent" />
-  }
-}
-
-export function adapterLabel(adapter: HarnessAgentAdapter | 'unknown'): string {
-  switch (adapter) {
-    case 'claude':
-      return 'Claude Code'
-    case 'codex':
-      return 'Codex'
-    case 'openclaw':
-      return 'OpenClaw'
-    default:
-      return 'Agent'
-  }
-}

packages/browseros-agent/apps/agent/entrypoints/app/agents/AgentList.tsx

@@ -1,176 +1,117 @@
-import { Loader2 } from 'lucide-react'
-import { type FC, useMemo } from 'react'
-import { AgentRowCard } from './AgentRowCard'
-import { AgentsEmptyState } from './AgentsEmptyState'
-import type {
-  HarnessAdapterDescriptor,
-  HarnessAgent,
-  HarnessAgentAdapter,
-} from './agent-harness-types'
-import type {
-  AgentAdapterHealth,
-  AgentRowData,
-} from './agent-row/agent-row.types'
+import { Bot, Cpu, Loader2, MessageSquare, Plus, Trash2 } from 'lucide-react'
+import type { FC } from 'react'
+import { Badge } from '@/components/ui/badge'
+import { Button } from '@/components/ui/button'
+import { Card, CardContent, CardHeader, CardTitle } from '@/components/ui/card'
 import type { AgentListItem } from './agents-page-types'
-import type { AgentLiveness } from './LivenessDot'
 
 interface AgentListProps {
   agents: AgentListItem[]
-  /** Optional per-agent activity metadata, keyed by `agentId`. */
-  activity?: Record<
-    string,
-    { status: AgentLiveness; lastUsedAt: number | null }
-  >
-  /** Lookup table from harness id → enriched agent record. */
-  harnessAgentLookup?: Map<string, HarnessAgent>
-  /** Adapter catalog (carries per-adapter health). */
-  adapters: HarnessAdapterDescriptor[]
   loading: boolean
   deletingAgentKey: string | null
+  onChatAgent: (agent: AgentListItem) => void
   onCreateAgent: () => void
   onDeleteAgent: (agent: AgentListItem) => void
-  onPinToggle: (agent: AgentListItem, next: boolean) => void
 }
 
 export const AgentList: FC<AgentListProps> = ({
   agents,
-  activity,
-  harnessAgentLookup,
-  adapters,
   loading,
   deletingAgentKey,
+  onChatAgent,
   onCreateAgent,
   onDeleteAgent,
-  onPinToggle,
 }) => {
-  const adapterHealth = useMemo(() => {
-    const map = new Map<HarnessAgentAdapter, AgentAdapterHealth>()
-    for (const adapter of adapters) {
-      if (adapter.health) {
-        map.set(adapter.id, {
-          healthy: adapter.health.healthy,
-          reason: adapter.health.reason,
-        })
-      }
-    }
-    return map
-  }, [adapters])
-
-  // Sort: pinned rows first, then most recently used, then never-used
-  // agents in id-stable order. The gateway's `main` agent stays
-  // pinned-to-top when never touched so a fresh install has an
-  // obvious starting point.
-  const ordered = useMemo(() => {
-    const withMeta = agents.map((agent) => {
-      const harness = harnessAgentLookup?.get(agent.agentId)
-      return {
-        agent,
-        pinned: harness?.pinned ?? false,
-        lastUsedAt: activity?.[agent.agentId]?.lastUsedAt ?? null,
-      }
-    })
-    return withMeta
-      .sort((a, b) => {
-        if (a.pinned !== b.pinned) return a.pinned ? -1 : 1
-        const aSeed = a.agent.agentId === 'main' && a.lastUsedAt === null
-        const bSeed = b.agent.agentId === 'main' && b.lastUsedAt === null
-        if (aSeed && !bSeed) return -1
-        if (!aSeed && bSeed) return 1
-        const aValue = a.lastUsedAt ?? -Infinity
-        const bValue = b.lastUsedAt ?? -Infinity
-        if (aValue !== bValue) return bValue - aValue
-        return a.agent.agentId.localeCompare(b.agent.agentId)
-      })
-      .map((entry) => entry.agent)
-  }, [activity, agents, harnessAgentLookup])
-
   if (loading && agents.length === 0) {
     return (
-      <div className="flex h-36 items-center justify-center rounded-xl border border-border border-dashed bg-card/50">
+      <div className="flex h-36 items-center justify-center rounded-lg border border-border/70">
         <Loader2 className="size-5 animate-spin text-muted-foreground" />
       </div>
     )
   }
 
   if (agents.length === 0) {
-    return <AgentsEmptyState onCreateAgent={onCreateAgent} />
+    return (
+      <Card>
+        <CardContent className="flex h-48 flex-col items-center justify-center gap-4 text-center">
+          <div className="flex size-10 items-center justify-center rounded-lg bg-muted text-muted-foreground">
+            <Bot className="size-5" />
+          </div>
+          <div className="space-y-1">
+            <h2 className="font-medium text-base">No agents</h2>
+            <p className="text-muted-foreground text-sm">
+              Create an OpenClaw, Claude Code, or Codex agent.
+            </p>
+          </div>
+          <Button variant="outline" onClick={onCreateAgent}>
+            <Plus className="mr-2 size-4" />
+            New Agent
+          </Button>
+        </CardContent>
+      </Card>
+    )
   }
 
   return (
     <div className="grid gap-3">
-      {ordered.map((agent) => {
-        const harness = harnessAgentLookup?.get(agent.agentId)
-        const adapter: HarnessAgentAdapter | 'unknown' =
-          harness?.adapter ?? inferAdapterFromLabel(agent.runtimeLabel)
-        const data = buildRowData({
-          agent,
-          adapter,
-          harness,
-          activity: activity?.[agent.agentId],
-          adapterHealth:
-            adapterHealth.get(adapter as HarnessAgentAdapter) ?? null,
-        })
-        return (
-          <AgentRowCard
-            key={agent.key}
-            data={data}
-            deleting={deletingAgentKey === agent.key}
-            onDelete={onDeleteAgent}
-            onPinToggle={onPinToggle}
-          />
-        )
-      })}
+      {agents.map((agent) => (
+        <Card key={agent.key} className="rounded-lg border-border/70">
+          <CardHeader className="flex flex-row items-center justify-between gap-4 py-3">
+            <div className="flex min-w-0 items-center gap-3">
+              <div className="flex size-10 shrink-0 items-center justify-center rounded-lg bg-muted text-muted-foreground">
+                {agent.source === 'openclaw' ? (
+                  <Cpu className="size-5" />
+                ) : (
+                  <Bot className="size-5" />
+                )}
+              </div>
+              <div className="min-w-0">
+                <CardTitle className="truncate text-base">
+                  {agent.name}
+                </CardTitle>
+                <div className="mt-1 flex flex-wrap items-center gap-2 text-muted-foreground text-xs">
+                  <Badge variant="outline" className="rounded-md">
+                    {agent.runtimeLabel}
+                  </Badge>
+                  <span>{agent.modelLabel}</span>
+                  <Badge variant="outline" className="rounded-md">
+                    main
+                  </Badge>
+                </div>
+                <p className="mt-1 truncate font-mono text-muted-foreground text-xs">
+                  {agent.detail}
+                </p>
+              </div>
+            </div>
+            <div className="flex shrink-0 items-center gap-1">
+              <Button
+                variant="ghost"
+                size="sm"
+                onClick={() => onChatAgent(agent)}
+                disabled={!agent.canChat}
+              >
+                <MessageSquare className="mr-1 size-4" />
+                Chat
+              </Button>
+              {agent.canDelete ? (
+                <Button
+                  variant="ghost"
+                  size="icon"
+                  title="Delete agent"
+                  onClick={() => onDeleteAgent(agent)}
+                  disabled={deletingAgentKey === agent.key}
+                >
+                  {deletingAgentKey === agent.key ? (
+                    <Loader2 className="size-4 animate-spin" />
+                  ) : (
+                    <Trash2 className="size-4 text-destructive" />
+                  )}
+                </Button>
+              ) : null}
+            </div>
+          </CardHeader>
+        </Card>
+      ))}
     </div>
   )
 }
-
-function inferAdapterFromLabel(label: string): HarnessAgentAdapter | 'unknown' {
-  const lower = label?.toLowerCase()
-  if (lower === 'claude code') return 'claude'
-  if (lower === 'codex') return 'codex'
-  if (lower === 'openclaw') return 'openclaw'
-  return 'unknown'
-}
-
-const ZERO_BUCKETS = (): number[] => Array.from({ length: 14 }, () => 0)
-
-function buildRowData(input: {
-  agent: AgentListItem
-  adapter: HarnessAgentAdapter | 'unknown'
-  harness: HarnessAgent | undefined
-  activity: { status: AgentLiveness; lastUsedAt: number | null } | undefined
-  adapterHealth: AgentAdapterHealth | null
-}): AgentRowData {
-  const { agent, adapter, harness, activity, adapterHealth } = input
-  return {
-    agent,
-    adapter,
-    modelLabel: deriveModelLabel(agent, harness),
-    reasoningEffort: harness?.reasoningEffort ?? null,
-    status: activity?.status ?? 'unknown',
-    lastUsedAt: activity?.lastUsedAt ?? harness?.lastUsedAt ?? null,
-    pinned: harness?.pinned ?? false,
-    cwd: harness?.cwd ?? null,
-    lastUserMessage: harness?.lastUserMessage ?? null,
-    tokens: harness?.tokens ?? null,
-    turnsByDay: harness?.turnsByDay ?? ZERO_BUCKETS(),
-    failedByDay: harness?.failedByDay ?? ZERO_BUCKETS(),
-    lastError: harness?.lastError ?? null,
-    lastErrorAt: harness?.lastErrorAt ?? null,
-    activeTurnId: harness?.activeTurnId ?? null,
-    adapterHealth,
-  }
-}
-
-function deriveModelLabel(
-  agent: AgentListItem,
-  harness: HarnessAgent | undefined,
-): string | null {
-  // Prefer the agent rail's modelLabel when meaningful; harness's
-  // modelId is a stable identifier but the rail's `modelLabel`
-  // already maps to a friendly display string.
-  if (agent.modelLabel && agent.modelLabel !== 'default') {
-    return agent.modelLabel
-  }
-  return harness?.modelId ?? null
-}

packages/browseros-agent/apps/agent/entrypoints/app/agents/AgentRowCard.tsx

@@ -1,99 +0,0 @@
-import type { FC } from 'react'
-import { cn } from '@/lib/utils'
-import { AgentActions } from './agent-row/AgentActions'
-import { AgentErrorPanel } from './agent-row/AgentErrorPanel'
-import { AgentLastMessage } from './agent-row/AgentLastMessage'
-import { AgentMetaRow } from './agent-row/AgentMetaRow'
-import { AgentSummaryChips } from './agent-row/AgentSummaryChips'
-import { AgentTile } from './agent-row/AgentTile'
-import { AgentTitleRow } from './agent-row/AgentTitleRow'
-import type {
-  AgentRowCallbacks,
-  AgentRowData,
-} from './agent-row/agent-row.types'
-
-interface AgentRowCardProps extends AgentRowCallbacks {
-  data: AgentRowData
-  /** Whether THIS agent is mid-delete; renders a spinner in the menu. */
-  deleting?: boolean
-}
-
-/**
- * Composition shell for the agent rail. Owns no state; sub-components
- * each handle their own micro-state (error-panel collapse, etc.) and
- * emit callbacks (delete, pin/unpin) for the page to act on.
- *
- * The whole card carries state — not just the tile — so the row's
- * border subtly tells the user what's going on at a glance:
- *   working → accent-orange border with a soft glow
- *   error   → destructive border
- *   idle    → muted border, lifts on hover
- */
-export const AgentRowCard: FC<AgentRowCardProps> = ({
-  data,
-  deleting,
-  onDelete,
-  onPinToggle,
-}) => {
-  return (
-    <div
-      className={cn(
-        // Layout-stable hover. No translate, no shadow change — both
-        // visibly perturb neighbouring rows. Only the border tint
-        // shifts on hover, and the rail's vertical rhythm stays
-        // exactly the same in every state.
-        'group rounded-xl border bg-card p-4 shadow-sm transition-colors',
-        data.status === 'working'
-          ? 'border-[var(--accent-orange)]/40'
-          : data.status === 'error'
-            ? 'border-destructive/40'
-            : 'border-border hover:border-[var(--accent-orange)]/30',
-      )}
-    >
-      <div className="flex items-start gap-4">
-        <AgentTile
-          adapter={data.adapter}
-          status={data.status}
-          lastUsedAt={data.lastUsedAt}
-        />
-
-        <div className="min-w-0 flex-1">
-          <AgentTitleRow
-            agent={data.agent}
-            status={data.status}
-            pinned={data.pinned}
-            turnsByDay={data.turnsByDay}
-            failedByDay={data.failedByDay}
-            onPinToggle={(next) => onPinToggle(data.agent, next)}
-          />
-
-          <AgentSummaryChips
-            adapter={data.adapter}
-            modelLabel={data.modelLabel}
-            reasoningEffort={data.reasoningEffort}
-            adapterHealth={data.adapterHealth}
-          />
-
-          <AgentLastMessage message={data.lastUserMessage} />
-
-          <AgentMetaRow lastUsedAt={data.lastUsedAt} tokens={data.tokens} />
-
-          {data.status === 'error' && data.lastError && (
-            <AgentErrorPanel
-              agentId={data.agent.agentId}
-              message={data.lastError}
-              errorAt={data.lastErrorAt}
-            />
-          )}
-        </div>
-
-        <AgentActions
-          agent={data.agent}
-          activeTurnId={data.activeTurnId}
-          deleting={deleting}
-          onDelete={onDelete}
-        />
-      </div>
-    </div>
-  )
-}

packages/browseros-agent/apps/agent/entrypoints/app/agents/AgentsEmptyState.tsx

@@ -1,32 +0,0 @@
-import { Bot, Plus } from 'lucide-react'
-import type { FC } from 'react'
-import { Button } from '@/components/ui/button'
-
-interface AgentsEmptyStateProps {
-  onCreateAgent: () => void
-}
-
-export const AgentsEmptyState: FC<AgentsEmptyStateProps> = ({
-  onCreateAgent,
-}) => {
-  return (
-    <div className="rounded-xl border border-border border-dashed bg-card/50 p-12 text-center">
-      <div className="mx-auto mb-4 flex h-12 w-12 items-center justify-center rounded-xl bg-[var(--accent-orange)]/10">
-        <Bot className="h-6 w-6 text-[var(--accent-orange)]" />
-      </div>
-      <h3 className="mb-1 font-semibold">No agents yet</h3>
-      <p className="mx-auto mb-4 max-w-sm text-muted-foreground text-sm">
-        Spin up an OpenClaw, Claude Code, or Codex agent to chat with, schedule,
-        or run in the background.
-      </p>
-      <Button
-        onClick={onCreateAgent}
-        variant="outline"
-        className="border-[var(--accent-orange)] bg-[var(--accent-orange)]/10 text-[var(--accent-orange)] hover:bg-[var(--accent-orange)]/20 hover:text-[var(--accent-orange)]"
-      >
-        <Plus className="mr-1.5 h-4 w-4" />
-        Create your first agent
-      </Button>
-    </div>
-  )
-}

packages/browseros-agent/apps/agent/entrypoints/app/agents/AgentsHeader.tsx

@@ -1,41 +0,0 @@
-import { Bot, Plus } from 'lucide-react'
-import type { FC } from 'react'
-import { Button } from '@/components/ui/button'
-
-interface AgentsHeaderProps {
-  onCreateAgent: () => void
-}
-
-/**
- * Mirrors the visual shape of `SoulHeader` and `ScheduledTasksHeader`
- * so the page reads as part of the same family. Loose lifecycle
- * controls that used to sit next to the title moved into
- * `GatewayStatusBar` — they're OpenClaw-specific and don't apply to
- * Claude/Codex agents.
- */
-export const AgentsHeader: FC<AgentsHeaderProps> = ({ onCreateAgent }) => {
-  return (
-    <div className="rounded-xl border border-border bg-card p-6 shadow-sm transition-all hover:shadow-md">
-      <div className="flex items-start gap-4">
-        <div className="flex h-12 w-12 shrink-0 items-center justify-center rounded-xl bg-[var(--accent-orange)]/10">
-          <Bot className="h-6 w-6 text-[var(--accent-orange)]" />
-        </div>
-        <div className="flex-1">
-          <h2 className="mb-1 font-semibold text-xl">Agents</h2>
-          <p className="text-muted-foreground text-sm">
-            OpenClaw, Claude Code, and Codex agents — chat, schedule, and run
-            them in the background.
-          </p>
-        </div>
-        <Button
-          onClick={onCreateAgent}
-          className="border-[var(--accent-orange)] bg-[var(--accent-orange)]/10 text-[var(--accent-orange)] hover:bg-[var(--accent-orange)]/20 hover:text-[var(--accent-orange)]"
-          variant="outline"
-        >
-          <Plus className="mr-1.5 h-4 w-4" />
-          New Agent
-        </Button>
-      </div>
-    </div>
-  )
-}

packages/browseros-agent/apps/agent/entrypoints/app/agents/AgentsPage.tsx

@@ -3,9 +3,8 @@ import { type FC, useMemo, useState } from 'react'
 import { useNavigate } from 'react-router'
 import { useLlmProviders } from '@/lib/llm-providers/useLlmProviders'
 import { AgentList } from './AgentList'
-import { AgentsHeader } from './AgentsHeader'
 import { AgentTerminal } from './AgentTerminal'
-import type { HarnessAgent, HarnessAgentAdapter } from './agent-harness-types'
+import type { HarnessAgentAdapter } from './agent-harness-types'
 import { createAgentPageActions } from './agents-page-actions'
 import {
   useDefaultAgentName,
@@ -30,9 +29,9 @@ import {
   toHarnessListItem,
   toOpenClawListItem,
 } from './agents-page-utils'
-import { GatewayStatusBar } from './GatewayStatusBar'
 import { NewAgentDialog } from './NewAgentDialog'
 import {
+  AgentsPageHeader,
   ControlPlaneAlert,
   GatewayStateCards,
   InlineErrorAlert,
@@ -44,45 +43,51 @@ import {
   useCreateHarnessAgent,
   useDeleteHarnessAgent,
   useHarnessAgents,
-  useUpdateHarnessAgent,
 } from './useAgents'
-import { useOpenClawAgents, useOpenClawMutations } from './useOpenClaw'
+import {
+  useOpenClawAgents,
+  useOpenClawMutations,
+  useOpenClawStatus,
+} from './useOpenClaw'
 
 export const AgentsPage: FC = () => {
   const navigate = useNavigate()
+  const {
+    status,
+    loading: statusLoading,
+    error: statusError,
+    refetch: refetchStatus,
+  } = useOpenClawStatus()
   const { providers, defaultProviderId } = useLlmProviders()
   const {
     adapters,
     loading: adaptersLoading,
     error: adaptersError,
+    refetch: refetchAdapters,
   } = useAgentAdapters()
 
-  // The harness listing now carries the gateway lifecycle snapshot
-  // alongside the agents — one polling source for everything the
-  // agents page renders. The legacy `/claw/status` poll is dead from
-  // this surface; the chat-panel layout still uses it for now.
-  const {
-    harnessAgents,
-    gateway: status,
-    loading: harnessAgentsLoading,
-    error: harnessAgentsError,
-  } = useHarnessAgents()
-
   const openClawAgentsEnabled =
     status?.status === 'running' && status.controlPlaneStatus === 'connected'
   const {
     agents: openClawAgents,
     loading: openClawAgentsLoading,
     error: openClawAgentsError,
+    refetch: refetchOpenClawAgents,
   } = useOpenClawAgents(openClawAgentsEnabled)
+  const {
+    harnessAgents,
+    loading: harnessAgentsLoading,
+    error: harnessAgentsError,
+    refetch: refetchHarnessAgents,
+  } = useHarnessAgents()
   const createHarnessAgent = useCreateHarnessAgent()
   const deleteHarnessAgent = useDeleteHarnessAgent()
-  const updateHarnessAgent = useUpdateHarnessAgent()
   const {
     setupOpenClaw,
     createAgent: createOpenClawAgent,
     deleteAgent: deleteOpenClawAgent,
     startOpenClaw,
+    stopOpenClaw,
     restartOpenClaw,
     reconnectOpenClaw,
     actionInProgress,
@@ -153,68 +158,42 @@ export const AgentsPage: FC = () => {
     openClawAgentsEnabled,
     openClawAgents,
   )
-  const agentListItems = useMemo(() => {
-    // Dual-created OpenClaw agents (and the backfilled `main`/orphans
-    // post Step 9) live in both `/claw/agents` and `/agents` under the
-    // same id. Prefer the harness entry — it carries adapter/model/
-    // reasoning/lastUsedAt/status that the chat path actually uses —
-    // and drop the legacy duplicate so the rail doesn't show every
-    // OpenClaw agent twice.
-    const harnessIds = new Set(harnessAgents.map((agent) => agent.id))
-    const dedupedOpenClawAgents = visibleOpenClawAgents.filter(
-      (agent) => !harnessIds.has(agent.agentId),
-    )
-    return [
-      ...dedupedOpenClawAgents.map((agent) =>
+  const agentListItems = useMemo(
+    () => [
+      ...visibleOpenClawAgents.map((agent) =>
         toOpenClawListItem(agent, openClawManageable),
       ),
       ...harnessAgents.map(toHarnessListItem),
-    ]
-  }, [harnessAgents, openClawManageable, visibleOpenClawAgents])
-  // Lookup map so AgentList can render adapter chips, reasoning, etc.
-  // Computed up here to keep all hooks above the early returns below.
-  const harnessAgentLookup = useMemo(() => {
-    const map = new Map<string, HarnessAgent>()
-    for (const agent of harnessAgents) map.set(agent.id, agent)
-    return map
-  }, [harnessAgents])
-  // Activity map keyed by agentId. Sourced from the harness listing's
-  // server-side enrichment (`status` + `lastUsedAt`). Legacy gateway
-  // agents that don't have a harness record yet (rare post-backfill)
-  // simply miss from the map and render with the default `unknown`
-  // dot until reconciliation picks them up.
-  const agentActivity = useMemo(() => {
-    const map: Record<
-      string,
-      {
-        status: 'working' | 'idle' | 'asleep' | 'error'
-        lastUsedAt: number | null
-      }
-    > = {}
-    for (const agent of harnessAgents) {
-      if (!agent.status) continue
-      map[agent.id] = {
-        status: agent.status,
-        lastUsedAt: agent.lastUsedAt ?? null,
-      }
-    }
-    return map
-  }, [harnessAgents])
+    ],
+    [harnessAgents, openClawManageable, visibleOpenClawAgents],
+  )
   const inlineError = getInlineError({
     lifecyclePending,
     pageError,
+    statusError,
     openClawAgentsError,
     adaptersError,
     harnessAgentsError,
   })
   const agentsLoading = getAgentsLoading({
+    statusLoading,
     adaptersLoading,
     harnessAgentsLoading,
+    openClawAgentsEnabled,
     openClawAgentsLoading,
   })
   const creatingAgent = creatingOpenClawAgent || createHarnessAgent.isPending
   const deletingAgent = deletingOpenClawAgent || deleteHarnessAgent.isPending
 
+  const refreshAll = async () => {
+    await Promise.all([
+      refetchStatus(),
+      refetchAdapters(),
+      refetchHarnessAgents(),
+      openClawAgentsEnabled ? refetchOpenClawAgents() : Promise.resolve(),
+    ])
+  }
+
   const handleHarnessAdapterChange = (adapter: HarnessAgentAdapter) => {
     const descriptor = adapters.find((entry) => entry.id === adapter)
     setHarnessAdapterId(adapter)
@@ -260,9 +239,7 @@ export const AgentsPage: FC = () => {
     )
   }
 
-  // First-paint loader: until the harness listing has resolved at
-  // least once we don't know which adapters / agents to render.
-  if (harnessAgentsLoading && !status) {
+  if (statusLoading && !status) {
     return (
       <div className="flex items-center justify-center py-20">
         <Loader2 className="size-6 animate-spin text-muted-foreground" />
@@ -278,18 +255,27 @@ export const AgentsPage: FC = () => {
   const recoveryDetail = status ? getRecoveryDetail(status) : null
   const controlPlaneCopy = getControlPlaneCopyForStatus(status)
 
-  // Bar only makes sense when the gateway is meaningfully alive AND
-  // there's at least one OpenClaw agent in the merged list. Hide it
-  // for Claude/Codex-only setups so the page stays uncluttered.
-  const showGatewayStatusBar =
-    status?.status === 'running' &&
-    (visibleOpenClawAgents.length > 0 ||
-      harnessAgents.some((agent) => agent.adapter === 'openclaw'))
-
   return (
     <div className="min-h-full bg-background px-6 py-8">
-      <div className="fade-in slide-in-from-bottom-5 mx-auto flex w-full max-w-5xl animate-in flex-col gap-6 duration-500">
-        <AgentsHeader onCreateAgent={() => setCreateOpen(true)} />
+      <div className="mx-auto flex w-full max-w-5xl flex-col gap-6">
+        <AgentsPageHeader
+          actionInProgress={actionInProgress}
+          controlPlaneBusy={gatewayUiState.controlPlaneBusy}
+          reconnecting={reconnecting}
+          status={status}
+          onCreateAgent={() => setCreateOpen(true)}
+          onOpenTerminal={() => setShowTerminal(true)}
+          onReconnect={() => {
+            void runWithPageErrorHandling(reconnectOpenClaw)
+          }}
+          onRefresh={() => void refreshAll()}
+          onRestart={() => {
+            void runWithPageErrorHandling(restartOpenClaw)
+          }}
+          onStop={() => {
+            void runWithPageErrorHandling(stopOpenClaw)
+          }}
+        />
 
         {lifecycleBanner ? <LifecycleAlert message={lifecycleBanner} /> : null}
 
@@ -329,39 +315,15 @@ export const AgentsPage: FC = () => {
           }}
         />
 
-        {showGatewayStatusBar ? (
-          <GatewayStatusBar
-            status={status}
-            actionInProgress={actionInProgress}
-            onOpenTerminal={() => setShowTerminal(true)}
-            onRestart={() => {
-              void runWithPageErrorHandling(restartOpenClaw)
-            }}
-          />
-        ) : null}
-
         <AgentList
           agents={agentListItems}
-          activity={agentActivity}
-          harnessAgentLookup={harnessAgentLookup}
-          adapters={adapters}
           loading={agentsLoading}
           deletingAgentKey={deletingAgent ? deletingAgentKey : null}
+          onChatAgent={(agent) => navigate(`/agents/${agent.agentId}`)}
           onCreateAgent={() => setCreateOpen(true)}
           onDeleteAgent={(agent) => {
             void handleDelete(agent)
           }}
-          onPinToggle={(agent, next) => {
-            // Optimistic mutation; harness-only — gateway-original
-            // OpenClaw entries are gated server-side via the harness
-            // backfill, so we only fire when the row maps to a
-            // harness agent record.
-            if (!harnessAgentLookup.has(agent.agentId)) return
-            updateHarnessAgent.mutate({
-              agentId: agent.agentId,
-              patch: { pinned: next },
-            })
-          }}
         />
 
