chore(eval): drop the 60-char truncation on grader expected/actual values

Some criteria check long strings (job descriptions, post bodies, etc.) —
truncating to 60 chars hides exactly the bytes you need to diff. The
viewer's reasoning area already has max-height + scroll + word-break so
long content scrolls; nothing renders worse for being full-length.

.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
@@ -44,19 +44,6 @@ jobs:
         working-directory: packages/browseros-agent
         run: bun install --ignore-scripts
 
-      - name: Install Claude Code CLI
-        working-directory: packages/browseros-agent/apps/eval
-        env:
-          EVAL_CONFIG: ${{ github.event.inputs.config || 'configs/legacy/browseros-agent-weekly.json' }}
-        run: |
-          if bun -e "const config = await Bun.file(process.env.EVAL_CONFIG).json(); process.exit(config.agent?.type === 'claude-code' ? 0 : 1)"; then
-            npm install -g @anthropic-ai/claude-code@2.1.119
-            echo "Claude Code CLI installed at $(command -v claude)"
-            claude --version
-          else
-            echo "Eval config does not use Claude Code; skipping Claude Code CLI install"
-          fi
-
       - 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.
@@ -75,14 +62,11 @@ 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 }}
-          AWS_REGION: ${{ secrets.AWS_REGION || 'us-west-2' }}
-          AWS_ACCESS_KEY_ID: ${{ secrets.AWS_ACCESS_KEY_ID }}
-          AWS_SECRET_ACCESS_KEY: ${{ secrets.AWS_SECRET_ACCESS_KEY }}
           CLAUDE_CODE_OAUTH_TOKEN: ${{ secrets.CLAUDE_CODE_OAUTH_TOKEN }}
           NOPECHA_API_KEY: ${{ secrets.NOPECHA_API_KEY }}
           BROWSEROS_BINARY: /usr/bin/browseros
@@ -90,29 +74,12 @@ jobs:
           # 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"
-          # Capture the run directory so report.html can be generated before the R2 publish step.
-          SUMMARY_PATH="$(find results -name summary.json -type f -print | sort | tail -n 1)"
-          if [ -z "$SUMMARY_PATH" ]; then
-            echo "No eval run summary found"
-            exit 1
-          fi
-          RUN_DIR="$(dirname "$SUMMARY_PATH")"
-          echo "EVAL_RUN_DIR=$RUN_DIR" >> "$GITHUB_ENV"
+          xvfb-run --auto-servernum --server-args="-screen 0 1440x900x24" bun run src/index.ts -c "$EVAL_CONFIG"
 
-      - name: Generate run analysis report
-        if: success()
-        working-directory: packages/browseros-agent/apps/eval
-        env:
-          CLAUDE_CODE_OAUTH_TOKEN: ${{ secrets.CLAUDE_CODE_OAUTH_TOKEN }}
-        run: |
-          echo "Generating run report for $EVAL_RUN_DIR"
-          bun scripts/generate-report.ts --input "$EVAL_RUN_DIR" --output "$EVAL_RUN_DIR/report.html"
-
-      - name: Publish eval run to R2
+      - name: Upload runs to R2
         if: success()
         working-directory: packages/browseros-agent/apps/eval
         env:
@@ -121,7 +88,10 @@ jobs:
           EVAL_R2_SECRET_ACCESS_KEY: ${{ secrets.EVAL_R2_SECRET_ACCESS_KEY }}
           EVAL_R2_BUCKET: ${{ secrets.EVAL_R2_BUCKET }}
           EVAL_R2_CDN_BASE_URL: ${{ secrets.EVAL_R2_CDN_BASE_URL }}
-        run: bun run src/index.ts publish --run "$EVAL_RUN_DIR" --target r2
+          EVAL_CONFIG: ${{ github.event.inputs.config || 'configs/browseros-agent-weekly.json' }}
+        run: |
+          CONFIG_NAME=$(basename "$EVAL_CONFIG" .json)
+          bun scripts/upload-run.ts "results/$CONFIG_NAME"
 
       - name: Generate trend report
         if: success()
@@ -136,7 +106,7 @@ jobs:
           EVAL_R2_CDN_BASE_URL: ${{ secrets.EVAL_R2_CDN_BASE_URL }}
         run: bun apps/eval/scripts/weekly-report.ts /tmp/eval-report.html
 
-      - name: Upload trend report as artifact
+      - name: Upload report as artifact
         if: success()
         uses: actions/upload-artifact@v4
         with:

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

@@ -1,62 +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
-    permissions:
-      contents: write
-      pull-requests: write
-    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: Open auto-merge PR if internal-docs has new commits
-        env:
-          GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
-        run: |
-          set -e
-
-          # Skip if submodule not yet configured (handoff window before someone adds it)
-          if ! git config --file .gitmodules --get-regexp '^submodule\..internal-docs\.path$' >/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
-
-          BRANCH="bot/sync-internal-docs-$(date -u +%Y%m%d-%H%M%S)"
-          git config user.name  "browseros-bot"
-          git config user.email "bot@browseros.ai"
-          git checkout -b "$BRANCH"
-          git add .internal-docs
-          git commit -m "chore: sync internal-docs submodule"
-          git push -u origin "$BRANCH"
-
-          PR_URL=$(gh pr create \
-            --base dev \
-            --head "$BRANCH" \
-            --title "chore: sync internal-docs submodule" \
-            --body "Automated bump of the \`.internal-docs\` submodule pointer. Auto-merging.")
-
-          gh pr merge "$PR_URL" --auto --squash --delete-branch

.github/workflows/test.yml

@@ -63,15 +63,15 @@ jobs:
             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: 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
 

.gitmodules

@@ -1,4 +0,0 @@
-[submodule ".internal-docs"]
-	path = .internal-docs
-	url = git@github.com:browseros-ai/internal-docs.git
-	branch = main

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

@@ -79,15 +79,14 @@ 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 +156,9 @@ 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

@@ -17,7 +17,7 @@ export function groupProviderOptions(
       ? [{ key: 'llm' as const, label: 'AI Providers', options: llm }]
       : []),
     ...(acp.length
-      ? [{ key: 'acp' as const, label: 'Agents', options: acp }]
+      ? [{ key: 'acp' as const, label: 'ACP Models', options: acp }]
       : []),
   ]
 }
@@ -26,25 +26,14 @@ export function getProviderSearchValue(
   provider: Provider,
   groupLabel: string,
 ): string {
-  return [
-    provider.id,
-    provider.name,
-    provider.type,
-    groupLabel,
-    provider.adapterName,
-    provider.modelLabel,
-  ]
+  return [provider.id, provider.name, provider.type, groupLabel]
     .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(' · ')
+  return provider.modelControl === 'best-effort'
+    ? 'ACP model · best effort'
+    : 'ACP model'
 }

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

@@ -16,26 +16,22 @@ const options: Provider[] = [
   },
   {
     kind: 'acp',
-    id: 'agent-claude-review',
-    name: 'Review Bot',
+    id: 'acp:claude:haiku:medium',
+    name: 'Claude Code Haiku',
     type: 'acp',
-    adapterName: 'Claude Code',
-    modelLabel: 'Haiku',
     modelControl: 'best-effort',
   },
   {
     kind: 'acp',
-    id: 'agent-codex-browser',
-    name: 'Browser Driver',
+    id: 'acp:codex:gpt-5.5:medium',
+    name: 'Codex GPT-5.5',
     type: 'acp',
-    adapterName: 'Codex',
-    modelLabel: 'GPT-5.5',
     modelControl: 'runtime-supported',
   },
 ]
 
 describe('groupProviderOptions', () => {
-  it('groups normal providers separately from created agents', () => {
+  it('groups normal providers separately from ACP models', () => {
     expect(groupProviderOptions(options)).toEqual([
       {
         key: 'llm',
@@ -44,7 +40,7 @@ describe('groupProviderOptions', () => {
       },
       {
         key: 'acp',
-        label: 'Agents',
+        label: 'ACP Models',
         options: [options[2], options[3]],
       },
     ])
@@ -52,21 +48,20 @@ describe('groupProviderOptions', () => {
 })
 
 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',
+  it('matches ACP group labels and item labels', () => {
+    expect(getProviderSearchValue(options[2], 'ACP Models')).toContain(
+      'ACP Models',
+    )
+    expect(getProviderSearchValue(options[2], 'ACP Models')).toContain(
+      'Claude Code Haiku',
     )
   })
 })
 
 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')
+  it('does not present best-effort ACP models as guaranteed routing', () => {
+    expect(getProviderSubtitle(options[2])).toBe('ACP model · best effort')
+    expect(getProviderSubtitle(options[3])).toBe('ACP model')
     expect(getProviderSubtitle(options[0])).toBeUndefined()
   })
 })

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

@@ -41,10 +41,7 @@ export const ChatProviderSelector: FC<
       <PopoverTrigger asChild>{children}</PopoverTrigger>
       <PopoverContent side="bottom" align="start" className="w-64 p-0">
         <Command>
-          <CommandInput
-            placeholder="Search providers or agents..."
-            className="h-9"
-          />
+          <CommandInput placeholder="Search models..." className="h-9" />
           <CommandList>
             <CommandEmpty>No provider found</CommandEmpty>
             {groups.map((group) => (

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

@@ -7,8 +7,5 @@ export interface Provider {
   name: string
   type: ChatProviderType
   kind: 'llm' | 'acp'
-  agentId?: string
-  adapterName?: string
-  modelLabel?: string
   modelControl?: 'runtime-supported' | 'best-effort'
 }