         <SetupOpenClawDialog

packages/browseros-agent/apps/agent/entrypoints/app/agents/GatewayStatusBar.tsx

@@ -1,206 +0,0 @@
-import { Loader2, RotateCcw, Terminal } from 'lucide-react'
-import type { FC, ReactNode } from 'react'
-import { Badge } from '@/components/ui/badge'
-import { Button } from '@/components/ui/button'
-import { Separator } from '@/components/ui/separator'
-import {
-  Tooltip,
-  TooltipContent,
-  TooltipProvider,
-  TooltipTrigger,
-} from '@/components/ui/tooltip'
-import { cn } from '@/lib/utils'
-import type { OpenClawStatus } from './useOpenClaw'
-
-interface GatewayStatusBarProps {
-  status: OpenClawStatus | null
-  /** Disabled while a gateway lifecycle mutation is mid-flight. */
-  actionInProgress: boolean
-  onOpenTerminal: () => void
-  onRestart: () => void
-}
-
-/**
- * Compact one-line status bar for the OpenClaw gateway. Renders the
- * lifecycle pills (Running / Control plane connected) plus a Terminal
- * escape hatch and a Restart Gateway action. Lives between the page
- * header and the agent list when at least one OpenClaw agent is in
- * the merged list; collapses to nothing for Claude/Codex-only setups.
- *
- * Status is sourced from `GET /agents`'s `gateway` field — the agents
- * page no longer polls `/claw/status` directly. One endpoint, one
- * 5s interval, no duplicate state.
- */
-export const GatewayStatusBar: FC<GatewayStatusBarProps> = ({
-  status,
-  actionInProgress,
-  onOpenTerminal,
-  onRestart,
-}) => {
-  if (!status) return null
-
-  const runningPill = pillForRuntimeStatus(status.status)
-  const controlPlanePill = pillForControlPlane(status.controlPlaneStatus)
-
-  return (
-    <div className="rounded-xl border border-border bg-card px-4 py-3 shadow-sm">
-      <div className="flex items-center gap-3 text-sm">
-        <span className="font-medium text-muted-foreground">
-          OpenClaw gateway
-        </span>
-        <Badge
-          variant={runningPill.variant}
-          className={cn('gap-1.5', runningPill.className)}
-        >
-          <span
-            className={cn(
-              'inline-block h-1.5 w-1.5 rounded-full',
-              runningPill.dot,
-            )}
-          />
-          {runningPill.label}
-        </Badge>
-        <Badge
-          variant={controlPlanePill.variant}
-          className={cn('gap-1.5', controlPlanePill.className)}
-        >
-          <span
-            className={cn(
-              'inline-block h-1.5 w-1.5 rounded-full',
-              controlPlanePill.dot,
-            )}
-          />
-          {controlPlanePill.label}
-        </Badge>
-        <Separator orientation="vertical" className="h-4" />
-        <WithTooltip label="Open a shell into the OpenClaw gateway container for raw CLI access (config edits, session inspection).">
-          <Button variant="ghost" size="sm" onClick={onOpenTerminal}>
-            <Terminal className="mr-1.5 h-3.5 w-3.5" />
-            Terminal
-          </Button>
-        </WithTooltip>
-        <WithTooltip label="Restart the OpenClaw gateway. Useful when the gateway is stuck or after editing provider config.">
-          <Button
-            variant="ghost"
-            size="sm"
-            onClick={onRestart}
-            disabled={actionInProgress}
-            className="ml-auto"
-          >
-            {actionInProgress ? (
-              <Loader2 className="mr-1.5 h-3.5 w-3.5 animate-spin" />
-            ) : (
-              <RotateCcw className="mr-1.5 h-3.5 w-3.5" />
-            )}
-            Restart Gateway
-          </Button>
-        </WithTooltip>
-      </div>
-    </div>
-  )
-}
-
-const WithTooltip: FC<{ label: string; children: ReactNode }> = ({
-  label,
-  children,
-}) => (
-  <TooltipProvider delayDuration={250}>
-    <Tooltip>
-      <TooltipTrigger asChild>{children}</TooltipTrigger>
-      <TooltipContent side="bottom" className="max-w-xs text-xs">
-        {label}
-      </TooltipContent>
-    </Tooltip>
-  </TooltipProvider>
-)
-
-type PillKind = {
-  variant: 'default' | 'secondary' | 'outline' | 'destructive'
-  label: string
-  dot: string
-  className?: string
-}
-
-function pillForRuntimeStatus(status: OpenClawStatus['status']): PillKind {
-  switch (status) {
-    case 'running':
-      return {
-        variant: 'secondary',
-        label: 'Running',
-        dot: 'bg-emerald-500',
-        className: 'bg-emerald-50 text-emerald-900 hover:bg-emerald-50',
-      }
-    case 'starting':
-      return {
-        variant: 'secondary',
-        label: 'Starting',
-        dot: 'bg-amber-500 animate-pulse',
-        className: 'bg-amber-50 text-amber-900 hover:bg-amber-50',
-      }
-    case 'stopped':
-      return {
-        variant: 'outline',
-        label: 'Stopped',
-        dot: 'bg-muted-foreground/40',
-      }
-    case 'error':
-      return {
-        variant: 'destructive',
-        label: 'Error',
-        dot: 'bg-destructive-foreground',
-      }
-    default:
-      return {
-        variant: 'outline',
-        label: 'Unknown',
-        dot: 'bg-muted-foreground/40',
-      }
-  }
-}
-
-function pillForControlPlane(
-  status: OpenClawStatus['controlPlaneStatus'],
-): PillKind {
-  switch (status) {
-    case 'connected':
-      return {
-        variant: 'secondary',
-        label: 'Control plane connected',
-        dot: 'bg-emerald-500',
-        className: 'bg-emerald-50 text-emerald-900 hover:bg-emerald-50',
-      }
-    case 'connecting':
-      return {
-        variant: 'secondary',
-        label: 'Connecting',
-        dot: 'bg-amber-500 animate-pulse',
-        className: 'bg-amber-50 text-amber-900 hover:bg-amber-50',
-      }
-    case 'reconnecting':
-      return {
-        variant: 'secondary',
-        label: 'Reconnecting',
-        dot: 'bg-amber-500 animate-pulse',
-        className: 'bg-amber-50 text-amber-900 hover:bg-amber-50',
-      }
-    case 'recovering':
-      return {
-        variant: 'secondary',
-        label: 'Recovering',
-        dot: 'bg-amber-500 animate-pulse',
-        className: 'bg-amber-50 text-amber-900 hover:bg-amber-50',
-      }
-    case 'failed':
-      return {
-        variant: 'destructive',
-        label: 'Needs attention',
-        dot: 'bg-destructive-foreground',
-      }
-    default:
-      return {
-        variant: 'outline',
-        label: 'Disconnected',
-        dot: 'bg-muted-foreground/40',
-      }
-  }
-}

packages/browseros-agent/apps/agent/entrypoints/app/agents/LivenessDot.tsx

@@ -1,83 +0,0 @@
-import type { FC } from 'react'
-import {
-  Tooltip,
-  TooltipContent,
-  TooltipProvider,
-  TooltipTrigger,
-} from '@/components/ui/tooltip'
-import { cn } from '@/lib/utils'
-
-export type AgentLiveness = 'working' | 'idle' | 'asleep' | 'error' | 'unknown'
-
-interface LivenessDotProps {
-  status: AgentLiveness
-  /**
-   * Optional human-friendly secondary line, e.g. "Idle for 4 min" or
-   * "Asleep — no activity for 22 min". When absent the tooltip just
-   * reads the status label.
-   */
-  detail?: string
-  className?: string
-}
-
-const VARIANT: Record<
-  AgentLiveness,
-  { dot: string; ring: string; label: string }
-> = {
-  working: {
-    // Animated amber pulse + soft halo so the eye catches an active
-    // agent in a long list without the dot screaming for attention.
-    dot: 'bg-amber-500 animate-pulse',
-    ring: 'ring-2 ring-amber-200',
-    label: 'Working on a turn',
-  },
-  idle: {
-    dot: 'bg-emerald-500',
-    ring: 'ring-2 ring-emerald-100',
-    label: 'Idle',
-  },
-  asleep: {
-    dot: 'bg-muted-foreground/40',
-    ring: 'ring-2 ring-muted',
-    label: 'Asleep',
-  },
-  error: {
-    dot: 'bg-destructive',
-    ring: 'ring-2 ring-destructive/30',
-    label: 'Attention',
-  },
-  unknown: {
-    dot: 'bg-muted-foreground/30',
-    ring: 'ring-2 ring-muted',
-    label: 'Status unknown',
-  },
-}
-
-export const LivenessDot: FC<LivenessDotProps> = ({
-  status,
-  detail,
-  className,
-}) => {
-  const variant = VARIANT[status]
-  return (
-    <TooltipProvider delayDuration={150}>
-      <Tooltip>
-        <TooltipTrigger asChild>
-          <span
-            role="img"
-            aria-label={detail ?? variant.label}
-            className={cn(
-              'inline-block h-3 w-3 rounded-full',
-              variant.dot,
-              variant.ring,
-              className,
-            )}
-          />
-        </TooltipTrigger>
-        <TooltipContent side="right" className="text-xs">
-          {detail ?? variant.label}
-        </TooltipContent>
-      </Tooltip>
-    </TooltipProvider>
-  )
-}

packages/browseros-agent/apps/agent/entrypoints/app/agents/NewAgentDialog.tsx

@@ -154,6 +154,7 @@ export const NewAgentDialog: FC<NewAgentDialogProps> = ({
                 <SelectValue />
               </SelectTrigger>
               <SelectContent>
+                <SelectItem value="openclaw">OpenClaw</SelectItem>
                 {adapters.map((adapter) => (
                   <SelectItem key={adapter.id} value={adapter.id}>
                     {adapter.name}

packages/browseros-agent/apps/agent/entrypoints/app/agents/agent-display.helpers.ts

@@ -1,107 +0,0 @@
-import type { AgentListItem } from './agents-page-types'
-import type { AgentLiveness } from './LivenessDot'
-
-/**
- * Display rules for the redesigned agent rows. Pure helpers — no React,
- * no API calls — so they're trivial to unit-test and the row card stays
- * focused on layout.
- */
-
-const UUID_PATTERN =
-  /^[0-9a-f]{8}-[0-9a-f]{4}-[0-9a-f]{4}-[0-9a-f]{4}-[0-9a-f]{12}$/i
-
-const OC_UUID_PATTERN =
-  /^oc-[0-9a-f]{8}-[0-9a-f]{4}-[0-9a-f]{4}-[0-9a-f]{4}-[0-9a-f]{12}$/i
-
-/**
- * The agent rail used to render whatever the gateway returned for `name`.
- * Post-migration that's frequently the agent's UUID — readable to nobody.
- * Prefer the explicit `name` when it differs meaningfully from the id;
- * otherwise fall back to a short prefix users can recognize on second
- * glance.
- */
-export function displayName(agent: AgentListItem): string {
-  const name = agent.name?.trim()
-  const id = agent.agentId
-  if (!name || name === id) {
-    if (OC_UUID_PATTERN.test(id)) return id.slice(0, 11) // "oc-XXXXXXXX"
-    if (UUID_PATTERN.test(id)) return id.slice(0, 8)
-    return id
-  }
-  return name
-}
-
-export function canDelete(agent: AgentListItem): boolean {
-  // The gateway's protected `main` agent must not be deletable. The
-  // server enforces this too, but disabling the menu item avoids users
-  // hitting an opaque 400.
-  if (agent.agentId === 'main') return false
-  return agent.canDelete
-}
-
-/**
- * Rename will be wired to a future `PATCH /agents/:id` endpoint. The
- * legacy `/claw/agents` create flow named the agent on the gateway via
- * the `name` field but the field isn't editable post-create today.
- */
-export function canRename(_agent: AgentListItem): boolean {
-  return false
-}
-
-/**
- * The detail line carries the agent's workspace path. The `detail`
- * field on AgentListItem already holds it for OpenClaw entries
- * (`/home/node/.openclaw/workspace-...`); for harness agents it's the
- * synthetic `<adapter>:main` marker that's not informative — hide it.
- */
-export function workspaceLabel(agent: AgentListItem): string | null {
-  if (!agent.detail) return null
-  if (/^(claude|codex|openclaw):main$/.test(agent.detail)) return null
-  return agent.detail
-}
-
-const ONE_MINUTE = 60_000
-const ONE_HOUR = 60 * ONE_MINUTE
-const ONE_DAY = 24 * ONE_HOUR
-
-/**
- * Lightweight relative-time formatter. We don't want to drag in
- * `dayjs/relativeTime` just for a few labels.
- */
-export function formatRelativeTime(epochMs: number | null): string {
-  if (epochMs === null || !Number.isFinite(epochMs)) return 'never'
-  const diff = Math.max(0, Date.now() - epochMs)
-  if (diff < ONE_MINUTE) return 'just now'
-  if (diff < ONE_HOUR) {
-    const m = Math.floor(diff / ONE_MINUTE)
-    return `${m} min ago`
-  }
-  if (diff < ONE_DAY) {
-    const h = Math.floor(diff / ONE_HOUR)
-    return h === 1 ? '1 hr ago' : `${h} hr ago`
-  }
-  const d = Math.floor(diff / ONE_DAY)
-  return d === 1 ? '1 day ago' : `${d} days ago`
-}
-
-/**
- * Tooltip-friendly description of a row's current liveness state.
- * Returns `undefined` when the state has nothing extra to add (e.g.
- * `unknown` with no timestamp).
- */
-export function livenessDetail(
-  status: AgentLiveness,
-  lastUsedAt: number | null | undefined,
-): string | undefined {
-  if (lastUsedAt == null) return undefined
-  const diffMin = Math.floor((Date.now() - lastUsedAt) / 60_000)
-  if (status === 'idle') return `Idle for ${Math.max(0, diffMin)} min`
-  if (status === 'asleep') {
-    if (diffMin < 60) return `Asleep — quiet for ${diffMin} min`
-    const hr = Math.floor(diffMin / 60)
-    return `Asleep — quiet for ${hr} hr`
-  }
-  if (status === 'working') return 'Working on a turn'
-  if (status === 'error') return 'Attention — last turn failed'
-  return undefined
-}

packages/browseros-agent/apps/agent/entrypoints/app/agents/agent-harness-types.ts

@@ -1,6 +1,6 @@
 import type { AgentEntry } from './useOpenClaw'
 
-export type HarnessAgentAdapter = 'claude' | 'codex' | 'openclaw'
+export type HarnessAgentAdapter = 'claude' | 'codex'
 
 export type AgentHarnessStreamEvent =
   | {
@@ -33,8 +33,6 @@ export type AgentHarnessStreamEvent =
       code?: string
     }
 
-export type HarnessAgentLiveness = 'working' | 'idle' | 'asleep' | 'error'
-
 export interface HarnessAgent {
   id: string
   name: string
@@ -45,54 +43,6 @@ export interface HarnessAgent {
   sessionKey: string
   createdAt: number
   updatedAt: number
-  /**
-   * Server-derived liveness state. When the listing endpoint hasn't
-   * been enriched yet (older deployments) this is undefined and the UI
-   * falls back to `unknown`.
-   */
-  status?: HarnessAgentLiveness
-  /**
-   * Wall-clock ms of the last persisted turn. `null` for never-used
-   * agents. Drives the recency sort and the "Last used X min ago" copy.
-   */
-  lastUsedAt?: number | null
-  /** Pinned agents float to the top of the list. Defaults to `false`. */
-  pinned?: boolean
-  /** First non-blank line of the most recent user message; null if none. */
-  lastUserMessage?: string | null
-  /** Working directory the agent runs in; null when no session record yet. */
-  cwd?: string | null
-  /** Cumulative + 7-day rolling token usage; null when no record. */
-  tokens?: {
-    last7d: { input: number; output: number; requestCount: number }
-    cumulative: { input: number; output: number }
-  } | null
-  turnsByDay?: number[]
-  failedByDay?: number[]
-  lastError?: string | null
-  lastErrorAt?: number | null
-  /** When non-null, an in-flight turn this row can be resumed from. */
-  activeTurnId?: string | null
-  /** Persistent FIFO queue of messages waiting for this agent. */
-  queue?: HarnessQueuedMessage[]
-}
-
-export interface HarnessQueuedMessageAttachment {
-  mediaType: string
-  data: string
-}
-
-export interface HarnessQueuedMessage {
-  id: string
-  createdAt: number
-  message: string
-  attachments?: ReadonlyArray<HarnessQueuedMessageAttachment>
-}
-
-export interface HarnessAdapterHealth {
-  healthy: boolean
-  reason?: string
-  checkedAt: number
 }
 
 export interface HarnessAdapterDescriptor {
@@ -103,7 +53,6 @@ export interface HarnessAdapterDescriptor {
   modelControl: 'runtime-supported' | 'best-effort'
   models: Array<{ id: string; label: string; recommended?: boolean }>
   reasoningEfforts: Array<{ id: string; label: string; recommended?: boolean }>
-  health?: HarnessAdapterHealth
 }
 
 export interface CreateHarnessAgentInput {
@@ -113,36 +62,19 @@ export interface CreateHarnessAgentInput {
   reasoningEffort?: string
 }
 
-export interface HarnessHistoryReasoning {
-  text: string
-  durationMs?: number
-}
-
-export interface HarnessHistoryToolCall {
-  toolCallId?: string
-  toolName: string
-  status: 'pending' | 'running' | 'completed' | 'failed'
-  input?: unknown
-  output?: unknown
-  error?: string
-  durationMs?: number
-}
-
-export interface HarnessHistoryEntry {
+export interface HarnessTranscriptEntry {
   id: string
   agentId: string
   sessionId: 'main'
   role: 'user' | 'assistant'
   text: string
   createdAt: number
-  reasoning?: HarnessHistoryReasoning
-  toolCalls?: HarnessHistoryToolCall[]
 }
 
 export interface HarnessAgentHistoryPage {
   agentId: string
   sessionId: 'main'
-  items: HarnessHistoryEntry[]
+  items: HarnessTranscriptEntry[]
 }
 
 export function mapHarnessAgentToEntry(agent: HarnessAgent): AgentEntry {

packages/browseros-agent/apps/agent/entrypoints/app/agents/agent-row/AgentActions.tsx

@@ -1,160 +0,0 @@
-import {
-  Copy,
-  Loader2,
-  MessageSquare,
-  MoreHorizontal,
-  Pencil,
-  RotateCcw,
-  Trash2,
-} from 'lucide-react'
-import type { FC } from 'react'
-import { useNavigate } from 'react-router'
-import { toast } from 'sonner'
-import { Button } from '@/components/ui/button'
-import {
-  DropdownMenu,
-  DropdownMenuContent,
-  DropdownMenuItem,
-  DropdownMenuSeparator,
-  DropdownMenuTrigger,
-} from '@/components/ui/dropdown-menu'
-import {
-  Tooltip,
-  TooltipContent,
-  TooltipProvider,
-  TooltipTrigger,
-} from '@/components/ui/tooltip'
-import {
-  canDelete as canDeleteAgent,
-  canRename as canRenameAgent,
-  displayName,
-} from '../agent-display.helpers'
-import type { AgentListItem } from '../agents-page-types'
-
-interface AgentActionsProps {
-  agent: AgentListItem
-  activeTurnId: string | null
-  deleting?: boolean
-  onDelete: (agent: AgentListItem) => void
-}
-
-/**
- * Single primary CTA per row: `Resume` (filled, accent-orange, with a
- * pulsing dot) when an active turn exists; otherwise `Chat` (outline).
- * Both navigate to the same place — the chat hook auto-attaches via
- * `/chat/active` when there's a live turn — but the row signals which
- * action the user is actually taking.
- */
-export const AgentActions: FC<AgentActionsProps> = ({
-  agent,
-  activeTurnId,
-  deleting,
-  onDelete,
-}) => {
-  const navigate = useNavigate()
-  const allowDelete = canDeleteAgent(agent)
-  const allowRename = canRenameAgent(agent)
-
-  const handleChat = () => navigate(`/agents/${agent.agentId}`)
-  const handleCopyId = async () => {
-    try {
-      await navigator.clipboard.writeText(agent.agentId)
-      toast.success('Agent id copied')
-    } catch {
-      toast.error('Could not copy agent id')
-    }
-  }
-
-  return (
-    <div className="flex shrink-0 items-center gap-1.5">
-      {activeTurnId ? (
-        <Button
-          variant="default"
-          size="sm"
-          onClick={handleChat}
-          className="gap-2 bg-[var(--accent-orange)] text-white shadow-sm hover:bg-[var(--accent-orange)]/90"
-        >
-          <span className="relative flex size-2">
-            <span className="absolute inline-flex h-full w-full animate-ping rounded-full bg-white/70 opacity-75" />
-            <span className="relative inline-flex size-2 rounded-full bg-white" />
-          </span>
-          Resume
-        </Button>
-      ) : (
-        <Button variant="outline" size="sm" onClick={handleChat}>
-          <MessageSquare className="mr-1.5 size-3" />
-          Chat
-        </Button>
-      )}
-      <DropdownMenu>
-        <DropdownMenuTrigger asChild>
-          <Button
-            variant="ghost"
-            size="icon"
-            aria-label={`More actions for ${displayName(agent)}`}
-            className="size-8 text-muted-foreground hover:text-foreground"
-          >
-            <MoreHorizontal className="size-4" />
-          </Button>
-        </DropdownMenuTrigger>
-        <DropdownMenuContent align="end" className="w-44">
-          <DropdownMenuItem onSelect={() => void handleCopyId()}>
-            <Copy className="mr-2 size-3.5" />
-            Copy id
-          </DropdownMenuItem>
-          <ComingSoonItem
-            icon={Pencil}
-            label="Rename"
-            disabled={!allowRename}
-          />
-          <ComingSoonItem icon={RotateCcw} label="Reset history" disabled />
-          <DropdownMenuSeparator />
-          <DropdownMenuItem
-            onSelect={() => onDelete(agent)}
-            disabled={!allowDelete || deleting}
-            className="text-destructive focus:text-destructive"
-          >
-            {deleting ? (
-              <Loader2 className="mr-2 size-3.5 animate-spin" />
-            ) : (
-              <Trash2 className="mr-2 size-3.5" />
-            )}
-            Delete
-          </DropdownMenuItem>
-        </DropdownMenuContent>
-      </DropdownMenu>
-    </div>
-  )
-}
-
-interface ComingSoonItemProps {
-  icon: typeof Pencil
-  label: string
-  disabled: boolean
-}
-
-const ComingSoonItem: FC<ComingSoonItemProps> = ({
-  icon: Icon,
-  label,
-  disabled,
-}) => {
-  const item = (
-    <DropdownMenuItem disabled className="text-muted-foreground">
-      <Icon className="mr-2 size-3.5" />
-      {label}
-    </DropdownMenuItem>
-  )
-  if (!disabled) return item
-  return (
-    <TooltipProvider delayDuration={300}>
-      <Tooltip>
-        <TooltipTrigger asChild>
-          <span className="block w-full">{item}</span>
-        </TooltipTrigger>
-        <TooltipContent side="left" className="text-xs">
-          {label} coming soon
-        </TooltipContent>
-      </Tooltip>
-    </TooltipProvider>
-  )
-}

packages/browseros-agent/apps/agent/entrypoints/app/agents/agent-row/AgentErrorPanel.tsx

@@ -1,96 +0,0 @@
-import { AlertTriangle, ChevronDown } from 'lucide-react'
-import { type FC, useEffect, useState } from 'react'
-import { Button } from '@/components/ui/button'
-import {
-  Collapsible,
-  CollapsibleContent,
-  CollapsibleTrigger,
-} from '@/components/ui/collapsible'
-import {
-  HoverCard,
-  HoverCardContent,
-  HoverCardTrigger,
-} from '@/components/ui/hover-card'
-import { cn } from '@/lib/utils'
-import { truncate } from './agent-row.helpers'
-
-interface AgentErrorPanelProps {
-  agentId: string
-  message: string
-  errorAt: number | null
-}
-
-const STORAGE_PREFIX = 'agent-row:lastErrorSeenAt:'
-const PREVIEW_CHARS = 200
-
-export const AgentErrorPanel: FC<AgentErrorPanelProps> = ({
-  agentId,
-  message,
-  errorAt,
-}) => {
-  const storageKey = `${STORAGE_PREFIX}${agentId}`
-  // Open if we've never seen this `errorAt` for this agent. Once the
-  // user collapses the panel (or refreshes after seeing it), we mark
-  // it seen so it doesn't re-pop on every poll.
-  const [open, setOpen] = useState<boolean>(() => {
-    if (typeof window === 'undefined' || !errorAt) return true
-    const seen = Number(window.localStorage.getItem(storageKey) ?? 0)
-    return !Number.isFinite(seen) || errorAt > seen
-  })
-
-  useEffect(() => {
-    if (!open && errorAt && typeof window !== 'undefined') {
-      window.localStorage.setItem(storageKey, String(errorAt))
-    }
-  }, [open, errorAt, storageKey])
-
-  const preview = truncate(message, PREVIEW_CHARS)
-  const truncated = preview.length < message.length
-
-  return (
-    <Collapsible open={open} onOpenChange={setOpen} className="mt-3">
-      <div className="flex items-center justify-between rounded-md border border-destructive/30 bg-destructive/5 px-3 py-2">
-        <div className="flex items-center gap-2 font-medium text-destructive text-xs">
-          <AlertTriangle className="size-3.5" />
-          Last error
-        </div>
-        <CollapsibleTrigger asChild>
-          <Button
-            variant="ghost"
-            size="sm"
-            className="h-6 px-2 text-muted-foreground"
-          >
-            <span className="text-xs">{open ? 'hide' : 'show'}</span>
-            <ChevronDown
-              className={cn(
-                'ml-1 size-3 transition-transform',
-                open && 'rotate-180',
-              )}
-            />
-          </Button>
-        </CollapsibleTrigger>
-      </div>
-      <CollapsibleContent>
-        <div className="mt-1 rounded-md border-destructive/30 border-x border-b bg-destructive/5 px-3 pb-2 text-xs">
-          {truncated ? (
-            <HoverCard openDelay={300}>
-              <HoverCardTrigger asChild>
-                <span className="cursor-default font-mono text-foreground/80">
-                  {preview}…
-                </span>
-              </HoverCardTrigger>
-              <HoverCardContent
-                side="bottom"
-                className="max-w-md whitespace-pre-wrap font-mono text-xs"
-              >
-                {message}
-              </HoverCardContent>
-            </HoverCard>
-          ) : (
-            <span className="font-mono text-foreground/80">{message}</span>
-          )}
-        </div>
-      </CollapsibleContent>
-    </Collapsible>
-  )
-}

packages/browseros-agent/apps/agent/entrypoints/app/agents/agent-row/AgentLastMessage.tsx