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

@@ -1,36 +1,179 @@
-import { ArrowLeft } from 'lucide-react'
+import { ArrowLeft, Bot, Home } from 'lucide-react'
 import { type FC, useEffect, useMemo, useRef } from 'react'
 import { Navigate, useNavigate, useParams, useSearchParams } from 'react-router'
 import { Button } from '@/components/ui/button'
-import type {
-  HarnessAgent,
-  HarnessAgentAdapter,
-} from '@/entrypoints/app/agents/agent-harness-types'
-import type { AgentAdapterHealth } from '@/entrypoints/app/agents/agent-row/agent-row.types'
 import {
-  cancelHarnessTurn,
-  useAgentAdapters,
-  useEnqueueHarnessMessage,
-  useHarnessAgents,
-  useRemoveHarnessQueuedMessage,
-  useUpdateHarnessAgent,
-} from '@/entrypoints/app/agents/useAgents'
-import type { AgentEntry } from '@/entrypoints/app/agents/useOpenClaw'
-import { AgentRail } from './AgentRail'
+  type AgentEntry,
+  getModelDisplayName,
+} from '@/entrypoints/app/agents/useOpenClaw'
+import { cn } from '@/lib/utils'
 import { useAgentCommandData } from './agent-command-layout'
 import { ClawChat } from './ClawChat'
-import { ConversationHeader } from './ConversationHeader'
 import { ConversationInput } from './ConversationInput'
 import {
   buildChatHistoryFromClawMessages,
   filterTurnsPersistedInHistory,
   flattenHistoryPages,
 } from './claw-chat-types'
-import { consumePendingInitialMessage } from './pending-initial-message'
-import { QueuePanel } from './QueuePanel'
 import { useAgentConversation } from './useAgentConversation'
 import { useHarnessChatHistory } from './useHarnessChatHistory'
 
+function StatusBadge({ status }: { status: string }) {
+  return (
+    <div className="inline-flex items-center gap-2 rounded-full border border-border/60 bg-card px-3 py-1 text-[11px] text-muted-foreground uppercase tracking-[0.18em]">
+      <span
+        className={cn(
+          'size-1.5 rounded-full',
+          status === 'Working on your request'
+            ? 'bg-amber-500'
+            : status === 'Ready'
+              ? 'bg-emerald-500'
+              : status === 'Offline'
+                ? 'bg-muted-foreground/50'
+                : 'bg-[var(--accent-orange)]',
+        )}
+      />
+      <span>{status}</span>
+    </div>
+  )
+}
+
+function AgentIdentity({
+  name,
+  meta,
+  className,
+}: {
+  name: string
+  meta: string
+  className?: string
+}) {
+  return (
+    <div className={cn('min-w-0', className)}>
+      <div className="truncate font-semibold text-[15px] leading-5">{name}</div>
+      <div className="truncate text-muted-foreground text-xs leading-5">
+        {meta}
+      </div>
+    </div>
+  )
+}
+
+function ConversationHeader({
+  agentName,
+  agentMeta,
+  status,
+  backLabel,
+  backTarget,
+  onGoHome,
+}: {
+  agentName: string
+  agentMeta: string
+  status: string
+  backLabel: string
+  backTarget: 'home' | 'page'
+  onGoHome: () => void
+}) {
+  const BackIcon = backTarget === 'home' ? Home : ArrowLeft
+
+  return (
+    <div className="flex h-14 items-center justify-between gap-4 border-border/50 border-b px-5">
+      <div className="flex min-w-0 items-center gap-3">
+        <Button
+          variant="ghost"
+          size="icon"
+          onClick={onGoHome}
+          className="size-8 rounded-xl lg:hidden"
+          title={backLabel}
+        >
+          <BackIcon className="size-4" />
+        </Button>
+        <div className="flex size-8 shrink-0 items-center justify-center rounded-xl bg-muted text-muted-foreground">
+          <Bot className="size-4" />
+        </div>
+        <AgentIdentity name={agentName} meta={agentMeta} />
+      </div>
+
+      <StatusBadge status={status} />
+    </div>
+  )
+}
+
+function AgentRailHeader({ onGoHome }: { onGoHome: () => void }) {
+  return (
+    <div className="hidden h-14 items-center border-border/50 border-r border-b bg-background/70 px-4 lg:flex">
+      <div className="flex min-w-0 items-center gap-3">
+        <Button
+          variant="ghost"
+          size="icon"
+          onClick={onGoHome}
+          className="size-8 rounded-xl"
+          title="Back to home"
+        >
+          <ArrowLeft className="size-4" />
+        </Button>
+        <div className="truncate font-semibold text-[15px] leading-5">
+          Agents
+        </div>
+      </div>
+    </div>
+  )
+}
+
+function AgentRailList({
+  activeAgentId,
+  agents,
+  onSelectAgent,
+}: {
+  activeAgentId: string
+  agents: AgentEntry[]
+  onSelectAgent: (entry: AgentEntry) => void
+}) {
+  return (
+    <aside className="hidden min-h-0 flex-col border-border/50 border-r bg-background/70 lg:flex">
+      <div className="styled-scrollbar min-h-0 flex-1 space-y-2 overflow-y-auto px-3 py-3">
+        {agents.map((entry) => {
+          const active = entry.agentId === activeAgentId
+          const modelName = getAgentEntryMeta(entry)
+
+          return (
+            <button
+              key={entry.agentId}
+              type="button"
+              onClick={() => onSelectAgent(entry)}
+              className={cn(
+                'w-full rounded-2xl border px-3 py-3 text-left transition-all',
+                active
+                  ? 'border-[var(--accent-orange)]/30 bg-[var(--accent-orange)]/8 shadow-sm'
+                  : 'border-transparent bg-transparent hover:border-border/60 hover:bg-card',
+              )}
+            >
+              <div className="flex items-center gap-3">
+                <div
+                  className={cn(
+                    'flex size-9 items-center justify-center rounded-xl',
+                    active
+                      ? 'bg-[var(--accent-orange)]/12 text-[var(--accent-orange)]'
+                      : 'bg-muted text-muted-foreground',
+                  )}
+                >
+                  <Bot className="size-4" />
+                </div>
+                <AgentIdentity name={entry.name} meta={modelName} />
+              </div>
+            </button>
+          )
+        })}
+      </div>
+    </aside>
+  )
+}
+
+function getAgentEntryMeta(agent: AgentEntry | undefined): string {
+  if (agent?.source === 'agent-harness') {
+    return getModelDisplayName(agent.model) ?? 'ACP agent'
+  }
+  return getModelDisplayName(agent?.model) ?? 'OpenClaw agent'
+}
+
 function AgentConversationController({
   agentId,
   initialMessage,
@@ -69,33 +212,15 @@ function AgentConversationController({
     [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 { turns, streaming, send } = useAgentConversation(agentId, {
     runtime: 'agent-harness',
     sessionKey: null,
     history: chatHistory,
-    activeTurnId,
     onComplete: () => {
       void harnessHistoryQuery.refetch()
     },
     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],
@@ -114,59 +239,32 @@ function AgentConversationController({
   sendRef.current = send
 
   useEffect(() => {
-    if (disabled || !historyReady) return
-
-    // Registry-first: when the user submitted at /home with
-    // attachments, the rich payload is here. URL `?q=` may also be
-    // present and is the text-only fallback path; the registry wins
-    // when both exist because it carries the binary attachments
-    // alongside the text.
-    const pending = consumePendingInitialMessage(agentId)
-    if (pending) {
-      // Mark the dedup ref so the text-only branch below doesn't
-      // re-fire on the same render.
-      if (initialMessageKey) {
-        initialMessageSentRef.current = initialMessageKey
-      }
-      onInitialMessageConsumedRef.current()
-      void sendRef.current({
-        text: pending.text,
-        attachments: pending.attachments.map((a) => a.payload),
-        attachmentPreviews: pending.attachments.map((a) => ({
-          id: a.id,
-          kind: a.kind,
-          mediaType: a.mediaType,
-          name: a.name,
-          dataUrl: a.dataUrl,
-        })),
-      })
-      return
-    }
-
     const query = initialMessage?.trim()
     if (!initialMessageKey) {
-      // Reset is safe even on the post-registry-fire re-run: consume
-      // is destructive, so the registry is already drained — there's
-      // nothing left for a third run to re-send.
       initialMessageSentRef.current = null
       return
     }
 
-    if (!query || initialMessageSentRef.current === initialMessageKey) {
+    if (
+      !query ||
+      initialMessageSentRef.current === initialMessageKey ||
+      disabled ||
+      !historyReady
+    ) {
       return
     }
 
     initialMessageSentRef.current = initialMessageKey
     onInitialMessageConsumedRef.current()
     void sendRef.current({ text: query })
-  }, [agentId, disabled, historyReady, initialMessage, initialMessageKey])
+  }, [disabled, historyReady, initialMessage, initialMessageKey])
 
   const handleSelectAgent = (entry: AgentEntry) => {
     navigate(`${agentPathPrefix}/${entry.agentId}`)
   }
 
   return (
-    <div className="flex min-h-0 flex-1 flex-col overflow-hidden">
+    <div className="flex min-h-0 flex-col overflow-hidden">
       <ClawChat
         agentName={agentName}
         historyMessages={historyMessages}
@@ -183,15 +281,7 @@ function AgentConversationController({
       />
 
       <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}
@@ -206,31 +296,14 @@ 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,
-                  attachments,
-                })
-                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}...`
-            }
+            placeholder={`Message ${agentName}...`}
           />
         </div>
       </div>
@@ -245,22 +318,6 @@ interface AgentCommandConversationProps {
   createAgentPath?: string
 }
 