@@ -1,35 +0,0 @@
-import { Quote } from 'lucide-react'
-import type { FC } from 'react'
-import { firstNonBlankLine, truncate } from './agent-row.helpers'
-
-interface AgentLastMessageProps {
-  message: string | null
-}
-
-const PREVIEW_CHARS = 110
-
-/**
- * Inline preview of the most recent user message. Renders as a quoted,
- * italic line so the row reads like a conversation snippet rather than
- * a label-and-value pair. No hover-card — opening the agent's chat is
- * the canonical way to read the full message.
- */
-export const AgentLastMessage: FC<AgentLastMessageProps> = ({ message }) => {
-  if (!message) {
-    return (
-      <p className="mt-1 text-muted-foreground/70 text-xs italic">
-        No messages yet — start a chat
-      </p>
-    )
-  }
-  const preview = truncate(firstNonBlankLine(message), PREVIEW_CHARS)
-  return (
-    <p className="mt-1.5 flex items-start gap-1.5 text-foreground/85 text-sm italic leading-snug">
-      <Quote
-        className="mt-1 size-3 shrink-0 text-muted-foreground/60"
-        aria-hidden
-      />
-      <span className="truncate">{preview}</span>
-    </p>
-  )
-}

packages/browseros-agent/apps/agent/entrypoints/app/agents/agent-row/AgentMetaRow.tsx

@@ -1,37 +0,0 @@
-import type { FC } from 'react'
-import { formatRelativeTime } from '../agent-display.helpers'
-import { AgentTokenSummary } from './AgentTokenSummary'
-import type { AgentTokenUsage } from './agent-row.types'
-
-interface AgentMetaRowProps {
-  lastUsedAt: number | null
-  tokens: AgentTokenUsage | null
-}
-
-/**
- * Bottom-of-row meta line. Intentionally sparse — last activity time
- * and lifetime tokens. CWD is no longer surfaced here because the path
- * the server happens to be running from isn't actionable; if a future
- * surface needs the cwd (chat panel, debug view) it reads from the
- * listing payload directly.
- */
-export const AgentMetaRow: FC<AgentMetaRowProps> = ({ lastUsedAt, tokens }) => {
-  const lastUsedLabel = formatRelativeTime(lastUsedAt)
-  const tokensTotal =
-    (tokens?.cumulative.input ?? 0) + (tokens?.cumulative.output ?? 0)
-  const showTokens = tokensTotal > 0
-
-  return (
-    <div className="mt-2 flex flex-wrap items-center gap-x-2 text-muted-foreground text-xs">
-      <span>{lastUsedLabel}</span>
-      {showTokens && (
-        <>
-          <span aria-hidden className="text-muted-foreground/50">
-            ·
-          </span>
-          <AgentTokenSummary tokens={tokens} />
-        </>
-      )}
-    </div>
-  )
-}

packages/browseros-agent/apps/agent/entrypoints/app/agents/agent-row/AgentSparkline.tsx

@@ -1,92 +0,0 @@
-import type { FC } from 'react'
-import {
-  HoverCard,
-  HoverCardContent,
-  HoverCardTrigger,
-} from '@/components/ui/hover-card'
-import { cn } from '@/lib/utils'
-import { formatLocalDate, ROW_BAR_COUNT } from './agent-row.helpers'
-
-interface AgentSparklineProps {
-  /** 14 entries, oldest → newest. Today's bucket is the last index. */
-  turnsByDay: number[]
-  /** Same length, same order. Failed turns counted separately. */
-  failedByDay: number[]
-  className?: string
-}
-
-const MIN_BAR_HEIGHT_PX = 2
-const MAX_BAR_HEIGHT_PX = 18
-
-export const AgentSparkline: FC<AgentSparklineProps> = ({
-  turnsByDay,
-  failedByDay,
-  className,
-}) => {
-  if (turnsByDay.length === 0 || turnsByDay.every((n) => n === 0)) return null
-  const max = Math.max(1, ...turnsByDay)
-
-  return (
-    <HoverCard openDelay={250}>
-      <HoverCardTrigger asChild>
-        <div
-          role="img"
-          aria-label={`Last ${ROW_BAR_COUNT} days of activity`}
-          className={cn('flex h-5 items-end gap-px', className)}
-        >
-          {turnsByDay.map((count, idx) => {
-            const ratio = count / max
-            const height = Math.max(
-              MIN_BAR_HEIGHT_PX,
-              Math.round(ratio * MAX_BAR_HEIGHT_PX),
-            )
-            const isToday = idx === ROW_BAR_COUNT - 1
-            const failed = failedByDay[idx] ?? 0
-            return (
-              <div
-                // biome-ignore lint/suspicious/noArrayIndexKey: fixed-length sparkline buckets keyed by day position
-                key={`bar-${idx}`}
-                className={cn(
-                  'w-1.5 rounded-sm',
-                  count === 0
-                    ? 'bg-muted-foreground/15'
-                    : failed > 0
-                      ? 'bg-destructive/50'
-                      : 'bg-[var(--accent-orange)]/50',
-                  isToday && 'ring-1 ring-foreground/30',
-                )}
-                style={{ height }}
-              />
-            )
-          })}
-        </div>
-      </HoverCardTrigger>
-      <HoverCardContent side="left" className="w-56 text-xs">
-        <div className="mb-2 font-medium text-sm">Last 14 days</div>
-        <ul className="space-y-0.5">
-          {turnsByDay.map((count, idx) => {
-            const failed = failedByDay[idx] ?? 0
-            const dayLabel = formatLocalDate(idx)
-            return (
-              <li
-                // biome-ignore lint/suspicious/noArrayIndexKey: fixed-length list keyed by day position
-                key={`day-${idx}`}
-                className="flex items-center justify-between text-muted-foreground"
-              >
-                <span>{dayLabel}</span>
-                <span>
-                  {count}
-                  {failed > 0 && (
-                    <span className="ml-1 text-destructive">
-                      ({failed} failed)
-                    </span>
-                  )}
-                </span>
-              </li>
-            )
-          })}
-        </ul>
-      </HoverCardContent>
-    </HoverCard>
-  )
-}

packages/browseros-agent/apps/agent/entrypoints/app/agents/agent-row/AgentSummaryChips.tsx

@@ -1,71 +0,0 @@
-import { TriangleAlert } from 'lucide-react'
-import type { FC } from 'react'
-import { Badge } from '@/components/ui/badge'
-import {
-  HoverCard,
-  HoverCardContent,
-  HoverCardTrigger,
-} from '@/components/ui/hover-card'
-import { cn } from '@/lib/utils'
-import { adapterLabel } from '../AdapterIcon'
-import type { HarnessAgentAdapter } from '../agent-harness-types'
-import type { AgentAdapterHealth } from './agent-row.types'
-
-interface AgentSummaryChipsProps {
-  adapter: HarnessAgentAdapter | 'unknown'
-  modelLabel: string | null
-  reasoningEffort: string | null
-  /** When unhealthy, the adapter label dims and a warning chip appears. */
-  adapterHealth: AgentAdapterHealth | null
-}
-
-/**
- * Adapter / model / reasoning summary line. Always rendered (so OpenClaw
- * rows that fall back to defaults still expose what they're set up to do)
- * and surfaces adapter-health *only when unhealthy* — keeping the calm
- * default state silent and reserving visual noise for things the user
- * needs to act on.
- */
-export const AgentSummaryChips: FC<AgentSummaryChipsProps> = ({
-  adapter,
-  modelLabel,
-  reasoningEffort,
-  adapterHealth,
-}) => {
-  const parts = [adapterLabel(adapter)]
-  if (modelLabel) parts.push(modelLabel)
-  if (reasoningEffort) parts.push(reasoningEffort)
-  const unhealthy = adapterHealth?.healthy === false
-  return (
-    <div
-      className={cn(
-        'flex items-center gap-1.5 text-muted-foreground text-xs',
-        unhealthy && 'text-muted-foreground/70',
-      )}
-    >
-      <span className="truncate">{parts.join(' · ')}</span>
-      {unhealthy && adapterHealth && (
-        <HoverCard openDelay={200}>
-          <HoverCardTrigger asChild>
-            <Badge
-              variant="outline"
-              className="h-5 cursor-default gap-1 border-amber-500/40 bg-amber-50 px-1.5 text-amber-900 hover:bg-amber-50"
-            >
-              <TriangleAlert className="size-2.5" />
-              <span className="font-normal">Unavailable</span>
-            </Badge>
-          </HoverCardTrigger>
-          <HoverCardContent side="right" className="w-72 text-sm">
-            <div className="font-medium">
-              {adapterLabel(adapter)} CLI not available
-            </div>
-            <div className="mt-1 text-muted-foreground text-xs">
-              {adapterHealth.reason ??
-                'Adapter binary missing on $PATH. Install it from the adapter docs to use this agent.'}
-            </div>
-          </HoverCardContent>
-        </HoverCard>
-      )}
-    </div>
-  )
-}

packages/browseros-agent/apps/agent/entrypoints/app/agents/agent-row/AgentTile.tsx

@@ -1,37 +0,0 @@
-import type { FC } from 'react'
-import { cn } from '@/lib/utils'
-import { AdapterIcon } from '../AdapterIcon'
-import { livenessDetail } from '../agent-display.helpers'
-import type { HarnessAgentAdapter } from '../agent-harness-types'
-import { type AgentLiveness, LivenessDot } from '../LivenessDot'
-
-export interface AgentTileProps {
-  adapter: HarnessAgentAdapter | 'unknown'
-  status: AgentLiveness
-  lastUsedAt: number | null
-}
-
-/**
- * Adapter glyph + a single liveness dot. Adapter health is no longer
- * surfaced here — it lives as an inline pill inside `AgentSummaryChips`
- * so the user isn't asked to disambiguate two dots on the same tile.
- */
-export const AgentTile: FC<AgentTileProps> = ({
-  adapter,
-  status,
-  lastUsedAt,
-}) => (
-  <div className="relative shrink-0">
-    <div className="flex h-12 w-12 items-center justify-center rounded-xl bg-muted text-muted-foreground">
-      <AdapterIcon adapter={adapter} className="h-6 w-6" />
-    </div>
-    <LivenessDot
-      status={status}
-      detail={livenessDetail(status, lastUsedAt)}
-      className={cn(
-        'absolute -right-0.5 -bottom-0.5',
-        status === 'working' && 'animate-pulse',
-      )}
-    />
-  </div>
-)

packages/browseros-agent/apps/agent/entrypoints/app/agents/agent-row/AgentTitleRow.tsx

@@ -1,55 +0,0 @@
-import type { FC } from 'react'
-import { Badge } from '@/components/ui/badge'
-import { displayName } from '../agent-display.helpers'
-import type { AgentListItem } from '../agents-page-types'
-import type { AgentLiveness } from '../LivenessDot'
-import { AgentSparkline } from './AgentSparkline'
-import { PinToggle } from './PinToggle'
-
-interface AgentTitleRowProps {
-  agent: AgentListItem
-  status: AgentLiveness
-  pinned: boolean
-  turnsByDay: number[]
-  failedByDay: number[]
-  onPinToggle: (next: boolean) => void
-}
-
-/**
- * Title strip: name + status badge + (right-aligned) sparkline. The
- * pin toggle sits trailing the title so the title always flushes left
- * regardless of pin state — moving the star left of the title indents
- * the row's first line off-axis from the model/preview/meta lines
- * below it. When unpinned and not hovered, the toggle is removed from
- * layout entirely so it reserves no space at all.
- */
-export const AgentTitleRow: FC<AgentTitleRowProps> = ({
-  agent,
-  status,
-  pinned,
-  turnsByDay,
-  failedByDay,
-  onPinToggle,
-}) => (
-  <div className="mb-1 flex items-center gap-2">
-    <span className="truncate font-semibold">{displayName(agent)}</span>
-    {status === 'working' && (
-      <Badge
-        variant="secondary"
-        className="bg-amber-50 text-amber-900 hover:bg-amber-50"
-      >
-        Working
-      </Badge>
-    )}
-    {status === 'asleep' && (
-      <Badge variant="outline" className="text-muted-foreground">
-        Asleep
-      </Badge>
-    )}
-    {status === 'error' && <Badge variant="destructive">Attention</Badge>}
-    <PinToggle pinned={pinned} onToggle={onPinToggle} />
-    <div className="ml-auto">
-      <AgentSparkline turnsByDay={turnsByDay} failedByDay={failedByDay} />
-    </div>
-  </div>
-)

packages/browseros-agent/apps/agent/entrypoints/app/agents/agent-row/AgentTokenSummary.tsx

@@ -1,63 +0,0 @@
-import type { FC } from 'react'
-import {
-  HoverCard,
-  HoverCardContent,
-  HoverCardTrigger,
-} from '@/components/ui/hover-card'
-import { Progress } from '@/components/ui/progress'
-import { formatTokens } from './agent-row.helpers'
-import type { AgentTokenUsage } from './agent-row.types'
-
-interface AgentTokenSummaryProps {
-  tokens: AgentTokenUsage | null
-}
-
-/**
- * Inline token total + a HoverCard breakdown. Surfaces lifetime tokens
- * (the only window we can compute reliably from the session record).
- * Per-window stats land in a follow-up once the activity ledger ships.
- */
-export const AgentTokenSummary: FC<AgentTokenSummaryProps> = ({ tokens }) => {
-  if (!tokens) return null
-  const { input, output } = tokens.cumulative
-  const total = input + output
-  if (total === 0) return null
-  const inputPct = (input / total) * 100
-
-  return (
-    <HoverCard openDelay={200}>
-      <HoverCardTrigger asChild>
-        <span className="cursor-default text-muted-foreground tabular-nums transition-colors hover:text-foreground">
-          {formatTokens(total)} tokens
-        </span>
-      </HoverCardTrigger>
-      <HoverCardContent side="top" align="end" className="w-72 text-sm">
-        <div className="mb-3 flex items-center justify-between">
-          <span className="font-medium">Lifetime tokens</span>
-          <span className="text-muted-foreground text-xs tabular-nums">
-            {formatTokens(total)} total
-          </span>
-        </div>
-
-        <div className="space-y-2">
-          <div className="flex items-center justify-between text-xs">
-            <span className="text-muted-foreground">Input</span>
-            <span className="tabular-nums">{formatTokens(input)}</span>
-          </div>
-          <Progress value={inputPct} className="h-1.5" />
-
-          <div className="mt-2 flex items-center justify-between text-xs">
-            <span className="text-muted-foreground">Output</span>
-            <span className="tabular-nums">{formatTokens(output)}</span>
-          </div>
-          <Progress value={100 - inputPct} className="h-1.5" />
-        </div>
-
-        <p className="mt-3 border-t pt-2 text-muted-foreground text-xs leading-snug">
-          Cumulative across every turn this agent has run. Per-window stats
-          arrive in a future release.
-        </p>
-      </HoverCardContent>
-    </HoverCard>
-  )
-}

packages/browseros-agent/apps/agent/entrypoints/app/agents/agent-row/PinToggle.tsx

@@ -1,60 +0,0 @@
-import { Star } from 'lucide-react'
-import type { FC } from 'react'
-import { Button } from '@/components/ui/button'
-import {
-  Tooltip,
-  TooltipContent,
-  TooltipProvider,
-  TooltipTrigger,
-} from '@/components/ui/tooltip'
-import { cn } from '@/lib/utils'
-
-interface PinToggleProps {
-  pinned: boolean
-  onToggle: (next: boolean) => void
-}
-
-/**
- * Trailing star toggle. The button is *always rendered* — only its
- * opacity changes between pinned/unpinned/hover states — so the title
- * row's height is constant. Hiding the slot via `display: none` would
- * collapse the row's vertical metrics on hover and shift every card
- * below in the rail.
- *
- * Placement is trailing the title (after the status badge) so the
- * title itself flushes left regardless of pin state — leading the
- * row with the star would indent the title relative to the model /
- * preview / meta lines beneath it.
- */
-export const PinToggle: FC<PinToggleProps> = ({ pinned, onToggle }) => (
-  <TooltipProvider delayDuration={300}>
-    <Tooltip>
-      <TooltipTrigger asChild>
-        <Button
-          variant="ghost"
-          size="icon"
-          className={cn(
-            'size-6 text-muted-foreground transition-opacity hover:text-foreground',
-            pinned ? 'opacity-100' : 'opacity-0 group-hover:opacity-100',
-          )}
-          aria-pressed={pinned}
-          aria-label={pinned ? 'Unpin agent' : 'Pin agent'}
-          onClick={(event) => {
-            event.stopPropagation()
-            onToggle(!pinned)
-          }}
-        >
-          <Star
-            className={cn(
-              'size-3.5',
-              pinned && 'fill-amber-400 text-amber-500',
-            )}
-          />
-        </Button>
-      </TooltipTrigger>
-      <TooltipContent side="top" className="text-xs">
-        {pinned ? 'Unpin' : 'Pin to top'}
-      </TooltipContent>
-    </Tooltip>
-  </TooltipProvider>
-)

packages/browseros-agent/apps/agent/entrypoints/app/agents/agent-row/agent-row.helpers.test.ts

@@ -1,73 +0,0 @@
-import { describe, expect, it } from 'bun:test'
-import {
-  firstNonBlankLine,
-  formatLocalDate,
-  formatTokens,
-  ROW_BAR_COUNT,
-  truncate,
-} from './agent-row.helpers'
-
-describe('formatTokens', () => {
-  it('renders zero / NaN as "0"', () => {
-    expect(formatTokens(0)).toBe('0')
-    expect(formatTokens(Number.NaN)).toBe('0')
-  })
-
-  it('renders sub-1K as integer', () => {
-    expect(formatTokens(142)).toBe('142')
-  })
-
-  it('renders K with one decimal under 10', () => {
-    expect(formatTokens(8_400)).toBe('8.4K')
-  })
-
-  it('drops the decimal at >=10K', () => {
-    expect(formatTokens(120_000)).toBe('120K')
-  })
-
-  it('renders M with one decimal under 10', () => {
-    expect(formatTokens(1_200_000)).toBe('1.2M')
-  })
-})
-
-describe('firstNonBlankLine', () => {
-  it('returns the first non-blank line', () => {
-    expect(firstNonBlankLine('\n\nhello\nworld')).toBe('hello')
-  })
-
-  it('skips USER_QUERY envelope tags', () => {
-    expect(firstNonBlankLine('<USER_QUERY>\nfix tests\n</USER_QUERY>')).toBe(
-      'fix tests',
-    )
-  })
-
-  it('falls back to the trimmed input when nothing matches', () => {
-    expect(firstNonBlankLine('   single   ')).toBe('single')
-  })
-})
-
-describe('truncate', () => {
-  it('returns input unchanged when within limit', () => {
-    expect(truncate('hello', 10)).toBe('hello')
-  })
-
-  it('appends an ellipsis when over limit', () => {
-    expect(truncate('hello world', 6)).toBe('hello…')
-  })
-})
-
-describe('formatLocalDate', () => {
-  const today = new Date('2026-04-30T12:00:00Z')
-
-  it('labels today and yesterday explicitly', () => {
-    expect(formatLocalDate(ROW_BAR_COUNT - 1, today)).toBe('today')
-    expect(formatLocalDate(ROW_BAR_COUNT - 2, today)).toBe('yesterday')
-  })
-
-  it('returns a "Mon D" format for older days', () => {
-    const label = formatLocalDate(0, today)
-    // "Apr 17" or "Apr 17," depending on locale; just assert it
-    // contains a month abbreviation and a day number.
-    expect(label).toMatch(/[A-Za-z]+ \d+/)
-  })
-})

packages/browseros-agent/apps/agent/entrypoints/app/agents/agent-row/agent-row.helpers.ts

@@ -1,64 +0,0 @@
-/**
- * Pure formatters consumed by row sub-components. Kept distinct from
- * `agent-display.helpers.ts` (page-level helpers) so the row internals
- * have an obvious single home.
- */
-
-const TOKEN_THRESHOLDS: Array<[number, string]> = [
-  [1_000_000, 'M'],
-  [1_000, 'K'],
-]
-
-/** `1.2M`, `820K`, `8.4K`, `142`, `0`. */
-export function formatTokens(n: number): string {
-  if (!Number.isFinite(n) || n <= 0) return '0'
-  for (const [threshold, suffix] of TOKEN_THRESHOLDS) {
-    if (n >= threshold) {
-      const value = n / threshold
-      const decimal = value < 10 ? value.toFixed(1) : value.toFixed(0)
-      return `${decimal}${suffix}`
-    }
-  }
-  return String(Math.round(n))
-}
-
-const USER_QUERY_OPEN = /^<USER_QUERY>$/i
-const USER_QUERY_CLOSE = /^<\/USER_QUERY>$/i
-
-/**
- * First non-blank line, with the BrowserOS user-system-prompt
- * `<USER_QUERY>` envelope tags stripped so previews don't show
- * structural noise.
- */
-export function firstNonBlankLine(text: string): string {
-  const lines = text.split('\n').map((line) => line.trim())
-  for (const line of lines) {
-    if (!line) continue
-    if (USER_QUERY_OPEN.test(line) || USER_QUERY_CLOSE.test(line)) continue
-    return line
-  }
-  return text.trim()
-}
-
-export function truncate(text: string, max: number): string {
-  if (text.length <= max) return text
-  return `${text.slice(0, max - 1).trimEnd()}…`
-}
-
-const SPARKLINE_DAYS = 14
-
-/**
- * "today" / "yesterday" / "Apr 17" — given an index 0..13 from
- * oldest → newest. `today` defaults to `new Date()` so callers don't
- * have to thread a clock through.
- */
-export function formatLocalDate(idx: number, today: Date = new Date()): string {
-  if (idx === SPARKLINE_DAYS - 1) return 'today'
-  if (idx === SPARKLINE_DAYS - 2) return 'yesterday'
-  const offset = SPARKLINE_DAYS - 1 - idx
-  const date = new Date(today)
-  date.setDate(date.getDate() - offset)
-  return date.toLocaleDateString(undefined, { month: 'short', day: 'numeric' })
-}
-
-export const ROW_BAR_COUNT = SPARKLINE_DAYS

packages/browseros-agent/apps/agent/entrypoints/app/agents/agent-row/agent-row.types.ts

@@ -1,51 +0,0 @@
-import type { HarnessAgentAdapter } from '../agent-harness-types'
-import type { AgentListItem } from '../agents-page-types'
-import type { AgentLiveness } from '../LivenessDot'
-
-/**
- * Window-bounded token usage. Server returns `null` when no session
- * record exists yet for the agent.
- */
-export interface AgentTokenUsage {
-  last7d: { input: number; output: number; requestCount: number }
-  cumulative: { input: number; output: number }
-}
-
-export interface AgentAdapterHealth {
-  healthy: boolean
-  reason?: string
-}
-
-/**
- * Everything an `AgentRowCard` needs to render. Mirrors the shape
- * `useHarnessAgents` exposes; the page assembles one entry per row in
- * `AgentList` and passes it down. Sub-components only see slices of
- * this object — no prop drilling beyond two levels.
- */
-export interface AgentRowData {
-  agent: AgentListItem
-  adapter: HarnessAgentAdapter | 'unknown'
-  modelLabel: string | null
-  reasoningEffort: string | null
-  status: AgentLiveness
-  lastUsedAt: number | null
-  pinned: boolean
-  cwd: string | null
-  lastUserMessage: string | null
-  tokens: AgentTokenUsage | null
-  /** 14 entries, oldest → newest. Today is the last index. */
-  turnsByDay: number[]
-  /** Same length and ordering as `turnsByDay`. */
-  failedByDay: number[]
-  lastError: string | null
-  lastErrorAt: number | null
-  /** When non-null, an in-flight turn this row can be resumed from. */
-  activeTurnId: string | null
-  /** Adapter-level health, shared across rows for the same adapter. */
-  adapterHealth: AgentAdapterHealth | null
-}
-
-export interface AgentRowCallbacks {
-  onDelete: (agent: AgentListItem) => void
-  onPinToggle: (agent: AgentListItem, next: boolean) => void
-}

packages/browseros-agent/apps/agent/entrypoints/app/agents/agents-page-utils.ts

@@ -138,20 +138,24 @@ export function getVisibleOpenClawAgents(
 }
 
 export function getAgentsLoading(input: {
+  statusLoading: boolean
   adaptersLoading: boolean
   harnessAgentsLoading: boolean
+  openClawAgentsEnabled: boolean
   openClawAgentsLoading: boolean
 }): boolean {
   return (
+    input.statusLoading ||
     input.adaptersLoading ||
     input.harnessAgentsLoading ||
-    input.openClawAgentsLoading
+    (input.openClawAgentsEnabled && input.openClawAgentsLoading)
   )
 }
 
 export function getInlineError(input: {
   lifecyclePending: boolean
   pageError: string | null
+  statusError: Error | null
   openClawAgentsError: Error | null
   adaptersError: Error | null
   harnessAgentsError: Error | null
@@ -159,6 +163,7 @@ export function getInlineError(input: {
   if (input.lifecyclePending) return null
   return (
     input.pageError ??
+    input.statusError?.message ??
     input.openClawAgentsError?.message ??
     input.adaptersError?.message ??
     input.harnessAgentsError?.message ??

packages/browseros-agent/apps/agent/entrypoints/app/agents/useAgents.ts

@@ -8,20 +8,8 @@ import {
   type HarnessAdapterDescriptor,
   type HarnessAgent,
   type HarnessAgentHistoryPage,
-  type HarnessQueuedMessage,
   mapHarnessAgentToEntry,
 } from './agent-harness-types'
-import type { OpenClawStatus } from './useOpenClaw'
-
-/**
- * Combined response shape of `GET /agents`. The page polls this once
- * and consumes both fields, replacing the dedicated `/claw/status`
- * poll the previous design carried.
- */
-interface HarnessAgentsResponse {
-  agents: HarnessAgent[]
-  gateway: OpenClawStatus | null
-}
 
 export type { AgentHarnessStreamEvent }
 
@@ -81,31 +69,21 @@ export function useHarnessAgents(enabled = true) {
     error: urlError,
   } = useAgentServerUrl()
 
-  const query = useQuery<HarnessAgentsResponse, Error>({
+  const query = useQuery<HarnessAgent[], Error>({
     queryKey: [AGENT_QUERY_KEYS.agents, baseUrl],
     queryFn: async () => {
-      const data = await agentsFetch<HarnessAgentsResponse>(
+      const data = await agentsFetch<{ agents: HarnessAgent[] }>(
         baseUrl as string,
         '/',
       )
-      return {
-        agents: data.agents ?? [],
-        gateway: data.gateway ?? null,
-      }
+      return data.agents ?? []
     },
     enabled: Boolean(baseUrl) && !urlLoading && enabled,
-    // Poll every 5s so the per-agent liveness state (working / idle /
-    // asleep / error) and last-used timestamps stay fresh without a
-    // websocket. `refetchIntervalInBackground: false` lets a hidden
-    // tab go quiet — react-query's default, made explicit.
-    refetchInterval: 5_000,
-    refetchIntervalInBackground: false,
   })
 
   return {
-    agents: (query.data?.agents ?? []).map(mapHarnessAgentToEntry),
-    harnessAgents: query.data?.agents ?? [],
-    gateway: query.data?.gateway ?? null,
+    agents: (query.data ?? []).map(mapHarnessAgentToEntry),
+    harnessAgents: query.data ?? [],
     loading: query.isLoading || urlLoading,
     error: query.error ?? urlError,
     refetch: query.refetch,
@@ -136,63 +114,6 @@ export function useCreateHarnessAgent() {
   })
 }
 
-/**
- * Apply a partial update to a harness agent. Used by the pin-toggle
- * star and (eventually) the inline rename UI. Optimistically writes
- * the patch into the listing query cache so the row updates instantly,
- * then rolls back if the server rejects the change.
- */
-export function useUpdateHarnessAgent() {
-  const { baseUrl, isLoading: urlLoading } = useAgentServerUrl()
-  const queryClient = useQueryClient()
-
-  return useMutation({
-    mutationFn: async (input: {
-      agentId: string
-      patch: { name?: string; pinned?: boolean }
-    }) => {
-      if (!baseUrl || urlLoading) {
-        throw new Error('BrowserOS agent server URL is not ready')
-      }
-      const data = await agentsFetch<{ agent: HarnessAgent }>(
-        baseUrl,
-        `/${encodeURIComponent(input.agentId)}`,
-        {
-          method: 'PATCH',
-          headers: { 'Content-Type': 'application/json' },
-          body: JSON.stringify(input.patch),
-        },
-      )
-      return data.agent
-    },
-    onMutate: async ({ agentId, patch }) => {
-      const queryKey = [AGENT_QUERY_KEYS.agents, baseUrl]
-      await queryClient.cancelQueries({ queryKey })
-      const previous = queryClient.getQueryData<HarnessAgentsResponse>(queryKey)
-      if (!previous) return { previous: undefined }
-      queryClient.setQueryData<HarnessAgentsResponse>(queryKey, {
-        ...previous,
-        agents: previous.agents.map((agent) =>
-          agent.id === agentId ? { ...agent, ...patch } : agent,
-        ),
-      })
-      return { previous }
-    },
-    onError: (_err, _vars, context) => {
-      if (!context?.previous) return
-      queryClient.setQueryData(
-        [AGENT_QUERY_KEYS.agents, baseUrl],
-        context.previous,
-      )
-    },
-    onSettled: async () => {
-      await queryClient.invalidateQueries({
-        queryKey: [AGENT_QUERY_KEYS.agents],
-      })
-    },
-  })
-}
-
 export function useDeleteHarnessAgent() {
   const { baseUrl, isLoading: urlLoading } = useAgentServerUrl()
   const queryClient = useQueryClient()
@@ -220,97 +141,16 @@ export async function chatWithHarnessAgent(
   agentId: string,
   message: string,
   signal?: AbortSignal,
-  attachments?: ReadonlyArray<unknown>,
 ): Promise<Response> {
   const baseUrl = await getAgentServerUrl()
   return fetch(`${baseUrl}/agents/${encodeURIComponent(agentId)}/chat`, {
     method: 'POST',
     headers: { 'Content-Type': 'application/json' },
-    body: JSON.stringify({
-      message,
-      ...(attachments && attachments.length > 0 ? { attachments } : {}),
-    }),
+    body: JSON.stringify({ message }),
     signal,
   })
 }
 
-/**
- * Subscribe to an existing turn (the server's `ActiveTurnRegistry`
- * decoupled the turn lifecycle from POST /chat). `lastSeq` lets the
- * client resume after a disconnect — the server replays buffered
- * frames with seq > lastSeq, then tails new ones.
- */
-export async function attachToHarnessTurn(
-  agentId: string,
-  options: { turnId?: string; lastSeq?: number; signal?: AbortSignal } = {},
-): Promise<Response> {
-  const baseUrl = await getAgentServerUrl()
-  const url = new URL(
-    `${baseUrl}/agents/${encodeURIComponent(agentId)}/chat/stream`,
-  )
-  if (options.turnId) url.searchParams.set('turnId', options.turnId)
-  const headers: Record<string, string> = {}
-  if (typeof options.lastSeq === 'number') {
-    headers['Last-Event-ID'] = String(options.lastSeq)
-  }
-  return fetch(url.toString(), { signal: options.signal, headers })
-}
-
-export interface HarnessActiveTurnInfo {
-  turnId: string
-  agentId: string
-  sessionId: 'main'
-  status: 'running' | 'done' | 'error' | 'cancelled'
-  lastSeq: number
-  startedAt: number
-  endedAt?: number
-  /** User message that kicked off the turn; null when not captured. */
-  prompt: string | null
-}
-
-/**
- * Discover an in-flight turn for an agent. Used on chat mount so the
- * UI reattaches instead of starting a new turn after a tab/refresh.
- */
-export async function fetchActiveHarnessTurn(
-  agentId: string,
-): Promise<HarnessActiveTurnInfo | null> {
-  const baseUrl = await getAgentServerUrl()
-  const response = await fetch(
-    `${baseUrl}/agents/${encodeURIComponent(agentId)}/chat/active`,
-  )
-  if (!response.ok) return null
-  const body = (await response.json()) as {
-    active: HarnessActiveTurnInfo | null
-  }
-  return body.active
-}
-
-/**
- * Stop button. Hits the explicit cancel endpoint instead of just
- * aborting the fetch (which now only detaches *this* subscriber from
- * the buffer; the underlying turn would otherwise keep running).
- */
-export async function cancelHarnessTurn(
-  agentId: string,
-  options: { turnId?: string; reason?: string } = {},
-): Promise<{ cancelled: boolean }> {
-  const baseUrl = await getAgentServerUrl()
-  const response = await fetch(
-    `${baseUrl}/agents/${encodeURIComponent(agentId)}/chat/cancel`,
-    {
-      method: 'POST',
-      headers: { 'Content-Type': 'application/json' },
-      body: JSON.stringify({
-        ...(options.turnId ? { turnId: options.turnId } : {}),
-        ...(options.reason ? { reason: options.reason } : {}),
-      }),
-    },
-  )
-  if (!response.ok) return { cancelled: false }
-  return (await response.json()) as { cancelled: boolean }
-}
-
 export async function fetchHarnessAgentHistory(
   agentId: string,
 ): Promise<HarnessAgentHistoryPage> {
@@ -320,145 +160,3 @@ export async function fetchHarnessAgentHistory(
     `/${encodeURIComponent(agentId)}/sessions/main/history`,
   )
 }
-
-export interface EnqueueMessageInput {
-  message: string
-  attachments?: ReadonlyArray<unknown>
-}
-
-export async function enqueueHarnessMessage(
-  agentId: string,
-  input: EnqueueMessageInput,
-): Promise<HarnessQueuedMessage> {
-  const baseUrl = await getAgentServerUrl()
-  const response = await fetch(
-    `${baseUrl}/agents/${encodeURIComponent(agentId)}/queue`,
-    {
-      method: 'POST',
-      headers: { 'Content-Type': 'application/json' },
-      body: JSON.stringify({
-        message: input.message,
-        ...(input.attachments && input.attachments.length > 0
-          ? { attachments: input.attachments }
-          : {}),
-      }),
-    },
-  )
-  if (!response.ok) {
-    let message = `Request failed with status ${response.status}`
-    try {
-      const body = (await response.json()) as { error?: string }
-      if (body.error) message = body.error
-    } catch {}
-    throw new Error(message)
-  }
-  const body = (await response.json()) as { queued: HarnessQueuedMessage }
-  return body.queued
-}
-
-export async function removeHarnessQueuedMessage(
-  agentId: string,
-  messageId: string,
-): Promise<{ removed: boolean }> {
-  const baseUrl = await getAgentServerUrl()
-  const response = await fetch(
-    `${baseUrl}/agents/${encodeURIComponent(agentId)}/queue/${encodeURIComponent(
-      messageId,
-    )}`,
-    { method: 'DELETE' },
-  )
-  if (!response.ok) return { removed: false }
-  return (await response.json()) as { removed: boolean }
-}
-
-/**
- * Optimistic enqueue: writes the new queued message into the listing
- * cache immediately so the queue panel reflects the change without
- * waiting for the next poll. Rolls back if the server rejects.
- */
-export function useEnqueueHarnessMessage() {
-  const { baseUrl } = useAgentServerUrl()
-  const queryClient = useQueryClient()
-
-  return useMutation({
-    mutationFn: async (input: { agentId: string } & EnqueueMessageInput) =>
-      enqueueHarnessMessage(input.agentId, input),
-    onMutate: async (input) => {
-      const queryKey = [AGENT_QUERY_KEYS.agents, baseUrl]
-      await queryClient.cancelQueries({ queryKey })
-      const previous = queryClient.getQueryData<HarnessAgentsResponse>(queryKey)
-      if (!previous) return { previous: undefined }
-      const optimistic: HarnessQueuedMessage = {
-        id: `optimistic-${Math.random().toString(36).slice(2, 10)}`,
-        createdAt: Date.now(),
-        message: input.message,
-      }
-      queryClient.setQueryData<HarnessAgentsResponse>(queryKey, {
-        ...previous,
-        agents: previous.agents.map((agent) =>
-          agent.id === input.agentId
-            ? { ...agent, queue: [...(agent.queue ?? []), optimistic] }
-            : agent,
-        ),
-      })
-      return { previous }
-    },
-    onError: (_err, _vars, context) => {
-      if (!context?.previous) return
-      queryClient.setQueryData(
-        [AGENT_QUERY_KEYS.agents, baseUrl],
-        context.previous,
-      )
-    },
-    onSettled: async () => {
-      await queryClient.invalidateQueries({
-        queryKey: [AGENT_QUERY_KEYS.agents],
-      })
-    },
-  })
-}
-
-/**
- * Optimistic queue removal mirror of `useEnqueueHarnessMessage`.
- */
-export function useRemoveHarnessQueuedMessage() {
-  const { baseUrl } = useAgentServerUrl()
-  const queryClient = useQueryClient()
-
-  return useMutation({
-    mutationFn: async (input: { agentId: string; messageId: string }) =>
-      removeHarnessQueuedMessage(input.agentId, input.messageId),
-    onMutate: async (input) => {
-      const queryKey = [AGENT_QUERY_KEYS.agents, baseUrl]
-      await queryClient.cancelQueries({ queryKey })
-      const previous = queryClient.getQueryData<HarnessAgentsResponse>(queryKey)
-      if (!previous) return { previous: undefined }
-      queryClient.setQueryData<HarnessAgentsResponse>(queryKey, {
-        ...previous,
-        agents: previous.agents.map((agent) =>
-          agent.id === input.agentId
-            ? {
-                ...agent,
-                queue: (agent.queue ?? []).filter(
-                  (entry) => entry.id !== input.messageId,
-                ),
-              }
-            : agent,
-        ),
-      })
-      return { previous }
-    },
-    onError: (_err, _vars, context) => {
-      if (!context?.previous) return
-      queryClient.setQueryData(
-        [AGENT_QUERY_KEYS.agents, baseUrl],
-        context.previous,
-      )
-    },
-    onSettled: async () => {
-      await queryClient.invalidateQueries({
-        queryKey: [AGENT_QUERY_KEYS.agents],
-      })
-    },
-  })
-}

packages/browseros-agent/apps/agent/entrypoints/app/agents/useOpenClaw.ts

@@ -1,4 +1,5 @@
 import { useMutation, useQuery, useQueryClient } from '@tanstack/react-query'
+import { getAgentServerUrl } from '@/lib/browseros/helpers'
 import { useAgentServerUrl } from '@/lib/browseros/useBrowserOSProviders'
 
 export interface AgentEntry {
@@ -318,3 +319,25 @@ export function buildChatHistoryFromTurns(
 
   return messages
 }
+
+export async function chatWithAgent(
+  agentId: string,
+  message: string,
+  sessionKey?: string,
+  history: OpenClawChatHistoryMessage[] = [],
+  signal?: AbortSignal,
+  attachments?: ReadonlyArray<unknown>,
+): Promise<Response> {
+  const baseUrl = await getAgentServerUrl()
+  return fetch(`${baseUrl}/claw/agents/${agentId}/chat`, {
+    method: 'POST',
+    headers: { 'Content-Type': 'application/json' },
+    body: JSON.stringify({
+      message,
+      sessionKey,
+      history,
+      ...(attachments && attachments.length > 0 ? { attachments } : {}),
+    }),
+    signal,
+  })
+}

packages/browseros-agent/apps/agent/entrypoints/app/scheduled-tasks/NewScheduledTaskDialog.tsx

@@ -164,17 +164,9 @@ export const NewScheduledTaskDialog: FC<NewScheduledTaskDialogProps> = ({
   const resolvedProvider: Provider | null = (() => {
     const id = selectedProviderId ?? defaultProviderId
     const found = providers.find((p) => p.id === id)
-    if (found) {
-      return {
-        kind: 'llm' as const,
-        id: found.id,
-        name: found.name,
-        type: found.type,
-      }
-    }
+    if (found) return { id: found.id, name: found.name, type: found.type }
     if (providers[0])
       return {
-        kind: 'llm' as const,
         id: providers[0].id,
         name: providers[0].name,
         type: providers[0].type,
@@ -183,7 +175,6 @@ export const NewScheduledTaskDialog: FC<NewScheduledTaskDialogProps> = ({
   })()
 
   const providerOptions: Provider[] = providers.map((p) => ({
-    kind: 'llm',
     id: p.id,
     name: p.name,
     type: p.type,

packages/browseros-agent/apps/agent/entrypoints/sidepanel/index/ChatHeader.tsx

@@ -1,4 +1,4 @@
-import { Bot, Github, History, Plus, SettingsIcon } from 'lucide-react'
+import { Github, History, Plus, SettingsIcon } from 'lucide-react'
 import type { FC } from 'react'
 import { Link, useLocation, useNavigate } from 'react-router'
 import { ChatProviderSelector } from '@/components/chat/ChatProviderSelector'
@@ -64,9 +64,7 @@ export const ChatHeader: FC<ChatHeaderProps> = ({
             className="group relative inline-flex cursor-pointer items-center gap-2 rounded-lg p-2 text-muted-foreground transition-colors hover:bg-muted/50 hover:text-foreground data-[state=open]:bg-accent"
             title="Change AI Provider"
           >
-            {selectedProvider.kind === 'acp' ? (
-              <Bot className="h-[18px] w-[18px]" />
-            ) : selectedProvider.type === 'browseros' ? (
+            {selectedProvider.type === 'browseros' ? (
               <BrowserOSIcon size={18} />
             ) : (
               <ProviderIcon

packages/browseros-agent/apps/agent/entrypoints/sidepanel/index/sidepanel-chat-targets.test.ts

@@ -1,258 +0,0 @@
-import { describe, expect, it } from 'bun:test'
-import type {
-  HarnessAdapterDescriptor,
-  HarnessAgent,
-} from '@/entrypoints/app/agents/agent-harness-types'
-import type { LlmProviderConfig } from '@/lib/llm-providers/types'
-import {
-  buildSidepanelChatTargets,
-  persistSidepanelChatTargetSelection,
-  resolveSidepanelChatTarget,
-  type SidepanelChatTargetSelection,
-  toLlmProviderConfig,
-} from './sidepanel-chat-targets'
-
-const timestamp = 1000
-
-const providers: LlmProviderConfig[] = [
-  {
-    id: 'browseros',
-    type: 'browseros',
-    name: 'BrowserOS',
-    baseUrl: 'https://api.browseros.com/v1',
-    modelId: 'browseros-auto',
-    supportsImages: true,
-    contextWindow: 200000,
-    temperature: 0.2,
-    createdAt: timestamp,
-    updatedAt: timestamp,
-  },
-  {
-    id: 'anthropic-sonnet',
-    type: 'anthropic',
-    name: 'Anthropic Sonnet',
-    modelId: 'claude-sonnet-4-6',
-    apiKey: 'sk-ant',
-    supportsImages: true,
-    contextWindow: 200000,
-    temperature: 0.2,
-    createdAt: timestamp,
-    updatedAt: timestamp,
-  },
-]
-
-const adapters: HarnessAdapterDescriptor[] = [
-  {
-    id: 'claude',
-    name: 'Claude Code',
-    defaultModelId: 'haiku',
-    defaultReasoningEffort: 'medium',
-    modelControl: 'best-effort',
-    models: [
-      { id: 'sonnet', label: 'Sonnet' },
-      { id: 'haiku', label: 'Haiku', recommended: true },
-    ],
-    reasoningEfforts: [
-      { id: 'medium', label: 'Medium', recommended: true },
-      { id: 'high', label: 'High' },
-    ],
-  },
-  {
-    id: 'codex',
-    name: 'Codex',
-    defaultModelId: 'gpt-5.5',
-    defaultReasoningEffort: 'medium',
-    modelControl: 'runtime-supported',
-    models: [{ id: 'gpt-5.5', label: 'GPT-5.5', recommended: true }],
-    reasoningEfforts: [{ id: 'medium', label: 'Medium', recommended: true }],
-  },
-  {
-    id: 'openclaw',
-    name: 'OpenClaw',
-    defaultModelId: 'default',
-    defaultReasoningEffort: 'medium',
-    modelControl: 'best-effort',
-    models: [],
-    reasoningEfforts: [
-      { id: 'medium', label: 'Medium', recommended: true },
-      { id: 'high', label: 'High' },
-    ],
-  },
-]
-
-const agents: HarnessAgent[] = [
-  {
-    id: 'agent-codex',
-    name: 'Review Bot',
-    adapter: 'codex',
-    modelId: 'gpt-5.5',
-    reasoningEffort: 'medium',
-    permissionMode: 'approve-all',
-    sessionKey: 'agent:agent-codex:main',
-    createdAt: timestamp,
-    updatedAt: timestamp,
-  },
-  {
-    id: 'agent-openclaw',
-    name: 'Research Claw',
-    adapter: 'openclaw',
-    modelId: 'default',
-    reasoningEffort: 'high',
-    permissionMode: 'approve-all',
-    sessionKey: 'agent:agent-openclaw:main',
-    createdAt: timestamp,
-    updatedAt: timestamp,
-  },
-]
-
-describe('buildSidepanelChatTargets', () => {
-  it('returns LLM targets plus one ACP target per persisted harness agent', () => {
-    const targets = buildSidepanelChatTargets({ providers, adapters, agents })
-
-    expect(targets.map((target) => target.id)).toEqual([
-      'browseros',
-      'anthropic-sonnet',
-      'agent-codex',
-      'agent-openclaw',
-    ])
-  })
-
-  it('does not emit catalog-only ACP targets without persisted agents', () => {
-    const targets = buildSidepanelChatTargets({
-      providers,
-      adapters,
-      agents: [],
-    })
-
-    expect(targets.map((target) => target.id)).toEqual([
-      'browseros',
-      'anthropic-sonnet',
-    ])
-  })
-
-  it('uses the created OpenClaw agent name instead of a generic adapter target', () => {
-    const targets = buildSidepanelChatTargets({ providers, adapters, agents })
-    const openclaw = targets.find((target) => target.id === 'agent-openclaw')
-
-    expect(openclaw).toMatchObject({
-      kind: 'acp',
-      id: 'agent-openclaw',
-      agentId: 'agent-openclaw',
-      adapter: 'openclaw',
-      adapterName: 'OpenClaw',
-      modelId: 'default',
-      modelLabel: 'default',
-      name: 'Research Claw',
-      modelControl: 'best-effort',
-      reasoningEffort: 'high',
-    })
-  })
-
-  it('preserves adapter metadata for created agent targets', () => {
-    const targets = buildSidepanelChatTargets({ providers, adapters, agents })
-    const codex = targets.find((target) => target.id === 'agent-codex')
-
-    expect(codex).toMatchObject({
-      kind: 'acp',
-      agentId: 'agent-codex',
-      adapter: 'codex',
-      adapterName: 'Codex',
-      modelId: 'gpt-5.5',
-      modelLabel: 'GPT-5.5',
-      modelControl: 'runtime-supported',
-      recommended: true,
-      reasoningEffort: 'medium',
-      reasoningEffortLabel: 'Medium',
-    })
-  })
-
-  it('still returns LLM targets when agents and adapters are unavailable', () => {
-    expect(
-      buildSidepanelChatTargets({ providers, adapters: [], agents: [] }),
-    ).toEqual([
-      {
-        kind: 'llm',
-        id: 'browseros',
-        name: 'BrowserOS',
-        type: 'browseros',
-        provider: providers[0],
-      },
-      {
-        kind: 'llm',
-        id: 'anthropic-sonnet',
-        name: 'Anthropic Sonnet',
-        type: 'anthropic',
-        provider: providers[1],
-      },
-    ])
-  })
-})
-
-describe('resolveSidepanelChatTarget', () => {
-  it('resolves selected LLM targets back to their provider config', () => {
-    const targets = buildSidepanelChatTargets({ providers, adapters, agents })
-    const resolved = resolveSidepanelChatTarget({
-      targets,
-      defaultProviderId: 'browseros',
-      selection: { kind: 'llm', id: 'anthropic-sonnet' },
-    })
-
-    expect(resolved?.kind).toBe('llm')
-    expect(toLlmProviderConfig(resolved)?.modelId).toBe('claude-sonnet-4-6')
-  })
-
-  it('falls back to the current default LLM provider when a persisted ACP target is stale', () => {
-    const targets = buildSidepanelChatTargets({
-      providers,
-      adapters,
-      agents: [],
-    })
-
-    expect(
-      resolveSidepanelChatTarget({
-        targets,
-        defaultProviderId: 'anthropic-sonnet',
-        selection: { kind: 'acp', id: 'agent-codex' },
-      }),
-    ).toMatchObject({
-      kind: 'llm',
-      id: 'anthropic-sonnet',
-    })
-  })
-
-  it('falls back when an old catalog-style ACP target id is persisted', () => {
-    const targets = buildSidepanelChatTargets({ providers, adapters, agents })
-
-    expect(
-      resolveSidepanelChatTarget({
-        targets,
-        defaultProviderId: 'anthropic-sonnet',
-        selection: { kind: 'acp', id: 'acp:codex:gpt-5.5:medium' },
-      }),
-    ).toMatchObject({
-      kind: 'llm',
-      id: 'anthropic-sonnet',
-    })
-  })
-})
-
-describe('persistSidepanelChatTargetSelection', () => {
-  it('stores only target identity and does not mutate LLM provider arrays', async () => {
-    let savedSelection: SidepanelChatTargetSelection | null = null
-    const originalProviders = providers.map((provider) => ({ ...provider }))
-    const targets = buildSidepanelChatTargets({ providers, adapters, agents })
-    const target = targets.find((candidate) => candidate.id === 'agent-codex')
-
-    await persistSidepanelChatTargetSelection(target, {
-      setValue: async (value) => {
-        savedSelection = value
-      },
-    })
-
-    expect(savedSelection as SidepanelChatTargetSelection | null).toEqual({
-      kind: 'acp',
-      id: 'agent-codex',
-    })
-    expect(providers).toEqual(originalProviders)
-  })
-})

packages/browseros-agent/apps/agent/entrypoints/sidepanel/index/sidepanel-chat-targets.ts