-function inferAdapterFromEntry(
-  entry: AgentEntry | undefined,
-): HarnessAgentAdapter | 'unknown' {
-  if (!entry) return 'unknown'
-  if (entry.source === 'agent-harness') {
-    // Harness entries don't carry the adapter on AgentEntry; the rail
-    // / header read the harness record directly. This branch only runs
-    // before the harness query resolves, so 'unknown' is correct — the
-    // tile's bot fallback renders until data arrives.
-    return 'unknown'
-  }
-  // OpenClaw-only entries (no harness shadow) are deprecated in
-  // practice but the rail still tolerates them.
-  return 'openclaw'
-}
-
 export const AgentCommandConversation: FC<AgentCommandConversationProps> = ({
   variant = 'command',
   backPath = '/home',
@@ -271,110 +328,60 @@ export const AgentCommandConversation: FC<AgentCommandConversationProps> = ({
   const [searchParams, setSearchParams] = useSearchParams()
   const navigate = useNavigate()
   const { agents } = useAgentCommandData()
-  const { harnessAgents } = useHarnessAgents()
-  const { adapters } = useAgentAdapters()
-  const updateAgent = useUpdateHarnessAgent()
-
   const shouldRedirectHome = !agentId
   const resolvedAgentId = agentId ?? ''
-  const harnessAgent = harnessAgents.find(
-    (entry) => entry.id === resolvedAgentId,
-  )
-  const entry = agents.find((item) => item.agentId === resolvedAgentId)
-  const fallbackName = entry?.name || resolvedAgentId || 'Agent'
-  const fallbackAdapter = inferAdapterFromEntry(entry)
+  const agent = agents.find((entry) => entry.agentId === resolvedAgentId)
+  const agentName = agent?.name || resolvedAgentId || 'Agent'
+  const agentMeta = getAgentEntryMeta(agent)
   const initialMessage = searchParams.get('q')
   const isPageVariant = variant === 'page'
   const backLabel = isPageVariant ? 'Back to agents' : 'Back to home'
 
-  const adapterHealth = useMemo<AgentAdapterHealth | null>(() => {
-    const adapterId = harnessAgent?.adapter
-    if (!adapterId) return null
-    const descriptor = adapters.find((item) => item.id === adapterId)
-    if (!descriptor?.health) return null
-    return {
-      healthy: descriptor.health.healthy,
-      reason: descriptor.health.reason,
-    }
-  }, [adapters, harnessAgent?.adapter])
-
   if (shouldRedirectHome) {
     return <Navigate to="/home" replace />
   }
 
-  const handleSelectHarnessAgent = (target: HarnessAgent) => {
-    navigate(`${agentPathPrefix}/${target.id}`)
+  const handleSelectAgent = (entry: AgentEntry) => {
+    navigate(`${agentPathPrefix}/${entry.agentId}`)
   }
 
-  const handlePinToggle = (target: HarnessAgent | null, next: boolean) => {
-    if (!target) return
-    updateAgent.mutate({
-      agentId: target.id,
-      patch: { pinned: next },
-    })
-  }
+  // 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'
 
   return (
     <div className="absolute inset-0 overflow-hidden bg-background md:pl-[theme(spacing.14)]">
-      <div className="mx-auto flex h-full w-full max-w-[1480px] flex-col">
-        {/* Shared top band — the rail's "Agents" header and the chat
-            header live on one row so they're aligned by construction. */}
-        <div className="flex shrink-0 items-stretch border-border/50 border-b">
-          <div className="hidden min-h-[60px] w-[288px] shrink-0 items-center gap-3 border-border/50 border-r px-4 lg:flex">
-            <Button
-              variant="ghost"
-              size="icon"
-              onClick={() => navigate(backPath)}
-              className="size-8 rounded-xl"
-              title="Back to home"
-            >
-              <ArrowLeft className="size-4" />
-            </Button>
-            <div className="truncate font-semibold text-[15px] leading-5">
-              Agents
-            </div>
-          </div>
-          <div className="min-w-0 flex-1">
-            <ConversationHeader
-              agent={harnessAgent ?? null}
-              fallbackName={fallbackName}
-              fallbackAdapter={fallbackAdapter}
-              adapterHealth={adapterHealth}
-              backLabel={backLabel}
-              backTarget={isPageVariant ? 'page' : 'home'}
-              onGoHome={() => navigate(backPath)}
-              onPinToggle={(next) =>
-                handlePinToggle(harnessAgent ?? null, next)
-              }
-            />
-          </div>
-        </div>
+      <div className="mx-auto grid h-full w-full max-w-[1480px] lg:grid-cols-[288px_minmax(0,1fr)] lg:grid-rows-[3.5rem_minmax(0,1fr)]">
+        <AgentRailHeader onGoHome={() => navigate(backPath)} />
 
-        {/* Body grid: rail list + chat. Both columns share the same
-            top edge (the band above) so headers can never drift. */}
-        <div className="grid min-h-0 flex-1 grid-rows-[minmax(0,1fr)] lg:grid-cols-[288px_minmax(0,1fr)]">
-          <AgentRail
-            agents={harnessAgents}
-            adapters={adapters}
-            activeAgentId={resolvedAgentId}
-            onSelectAgent={handleSelectHarnessAgent}
-            onPinToggle={(target, next) => handlePinToggle(target, next)}
-          />
+        <ConversationHeader
+          agentName={agentName}
+          agentMeta={agentMeta}
+          status={statusCopy}
+          backLabel={backLabel}
+          backTarget={isPageVariant ? 'page' : 'home'}
+          onGoHome={() => navigate(backPath)}
+        />
 