@@ -1,178 +0,0 @@
-import type {
-  HarnessAdapterDescriptor,
-  HarnessAgent,
-  HarnessAgentAdapter,
-} from '@/entrypoints/app/agents/agent-harness-types'
-import type { LlmProviderConfig, ProviderType } from '@/lib/llm-providers/types'
-
-export type SidepanelTargetKind = 'llm' | 'acp'
-
-export type SidepanelChatTarget =
-  | {
-      kind: 'llm'
-      id: string
-      name: string
-      type: ProviderType
-      provider: LlmProviderConfig
-    }
-  | {
-      kind: 'acp'
-      id: string
-      name: string
-      type: 'acp'
-      agentId: string
-      adapter: HarnessAgentAdapter
-      adapterName: string
-      modelId: string
-      modelLabel: string
-      modelControl: HarnessAdapterDescriptor['modelControl']
-      recommended?: boolean
-      reasoningEffort: string
-      reasoningEffortLabel?: string
-    }
-
-export type SidepanelChatTargetSelection = Pick<
-  SidepanelChatTarget,
-  'kind' | 'id'
->
-
-interface BuildSidepanelChatTargetsInput {
-  providers: LlmProviderConfig[]
-  adapters: HarnessAdapterDescriptor[]
-  agents?: HarnessAgent[]
-}
-
-interface ResolveSidepanelChatTargetInput {
-  targets: SidepanelChatTarget[]
-  defaultProviderId: string
-  selection?: SidepanelChatTargetSelection | null
-}
-
-interface SidepanelChatTargetSelectionWriter {
-  setValue(value: SidepanelChatTargetSelection | null): Promise<void>
-}
-
-interface SidepanelChatTargetSelectionReader {
-  getValue(): Promise<SidepanelChatTargetSelection | null>
-}
-
-type SidepanelChatTargetSelectionStore = SidepanelChatTargetSelectionReader &
-  SidepanelChatTargetSelectionWriter
-
-let sidepanelChatTargetSelectionStorage:
-  | SidepanelChatTargetSelectionStore
-  | undefined
-
-export function buildSidepanelChatTargets({
-  providers,
-  adapters,
-  agents = [],
-}: BuildSidepanelChatTargetsInput): SidepanelChatTarget[] {
-  return [
-    ...providers.map(toLlmTarget),
-    ...agents.map((agent) => toAcpTargetForAgent(agent, adapters)),
-  ]
-}
-
-function toAcpTargetForAgent(
-  agent: HarnessAgent,
-  adapters: HarnessAdapterDescriptor[],
-): SidepanelChatTarget {
-  const adapter = adapters.find((entry) => entry.id === agent.adapter)
-  const modelId = agent.modelId ?? adapter?.defaultModelId ?? 'default'
-  const reasoningEffort =
-    agent.reasoningEffort ?? adapter?.defaultReasoningEffort ?? 'medium'
-  const model = adapter?.models.find((entry) => entry.id === modelId)
-  const reasoning = adapter?.reasoningEfforts.find(
-    (effort) => effort.id === reasoningEffort,
-  )
-
-  return {
-    kind: 'acp',
-    id: agent.id,
-    name: agent.name,
-    type: 'acp',
-    agentId: agent.id,
-    adapter: agent.adapter,
-    adapterName: adapter?.name ?? formatAdapterName(agent.adapter),
-    modelId,
-    modelLabel: model?.label ?? modelId,
-    modelControl: adapter?.modelControl ?? 'best-effort',
-    recommended: model?.recommended,
-    reasoningEffort,
-    reasoningEffortLabel: reasoning?.label,
-  }
-}
-
-function formatAdapterName(adapter: HarnessAgentAdapter): string {
-  if (adapter === 'claude') return 'Claude Code'
-  if (adapter === 'codex') return 'Codex'
-  if (adapter === 'openclaw') return 'OpenClaw'
-  return adapter
-}
-
-export function resolveSidepanelChatTarget({
-  targets,
-  defaultProviderId,
-  selection,
-}: ResolveSidepanelChatTargetInput): SidepanelChatTarget | undefined {
-  if (selection) {
-    const selected = targets.find(
-      (target) => target.kind === selection.kind && target.id === selection.id,
-    )
-    if (selected) return selected
-  }
-
-  return (
-    targets.find(
-      (target) => target.kind === 'llm' && target.id === defaultProviderId,
-    ) ?? targets.find((target) => target.kind === 'llm')
-  )
-}
-
-export function toLlmProviderConfig(
-  target: SidepanelChatTarget | undefined,
-): LlmProviderConfig | undefined {
-  return target?.kind === 'llm' ? target.provider : undefined
-}
-
-export async function persistSidepanelChatTargetSelection(
-  target: SidepanelChatTarget | undefined,
-  store?: SidepanelChatTargetSelectionWriter,
-): Promise<void> {
-  const targetStore = store ?? (await getSidepanelChatTargetSelectionStorage())
-  await targetStore.setValue(
-    target ? { kind: target.kind, id: target.id } : null,
-  )
-}
-
-export async function loadSidepanelChatTargetSelection(
-  store?: SidepanelChatTargetSelectionReader,
-): Promise<SidepanelChatTargetSelection | null> {
-  const targetStore = store ?? (await getSidepanelChatTargetSelectionStorage())
-  return targetStore.getValue()
-}
-
-function toLlmTarget(provider: LlmProviderConfig): SidepanelChatTarget {
-  return {
-    kind: 'llm',
-    id: provider.id,
-    name: provider.name,
-    type: provider.type,
-    provider,
-  }
-}
-
-async function getSidepanelChatTargetSelectionStorage(): Promise<SidepanelChatTargetSelectionStore> {
-  if (sidepanelChatTargetSelectionStorage) {
-    return sidepanelChatTargetSelectionStorage
-  }
-
-  const { storage } = await import('@wxt-dev/storage')
-  sidepanelChatTargetSelectionStorage =
-    storage.defineItem<SidepanelChatTargetSelection | null>(
-      'local:sidepanel-chat-target-selection',
-      { fallback: null },
-    )
-  return sidepanelChatTargetSelectionStorage
-}

packages/browseros-agent/apps/agent/entrypoints/sidepanel/index/useChatRefs.ts

@@ -1,21 +1,9 @@
-import { useCallback, useEffect, useMemo, useRef, useState } from 'react'
+import { useEffect, useRef } from 'react'
 import useDeepCompareEffect from 'use-deep-compare-effect'
-import {
-  useAgentAdapters,
-  useHarnessAgents,
-} from '@/entrypoints/app/agents/useAgents'
 import type { LlmProviderConfig } from '@/lib/llm-providers/types'
 import { useLlmProviders } from '@/lib/llm-providers/useLlmProviders'
 import { type McpServer, useMcpServers } from '@/lib/mcp/mcpServerStorage'
 import { usePersonalization } from '@/lib/personalization/personalizationStorage'
-import {
-  buildSidepanelChatTargets,
-  loadSidepanelChatTargetSelection,
-  persistSidepanelChatTargetSelection,
-  resolveSidepanelChatTarget,
-  type SidepanelChatTarget,
-  type SidepanelChatTargetSelection,
-} from './sidepanel-chat-targets'
 
 const constructMcpServers = (servers: McpServer[]) => {
   return servers
@@ -35,53 +23,14 @@ const constructCustomServers = (servers: McpServer[]) => {
 export const useChatRefs = () => {
   const { servers: mcpServers } = useMcpServers()
   const {
-    providers: llmProviders,
     selectedProvider: selectedLlmProvider,
-    setDefaultProvider,
     isLoading: isLoadingProviders,
   } = useLlmProviders()
-  const { adapters, loading: isLoadingAdapters } = useAgentAdapters()
-  const { harnessAgents, loading: isLoadingAgents } = useHarnessAgents()
   const { personalization } = usePersonalization()
-  const [targetSelection, setTargetSelection] =
-    useState<SidepanelChatTargetSelection | null>(null)
-
-  useEffect(() => {
-    let cancelled = false
-    loadSidepanelChatTargetSelection().then((selection) => {
-      if (!cancelled) setTargetSelection(selection)
-    })
-    return () => {
-      cancelled = true
-    }
-  }, [])
-
-  const chatTargets = useMemo(
-    () =>
-      buildSidepanelChatTargets({
-        providers: llmProviders,
-        adapters,
-        agents: harnessAgents,
-      }),
-    [llmProviders, adapters, harnessAgents],
-  )
-
-  const selectedChatTarget = useMemo(
-    () =>
-      resolveSidepanelChatTarget({
-        targets: chatTargets,
-        defaultProviderId: selectedLlmProvider?.id ?? llmProviders[0]?.id ?? '',
-        selection: targetSelection,
-      }),
-    [chatTargets, llmProviders, selectedLlmProvider, targetSelection],
-  )
 
   const selectedLlmProviderRef = useRef<LlmProviderConfig | null>(
     selectedLlmProvider,
   )
-  const selectedChatTargetRef = useRef<SidepanelChatTarget | undefined>(
-    selectedChatTarget,
-  )
   const enabledMcpServersRef = useRef(constructMcpServers(mcpServers))
   const enabledCustomServersRef = useRef(constructCustomServers(mcpServers))
   const personalizationRef = useRef(personalization)
@@ -92,36 +41,16 @@ export const useChatRefs = () => {
     enabledCustomServersRef.current = constructCustomServers(mcpServers)
   }, [selectedLlmProvider, mcpServers])
 
-  useEffect(() => {
-    selectedChatTargetRef.current = selectedChatTarget
-  }, [selectedChatTarget])
-
   useEffect(() => {
     personalizationRef.current = personalization
   }, [personalization])
 
-  const selectChatTarget = useCallback(
-    async (target: SidepanelChatTarget | undefined) => {
-      selectedChatTargetRef.current = target
-      setTargetSelection(target ? { kind: target.kind, id: target.id } : null)
-      await persistSidepanelChatTargetSelection(target)
-    },
-    [],
-  )
-
   return {
     selectedLlmProviderRef,
-    selectedChatTargetRef,
     enabledMcpServersRef,
     enabledCustomServersRef,
     personalizationRef,
-    llmProviders,
-    setDefaultProvider,
-    chatTargets,
-    selectedChatTarget,
-    selectChatTarget,
     selectedLlmProvider,
-    isLoadingProviders:
-      isLoadingProviders || isLoadingAdapters || isLoadingAgents,
+    isLoadingProviders,
   }
 }

packages/browseros-agent/apps/agent/entrypoints/sidepanel/index/useChatSession.test.ts

@@ -1,153 +0,0 @@
-import { describe, expect, it } from 'bun:test'
-import type { LlmProviderConfig } from '@/lib/llm-providers/types'
-import type { ChatMode } from './chatTypes'
-import type { SidepanelChatTarget } from './sidepanel-chat-targets'
-import { buildSidepanelPreparedSendMessagesRequest } from './useChatSessionRequest'
-
-const conversationId = '00000000-0000-4000-8000-000000000001'
-
-describe('buildSidepanelPreparedSendMessagesRequest', () => {
-  it('keeps LLM targets on the existing /chat request body', () => {
-    const request = buildSidepanelPreparedSendMessagesRequest({
-      agentServerUrl: 'http://127.0.0.1:5151',
-      target: llmTarget,
-      fallbackProvider,
-      message: 'Summarize this page',
-      ...commonRequestInput(),
-    })
-
-    expect(request.api).toBe('http://127.0.0.1:5151/chat')
-    expect(request.body).toMatchObject({
-      message: 'Summarize this page',
-      conversationId,
-      provider: 'browseros',
-      providerType: 'browseros',
-      providerName: 'BrowserOS',
-      model: 'gpt-5',
-      mode: 'agent',
-      browserContext: {
-        activeTab: { id: 10, url: 'https://example.com', title: 'Example' },
-        enabledMcpServers: ['slack'],
-      },
-      userSystemPrompt: 'Be concise',
-      userWorkingDir: '/tmp/work',
-      previousConversation: [{ role: 'assistant', content: 'Prior answer' }],
-      selectedText: 'selected text',
-      selectedTextSource: {
-        url: 'https://example.com',
-        title: 'Example',
-      },
-    })
-  })
-
-  it('sends created-agent targets to the agent-id sidepanel route', () => {
-    const request = buildSidepanelPreparedSendMessagesRequest({
-      agentServerUrl: 'http://127.0.0.1:5151',
-      target: acpTarget,
-      fallbackProvider,
-      message: 'Inspect the current tab',
-      approvalResponses: [
-        { approvalId: 'approval-1', approved: true, reason: 'ok' },
-      ],
-      ...commonRequestInput(),
-    })
-
-    expect(request.api).toBe(
-      'http://127.0.0.1:5151/agents/agent-codex/sidepanel/chat',
-    )
-    expect(request.body).toEqual({
-      conversationId,
-      message: 'Inspect the current tab',
-      browserContext: {
-        activeTab: { id: 10, url: 'https://example.com', title: 'Example' },
-        enabledMcpServers: ['slack'],
-      },
-      userSystemPrompt: 'Be concise',
-      userWorkingDir: '/tmp/work',
-      selectedText: 'selected text',
-      selectedTextSource: {
-        url: 'https://example.com',
-        title: 'Example',
-      },
-    })
-  })
-
-  it('keeps tool approval retry payloads scoped to LLM chat', () => {
-    const request = buildSidepanelPreparedSendMessagesRequest({
-      agentServerUrl: 'http://127.0.0.1:5151',
-      target: llmTarget,
-      fallbackProvider,
-      approvalResponses: [
-        { approvalId: 'approval-1', approved: false, reason: 'no' },
-      ],
-      ...commonRequestInput(),
-    })
-
-    expect(request.api).toBe('http://127.0.0.1:5151/chat')
-    expect(request.body).toMatchObject({
-      message: '',
-      toolApprovalResponses: [
-        { approvalId: 'approval-1', approved: false, reason: 'no' },
-      ],
-    })
-  })
-})
-
-function commonRequestInput() {
-  return {
-    conversationId,
-    mode: 'agent' as ChatMode,
-    browserContext: {
-      activeTab: { id: 10, url: 'https://example.com', title: 'Example' },
-      enabledMcpServers: ['slack'],
-    },
-    userSystemPrompt: 'Be concise',
-    userWorkingDir: '/tmp/work',
-    previousConversation: [
-      { role: 'assistant' as const, content: 'Prior answer' },
-    ],
-    declinedApps: ['gmail'],
-    aclRules: [{ id: 'rule-1', sitePattern: '*://*/*', enabled: true }],
-    selectedText: 'selected text',
-    selectedTextSource: {
-      url: 'https://example.com',
-      title: 'Example',
-    },
-    toolApprovalConfig: { categories: { navigation: true } },
-  }
-}
-
-const fallbackProvider: LlmProviderConfig = {
-  id: 'browseros',
-  type: 'browseros',
-  name: 'BrowserOS',
-  modelId: 'gpt-5',
-  supportsImages: true,
-  contextWindow: 128000,
-  temperature: 0.7,
-  createdAt: 1000,
-  updatedAt: 1000,
-}
-
-const llmTarget: SidepanelChatTarget = {
-  kind: 'llm',
-  id: fallbackProvider.id,
-  name: fallbackProvider.name,
-  type: fallbackProvider.type,
-  provider: fallbackProvider,
-}
-
-const acpTarget: SidepanelChatTarget = {
-  kind: 'acp',
-  id: 'agent-codex',
-  name: 'Review bot',
-  type: 'acp',
-  agentId: 'agent-codex',
-  adapter: 'codex',
-  adapterName: 'Codex',
-  modelId: 'gpt-5.5',
-  modelLabel: 'GPT-5.5',
-  modelControl: 'best-effort',
-  reasoningEffort: 'medium',
-  reasoningEffortLabel: 'Medium',
-}

packages/browseros-agent/apps/agent/entrypoints/sidepanel/index/useChatSession.ts

@@ -26,14 +26,15 @@ import { useInvalidateCredits } from '@/lib/credits/useCredits'
 import { declinedAppsStorage } from '@/lib/declined-apps/storage'
 import { useGraphqlQuery } from '@/lib/graphql/useGraphqlQuery'
 import { createDefaultBrowserOSProvider } from '@/lib/llm-providers/storage'
-import type {
-  ApprovalResponseData,
-  ChatRequestBrowserContext,
+import { useLlmProviders } from '@/lib/llm-providers/useLlmProviders'
+import {
+  type ApprovalResponseData,
+  buildChatRequestBody,
+  type ChatRequestBrowserContext,
 } from '@/lib/messaging/server/buildChatRequestBody'
 import { track } from '@/lib/metrics/track'
 import { searchActionsStorage } from '@/lib/search-actions/searchActionsStorage'
 import { selectedTextStorage } from '@/lib/selected-text/selectedTextStorage'
-import { sentry } from '@/lib/sentry/sentry'
 import { stopAgentStorage } from '@/lib/stop-agent/stop-agent-storage'
 import {
   type ApprovalResponse,
@@ -51,12 +52,7 @@ import {
 import { selectedWorkspaceStorage } from '@/lib/workspace/workspace-storage'
 import type { ChatMode } from './chatTypes'
 import { GetConversationWithMessagesDocument } from './graphql/chatSessionDocument'
-import { toLlmProviderConfig } from './sidepanel-chat-targets'
 import { useChatRefs } from './useChatRefs'
-import {
-  buildSidepanelPreparedSendMessagesRequest,
-  toProviderOption,
-} from './useChatSessionRequest'
 import { useExecutionHistoryTracker } from './useExecutionHistoryTracker'
 import { useNotifyActiveTab } from './useNotifyActiveTab'
 import { useRemoteConversationSave } from './useRemoteConversationSave'
@@ -190,19 +186,16 @@ const buildRequestBrowserContext = ({
 export const useChatSession = (options?: ChatSessionOptions) => {
   const {
     selectedLlmProviderRef,
-    selectedChatTargetRef,
     enabledMcpServersRef,
     enabledCustomServersRef,
     personalizationRef,
-    setDefaultProvider,
-    chatTargets,
-    selectedChatTarget,
-    selectChatTarget,
     selectedLlmProvider,
     isLoadingProviders,
   } = useChatRefs()
   const invalidateCredits = useInvalidateCredits()
 
+  const { providers: llmProviders, setDefaultProvider } = useLlmProviders()
+
   const {
     baseUrl: agentServerUrl,
     isLoading: isLoadingAgentUrl,
@@ -225,7 +218,11 @@ export const useChatSession = (options?: ChatSessionOptions) => {
     agentUrlRef.current = agentServerUrl
   }, [agentServerUrl])
 
-  const providers: Provider[] = chatTargets.map(toProviderOption)
+  const providers: Provider[] = llmProviders.map((p) => ({
+    id: p.id,
+    name: p.name,
+    type: p.type,
+  }))
 
   const [mode, setMode] = useState<ChatMode>('agent')
   const [textToAction, setTextToAction] = useState<Map<string, ChatAction>>(
@@ -327,8 +324,15 @@ export const useChatSession = (options?: ChatSessionOptions) => {
     textToActionRef.current = textToAction
   }, [mode, textToAction])
 
-  const selectedProvider = selectedChatTarget
-    ? toProviderOption(selectedChatTarget)
+  const selectedProvider = selectedLlmProvider
+    ? {
+        id: selectedLlmProvider.id,
+        name: selectedLlmProvider.name,
+        type:
+          selectedLlmProvider.id === 'browseros'
+            ? ('browseros' as const)
+            : selectedLlmProvider.type,
+      }
     : providers[0]
 
   const {
@@ -342,8 +346,7 @@ export const useChatSession = (options?: ChatSessionOptions) => {
   } = useChat({
     transport: new DefaultChatTransport({
       prepareSendMessagesRequest: async ({ messages }) => {
-        const target = selectedChatTargetRef.current
-        const fallbackProvider =
+        const provider =
           selectedLlmProviderRef.current ?? createDefaultBrowserOSProvider()
         const activeTabsList = await chrome.tabs.query({
           active: true,
@@ -392,46 +395,51 @@ export const useChatSession = (options?: ChatSessionOptions) => {
           personalizationRef.current,
         )
 
-        const commonRequest = {
-          conversationId: conversationIdRef.current,
-          mode: currentMode,
-          browserContext: requestBrowserContext,
-          userSystemPrompt,
-          userWorkingDir: workingDirRef.current,
-          previousConversation,
-          declinedApps,
-          aclRules: enabledAclRules,
-          toolApprovalConfig: approvalConfig,
-        }
-
-        const approvalResponses =
-          target?.kind === 'acp' ? null : extractApprovalResponses(messages)
+        const approvalResponses = extractApprovalResponses(messages)
         if (approvalResponses) {
-          return buildSidepanelPreparedSendMessagesRequest({
-            agentServerUrl: agentUrlRef.current ?? undefined,
-            target,
-            fallbackProvider,
-            ...commonRequest,
-            approvalResponses,
-          })
+          return {
+            api: `${agentUrlRef.current}/chat`,
+            body: buildChatRequestBody({
+              conversationId: conversationIdRef.current,
+              provider,
+              mode: currentMode,
+              browserContext: requestBrowserContext,
+              userSystemPrompt,
+              userWorkingDir: workingDirRef.current,
+              previousConversation,
+              declinedApps,
+              aclRules: enabledAclRules,
+              toolApprovalConfig: approvalConfig,
+              toolApprovalResponses: approvalResponses,
+            }),
+          }
         }
 
         const message = getLastMessageText(messages)
 
-        const result = buildSidepanelPreparedSendMessagesRequest({
-          agentServerUrl: agentUrlRef.current ?? undefined,
-          target,
-          fallbackProvider,
-          message,
-          ...commonRequest,
-          selectedText: activeTabSelection?.text,
-          selectedTextSource: activeTabSelection
-            ? {
-                url: activeTabSelection.url,
-                title: activeTabSelection.title,
-              }
-            : undefined,
-        })
+        const result = {
+          api: `${agentUrlRef.current}/chat`,
+          body: buildChatRequestBody({
+            message,
+            conversationId: conversationIdRef.current,
+            provider,
+            mode: currentMode,
+            browserContext: requestBrowserContext,
+            userSystemPrompt,
+            userWorkingDir: workingDirRef.current,
+            previousConversation,
+            declinedApps,
+            aclRules: enabledAclRules,
+            selectedText: activeTabSelection?.text,
+            selectedTextSource: activeTabSelection
+              ? {
+                  url: activeTabSelection.url,
+                  title: activeTabSelection.title,
+                }
+              : undefined,
+            toolApprovalConfig: approvalConfig,
+          }),
+        }
 
         // Track which tab's selection was sent so we can clear it on success
         pendingSelectionTabKeyRef.current =
@@ -443,7 +451,7 @@ export const useChatSession = (options?: ChatSessionOptions) => {
     sendAutomaticallyWhen: () => {
       if (approvalJustRespondedRef.current) {
         approvalJustRespondedRef.current = false
-        return selectedChatTargetRef.current?.kind !== 'acp'
+        return true
       }
       return false
     },
@@ -678,22 +686,10 @@ export const useChatSession = (options?: ChatSessionOptions) => {
   }, [dispatchMessage, isIntegrationsSynced])
 
   const sendMessage = (params: { text: string; action?: ChatAction }) => {
-    const target = selectedChatTargetRef.current
-    const llmTargetProvider = toLlmProviderConfig(target)
-    const agentTarget = target?.kind === 'acp' ? target : undefined
     track(MESSAGE_SENT_EVENT, {
       mode,
-      provider_id:
-        agentTarget?.agentId ??
-        llmTargetProvider?.id ??
-        selectedLlmProvider?.id,
-      provider_type: agentTarget ? 'acp' : llmTargetProvider?.type,
-      agent_id: agentTarget?.agentId,
-      adapter: agentTarget?.adapter,
-      model:
-        agentTarget?.modelId ??
-        llmTargetProvider?.modelId ??
-        selectedLlmProvider?.modelId,
+      provider_type: selectedLlmProvider?.type,
+      model: selectedLlmProvider?.modelId,
     })
 
     if (!isIntegrationsSyncedRef.current) {
@@ -745,54 +741,14 @@ export const useChatSession = (options?: ChatSessionOptions) => {
     addToolApprovalResponse(params)
   }
 
-  const resetConversationState = () => {
-    stop()
-    void finishExecutionTask({ isAbort: true })
-    setConversationId(crypto.randomUUID())
-    setMessages([])
-    setTextToAction(new Map())
-    setLiked({})
-    setDisliked({})
-    setRestoredConversationId(null)
-    resetRemoteConversation()
-  }
-
   const handleSelectProvider = (provider: Provider) => {
-    const target = chatTargets.find(
-      (candidate) =>
-        candidate.id === provider.id && candidate.kind === provider.kind,
-    )
-    if (!target) return
-
-    const previousTarget = selectedChatTargetRef.current
+    const fullProvider = llmProviders.find((p) => p.id === provider.id)
     track(PROVIDER_SELECTED_EVENT, {
-      provider_id: target.id,
-      provider_type: target.kind === 'acp' ? 'acp' : target.type,
-      model_id:
-        target.kind === 'acp' ? target.modelId : target.provider.modelId,
-      agent_id: target.kind === 'acp' ? target.agentId : undefined,
-      adapter: target.kind === 'acp' ? target.adapter : undefined,
+      provider_id: provider.id,
+      provider_type: provider.type,
+      model_id: fullProvider?.modelId,
     })
-
-    void selectChatTarget(target).catch((error) => {
-      sentry.captureException(error, {
-        extra: {
-          message: 'Failed to persist sidepanel chat target selection',
-          targetId: target.id,
-          targetKind: target.kind,
-        },
-      })
-    })
-    if (target.kind === 'llm') setDefaultProvider(target.provider.id)
-
-    if (
-      previousTarget &&
-      (previousTarget.kind !== target.kind ||
-        previousTarget.id !== target.id) &&
-      messagesRef.current.length > 0
-    ) {
-      resetConversationState()
-    }
+    setDefaultProvider(provider.id)
   }
 
   const getActionForMessage = (message: UIMessage) => {
@@ -806,7 +762,15 @@ export const useChatSession = (options?: ChatSessionOptions) => {
 
   const resetConversation = () => {
     track(CONVERSATION_RESET_EVENT, { message_count: messages.length })
-    resetConversationState()
+    stop()
+    void finishExecutionTask({ isAbort: true })
+    setConversationId(crypto.randomUUID())
+    setMessages([])
+    setTextToAction(new Map())
+    setLiked({})
+    setDisliked({})
+    setRestoredConversationId(null)
+    resetRemoteConversation()
   }
 
   const isRestoringConversation =

packages/browseros-agent/apps/agent/entrypoints/sidepanel/index/useChatSessionRequest.ts

@@ -1,74 +0,0 @@
-import type { Provider } from '../../../components/chat/chatComponentTypes'
-import type { LlmProviderConfig } from '../../../lib/llm-providers/types'
-import {
-  type ApprovalResponseData,
-  buildChatRequestBody,
-} from '../../../lib/messaging/server/buildChatRequestBody'
-import {
-  type SidepanelChatTarget,
-  toLlmProviderConfig,
-} from './sidepanel-chat-targets'
-
-type LlmChatRequestBodyInput = Parameters<typeof buildChatRequestBody>[0]
-
-type CommonSidepanelRequestInput = Omit<
-  LlmChatRequestBodyInput,
-  'provider' | 'message' | 'toolApprovalResponses' | 'isScheduledTask'
->
-
-interface BuildSidepanelPreparedSendMessagesRequestInput
-  extends CommonSidepanelRequestInput {
-  agentServerUrl: string | undefined
-  target: SidepanelChatTarget | undefined
-  fallbackProvider: LlmProviderConfig
-  message?: string
-  approvalResponses?: ApprovalResponseData[] | null
-}
-
-export function buildSidepanelPreparedSendMessagesRequest({
-  agentServerUrl,
-  target,
-  fallbackProvider,
-  message,
-  approvalResponses,
-  ...common
-}: BuildSidepanelPreparedSendMessagesRequestInput) {
-  if (target?.kind === 'acp') {
-    return {
-      api: `${agentServerUrl}/agents/${encodeURIComponent(target.agentId)}/sidepanel/chat`,
-      body: {
-        conversationId: common.conversationId,
-        message: message ?? '',
-        browserContext: common.browserContext,
-        userSystemPrompt: common.userSystemPrompt,
-        userWorkingDir: common.userWorkingDir,
-        selectedText: common.selectedText,
-        selectedTextSource: common.selectedTextSource,
-      },
-    }
-  }
-
-  const provider = toLlmProviderConfig(target) ?? fallbackProvider
-  return {
-    api: `${agentServerUrl}/chat`,
-    body: buildChatRequestBody({
-      ...common,
-      provider,
-      message,
-      toolApprovalResponses: approvalResponses ?? undefined,
-    }),
-  }
-}
-
-export function toProviderOption(target: SidepanelChatTarget): Provider {
-  return {
-    id: target.id,
-    name: target.name,
-    type: target.type,
-    kind: target.kind,
-    agentId: target.kind === 'acp' ? target.agentId : undefined,
-    adapterName: target.kind === 'acp' ? target.adapterName : undefined,
-    modelLabel: target.kind === 'acp' ? target.modelLabel : undefined,
-    modelControl: target.kind === 'acp' ? target.modelControl : undefined,
-  }
-}

packages/browseros-agent/apps/agent/lib/agent-conversations/types.ts

@@ -59,3 +59,15 @@ export interface AgentConversation {
   createdAt: number
   updatedAt: number
 }
+
+export interface AgentCardData {
+  agentId: string
+  name: string
+  model?: string
+  status: 'idle' | 'working' | 'error'
+  lastMessage?: string
+  lastMessageTimestamp?: number
+  activitySummary?: string
+  currentTool?: string
+  costUsd?: number
+}

packages/browseros-agent/apps/agent/lib/sse.ts

@@ -2,75 +2,29 @@ function isAbortError(error: unknown): boolean {
   return error instanceof DOMException && error.name === 'AbortError'
 }
 
-export interface ParsedSSEEvent<T> {
-  data: T
-  /** Numeric `id:` line on the same SSE event, if any. */
-  seq?: number
-}
-
 export function parseSSELines<T>(buffer: string): {
-  events: ParsedSSEEvent<T>[]
+  events: T[]
   remainder: string
 } {
-  // SSE events are separated by blank lines. Buffer lines until we hit
-  // a blank, then assemble each event. Lines we recognise: `id: <n>`
-  // and `data: <payload>`. Everything else is ignored.
-  const events: ParsedSSEEvent<T>[] = []
   const lines = buffer.split('\n')
-  // Find the last blank-line boundary; everything after it is the
-  // remainder (next event partially received).
-  let lastBoundary = -1
-  for (let i = lines.length - 1; i >= 0; i--) {
-    if (lines[i] === '') {
-      lastBoundary = i
-      break
-    }
-  }
-  const completeLines = lastBoundary >= 0 ? lines.slice(0, lastBoundary) : []
-  const remainder =
-    lastBoundary >= 0 ? lines.slice(lastBoundary + 1).join('\n') : buffer
+  const remainder = lines.pop() ?? ''
+  const events: T[] = []
 
-  let currentSeq: number | undefined
-  let currentData: string | null = null
-  const flush = () => {
-    if (currentData != null && currentData !== '[DONE]') {
-      try {
-        events.push({
-          data: JSON.parse(currentData) as T,
-          seq: currentSeq,
-        })
-      } catch {
-        // ignore
-      }
-    }
-    currentSeq = undefined
-    currentData = null
+  for (const line of lines) {
+    if (!line.startsWith('data: ')) continue
+    const payload = line.slice(6)
+    if (payload === '[DONE]') continue
+    try {
+      events.push(JSON.parse(payload) as T)
+    } catch {}
   }
 
-  for (const line of completeLines) {
-    if (line === '') {
-      flush()
-      continue
-    }
-    if (line.startsWith('id: ')) {
-      const n = Number.parseInt(line.slice(4).trim(), 10)
-      if (Number.isFinite(n)) currentSeq = n
-      continue
-    }
-    if (line.startsWith('data: ')) {
-      currentData = line.slice(6)
-    }
-  }
-  // Catch a complete trailing event with no terminating blank line —
-  // shouldn't happen in well-formed SSE, but be tolerant.
-  flush()
-
   return { events, remainder }
 }
 
 export async function consumeSSEStream<T>(
   response: Response,
-  onEvent: (event: T, meta: { seq?: number }) => void,
+  onEvent: (event: T) => void,
   signal?: AbortSignal,
 ): Promise<void> {
   const reader = response.body?.getReader()
@@ -95,7 +49,7 @@ export async function consumeSSEStream<T>(
       buffer = remainder
 
       for (const event of events) {
-        onEvent(event.data, { seq: event.seq })
+        onEvent(event)
       }
     }
   } catch (error) {
@@ -110,7 +64,7 @@ export async function consumeSSEStream<T>(
     if (buffer) {
       const { events } = parseSSELines<T>(buffer)
       for (const event of events) {
-        onEvent(event.data, { seq: event.seq })
+        onEvent(event)
       }
     }
   }

packages/browseros-agent/apps/agent/package.json

@@ -9,7 +9,6 @@
     "build": "bun run codegen && wxt build",
     "build:dev": "bun --env-file=.env.development wxt build --mode development",
     "zip": "wxt zip",
-    "test": "bun run ../../scripts/run-bun-test.ts ./apps/agent",
     "compile": "bun --env-file=.env.development wxt prepare && tsgo --noEmit",
     "lint": "bunx biome check",
     "typecheck": "bun --env-file=.env.development wxt prepare && tsgo --noEmit",

packages/browseros-agent/apps/eval/.env.example

@@ -1,51 +0,0 @@
-# Copy to .env.development for local eval runs.
-
-# Provider keys used by existing config files.
-OPENROUTER_API_KEY=
-FIREWORKS_API_KEY=
-ANTHROPIC_API_KEY=
-OPENAI_API_KEY=
-GOOGLE_GENERATIVE_AI_API_KEY=
-
-# Claude Agent SDK token used by performance_grader.
-CLAUDE_CODE_OAUTH_TOKEN=
-
-# Suite-mode model selection.
-EVAL_VARIANT=local
-EVAL_AGENT_PROVIDER=openai-compatible
-EVAL_AGENT_MODEL=
-EVAL_AGENT_API_KEY=
-EVAL_AGENT_BASE_URL=
-EVAL_AGENT_SUPPORTS_IMAGES=true
-
-# Optional suite-mode executor override for orchestrator suites.
-EVAL_EXECUTOR_MODEL=
-EVAL_EXECUTOR_API_KEY=
-EVAL_EXECUTOR_BASE_URL=
-
-# Clado visual action executor.
-CLADO_ACTION_MODEL=
-CLADO_ACTION_API_KEY=
-CLADO_ACTION_BASE_URL=
-# Backward-compatible alias used by older local scripts.
-CLADO_ACTION_URL=
-
-# BrowserOS runner.
-BROWSEROS_BINARY=/Applications/BrowserOS.app/Contents/MacOS/BrowserOS
-BROWSEROS_SERVER_URL=http://127.0.0.1:9110
-BROWSEROS_SERVER_LOG_DIR=/tmp/browseros-server-logs
-BROWSEROS_CONFIG_URL=
-
-# Captcha solver extension.
-NOPECHA_API_KEY=
-
-# WebArena-Infinity.
-WEBARENA_INFINITY_DIR=
-INFINITY_APP_URL=
-
-# R2 publishing and weekly report.
-EVAL_R2_ACCOUNT_ID=
-EVAL_R2_ACCESS_KEY_ID=
-EVAL_R2_SECRET_ACCESS_KEY=
-EVAL_R2_BUCKET=browseros-eval
-EVAL_R2_CDN_BASE_URL=https://eval.browseros.com

packages/browseros-agent/apps/eval/README.md

@@ -14,7 +14,6 @@ Evaluation framework for BrowserOS browser automation agents. Runs tasks from st
 
 ```bash
 cd apps/eval
-cp .env.example .env.development
 # Edit .env.development with your keys, then:
 bun run eval
 ```
@@ -24,55 +23,11 @@ Opens the eval dashboard at `http://localhost:9900` in config mode. From there:
 ### CLI mode
 
 ```bash
-bun run eval -c configs/legacy/browseros-agent-weekly.json
-bun run eval suite --config configs/legacy/browseros-agent-weekly.json --publish r2
+bun run eval -c configs/browseros-agent-weekly.json
 ```
 
 Runs immediately. Dashboard still available at `http://localhost:9900` for live progress.
 
-The `suite` command is the workflow-compatible full loop: execute tasks, run graders, write artifacts, and optionally publish to R2. The old `-c` form remains supported during migration.
-
-```bash
-bun run eval run --config configs/legacy/browseros-agent-weekly.json
-bun run eval suite --suite configs/suites/agisdk-daily-10.json --variant kimi-fireworks --publish r2
-bun run eval grade --run results/browseros-agent-weekly/2026-04-29-1430
-bun run eval publish --run results/browseros-agent-weekly/2026-04-29-1430 --target r2
-```
-
-Config files live in two groups:
-
-```txt
-configs/legacy/  # Complete EvalConfig files used by older workflows and the dashboard
-configs/suites/  # Suite definitions; model/provider comes from CLI flags or env
-```
-
-Suite mode takes model settings from CLI flags first, then env:
-
-```bash
-EVAL_VARIANT=kimi-fireworks \
-EVAL_AGENT_PROVIDER=openai-compatible \
-EVAL_AGENT_MODEL=accounts/fireworks/models/kimi-k2p5 \
-EVAL_AGENT_API_KEY=$FIREWORKS_API_KEY \
-EVAL_AGENT_BASE_URL=https://api.fireworks.ai/inference/v1 \
-bun run eval suite --suite configs/suites/agisdk-daily-10.json --publish r2
-```
-
-### Suites and variants
-
-A **suite** is what we run: the task dataset, graders, worker count, timeout, and browser settings. For example, `agisdk-daily-10` means "run these 10 AGI SDK tasks and grade them with `agisdk_state_diff`."
-
-A **variant** is the model setup we are testing on that suite. `EVAL_VARIANT` is just the human-readable name for that setup. The actual model connection still comes from `EVAL_AGENT_PROVIDER`, `EVAL_AGENT_MODEL`, `EVAL_AGENT_API_KEY`, and `EVAL_AGENT_BASE_URL`.
-
-This lets us run the same suite against multiple model setups without copying the benchmark config:
-
-```txt
-agisdk-daily-10 + kimi-fireworks
-agisdk-daily-10 + claude-sonnet
-agisdk-daily-10 + clado-action-000159
-```
-
-For `orchestrator-executor` suites, there can also be an executor model/backend. The `EVAL_AGENT_*` vars describe the main agent or orchestrator. The optional `EVAL_EXECUTOR_*` or `CLADO_ACTION_*` vars describe the delegated executor.
-
 ## Agent types
 
 | Type | Description |
@@ -111,9 +66,9 @@ The orchestrator works with any LLM provider. The executor can be another LLM, o
     },
     "executor": {
       "provider": "clado-action",
-      "model": "Qwen3.5-35B-A3B-action-000159-merged",
+      "model": "qwen3-vl-30b-a3b-instruct",
       "apiKey": "",
-      "baseUrl": "https://clado-ai--clado-browseros-action-000159-merged-actionmod-f4a6ef.modal.run"
+      "baseUrl": "https://clado-ai--clado-browseros-action-actionmodel-generate.modal.run"
     }
   }
 }
@@ -141,20 +96,6 @@ The `apiKey` field supports two formats:
 - **Env var name**: `"OPENAI_API_KEY"` — resolved from `.env.development` at runtime
 - **Direct value**: `"sk-xxxxx"` — used as-is (not recommended)
 
-### Environment variables
-
-| Variable | Used for |
-|----------|----------|
-| `EVAL_AGENT_PROVIDER`, `EVAL_AGENT_MODEL`, `EVAL_AGENT_API_KEY`, `EVAL_AGENT_BASE_URL`, `EVAL_AGENT_SUPPORTS_IMAGES` | Suite variant model selection |
-| `FIREWORKS_API_KEY`, `OPENROUTER_API_KEY`, `ANTHROPIC_API_KEY`, provider-specific keys | Config-file or provider-backed model calls |
-| `EVAL_EXECUTOR_MODEL`, `EVAL_EXECUTOR_API_KEY`, `EVAL_EXECUTOR_BASE_URL` | Suite-mode orchestrator executor override |
-| `CLADO_ACTION_MODEL`, `CLADO_ACTION_API_KEY`, `CLADO_ACTION_BASE_URL` | Clado executor defaults |
-| `BROWSEROS_BINARY` | BrowserOS binary path in CI/local smoke runs |
-| `BROWSEROS_SERVER_URL` | Optional grader MCP URL override |
-| `WEBARENA_INFINITY_DIR` | Local WebArena-Infinity checkout for Infinity tasks |
-| `NOPECHA_API_KEY` | CAPTCHA solver extension |
-| `EVAL_R2_ACCOUNT_ID`, `EVAL_R2_ACCESS_KEY_ID`, `EVAL_R2_SECRET_ACCESS_KEY`, `EVAL_R2_BUCKET`, `EVAL_R2_CDN_BASE_URL` | R2 upload and viewer URL |
-
 ### Supported providers
 
 | Provider | `provider` value | Requires `baseUrl` |
@@ -169,22 +110,6 @@ The `apiKey` field supports two formats:
 | Ollama | `ollama` | No |
 | Clado Action (executor only) | `clado-action` | Yes |
 
-### R2 publishing
-
-`suite --config ... --publish r2` and `publish --target r2` upload the run artifacts plus `viewer.html` to the viewer-compatible R2 layout:
-
-```bash
-export EVAL_R2_ACCOUNT_ID=...
-export EVAL_R2_ACCESS_KEY_ID=...
-export EVAL_R2_SECRET_ACCESS_KEY=...
-export EVAL_R2_BUCKET=browseros-eval
-export EVAL_R2_CDN_BASE_URL=https://eval.browseros.com
-```
-
-`EVAL_R2_CDN_BASE_URL` must be a public R2 custom domain, `r2.dev` URL, or Worker URL. Do not set it to the private `*.r2.cloudflarestorage.com` S3 API endpoint.
-
-Published runs are available at `EVAL_R2_CDN_BASE_URL/viewer.html?run=<run-id>`.
-
 ### BrowserOS infrastructure
 
 ```json
@@ -212,12 +137,10 @@ Each worker gets its own Chrome instance. Worker N uses `base_port + N` for CDP
 
 | File | Tasks | Description |
 |------|-------|-------------|
-| `agisdk-daily-10.jsonl` | 10 | Daily AGI SDK / REAL Bench subset |
 | `webvoyager.jsonl` | 643 | Full WebVoyager benchmark |
 | `mind2web.jsonl` | 300 | Online-Mind2Web |
 | `webbench-{0,1,2}of4-50.jsonl` | 50 each | WebBench shards (50-task subsets) |
-| `agisdk-real-smoke.jsonl` | 1 | AGI SDK / REAL Bench smoke task |
-| `agisdk-real.jsonl` | 36 | AGI SDK / REAL Bench (action-only tasks) |
+| `agisdk-real.jsonl` | 40 | AGI SDK / REAL Bench (action-only tasks) |
 | `webarena-infinity-hard-50.jsonl` | 50 | WebArena-Infinity hard set |
 | `browsecomp-medium-hard-50.jsonl` | 50 | BrowseComp medium-hard |
 | `browsecomp-very-hard-50.jsonl` | 50 | BrowseComp very-hard |
@@ -244,47 +167,14 @@ results/
   browseros-agent-weekly/
     2026-04-29-1430/
       Amazon--0/
-        attempt.json          # Stable attempt summary for viewer/reporting
         metadata.json         # Task result, timing, grader scores
-        grades.json           # Compact grader results
         messages.jsonl         # Full message log
-        grader-artifacts/      # Grader-specific inputs/outputs/stderr
         screenshots/
           001.png              # Step-by-step screenshots
           002.png
       summary.json             # Aggregate pass rates
 ```
 
-R2 publishing preserves the task files under `runs/<run-id>/...`, writes `runs/<run-id>/manifest.json`, and uploads `viewer.html` at the bucket root. The viewer URL is `EVAL_R2_CDN_BASE_URL/viewer.html?run=<run-id>`.
-
-### R2 viewer manifest
-
-`runs/<run-id>/manifest.json` is the source of truth for the public viewer. New manifests include `schemaVersion: 2` and each task includes explicit artifact paths:
-
-```json
-{
-  "schemaVersion": 2,
-  "runId": "agisdk-real-smoke-2026-04-30-0000",
-  "tasks": [
-    {
-      "queryId": "agisdk-dashdish-10",
-      "paths": {
-        "metadata": "tasks/agisdk-dashdish-10/metadata.json",
-        "messages": "tasks/agisdk-dashdish-10/messages.jsonl",
-        "grades": "tasks/agisdk-dashdish-10/grades.json",
-        "trace": "tasks/agisdk-dashdish-10/trace.jsonl",
-        "screenshots": "tasks/agisdk-dashdish-10/screenshots",
-        "graderArtifacts": "tasks/agisdk-dashdish-10/grader-artifacts"
-      }
-    }
-  ]
-}
-```
-
-The static viewer uses `task.paths` when present. Older uploaded runs without `schemaVersion` or `task.paths` still work through the legacy inferred layout: `runs/<run-id>/<task-id>/metadata.json`, `messages.jsonl`, and `screenshots/<n>.png`.
-
-Manifest paths are stable artifact locations, not a guarantee that every optional artifact exists for every task. For example, `attempt.json`, `trace.jsonl`, or grader artifact directories may be absent when that artifact was not produced by the run.
-
 ## Troubleshooting
 
 **BrowserOS not found**: Expects `/Applications/BrowserOS.app/Contents/MacOS/BrowserOS`. Set `BROWSEROS_BINARY` to override.

packages/browseros-agent/apps/eval/configs/agisdk-real-smoke.json

@@ -7,8 +7,8 @@
     "baseUrl": "https://openrouter.ai/api/v1",
     "supportsImages": true
   },
-  "dataset": "../../data/agisdk-real-smoke.jsonl",
-  "num_workers": 1,
+  "dataset": "../data/agisdk-real.jsonl",
+  "num_workers": 10,
   "restart_server_per_task": true,
   "browseros": {
     "server_url": "http://127.0.0.1:9110",

packages/browseros-agent/apps/eval/configs/browseros-agent-weekly.json

@@ -7,7 +7,7 @@
     "baseUrl": "https://openrouter.ai/api/v1",
     "supportsImages": true
   },
-  "dataset": "../../data/webbench-2of4-50.jsonl",
+  "dataset": "../data/webbench-2of4-50.jsonl",
   "num_workers": 10,
   "restart_server_per_task": true,
   "browseros": {

packages/browseros-agent/apps/eval/configs/browseros-oe-agent-weekly.json

@@ -14,7 +14,7 @@
       "baseUrl": "https://api.fireworks.ai/inference/v1"
     }
   },
-  "dataset": "../../data/webbench-2of4-50.jsonl",
+  "dataset": "../data/webbench-2of4-50.jsonl",
   "num_workers": 10,
   "restart_server_per_task": true,
   "browseros": {

packages/browseros-agent/apps/eval/configs/browseros-oe-clado-weekly.json

@@ -9,12 +9,12 @@
     },
     "executor": {
       "provider": "clado-action",
-      "model": "Qwen3.5-35B-A3B-action-000159-merged",
+      "model": "qwen3-vl-30b-a3b-instruct",
       "apiKey": "",
-      "baseUrl": "https://clado-ai--clado-browseros-action-000159-merged-actionmod-f4a6ef.modal.run"
+      "baseUrl": "https://clado-ai--clado-browseros-action-actionmodel-generate.modal.run"
     }
   },
-  "dataset": "../../data/agisdk-real.jsonl",
+  "dataset": "../data/webbench-2of4-50.jsonl",
   "num_workers": 10,
   "restart_server_per_task": true,
   "browseros": {
@@ -23,11 +23,11 @@
     "base_server_port": 9110,
     "base_extension_port": 9310,
     "load_extensions": false,
-    "headless": true
+    "headless": false
   },
   "captcha": {
     "api_key_env": "NOPECHA_API_KEY"
   },
-  "graders": ["agisdk_state_diff"],
+  "graders": ["performance_grader"],
   "timeout_ms": 1800000
 }

packages/browseros-agent/apps/eval/configs/infinity-hard-50.json

@@ -7,7 +7,7 @@
     "baseUrl": "https://openrouter.ai/api/v1",
     "supportsImages": true
   },
-  "dataset": "../../data/webarena-infinity-hard-50.jsonl",
+  "dataset": "../data/webarena-infinity-hard-50.jsonl",
   "num_workers": 10,
   "restart_server_per_task": true,
   "browseros": {

packages/browseros-agent/apps/eval/configs/legacy/agisdk-real.json

@@ -1,26 +0,0 @@
-{
-  "agent": {
-    "type": "single",
-    "provider": "openai-compatible",
-    "model": "accounts/fireworks/models/kimi-k2p5",
-    "apiKey": "FIREWORKS_API_KEY",
-    "baseUrl": "https://api.fireworks.ai/inference/v1",
-    "supportsImages": true
-  },
-  "dataset": "../../data/agisdk-real.jsonl",
-  "num_workers": 4,
-  "restart_server_per_task": true,
-  "browseros": {
-    "server_url": "http://127.0.0.1:9110",
-    "base_cdp_port": 9010,
-    "base_server_port": 9110,
-    "base_extension_port": 9310,
-    "load_extensions": false,
-    "headless": false
-  },
-  "captcha": {
-    "api_key_env": "NOPECHA_API_KEY"
-  },
-  "graders": ["agisdk_state_diff"],
-  "timeout_ms": 1800000
-}

packages/browseros-agent/apps/eval/configs/suites/agisdk-daily-10.json

@@ -1,22 +0,0 @@
-{
-  "id": "agisdk-daily-10",
-  "dataset": "../../data/agisdk-daily-10.jsonl",
-  "agent": {
-    "type": "single"
-  },
-  "graders": ["agisdk_state_diff"],
-  "workers": 1,
-  "restartBrowserPerTask": true,
-  "timeoutMs": 1800000,
-  "browseros": {
-    "server_url": "http://127.0.0.1:9110",
-    "base_cdp_port": 9010,
-    "base_server_port": 9110,
-    "base_extension_port": 9310,
-    "load_extensions": false,
-    "headless": true
-  },
-  "captcha": {
-    "api_key_env": "NOPECHA_API_KEY"
-  }
-}

packages/browseros-agent/apps/eval/configs/suites/agisdk-real-smoke.json

@@ -1,22 +0,0 @@
-{
-  "id": "agisdk-real-smoke",
-  "dataset": "../../data/agisdk-real-smoke.jsonl",
-  "agent": {
-    "type": "single"
-  },
-  "graders": ["agisdk_state_diff"],
-  "workers": 1,
-  "restartBrowserPerTask": true,
-  "timeoutMs": 1800000,
-  "browseros": {
-    "server_url": "http://127.0.0.1:9110",
-    "base_cdp_port": 9010,
-    "base_server_port": 9110,
-    "base_extension_port": 9310,
-    "load_extensions": false,
-    "headless": false
-  },
-  "captcha": {
-    "api_key_env": "NOPECHA_API_KEY"
-  }
-}

packages/browseros-agent/apps/eval/configs/suites/agisdk-real.json

@@ -1,22 +0,0 @@
-{
-  "id": "agisdk-real",
-  "dataset": "../../data/agisdk-real.jsonl",
-  "agent": {
-    "type": "single"
-  },
-  "graders": ["agisdk_state_diff"],
-  "workers": 1,
-  "restartBrowserPerTask": true,
-  "timeoutMs": 1800000,
-  "browseros": {
-    "server_url": "http://127.0.0.1:9110",
-    "base_cdp_port": 9010,
-    "base_server_port": 9110,
-    "base_extension_port": 9310,
-    "load_extensions": false,
-    "headless": false
-  },
-  "captcha": {
-    "api_key_env": "NOPECHA_API_KEY"
-  }
-}

packages/browseros-agent/apps/eval/configs/test-mind2web.json

@@ -5,7 +5,7 @@
     "model": "openai/gpt-4.1",
     "apiKey": "OPENROUTER_API_KEY"
   },
-  "dataset": "../../data/mind2web.jsonl",
+  "dataset": "../data/mind2web.jsonl",
   "num_workers": 5,
   "restart_server_per_task": true,
   "browseros": {

packages/browseros-agent/apps/eval/configs/test-webvoyager.json

@@ -7,7 +7,7 @@
     "baseUrl": "https://api.fireworks.ai/inference/v1",
     "supportsImages": true
   },
-  "dataset": "../../data/webvoyager.jsonl",
+  "dataset": "../data/webvoyager.jsonl",
   "num_workers": 3,
   "restart_server_per_task": true,
   "browseros": {

packages/browseros-agent/apps/eval/data/agisdk-daily-10.jsonl

@@ -1,10 +0,0 @@
-{"query_id": "agisdk-dashdish-10", "dataset": "agisdk-real", "query": "Place an order from \"Souvla\" for a \"Medium Classic Cheeseburger\" and a \"Small Bacon Double Cheeseburger\" with \"Standard Delivery\" as the method with the default charged options.", "graders": ["agisdk_state_diff"], "start_url": "https://evals-dashdish.vercel.app", "metadata": {"original_task_id": "dashdish-10", "website": "DashDish", "category": "agisdk-real", "additional": {"agisdk_task_id": "dashdish-10", "challenge_type": "action", "difficulty": "hard", "similar_to": "Doordash"}}}
-{"query_id": "agisdk-fly-unified-5", "dataset": "agisdk-real", "query": "Find me the cheapest fare for a flight from Orlando to Milwaukee on December 5th, 2024 and book it.\nPassenger: John Doe\nDate of Birth: 01/01/1990\nSex: Male\nSeat Selection: No\nPayment: Credit Card (378342143523967), Exp: 12/30, Security Code: 420 Address: 123 Main St, San Francisco, CA, 94105, USA, Phone: 555-123-4567, Email: johndoe@example.com.", "graders": ["agisdk_state_diff"], "start_url": "https://evals-fly-unified.vercel.app", "metadata": {"original_task_id": "fly-unified-5", "website": "Fly Unified", "category": "agisdk-real", "additional": {"agisdk_task_id": "fly-unified-5", "challenge_type": "retrieval-action", "difficulty": "medium", "similar_to": "United Airlines"}}}
-{"query_id": "agisdk-udriver-10", "dataset": "agisdk-real", "query": "Order me a ride for 4pm, I'll be at the de Young muesum headed to the Waterbar, fanciest option possible please.", "graders": ["agisdk_state_diff"], "start_url": "https://evals-udriver.vercel.app", "metadata": {"original_task_id": "udriver-10", "website": "UDriver", "category": "agisdk-real", "additional": {"agisdk_task_id": "udriver-10", "challenge_type": "action", "difficulty": "hard", "similar_to": "Uber"}}}
-{"query_id": "agisdk-udriver-9", "dataset": "agisdk-real", "query": "Book me a ride from the thai restaurant I last took a ride to for later today at 2pm, I'll be at 333 Apartments on Fremont", "graders": ["agisdk_state_diff"], "start_url": "https://evals-udriver.vercel.app", "metadata": {"original_task_id": "udriver-9", "website": "UDriver", "category": "agisdk-real", "additional": {"agisdk_task_id": "udriver-9", "challenge_type": "retrieval-action", "difficulty": "hard", "similar_to": "Uber"}}}
-{"query_id": "agisdk-topwork-4", "dataset": "agisdk-real", "query": "Create a job post for a UI/UX Designer with expertise in Figma, Sketch, and Adobe Creative Suite, including project details, timeline, and required skills (Wireframing, Prototyping, Responsive Design).", "graders": ["agisdk_state_diff"], "start_url": "https://evals-topwork.vercel.app", "metadata": {"original_task_id": "topwork-4", "website": "TopWork", "category": "agisdk-real", "additional": {"agisdk_task_id": "topwork-4", "challenge_type": "action", "difficulty": "medium", "similar_to": "Upwork"}}}
-{"query_id": "agisdk-gocalendar-4", "dataset": "agisdk-real", "query": "Change the \"Team Check-In\" event on July 18, 2024, name to \"Project Kickoff\" and update the location to \"Zoom\"", "graders": ["agisdk_state_diff"], "start_url": "https://evals-gocalendar.vercel.app", "metadata": {"original_task_id": "gocalendar-4", "website": "GoCalendar", "category": "agisdk-real", "additional": {"agisdk_task_id": "gocalendar-4", "challenge_type": "action", "difficulty": "medium", "similar_to": "Google Calendar"}}}
-{"query_id": "agisdk-staynb-6", "dataset": "agisdk-real", "query": "Find and book the stay with the best value for money (cheapest stay with the best reviews) for 1 day. For fields you don't know the answer for, just fill them in with anything of your choice.", "graders": ["agisdk_state_diff"], "start_url": "https://evals-staynb.vercel.app", "metadata": {"original_task_id": "staynb-6", "website": "StayNB", "category": "agisdk-real", "additional": {"agisdk_task_id": "staynb-6", "challenge_type": "retrieval-action", "difficulty": "medium", "similar_to": "Airbnb"}}}
-{"query_id": "agisdk-udriver-11", "dataset": "agisdk-real", "query": "I need to go from Pacific Catch on Chestnut back home to 333 Fremont now. If the fancy version is within ten dollars of the regular one, book that.", "graders": ["agisdk_state_diff"], "start_url": "https://evals-udriver.vercel.app", "metadata": {"original_task_id": "udriver-11", "website": "UDriver", "category": "agisdk-real", "additional": {"agisdk_task_id": "udriver-11", "challenge_type": "action", "difficulty": "hard", "similar_to": "Uber"}}}
-{"query_id": "agisdk-networkin-5", "dataset": "agisdk-real", "query": "Send a connection request to John Smith.", "graders": ["agisdk_state_diff"], "start_url": "https://evals-networkin.vercel.app", "metadata": {"original_task_id": "networkin-5", "website": "Networkin", "category": "agisdk-real", "additional": {"agisdk_task_id": "networkin-5", "challenge_type": "action", "difficulty": "easy", "similar_to": "LinkedIn"}}}
-{"query_id": "agisdk-zilloft-6", "dataset": "agisdk-real", "query": "Select a property listed in San Francisco as \"Condos\" within a price range under $300,000 and request a tour for tomorrow at 4:00 PM. Use these contact details: Name: Sarah Brown, Email: sarahbrown@example.com, Phone: 555-987-6543.", "graders": ["agisdk_state_diff"], "start_url": "https://evals-zilloft.vercel.app", "metadata": {"original_task_id": "zilloft-6", "website": "Zilloft", "category": "agisdk-real", "additional": {"agisdk_task_id": "zilloft-6", "challenge_type": "action", "difficulty": "medium", "similar_to": "Zillow"}}}