-          <div className="flex h-full min-h-0 flex-col overflow-hidden">
-            <AgentConversationController
-              key={resolvedAgentId}
-              agentId={resolvedAgentId}
-              agents={agents}
-              initialMessage={initialMessage}
-              onInitialMessageConsumed={() =>
-                setSearchParams({}, { replace: true })
-              }
-              agentPathPrefix={agentPathPrefix}
-              createAgentPath={createAgentPath}
-            />
-          </div>
-        </div>
+        <AgentRailList
+          activeAgentId={resolvedAgentId}
+          agents={agents}
+          onSelectAgent={handleSelectAgent}
+        />
+
+        <AgentConversationController
+          key={resolvedAgentId}
+          agentId={resolvedAgentId}
+          agents={agents}
+          initialMessage={initialMessage}
+          onInitialMessageConsumed={() =>
+            setSearchParams({}, { replace: true })
+          }
+          agentPathPrefix={agentPathPrefix}
+          createAgentPath={createAgentPath}
+        />
       </div>
     </div>
   )

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

@@ -1,29 +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,
-  type ConversationInputSendInput,
-} from './ConversationInput'
-import { orderHomeAgents } from './home-agent-card.helpers'
-import { setPendingInitialMessage } from './pending-initial-message'
+import { ConversationInput } from './ConversationInput'
+import { buildAgentCardData } from './useAgentCardData'
 
 function EmptyAgentsState({ onOpenAgents }: { onOpenAgents: () => void }) {
   return (
@@ -49,13 +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
 }) {
@@ -81,7 +68,6 @@ function RecentThreads({
       </div>
       <AgentCardDock
         agents={agents}
-        adapters={adapters}
         activeAgentId={activeAgentId ?? undefined}
         onSelectAgent={onSelectAgent}
         onCreateAgent={onOpenAgents}
@@ -93,46 +79,28 @@ 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: ConversationInputSendInput) => {
+  const handleSend = (input: { text: string }) => {
     if (!selectedAgentId) return
-    // Stash text + attachments in the in-memory registry. Text also
-    // travels in `?q=` so a hard refresh / shareable URL still works
-    // for text-only prompts; attachments are registry-only because a
-    // multi-megabyte dataUrl can't ride a URL search param. The chat
-    // screen prefers the registry when both are present.
-    setPendingInitialMessage({
-      agentId: selectedAgentId,
-      text: input.text,
-      attachments: input.attachments,
-      createdAt: Date.now(),
-    })
     navigate(
       `/home/agents/${selectedAgentId}?q=${encodeURIComponent(input.text)}`,
     )
@@ -142,7 +110,7 @@ export const AgentCommandHome: FC = () => {
     setSelectedAgentId(agent.agentId)
   }
 
-  const selectedAgent = legacyAgents.find(
+  const selectedAgent = agents.find(
     (agent) => agent.agentId === selectedAgentId,
   )
   const selectedAgentReady = selectedAgent
@@ -150,15 +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">
@@ -174,7 +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}
@@ -182,10 +148,10 @@ export const AgentCommandHome: FC = () => {
                   streaming={false}
                   disabled={!selectedAgentReady}
                   status={selectedAgentStatus}
-                  attachmentsEnabled={true}
+                  attachmentsEnabled={false}
                   placeholder={
                     selectedAgentReady
-                      ? `Ask ${selectedAgentName} to handle a task...`
+                      ? `Ask ${selectedCard?.name ?? 'your agent'} to handle a task...`
                       : 'Agent runtime is not running...'
                   }
                 />
@@ -196,8 +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/AgentRail.tsx

@@ -1,65 +0,0 @@
-import { type FC, useMemo } from 'react'
-import type {
-  HarnessAdapterDescriptor,
-  HarnessAgent,
-  HarnessAgentAdapter,
-} from '@/entrypoints/app/agents/agent-harness-types'
-import type { AgentAdapterHealth } from '@/entrypoints/app/agents/agent-row/agent-row.types'
-import { orderAgentsByPinThenRecency } from '@/entrypoints/app/agents/agents-list-order'
-import { AgentRailRow } from './AgentRailRow'
-
-interface AgentRailProps {
-  agents: HarnessAgent[]
-  adapters: HarnessAdapterDescriptor[]
-  activeAgentId: string
-  onSelectAgent: (agent: HarnessAgent) => void
-  onPinToggle: (agent: HarnessAgent, next: boolean) => void
-}
-
-/**
- * Left-column scrollable list of agents. The "Agents" label + back
- * button live in the shared top band above (so the rail header and
- * the chat header sit on a single aligned strip rather than as two
- * separately-sized headers per column). Sort matches `/agents`:
- * pinned-first → recency, so the rail doesn't reshuffle as turns
- * transition every 5 s.
- */
-export const AgentRail: FC<AgentRailProps> = ({
-  agents,
-  adapters,
-  activeAgentId,
-  onSelectAgent,
-  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])
-
-  const ordered = useMemo(() => orderAgentsByPinThenRecency(agents), [agents])
-
-  return (
-    <aside className="hidden min-h-0 flex-col border-border/50 border-r bg-background/70 lg:flex">
-      <div className="styled-scrollbar min-h-0 flex-1 space-y-1.5 overflow-y-auto px-3 py-3">
-        {ordered.map((agent) => (
-          <AgentRailRow
-            key={agent.id}
-            agent={agent}
-            active={agent.id === activeAgentId}
-            adapterHealth={adapterHealth.get(agent.adapter) ?? null}
-            onSelect={() => onSelectAgent(agent)}
-            onPinToggle={(next) => onPinToggle(agent, next)}
-          />
-        ))}
-      </div>
-    </aside>
-  )
-}