packages/browseros-agent/apps/eval/data/agisdk-real-smoke.jsonl

@@ -1 +0,0 @@
-{"query_id": "agisdk-dashdish-10", "dataset": "agisdk-real", "query": "Place an order from \"Souvla\" for a \"Medium Classic Cheeseburger\" and a \"Small Bacon Double Cheeseburger\" with \"Standard Delivery\" as the method with the default charged options.", "graders": ["agisdk_state_diff"], "start_url": "https://evals-dashdish.vercel.app", "metadata": {"original_task_id": "dashdish-10", "website": "DashDish", "category": "agisdk-real", "additional": {"agisdk_task_id": "dashdish-10", "challenge_type": "action", "difficulty": "hard", "similar_to": "Doordash"}}}

packages/browseros-agent/apps/eval/data/agisdk-real.jsonl

@@ -32,5 +32,9 @@
 {"query_id": "agisdk-networkin-10", "dataset": "agisdk-real", "query": "Generate a polite follow-up message for a previous unanswered chat, starting with \"Following up on\".", "graders": ["agisdk_state_diff"], "start_url": "https://evals-networkin.vercel.app", "metadata": {"original_task_id": "networkin-10", "website": "Networkin", "category": "agisdk-real", "additional": {"agisdk_task_id": "networkin-10", "challenge_type": "action", "difficulty": "medium", "similar_to": "LinkedIn"}}}
 {"query_id": "agisdk-gomail-3", "dataset": "agisdk-real", "query": "Compose a new email to jonathan.smith@example.com with the subject \"Meeting Notes\" and body \"Please find the meeting notes attached.\"", "graders": ["agisdk_state_diff"], "start_url": "https://evals-gomail.vercel.app", "metadata": {"original_task_id": "gomail-3", "website": "GoMail", "category": "agisdk-real", "additional": {"agisdk_task_id": "gomail-3", "challenge_type": "action", "difficulty": "easy", "similar_to": "Gmail"}}}
 {"query_id": "agisdk-udriver-6", "dataset": "agisdk-real", "query": "Me and 4 friends need a ride from the Palace Hotel to dinner at Osha Thai leaving now", "graders": ["agisdk_state_diff"], "start_url": "https://evals-udriver.vercel.app", "metadata": {"original_task_id": "udriver-6", "website": "UDriver", "category": "agisdk-real", "additional": {"agisdk_task_id": "udriver-6", "challenge_type": "action", "difficulty": "hard", "similar_to": "Uber"}}}
+{"query_id": "agisdk-staynb-9", "dataset": "agisdk-real", "query": "Book a stay with the maximum number of guests supported. For fields you don't know the answer for, just fill them in with anything of your choice.", "graders": ["agisdk_state_diff"], "start_url": "https://evals-staynb.vercel.app", "metadata": {"original_task_id": "staynb-9", "website": "StayNB", "category": "agisdk-real", "additional": {"agisdk_task_id": "staynb-9", "challenge_type": "action", "difficulty": "hard", "similar_to": "Airbnb"}}}
 {"query_id": "agisdk-zilloft-3", "dataset": "agisdk-real", "query": "Find a home in San Diego priced under $150,000 with at least 2 bedrooms and request a tour. Use these details: Contact Name: John Doe, Email: johndoe@example.com, Phone: 555-123-4567, Tour Time: 2:00 PM, Tour Date: First available.", "graders": ["agisdk_state_diff"], "start_url": "https://evals-zilloft.vercel.app", "metadata": {"original_task_id": "zilloft-3", "website": "Zilloft", "category": "agisdk-real", "additional": {"agisdk_task_id": "zilloft-3", "challenge_type": "retrieval-action", "difficulty": "easy", "similar_to": "Zillow"}}}
 {"query_id": "agisdk-fly-unified-6", "dataset": "agisdk-real", "query": "Reserve me a seat for the flight from Austin to Pittsburgh departing on December 11th, 2024 at 8:00 in Basic Economy.\nPassenger: Alice Brown\nDate of Birth: 05/20/1992\nSex: Female\nSeat Selection: Yes (Aisle seat)\nPayment: Credit Card (378342143523967), Exp: 09/27, security code: 332 Address: 789 Pine St, Los Angeles, CA, 90012, USA, Phone: 555-456-7890, Email: alicebrown@example.com.", "graders": ["agisdk_state_diff"], "start_url": "https://evals-fly-unified.vercel.app", "metadata": {"original_task_id": "fly-unified-6", "website": "Fly Unified", "category": "agisdk-real", "additional": {"agisdk_task_id": "fly-unified-6", "challenge_type": "action", "difficulty": "medium", "similar_to": "United Airlines"}}}
+{"query_id": "agisdk-opendining-3", "dataset": "agisdk-real", "query": "Book a table at \"The Royal Dine\" for a party of 4 on July 20, 2024, at 7 PM. For fields you don't know the answer for, just fill them in with anything of your choice.", "graders": ["agisdk_state_diff"], "start_url": "https://evals-opendining.vercel.app", "metadata": {"original_task_id": "opendining-3", "website": "OpenDining", "category": "agisdk-real", "additional": {"agisdk_task_id": "opendining-3", "challenge_type": "action", "difficulty": "easy", "similar_to": "OpenTable"}}}
+{"query_id": "agisdk-gocalendar-7", "dataset": "agisdk-real", "query": "Reschedule the \"Morning Coffee with sister\" event from July 18, 2024, at 9 AM to July 19, 2024, at 10AM using drag-and-drop functionality", "graders": ["agisdk_state_diff"], "start_url": "https://evals-gocalendar.vercel.app", "metadata": {"original_task_id": "gocalendar-7", "website": "GoCalendar", "category": "agisdk-real", "additional": {"agisdk_task_id": "gocalendar-7", "challenge_type": "action", "difficulty": "medium", "similar_to": "Google Calendar"}}}
+{"query_id": "agisdk-staynb-5", "dataset": "agisdk-real", "query": "Use the search bar to look for a stay. For the \"Where\" section, use the \"Search by region\" popover and select \"Europe\". Set the check-in date to October 13th and the check-out date to October 23rd. For the \"Who\" section, select 1 infant, 2 children, and 2 adults. Press the search button, select the first stay, and book it.", "graders": ["agisdk_state_diff"], "start_url": "https://evals-staynb.vercel.app", "metadata": {"original_task_id": "staynb-5", "website": "StayNB", "category": "agisdk-real", "additional": {"agisdk_task_id": "staynb-5", "challenge_type": "action", "difficulty": "medium", "similar_to": "Airbnb"}}}

packages/browseros-agent/apps/eval/package.json

@@ -5,7 +5,6 @@
   "type": "module",
   "scripts": {
     "eval": "bun --env-file=.env.development run src/index.ts",
-    "test": "bun run ../../scripts/run-bun-test.ts ./apps/eval/tests",
     "typecheck": "tsc --noEmit"
   },
   "dependencies": {

packages/browseros-agent/apps/eval/scripts/build-agisdk-dataset.py

@@ -64,37 +64,6 @@ EXCLUDED_TASKS = {
     # the grader's first criterion (search history contains "stanford") was
     # never triggered server-side. Eval-site bug.
     "networkin-9",
-    # Goal text instructs "move event to July 19, 10 AM" but the grader expects
-    # `eventsDiff.updated.*.start == "2024-07-18T17:00:00Z"` (= July 18, 10 AM
-    # PDT — same day, 1 hour shift). Goal contradicts grader: following the
-    # goal yields July 19 timestamps; satisfying the grader requires ignoring
-    # the explicit "to July 19" instruction. Confirmed via 8-trial deep-dive:
-    # never passed even after the Phase 2 HTML5 dnd dispatch fix made the drag
-    # actually populate `eventsDiff.updated` (now produces July 19 values, but
-    # grader rejects them).
-    "gocalendar-7",
-    # Grader hardcodes literal year strings `'Oct 13 2025'` / `'Oct 23 2025'`
-    # in checkin/checkout criteria. Today is 2026, and the staynb date picker
-    # interprets bare "Oct 13" as the most recent past instance — currently
-    # 2024, not 2025. Even a perfectly-acting agent cannot produce a booking
-    # whose persisted date contains "2025". Confirmed via 8 trials, 0 passes.
-    "staynb-5",
-    # Goal says "maximum number of guests supported"; grader expects the very
-    # specific string "32 Guests, 16 Infants" — which requires the agent to
-    # know that (a) Adults+Children sum into the displayed "Guests" count,
-    # (b) Infants render separately, (c) Pets are excluded, (d) per-category
-    # cap is 16 despite no UI affordance signalling it. None of this is in
-    # the prompt. 8 trials, 0 passes; even Opus 4.6 stopped at 16 (one
-    # category maxed). Task is under-specified relative to grader expectation.
-    "staynb-9",
-    # Grader requires `contains(booking.date, '2024-07-20')` but the eval-site
-    # date picker is a React-controlled textbox that the agent's `fill` tool
-    # frequently no-ops on. 3 of 8 trials passed (when fill happened to stick),
-    # 5 failed with `actual_value: False` (booking persisted with the eval-site
-    # default search date, not Jul 20). Effectively a coin-flip task that
-    # exercises tool-fidelity flakiness rather than agent capability —
-    # contributes noise, not signal. Excluding for eval reliability.
-    "opendining-3",
 }
 
 # Far-future replacement used by `freshen_goal_dates` when a task's hardcoded

packages/browseros-agent/apps/eval/scripts/test-clado-api.ts

@@ -1,73 +1,34 @@
 /**
- * Smoke-test for the Clado BrowserOS Action endpoint.
- *
- * Health-checks the model, then runs a generate call and prints every
- * field the new contract documents (action, coordinates, text, key,
- * direction, scroll/drag fields, wait, end+final_answer, thinking,
- * parse_error, raw_response).
+ * Test script for Clado API endpoints (grounding + action models)
  *
  * Usage:
  *   bun apps/eval/scripts/test-clado-api.ts [screenshot-path]
  *
- * If no screenshot path is given, captures one over MCP from a
- * running BrowserOS server (default http://127.0.0.1:9110, override
- * with BROWSEROS_URL).
- *
- * Cold start can take ~5 minutes; the script waits up to 6.
+ * If no screenshot provided, captures one from a running BrowserOS server.
  */
 
 import { readFile } from 'node:fs/promises'
 import { resolve } from 'node:path'
 
 const ACTION_URL =
-  'https://clado-ai--clado-browseros-action-000159-merged-actionmod-f4a6ef.modal.run'
+  'https://clado-ai--clado-browseros-action-actionmodel-generate.modal.run'
 const ACTION_HEALTH_URL =
-  'https://clado-ai--clado-browseros-action-000159-merged-actionmod-5e5033.modal.run'
+  'https://clado-ai--clado-browseros-action-actionmodel-health.modal.run'
+const GROUNDING_URL =
+  'https://clado-ai--clado-browseros-grounding-groundingmodel-generate.modal.run'
+const GROUNDING_HEALTH_URL =
+  'https://clado-ai--clado-browseros-grounding-groundingmodel-health.modal.run'
 
-const COLD_START_BUDGET_MS = 360_000 // 6 min — Clado cold start is ~5 min
-const COLD_START_WARN_MS = 30_000
-
-interface CladoResponse {
-  action?: string | null
-  thinking?: string | null
-  raw_response?: string
-  parse_error?: string | null
-  inference_time_seconds?: number
-  x?: number
-  y?: number
-  text?: string
-  key?: string
-  direction?: string
-  amount?: number
-  startX?: number
-  startY?: number
-  endX?: number
-  endY?: number
-  time?: number
-  final_answer?: string | null
-}
-
-async function checkHealth(): Promise<boolean> {
-  console.log(`\n--- Action model health ---`)
-  console.log(`  URL:   ${ACTION_HEALTH_URL}`)
-  console.log(
-    `  Note:  cold start can take ~5 min; waiting up to ${COLD_START_BUDGET_MS / 1000}s.`,
-  )
+async function checkHealth(name: string, url: string): Promise<boolean> {
+  console.log(`\n--- ${name} health check ---`)
+  console.log(`  URL: ${url}`)
   const start = performance.now()
-  const warn = setTimeout(() => {
-    console.log(
-      `  ...still waiting (${COLD_START_WARN_MS / 1000}s in) — model is likely cold-starting on Modal.`,
-    )
-  }, COLD_START_WARN_MS)
-
   try {
-    const resp = await fetch(ACTION_HEALTH_URL, {
-      signal: AbortSignal.timeout(COLD_START_BUDGET_MS),
-    })
+    const resp = await fetch(url, { signal: AbortSignal.timeout(30_000) })
     const elapsed = ((performance.now() - start) / 1000).toFixed(2)
     const body = await resp.text()
     console.log(`  Status: ${resp.status} (${elapsed}s)`)
-    console.log(`  Body:   ${body.slice(0, 400)}`)
+    console.log(`  Body: ${body.slice(0, 200)}`)
     return resp.ok
   } catch (err) {
     const elapsed = ((performance.now() - start) / 1000).toFixed(2)
@@ -75,34 +36,63 @@ async function checkHealth(): Promise<boolean> {
       `  FAILED (${elapsed}s): ${err instanceof Error ? err.message : err}`,
     )
     return false
-  } finally {
-    clearTimeout(warn)
   }
 }
 
-async function generate(
-  label: string,
+async function testGenerate(
+  name: string,
+  url: string,
   payload: Record<string, unknown>,
-): Promise<CladoResponse | null> {
-  console.log(`\n--- ${label} ---`)
-  console.log(`  URL:         ${ACTION_URL}`)
+): Promise<Record<string, unknown> | null> {
+  console.log(`\n--- ${name} generate ---`)
+  console.log(`  URL: ${url}`)
   console.log(`  Instruction: ${payload.instruction}`)
   console.log(
-    `  Image size:  ${((payload.image_base64 as string).length / 1024).toFixed(0)} KB (base64)`,
+    `  Image size: ${((payload.image_base64 as string).length / 1024).toFixed(0)} KB (base64)`,
   )
-  if (payload.history && payload.history !== 'None') {
-    console.log(`  History:     ${payload.history}`)
-  }
+  if (payload.history) console.log(`  History: ${payload.history}`)
 
   const start = performance.now()
-  let resp: Response
   try {
-    resp = await fetch(ACTION_URL, {
+    const resp = await fetch(url, {
       method: 'POST',
       headers: { 'Content-Type': 'application/json' },
       body: JSON.stringify(payload),
-      signal: AbortSignal.timeout(COLD_START_BUDGET_MS),
+      signal: AbortSignal.timeout(120_000),
     })
+    const elapsed = ((performance.now() - start) / 1000).toFixed(2)
+
+    if (!resp.ok) {
+      const body = await resp.text()
+      console.log(`  FAILED: HTTP ${resp.status} (${elapsed}s)`)
+      console.log(`  Body: ${body.slice(0, 400)}`)
+      return null
+    }
+
+    const result = (await resp.json()) as Record<string, unknown>
+    console.log(`  Status: ${resp.status} (${elapsed}s)`)
+    console.log(`  Action: ${result.action}`)
+    if (result.x !== null && result.x !== undefined)
+      console.log(`  Coordinates: (${result.x}, ${result.y})`)
+    if (result.text)
+      console.log(`  Text: ${(result.text as string).slice(0, 100)}`)
+    if (result.key) console.log(`  Key: ${result.key}`)
+    if (result.inference_time_seconds)
+      console.log(`  Inference: ${result.inference_time_seconds}s`)
+
+    // Show thinking if present
+    const raw = result.raw_response as string | undefined
+    if (raw) {
+      const thinkMatch = raw.match(/<thinking>([\s\S]*?)<\/thinking>/)
+      if (thinkMatch) {
+        const thinking = thinkMatch[1].trim()
+        console.log(
+          `  Thinking: ${thinking.slice(0, 200)}${thinking.length > 200 ? '...' : ''}`,
+        )
+      }
+    }
+
+    return result
   } catch (err) {
     const elapsed = ((performance.now() - start) / 1000).toFixed(2)
     console.log(
@@ -110,50 +100,6 @@ async function generate(
     )
     return null
   }
-  const elapsed = ((performance.now() - start) / 1000).toFixed(2)
-
-  if (!resp.ok) {
-    const body = await resp.text()
-    console.log(`  HTTP ${resp.status} ${resp.statusText} (${elapsed}s)`)
-    console.log(`  Body: ${body.slice(0, 400)}`)
-    return null
-  }
-
-  const result = (await resp.json()) as CladoResponse
-  console.log(`  HTTP ${resp.status} (${elapsed}s)`)
-  console.log(`  action:                ${result.action ?? 'null'}`)
-  if (result.parse_error) {
-    console.log(`  parse_error:           ${result.parse_error}`)
-  }
-  if (result.thinking) {
-    const trimmed = result.thinking.replace(/\s+/g, ' ').trim()
-    console.log(
-      `  thinking:              ${trimmed.slice(0, 240)}${trimmed.length > 240 ? '…' : ''}`,
-    )
-  }
-  if (typeof result.x === 'number' || typeof result.y === 'number') {
-    console.log(`  x, y:                  ${result.x}, ${result.y}`)
-  }
-  if (typeof result.text === 'string')
-    console.log(`  text:                  ${result.text.slice(0, 120)}`)
-  if (typeof result.key === 'string')
-    console.log(`  key:                   ${result.key}`)
-  if (typeof result.direction === 'string')
-    console.log(`  direction:             ${result.direction}`)
-  if (typeof result.amount === 'number')
-    console.log(`  amount:                ${result.amount}`)
-  if (typeof result.startX === 'number' || typeof result.endX === 'number') {
-    console.log(
-      `  drag:                  (${result.startX}, ${result.startY}) → (${result.endX}, ${result.endY})`,
-    )
-  }
-  if (typeof result.time === 'number')
-    console.log(`  time:                  ${result.time}s`)
-  if (result.final_answer)
-    console.log(`  final_answer:          ${result.final_answer.slice(0, 240)}`)
-  if (typeof result.inference_time_seconds === 'number')
-    console.log(`  inference_time_seconds: ${result.inference_time_seconds}`)
-  return result
 }
 
 async function loadScreenshot(path?: string): Promise<string> {
@@ -164,9 +110,10 @@ async function loadScreenshot(path?: string): Promise<string> {
     return data.toString('base64')
   }
 
+  // Try to capture from a running BrowserOS server
   const serverUrl = process.env.BROWSEROS_URL || 'http://127.0.0.1:9110'
   console.log(
-    `No screenshot path provided. Capturing from ${serverUrl} via MCP...`,
+    `No screenshot path provided. Trying to capture from ${serverUrl}...`,
   )
 
   const { Client } = await import('@modelcontextprotocol/sdk/client/index.js')
@@ -187,101 +134,82 @@ async function loadScreenshot(path?: string): Promise<string> {
       arguments: { format: 'png', page: 1 },
     })) as { content: Array<{ type: string; data?: string }> }
 
-    const image = result.content?.find((c) => c.type === 'image')
-    if (!image?.data)
-      throw new Error('No image data in take_screenshot response')
+    const imageContent = result.content?.find((c) => c.type === 'image')
+    if (!imageContent?.data)
+      throw new Error('No image data in screenshot response')
 
     console.log(
-      `Captured screenshot (${(image.data.length / 1024).toFixed(0)} KB base64)`,
+      `Captured screenshot (${(imageContent.data.length / 1024).toFixed(0)} KB base64)`,
     )
-    return image.data
+    return imageContent.data
   } finally {
     try {
       await transport.close()
-    } catch {
-      /* ignore */
-    }
+    } catch {}
   }
 }
 
-function summarize(history: CladoResponse[]): string {
-  if (history.length === 0) return 'None'
-  return history
-    .map((h) => {
-      switch (h.action) {
-        case 'click':
-        case 'double_click':
-        case 'right_click':
-        case 'hover':
-          return `${h.action}(${h.x}, ${h.y})`
-        case 'type':
-          return `type(${JSON.stringify(h.text ?? '')})`
-        case 'press_key':
-          return `press_key(${JSON.stringify(h.key ?? '')})`
-        case 'scroll':
-          return `scroll(${h.direction ?? 'down'})`
-        case 'drag':
-          return `drag(${h.startX},${h.startY} -> ${h.endX},${h.endY})`
-        case 'wait':
-          return `wait(${h.time ?? 1}s)`
-        case 'end':
-          return 'end()'
-        default:
-          return h.action ?? 'invalid'
-      }
-    })
-    .join(' -> ')
-}
-
 async function main() {
-  console.log('=== Clado action endpoint smoke test ===')
+  const screenshotPath = process.argv[2]
 
-  const healthy = await checkHealth()
-  if (!healthy) {
-    console.log('\nHealth check failed. Exiting.')
+  console.log('=== Clado API Test ===\n')
+
+  // Health checks (parallel)
+  const [actionHealthy, groundingHealthy] = await Promise.all([
+    checkHealth('Action Model', ACTION_HEALTH_URL),
+    checkHealth('Grounding Model', GROUNDING_HEALTH_URL),
+  ])
+
+  if (!actionHealthy && !groundingHealthy) {
+    console.log('\nBoth endpoints are down. Exiting.')
     process.exit(1)
   }
 
+  // Load screenshot
   let imageBase64: string
   try {
-    imageBase64 = await loadScreenshot(process.argv[2])
+    imageBase64 = await loadScreenshot(screenshotPath)
   } catch (err) {
     console.log(
       `\nFailed to load screenshot: ${err instanceof Error ? err.message : err}`,
     )
     console.log(
-      'Pass a path: bun apps/eval/scripts/test-clado-api.ts path/to/screenshot.png',
+      'Provide a screenshot path: bun apps/eval/scripts/test-clado-api.ts path/to/screenshot.png',
     )
     process.exit(1)
   }
 
-  const history: CladoResponse[] = []
+  const instruction = 'Click on the search button or search bar'
 
-  // Step 1: open task — let the model decide what to do.
-  const step1 = await generate('Step 1: cold task', {
-    instruction: 'Find the search bar and click it',
-    image_base64: imageBase64,
-    history: 'None',
-  })
-  if (step1?.action) history.push(step1)
-
-  // Step 2: continuation with history, asks for typing.
-  if (step1?.action) {
-    const step2 = await generate('Step 2: with history', {
-      instruction: 'Type "hello world" into the search bar',
+  // Test grounding model
+  if (groundingHealthy) {
+    await testGenerate('Grounding Model', GROUNDING_URL, {
+      instruction,
       image_base64: imageBase64,
-      history: summarize(history),
     })
-    if (step2?.action) history.push(step2)
+  } else {
+    console.log('\nSkipping grounding model (unhealthy)')
   }
 
-  // Step 3: ask for end with a final answer to exercise that field.
-  await generate('Step 3: ask for end+final_answer', {
-    instruction:
-      'You have completed the task. Reply with end() and final_answer="done".',
-    image_base64: imageBase64,
-    history: summarize(history),
-  })
+  // Test action model (no history)
+  if (actionHealthy) {
+    const result = await testGenerate('Action Model (step 1)', ACTION_URL, {
+      instruction,
+      image_base64: imageBase64,
+      history: 'None',
+    })
+
+    // Test action model with history (simulate multi-turn)
+    if (result && result.action === 'click') {
+      await testGenerate('Action Model (step 2, with history)', ACTION_URL, {
+        instruction: 'Type "hello world" in the search bar',
+        image_base64: imageBase64,
+        history: `click(${result.x}, ${result.y})`,
+      })
+    }
+  } else {
+    console.log('\nSkipping action model (unhealthy)')
+  }
 
   console.log('\n=== Done ===')
 }

packages/browseros-agent/apps/eval/scripts/upload-run.ts

@@ -1,43 +1,349 @@
-#!/usr/bin/env bun
-
 /**
  * Upload eval runs to R2.
  *
  * Two modes:
  *   bun scripts/upload-run.ts results/browseros-agent-weekly/2026-03-21-1730
+ *       → uploads that specific run
+ *
  *   bun scripts/upload-run.ts results/browseros-agent-weekly
+ *       → finds all timestamped subfolders, uploads any not yet in R2
+ *
+ * Env vars: EVAL_R2_ACCOUNT_ID, EVAL_R2_ACCESS_KEY_ID, EVAL_R2_SECRET_ACCESS_KEY
+ *           EVAL_R2_BUCKET (default: browseros-eval)
+ *           EVAL_R2_CDN_BASE_URL (default: https://eval.browseros.com)
  */
 
+import { readdir, readFile, stat } from 'node:fs/promises'
+import { basename, dirname, extname, join } from 'node:path'
 import {
-  loadR2ConfigFromEnv,
-  R2Publisher,
-} from '../src/publishing/r2-publisher'
+  GetObjectCommand,
+  PutObjectCommand,
+  S3Client,
+} from '@aws-sdk/client-s3'
 
-async function main(): Promise<void> {
-  const inputDir = process.argv[2]
-  if (!inputDir) {
-    throw new Error(
-      'Usage:\n' +
-        '  bun scripts/upload-run.ts results/config-name/2026-03-21-1730\n' +
-        '  bun scripts/upload-run.ts results/config-name',
+const CONCURRENCY = 20
+
+const CONTENT_TYPES: Record<string, string> = {
+  '.json': 'application/json',
+  '.jsonl': 'application/x-ndjson',
+  '.png': 'image/png',
+}
+
+interface R2Config {
+  accountId: string
+  accessKeyId: string
+  secretAccessKey: string
+  bucket: string
+  cdnBaseUrl: string
+}
+
+function loadConfig(): R2Config {
+  const accountId = process.env.EVAL_R2_ACCOUNT_ID
+  const accessKeyId = process.env.EVAL_R2_ACCESS_KEY_ID
+  const secretAccessKey = process.env.EVAL_R2_SECRET_ACCESS_KEY
+
+  if (!accountId || !accessKeyId || !secretAccessKey) {
+    console.error(
+      'Missing required env vars: EVAL_R2_ACCOUNT_ID, EVAL_R2_ACCESS_KEY_ID, EVAL_R2_SECRET_ACCESS_KEY',
     )
+    process.exit(1)
   }
 