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

@@ -1,102 +0,0 @@
-import type { FC } from 'react'
-import { Badge } from '@/components/ui/badge'
-import { adapterLabel } from '@/entrypoints/app/agents/AdapterIcon'
-import type { HarnessAgent } from '@/entrypoints/app/agents/agent-harness-types'
-import { AgentSummaryChips } from '@/entrypoints/app/agents/agent-row/AgentSummaryChips'
-import { AgentTile } from '@/entrypoints/app/agents/agent-row/AgentTile'
-import type { AgentAdapterHealth } from '@/entrypoints/app/agents/agent-row/agent-row.types'
-import { PinToggle } from '@/entrypoints/app/agents/agent-row/PinToggle'
-import { cn } from '@/lib/utils'
-
-interface AgentRailRowProps {
-  agent: HarnessAgent
-  active: boolean
-  adapterHealth: AgentAdapterHealth | null
-  onSelect: () => void
-  onPinToggle: (next: boolean) => void
-}
-
-/**
- * Compact rail row for the chat-screen sidebar. Slims `<AgentRowCard>`
- * down to the essentials that fit a ~280 px rail: tile + name + status
- * badge + pin star, with the adapter / model / reasoning chips on a
- * second line. Token totals, sparkline, last-message preview all stay
- * on the `/agents` page where rows are full-width.
- */
-export const AgentRailRow: FC<AgentRailRowProps> = ({
-  agent,
-  active,
-  adapterHealth,
-  onSelect,
-  onPinToggle,
-}) => {
-  const status = agent.status ?? 'unknown'
-  const lastUsedAt = agent.lastUsedAt ?? null
-  const pinned = agent.pinned ?? false
-  return (
-    <button
-      type="button"
-      onClick={onSelect}
-      className={cn(
-        'group w-full rounded-2xl border px-3 py-3 text-left transition-colors',
-        active
-          ? 'border-[var(--accent-orange)]/30 bg-[var(--accent-orange)]/8'
-          : 'border-transparent bg-transparent hover:border-border/60 hover:bg-card',
-      )}
-    >
-      <div className="flex min-w-0 items-start gap-3">
-        <AgentTile
-          adapter={agent.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-[14px] leading-5">
-              {agent.name}
-            </span>
-            {status === 'working' && (
-              <Badge
-                variant="secondary"
-                className="h-5 bg-amber-50 px-1.5 text-[10px] text-amber-900 hover:bg-amber-50"
-              >
-                Working
-              </Badge>
-            )}
-            {status === 'asleep' && (
-              <Badge
-                variant="outline"
-                className="h-5 px-1.5 text-[10px] text-muted-foreground"
-              >
-                Asleep
-              </Badge>
-            )}
-            {status === 'error' && (
-              <Badge variant="destructive" className="h-5 px-1.5 text-[10px]">
-                Attention
-              </Badge>
-            )}
-            <div className="ml-auto">
-              <PinToggle pinned={pinned} onToggle={onPinToggle} />
-            </div>
-          </div>
-          <AgentSummaryChips
-            adapter={agent.adapter}
-            modelLabel={agent.modelId ?? null}
-            reasoningEffort={agent.reasoningEffort ?? null}
-            adapterHealth={adapterHealth}
-          />
-        </div>
-      </div>
-    </button>
-  )
-}
-
-/**
- * Tooltip-only label helper kept exported in case the tile row needs to
- * show "Codex agent" or similar in a future state. Inlined fallback for
- * the rare `unknown` adapter rendering path.
- */
-export function railRowAdapterLabel(agent: HarnessAgent): string {
-  return adapterLabel(agent.adapter)
-}