-  const publisher = new R2Publisher({ config: loadR2ConfigFromEnv() })
-  const result = await publisher.publishPath(inputDir)
-  for (const run of result.uploadedRuns) {
-    console.log(`Uploaded ${run.uploadedFiles} files for ${run.runId}`)
-    console.log(run.viewerUrl)
+  return {
+    accountId,
+    accessKeyId,
+    secretAccessKey,
+    bucket: process.env.EVAL_R2_BUCKET || 'browseros-eval',
+    cdnBaseUrl: (
+      process.env.EVAL_R2_CDN_BASE_URL || 'https://eval.browseros.com'
+    ).replace(/\/+$/, ''),
   }
-  for (const runId of result.skippedRuns) {
-    console.log(`${runId}: already uploaded, skipping`)
-  }
-  console.log(
-    `Done. Uploaded ${result.uploadedRuns.length} run(s), skipped ${result.skippedRuns.length}.`,
+}
+
+function createClient(config: R2Config): S3Client {
+  return new S3Client({
+    region: 'auto',
+    endpoint: `https://${config.accountId}.r2.cloudflarestorage.com`,
+    credentials: {
+      accessKeyId: config.accessKeyId,
+      secretAccessKey: config.secretAccessKey,
+    },
+  })
+}
+
+async function upload(
+  client: S3Client,
+  bucket: string,
+  key: string,
+  body: Buffer,
+  contentType: string,
+) {
+  await client.send(
+    new PutObjectCommand({
+      Bucket: bucket,
+      Key: key,
+      Body: body,
+      ContentType: contentType,
+    }),
   )
 }
 
-main().catch((error) => {
-  console.error(error instanceof Error ? error.message : String(error))
-  process.exit(1)
-})
+async function collectFiles(dir: string): Promise<string[]> {
+  const files: string[] = []
+  const entries = await readdir(dir, { withFileTypes: true })
+  for (const entry of entries) {
+    const full = join(dir, entry.name)
+    if (entry.isDirectory()) {
+      files.push(...(await collectFiles(full)))
+    } else {
+      files.push(full)
+    }
+  }
+  return files
+}
+
+async function runPool<T>(
+  items: T[],
+  concurrency: number,
+  fn: (item: T) => Promise<void>,
+) {
+  let i = 0
+  const workers = Array.from({ length: concurrency }, async () => {
+    while (i < items.length) {
+      const idx = i++
+      await fn(items[idx])
+    }
+  })
+  await Promise.all(workers)
+}
+
+// Check if a run has already been uploaded to R2
+async function isUploaded(
+  client: S3Client,
+  bucket: string,
+  runId: string,
+): Promise<boolean> {
+  try {
+    await client.send(
+      new GetObjectCommand({
+        Bucket: bucket,
+        Key: `runs/${runId}/manifest.json`,
+      }),
+    )
+    return true
+  } catch {
+    return false
+  }
+}
+
+// Detect if a directory is a run dir (has task subdirs with metadata.json)
+// vs a config dir (has timestamped subdirs like 2026-03-21-1730/)
+async function isRunDir(dir: string): Promise<boolean> {
+  const entries = await readdir(dir, { withFileTypes: true })
+  const subdirs = entries.filter((e) => e.isDirectory())
+  for (const subdir of subdirs) {
+    const metaPath = join(dir, subdir.name, 'metadata.json')
+    const metaStat = await stat(metaPath).catch(() => null)
+    if (metaStat?.isFile()) return true
+  }
+  return false
+}
+
+async function uploadSingleRun(
+  runDir: string,
+  runId: string,
+  r2Config: R2Config,
+  client: S3Client,
+): Promise<void> {
+  const taskDirs = await readdir(runDir, { withFileTypes: true })
+  const taskEntries = taskDirs.filter((d) => d.isDirectory())
+
+  if (taskEntries.length === 0) {
+    console.warn(`  No task subdirectories in ${runId}, skipping`)
+    return
+  }
+
+  const manifestTasks: Record<string, unknown>[] = []
+  const jobs: { key: string; filePath: string; contentType: string }[] = []
+
+  // Extract agent config from first task
+  let agentConfig: Record<string, unknown> | undefined
+  let dataset: string | undefined
+
+  for (const taskDir of taskEntries) {
+    const taskId = taskDir.name
+    const taskPath = join(runDir, taskId)
+    const metaPath = join(taskPath, 'metadata.json')
+
+    let meta: Record<string, unknown> = {}
+    try {
+      meta = JSON.parse(await readFile(metaPath, 'utf-8'))
+    } catch {
+      continue
+    }
+
+    if (!agentConfig && meta.agent_config)
+      agentConfig = meta.agent_config as Record<string, unknown>
+    if (!dataset && meta.dataset) dataset = meta.dataset as string
+
+    const files = await collectFiles(taskPath)
+    let screenshotCount = 0
+
+    for (const file of files) {
+      const relative = file.slice(taskPath.length + 1)
+      const ext = extname(file)
+      if (relative.startsWith('screenshots/') && ext === '.png')
+        screenshotCount++
+
+      jobs.push({
+        key: `runs/${runId}/${taskId}/${relative}`,
+        filePath: file,
+        contentType: CONTENT_TYPES[ext] || 'application/octet-stream',
+      })
+    }
+
+    manifestTasks.push({
+      queryId: meta.query_id || taskId,
+      query: meta.query || '',
+      startUrl: meta.start_url || '',
+      status:
+        meta.termination_reason === 'completed'
+          ? 'completed'
+          : meta.termination_reason || 'unknown',
+      durationMs: meta.total_duration_ms || 0,
+      screenshotCount: (meta.screenshot_count as number) || screenshotCount,
+      graderResults: meta.grader_results || {},
+    })
+  }
+
+  if (manifestTasks.length === 0) {
+    console.warn(`  No completed tasks in ${runId}, skipping`)
+    return
+  }
+
+  console.log(
+    `  Uploading ${jobs.length} files across ${manifestTasks.length} tasks...`,
+  )
+
+  let uploaded = 0
+  await runPool(jobs, CONCURRENCY, async (job) => {
+    const body = await readFile(job.filePath)
+    await upload(client, r2Config.bucket, job.key, body, job.contentType)
+    uploaded++
+    if (uploaded % 50 === 0 || uploaded === jobs.length) {
+      console.log(`    ${uploaded}/${jobs.length}`)
+    }
+  })
+
+  // Read summary.json if it exists
+  let summaryData: Record<string, unknown> | undefined
+  try {
+    summaryData = JSON.parse(
+      await readFile(join(runDir, 'summary.json'), 'utf-8'),
+    )
+  } catch {}
+
+  // Upload manifest
+  const manifest = {
+    runId,
+    uploadedAt: new Date().toISOString(),
+    agentConfig,
+    dataset,
+    summary: summaryData
+      ? {
+          passRate: summaryData.passRate,
+          avgDurationMs: summaryData.avgDurationMs,
+        }
+      : undefined,
+    tasks: manifestTasks,
+  }
+  const manifestBody = Buffer.from(JSON.stringify(manifest, null, 2))
+  await upload(
+    client,
+    r2Config.bucket,
+    `runs/${runId}/manifest.json`,
+    manifestBody,
+    'application/json',
+  )
+
+  // Upload viewer.html to bucket root
+  const viewerPath = join(
+    import.meta.dir,
+    '..',
+    'src',
+    'dashboard',
+    'viewer.html',
+  )
+  const viewerBody = await readFile(viewerPath)
+  await upload(client, r2Config.bucket, 'viewer.html', viewerBody, 'text/html')
+
+  console.log(`  Uploaded ${uploaded + 2} files`)
+  console.log(`  ${r2Config.cdnBaseUrl}/viewer.html?run=${runId}`)
+}
+
+async function main() {
+  const inputDir = process.argv[2]
+  if (!inputDir) {
+    console.error(
+      'Usage:\n' +
+        '  bun scripts/upload-run.ts results/config-name/2026-03-21-1730  (specific run)\n' +
+        '  bun scripts/upload-run.ts results/config-name                   (all un-uploaded runs)',
+    )
+    process.exit(1)
+  }
+
+  const dirStat = await stat(inputDir).catch(() => null)
+  if (!dirStat?.isDirectory()) {
+    console.error(`Not a directory: ${inputDir}`)
+    process.exit(1)
+  }
+
+  const r2Config = loadConfig()
+  const client = createClient(r2Config)
+
+  if (await isRunDir(inputDir)) {
+    // Single run: results/config-name/2026-03-21-1730
+    const timestamp = basename(inputDir)
+    const configName = basename(dirname(inputDir))
+    const runId = `${configName}-${timestamp}`
+    console.log(`Uploading run: ${runId}`)
+    await uploadSingleRun(inputDir, runId, r2Config, client)
+  } else {
+    // Config dir: results/config-name/ — upload all un-uploaded runs
+    const configName = basename(inputDir)
+    const entries = await readdir(inputDir, { withFileTypes: true })
+    const runDirs = entries
+      .filter((e) => e.isDirectory())
+      .map((e) => e.name)
+      .sort()
+
+    if (runDirs.length === 0) {
+      console.error('No run subdirectories found')
+      process.exit(1)
+    }
+
+    console.log(
+      `Found ${runDirs.length} runs for config "${configName}", checking R2...`,
+    )
+
+    let uploadedCount = 0
+    for (const dir of runDirs) {
+      const runId = `${configName}-${dir}`
+      const alreadyUploaded = await isUploaded(client, r2Config.bucket, runId)
+      if (alreadyUploaded) {
+        console.log(`  ${runId}: already uploaded, skipping`)
+        continue
+      }
+
+      console.log(`  ${runId}: uploading...`)
+      await uploadSingleRun(join(inputDir, dir), runId, r2Config, client)
+      uploadedCount++
+    }
+
+    console.log(
+      `\nDone. Uploaded ${uploadedCount} new run(s), ${runDirs.length - uploadedCount} already in R2.`,
+    )
+  }
+}
+
+main()

packages/browseros-agent/apps/eval/scripts/weekly-report.ts

@@ -24,11 +24,45 @@ import {
   PutObjectCommand,
   S3Client,
 } from '@aws-sdk/client-s3'
-import {
-  buildRunSummaries,
-  type ReportManifest,
-  type RunSummary,
-} from '../src/reporting/run-summary'
+
+interface ManifestTask {
+  queryId: string
+  query: string
+  status: string
+  durationMs: number
+  screenshotCount: number
+  graderResults: Record<string, { pass: boolean; score: number }>
+}
+
+interface Manifest {
+  runId: string
+  uploadedAt: string
+  agentConfig?: { type?: string; model?: string }
+  dataset?: string
+  summary?: { passRate?: number; avgDurationMs?: number }
+  tasks: ManifestTask[]
+}
+
+interface RunSummary {
+  runId: string
+  configName: string
+  date: string
+  avgScore: number
+  total: number
+  completed: number
+  failed: number
+  timeout: number
+  avgDurationMs: number
+  model: string
+  dataset: string
+  agentType: string
+}
+
+const PASS_FAIL_GRADER_ORDER = [
+  'agisdk_state_diff',
+  'infinity_state',
+  'performance_grader',
+]
 
 function requireEnv(name: string): string {
   const value = process.env[name]
@@ -53,7 +87,7 @@ const client = new S3Client({
 // Step 1: List all manifest.json files in runs/
 console.log('Scanning R2 for eval runs...')
 
-const manifests: ReportManifest[] = []
+const manifests: Manifest[] = []
 let continuationToken: string | undefined
 
 do {
@@ -93,9 +127,64 @@ if (manifests.length === 0) {
 }
 
 // Step 2: Build run summaries
-const runs: RunSummary[] = buildRunSummaries(manifests)
+const runs: RunSummary[] = manifests
+  .map((m) => {
+    const total = m.tasks.length
+    const completed = m.tasks.filter((t) => t.status === 'completed').length
+    const failed = m.tasks.filter((t) => t.status === 'failed').length
+    const timeout = m.tasks.filter((t) => t.status === 'timeout').length
+
+    let scoredCount = 0
+    let scoreSum = 0
+    for (const task of m.tasks) {
+      if (!task.graderResults) continue
+      for (const name of PASS_FAIL_GRADER_ORDER) {
+        if (task.graderResults[name]) {
+          scoredCount++
+          scoreSum += task.graderResults[name].score ?? 0
+          break
+        }
+      }
+    }
+
+    const avgScore = scoredCount > 0 ? (scoreSum / scoredCount) * 100 : 0
+    const durations = m.tasks
+      .filter((t) => t.durationMs > 0)
+      .map((t) => t.durationMs)
+    const avgDurationMs =
+      durations.length > 0
+        ? durations.reduce((a, b) => a + b, 0) / durations.length
+        : 0
+
+    const date = m.uploadedAt
+      ? `${m.uploadedAt.split('T')[0]} ${m.uploadedAt.split('T')[1]?.slice(0, 5) || ''}`
+      : m.runId.slice(0, 15)
+
+    const model = m.agentConfig?.model || 'unknown'
+    const dataset = m.dataset || m.runId
+    const agentType = m.agentConfig?.type || 'unknown'
+
+    const configName = extractConfigName(m.runId)
+    return {
+      runId: m.runId,
+      configName,
+      date,
+      avgScore,
+      total,
+      completed,
+      failed,
+      timeout,
+      avgDurationMs,
+      model,
+      dataset,
+      agentType,
+    }
+  })
+  .sort((a, b) => a.date.localeCompare(b.date))
 
 // Step 3: Identify unique config groups
+// runId can be "ci-weekly" (old) or "ci-weekly-2026-03-21-1730" (timestamped)
+// Extract config name by stripping the date-time suffix pattern
 function escHtml(s: string): string {
   return s
     .replace(/&/g, '&amp;')
@@ -104,6 +193,12 @@ function escHtml(s: string): string {
     .replace(/"/g, '&quot;')
 }
 
+function extractConfigName(runId: string): string {
+  // "browseros-agent-weekly-2026-03-21-1730" → "browseros-agent-weekly"
+  // "ci-weekly" → "ci-weekly" (no timestamp, old format)
+  return runId.replace(/-\d{4}-\d{2}-\d{2}-\d{4}$/, '')
+}
+
 const configGroups = [...new Set(runs.map((r) => r.configName))]
 const defaultConfig = configGroups.includes('ci-weekly')
   ? 'ci-weekly'

packages/browseros-agent/apps/eval/src/agents/orchestrated/backends/clado/clado-actions.ts

@@ -1,191 +0,0 @@
-import type {
-  CladoAction,
-  CladoActionResponse,
-  RawCladoActionPayload,
-} from './types'
-
-/** Parses Clado's structured response plus any raw `<answer>` blocks into executable actions. */
-export function parseCladoActions(
-  prediction: CladoActionResponse,
-): CladoAction[] {
-  const actionFromField =
-    typeof prediction.action === 'string' ? prediction.action : null
-
-  const rawActions = parseCladoActionsFromRawResponse(prediction.raw_response)
-  const primaryFromRaw = rawActions[0] ?? null
-  const mergedPrimary = {
-    ...primaryFromRaw,
-    ...prediction,
-    action: actionFromField ?? primaryFromRaw?.action,
-  }
-
-  const normalized: CladoAction[] = []
-  const primary = normalizeCladoActionPayload(mergedPrimary)
-  if (primary) normalized.push(primary)
-
-  for (const candidate of rawActions.slice(1)) {
-    const parsed = normalizeCladoActionPayload(candidate)
-    if (!parsed) continue
-    const prev = normalized[normalized.length - 1]
-    if (
-      !prev ||
-      getCladoActionSignature(prev) !== getCladoActionSignature(parsed)
-    ) {
-      normalized.push(parsed)
-    }
-  }
-
-  return normalized
-}
-
-export function normalizeCladoActionPayload(
-  payload: RawCladoActionPayload,
-): CladoAction | null {
-  if (!payload.action || typeof payload.action !== 'string') {
-    return null
-  }
-  return {
-    action: payload.action,
-    x: typeof payload.x === 'number' ? payload.x : undefined,
-    y: typeof payload.y === 'number' ? payload.y : undefined,
-    text: typeof payload.text === 'string' ? payload.text : undefined,
-    key: typeof payload.key === 'string' ? payload.key : undefined,
-    direction:
-      typeof payload.direction === 'string' ? payload.direction : undefined,
-    startX: typeof payload.startX === 'number' ? payload.startX : undefined,
-    startY: typeof payload.startY === 'number' ? payload.startY : undefined,
-    endX: typeof payload.endX === 'number' ? payload.endX : undefined,
-    endY: typeof payload.endY === 'number' ? payload.endY : undefined,
-    amount: typeof payload.amount === 'number' ? payload.amount : undefined,
-    time: typeof payload.time === 'number' ? payload.time : undefined,
-    final_answer:
-      typeof payload.final_answer === 'string'
-        ? payload.final_answer
-        : undefined,
-  }
-}
-
-export function parseCladoActionsFromRawResponse(
-  rawResponse: string | undefined,
-): RawCladoActionPayload[] {
-  if (!rawResponse) return []
-  const matches = [
-    ...rawResponse.matchAll(/<answer>\s*([\s\S]*?)\s*<\/answer>/gi),
-  ]
-  const parsed: RawCladoActionPayload[] = []
-  for (const match of matches) {
-    try {
-      parsed.push(JSON.parse(match[1]) as RawCladoActionPayload)
-    } catch {
-      // Ignore malformed answer blocks so one bad block does not drop the whole prediction.
-    }
-  }
-  return parsed
-}
-
-export function extractCladoThinking(
-  rawResponse: string | undefined,
-): string | undefined {
-  if (!rawResponse) return undefined
-  const matches = [
-    ...rawResponse.matchAll(/<thinking>\s*([\s\S]*?)\s*<\/thinking>/gi),
-  ]
-  if (matches.length === 0) return undefined
-
-  const merged = matches
-    .map((match) => match[1]?.replace(/\s+/g, ' ').trim() ?? '')
-    .filter((value) => value.length > 0)
-    .join(' ')
-
-  if (!merged) return undefined
-  return merged
-}
-
-export function summarizeCladoPrediction(
-  prediction: CladoActionResponse,
-): Record<string, unknown> {
-  const preview =
-    typeof prediction.raw_response === 'string' &&
-    prediction.raw_response.length > 0
-      ? prediction.raw_response.slice(0, 240)
-      : undefined
-
-  return {
-    action: prediction.action,
-    x: prediction.x,
-    y: prediction.y,
-    text: prediction.text,
-    key: prediction.key,
-    direction: prediction.direction,
-    startX: prediction.startX,
-    startY: prediction.startY,
-    endX: prediction.endX,
-    endY: prediction.endY,
-    amount: prediction.amount,
-    time: prediction.time,
-    inference_time_seconds: prediction.inference_time_seconds,
-    raw_response_preview: preview,
-  }
-}
-
-export function getCladoActionSignature(action: CladoAction): string {
-  switch (action.action) {
-    case 'click':
-    case 'double_click':
-    case 'right_click':
-    case 'hover':
-      return `${action.action}:${action.x ?? 'x'}:${action.y ?? 'y'}`
-    case 'type':
-      return `${action.action}:${(action.text ?? '').slice(0, 16)}`
-    case 'press_key':
-      return `${action.action}:${action.key ?? 'key'}`
-    case 'scroll':
-      return `${action.action}:${action.direction ?? 'down'}:${action.amount ?? 500}`
-    case 'drag':
-      return `${action.action}:${action.startX}:${action.startY}:${action.endX}:${action.endY}`
-    case 'wait':
-      return `${action.action}:${action.time ?? 1}`
-    case 'end':
-      return action.final_answer
-        ? `end(${action.final_answer.slice(0, 32)})`
-        : 'end()'
-    case 'invalid':
-      return `invalid(${(action.text ?? '').slice(0, 40)})`
-    default:
-      return action.action
-  }
-}
-
-export function formatCladoHistory(actions: CladoAction[]): string {
-  if (actions.length === 0) return 'None'
-
-  const parts = actions.map((action) => {
-    switch (action.action) {
-      case 'click':
-      case 'double_click':
-      case 'right_click':
-      case 'hover':
-        return `${action.action}(${Math.round(action.x ?? 500)}, ${Math.round(action.y ?? 500)})`
-      case 'type': {
-        const text = (action.text ?? '').replace(/'/g, "\\'")
-        return `type('${text}')`
-      }
-      case 'press_key':
-        return `press_key('${action.key ?? 'Enter'}')`
-      case 'scroll':
-        return `scroll(${action.direction ?? 'down'})`
-      case 'drag':
-        return `drag(${Math.round(action.startX ?? 500)},${Math.round(action.startY ?? 500)} -> ${Math.round(action.endX ?? 500)},${Math.round(action.endY ?? 500)})`
-      case 'wait':
-        return `wait(${Math.round(action.time ?? 1)}s)`
-      case 'end':
-        return 'end()'
-      case 'invalid':
-        return 'invalid()'
-      default:
-        return action.action
-    }
-  })
-
-  return parts.join(' -> ')
-}

packages/browseros-agent/apps/eval/src/agents/orchestrated/backends/clado/clado-browser-driver.ts

@@ -1,123 +0,0 @@
-import {
-  CLADO_PAGE_SCOPED_TOOLS,
-  type CladoActionPoint,
-  type CladoViewport,
-} from './types'
-
-export function clampCladoNormalizedCoordinate(value: number): number {
-  return Math.min(999, Math.max(0, Math.round(value)))
-}
-
-/** Converts Clado's 0-1000 normalized coordinate space into BrowserOS viewport pixels. */
-export function resolveCladoPoint(
-  viewport: CladoViewport,
-  normalizedX: number | undefined,
-  normalizedY: number | undefined,
-): CladoActionPoint {
-  const nx = clampCladoNormalizedCoordinate(normalizedX ?? 500)
-  const ny = clampCladoNormalizedCoordinate(normalizedY ?? 500)
-
-  return {
-    x: Math.round((nx / 1000) * viewport.width),
-    y: Math.round((ny / 1000) * viewport.height),
-  }
-}
-
-/** Adapts Clado action tool arguments to the BrowserOS MCP tool argument contract. */
-export function prepareCladoToolArgs(
-  toolName: string,
-  args: Record<string, unknown>,
-  pageId: number,
-): Record<string, unknown> {
-  const prepared: Record<string, unknown> = { ...args }
-
-  if (
-    toolName === 'evaluate_script' &&
-    typeof prepared.function === 'string' &&
-    prepared.expression === undefined
-  ) {
-    prepared.expression = toCladoEvaluateExpression(prepared.function)
-    delete prepared.function
-  }
-
-  if (
-    toolName === 'click_at' &&
-    typeof prepared.dblClick === 'boolean' &&
-    prepared.clickCount === undefined
-  ) {
-    prepared.clickCount = prepared.dblClick ? 2 : 1
-    delete prepared.dblClick
-  }
-
-  if (
-    CLADO_PAGE_SCOPED_TOOLS.has(toolName) &&
-    typeof prepared.page !== 'number'
-  ) {
-    prepared.page = pageId
-  }
-
-  return prepared
-}
-
-export function toCladoEvaluateExpression(rawFunction: unknown): string {
-  const source = String(rawFunction).trim()
-  if (source.startsWith('() =>') || source.startsWith('async () =>')) {
-    return `(${source})()`
-  }
-  if (source.startsWith('function')) {
-    return `(${source})()`
-  }
-  return source
-}
-
-export function normalizeCladoPressKey(key: string | undefined): string {
-  const raw = (key ?? '').trim()
-  if (!raw) throw new Error('press_key action missing key field')
-
-  const map: Record<string, string> = {
-    'C-a': 'Control+A',
-    'C-c': 'Control+C',
-    'C-v': 'Control+V',
-    'C-x': 'Control+X',
-    'C-z': 'Control+Z',
-    'C-y': 'Control+Y',
-    'C-s': 'Control+S',
-    'C-t': 'Control+T',
-    'C-w': 'Control+W',
-    'C-h': 'Control+H',
-    'C-f': 'Control+F',
-    'C-+': 'Control++',
-    'C--': 'Control+-',
-    'C-tab': 'Control+Tab',
-    'C-S-tab': 'Control+Shift+Tab',
-    'C-S-n': 'Control+Shift+N',
-    'C-down': 'Control+ArrowDown',
-    'M-a': 'Meta+A',
-    'M-c': 'Meta+C',
-    'M-v': 'Meta+V',
-    'M-x': 'Meta+X',
-    'M-f4': 'Alt+F4',
-  }
-  return map[raw] ?? raw
-}
-
-export function normalizeCladoDirection(
-  direction: string | undefined,
-): 'up' | 'down' | 'left' | 'right' {
-  if (
-    direction === 'up' ||
-    direction === 'down' ||
-    direction === 'left' ||
-    direction === 'right'
-  ) {
-    return direction
-  }
-  return 'down'
-}
-
-export function normalizeCladoScrollAmount(amount: number | undefined): number {
-  if (typeof amount !== 'number') return 500
-  if (amount <= 0) return 100
-  const clamped = Math.min(amount, 1000)
-  return Math.max(100, Math.round((clamped / 1000) * 900))
-}

packages/browseros-agent/apps/eval/src/agents/orchestrated/backends/clado/clado-client.ts

@@ -1,68 +0,0 @@
-import { CLADO_REQUEST_TIMEOUT_MS } from '../../../../constants'
-import { formatCladoHistory } from './clado-actions'
-import type { CladoAction, CladoActionResponse } from './types'
-
-export interface CladoActionClientOptions {
-  baseUrl?: string
-  apiKey?: string
-}
-
-export interface CladoActionPredictionInput {
-  instruction: string
-  imageBase64: string
-  actionHistory: CladoAction[]
-  signal?: AbortSignal
-}
-
-/** Calls the Clado action model without exposing credentials in process arguments or artifacts. */
-export class CladoActionClient {
-  constructor(private readonly options: CladoActionClientOptions) {}
-
-  async requestActionPrediction(
-    input: CladoActionPredictionInput,
-  ): Promise<CladoActionResponse> {
-    if (!this.options.baseUrl) {
-      throw new Error('executor.baseUrl must be set for clado-action provider')
-    }
-
-    const requestController = new AbortController()
-    const onAbort = () => requestController.abort()
-    input.signal?.addEventListener('abort', onAbort, { once: true })
-
-    const timeoutHandle = setTimeout(() => {
-      requestController.abort()
-    }, CLADO_REQUEST_TIMEOUT_MS)
-
-    try {
-      const headers: Record<string, string> = {
-        'Content-Type': 'application/json',
-      }
-      if (this.options.apiKey) {
-        headers.Authorization = `Bearer ${this.options.apiKey}`
-      }
-
-      const response = await fetch(this.options.baseUrl, {
-        method: 'POST',
-        headers,
-        body: JSON.stringify({
-          instruction: input.instruction,
-          image_base64: input.imageBase64,
-          history: formatCladoHistory(input.actionHistory),
-        }),
-        signal: requestController.signal,
-      })
-
-      if (!response.ok) {
-        const body = await response.text()
-        throw new Error(
-          `HTTP ${response.status} ${response.statusText}: ${body.slice(0, 400)}`,
-        )
-      }
-
-      return (await response.json()) as CladoActionResponse
-    } finally {
-      clearTimeout(timeoutHandle)
-      input.signal?.removeEventListener('abort', onAbort)
-    }
-  }
-}