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

@@ -1,179 +0,0 @@
-import { ArrowLeft, Home } from 'lucide-react'
-import type { FC } from 'react'
-import { Badge } from '@/components/ui/badge'
-import { Button } from '@/components/ui/button'
-import { formatRelativeTime } from '@/entrypoints/app/agents/agent-display.helpers'
-import type { HarnessAgent } from '@/entrypoints/app/agents/agent-harness-types'
-import { AgentSummaryChips } from '@/entrypoints/app/agents/agent-row/AgentSummaryChips'
-import { formatTokens } from '@/entrypoints/app/agents/agent-row/agent-row.helpers'
-import type { AgentAdapterHealth } from '@/entrypoints/app/agents/agent-row/agent-row.types'
-import { PinToggle } from '@/entrypoints/app/agents/agent-row/PinToggle'
-import type { AgentLiveness } from '@/entrypoints/app/agents/LivenessDot'
-import { cn } from '@/lib/utils'
-
-interface ConversationHeaderProps {
-  agent: HarnessAgent | null
-  fallbackName: string
-  fallbackAdapter: 'claude' | 'codex' | 'openclaw' | 'unknown'
-  adapterHealth: AgentAdapterHealth | null
-  backLabel: string
-  backTarget: 'home' | 'page'
-  onGoHome: () => void
-  onPinToggle: (next: boolean) => void
-}
-
-/**
- * Strip above the chat. Mirrors the `/agents` row card's title row +
- * summary chips so the user gets adapter health, pin state, and status
- * at a glance — but adds the meta line (last used · lifetime tokens ·
- * queued) that's specific to this surface.
- *
- * The mobile `lg:hidden` Back button is preserved so the small-screen
- * collapse keeps a navigable header without a sidebar.
- */
-export const ConversationHeader: FC<ConversationHeaderProps> = ({
-  agent,
-  fallbackName,
-  fallbackAdapter,
-  adapterHealth,
-  backLabel,
-  backTarget,
-  onGoHome,
-  onPinToggle,
-}) => {
-  const BackIcon = backTarget === 'home' ? Home : ArrowLeft
-  const adapter = agent?.adapter ?? fallbackAdapter
-  const status: AgentLiveness = agent?.status ?? 'unknown'
-  const lastUsedAt = agent?.lastUsedAt ?? null
-  const pinned = agent?.pinned ?? false
-  const queueCount = agent?.queue?.length ?? 0
-  const tokens = agent?.tokens ?? null
-  const lifetimeTotal = tokens
-    ? tokens.cumulative.input + tokens.cumulative.output
-    : 0
-
-  const metaParts: string[] = []
-  if (lastUsedAt !== null) metaParts.push(formatRelativeTime(lastUsedAt))
-  if (lifetimeTotal > 0) metaParts.push(`${formatTokens(lifetimeTotal)} tokens`)
-  if (queueCount > 0) {
-    metaParts.push(queueCount === 1 ? '1 queued' : `${queueCount} queued`)
-  }
-
-  return (
-    <div className="flex min-h-[60px] shrink-0 items-center justify-between gap-4 px-5 py-2.5">
-      <div className="flex min-w-0 items-center gap-3">
-        <Button
-          variant="ghost"
-          size="icon"
-          onClick={onGoHome}
-          className="size-8 shrink-0 rounded-xl lg:hidden"
-          title={backLabel}
-        >
-          <BackIcon className="size-4" />
-        </Button>
-        <div className="group min-w-0 flex-1">
-          <div className="flex items-center gap-2">
-            <span className="truncate font-semibold text-[15px] leading-6">
-              {agent?.name || fallbackName}
-            </span>
-            {agent ? (
-              <PinToggle pinned={pinned} onToggle={onPinToggle} />
-            ) : null}
-          </div>
-          <div className="mt-0.5 flex items-center gap-2">
-            <AgentSummaryChips
-              adapter={adapter}
-              modelLabel={agent?.modelId ?? null}
-              reasoningEffort={agent?.reasoningEffort ?? null}
-              adapterHealth={adapterHealth}
-            />
-          </div>
-        </div>
-      </div>
-      <div className="flex shrink-0 flex-col items-end gap-1">
-        <StatusPill
-          status={status}
-          hasActiveTurn={Boolean(agent?.activeTurnId)}
-        />
-        <div className="flex h-4 items-center text-[11px] text-muted-foreground">
-          <span className="truncate">
-            {metaParts.length > 0 ? metaParts.join(' · ') : '\u00A0'}
-          </span>
-        </div>
-      </div>
-    </div>
-  )
-}
-
-interface StatusPillProps {
-  status: AgentLiveness
-  hasActiveTurn: boolean
-}
-
-/**
- * Working / Asleep / Attention all get distinctive styling; idle keeps
- * the legacy emerald `Ready` pill so the default state is visually
- * calm. Defensive working: `idle + activeTurnId` falls through to the
- * working pill since the server says a turn is in flight.
- */
-const StatusPill: FC<StatusPillProps> = ({ status, hasActiveTurn }) => {
-  const effective: AgentLiveness =
-    status === 'idle' && hasActiveTurn ? 'working' : status
-
-  const base =
-    'inline-flex items-center gap-2 rounded-full border px-3 py-0.5 text-[11px] uppercase tracking-[0.18em]'
-
-  if (effective === 'working') {
-    return (
-      <Badge
-        variant="secondary"
-        className={cn(
-          base,
-          'border-amber-200 bg-amber-50 text-amber-900 hover:bg-amber-50',
-        )}
-      >
-        <span className="size-1.5 animate-pulse rounded-full bg-amber-500" />
-        Working
-      </Badge>
-    )
-  }
-  if (effective === 'asleep') {
-    return (
-      <Badge variant="outline" className={cn(base, 'text-muted-foreground')}>
-        <span className="size-1.5 rounded-full bg-muted-foreground/50" />
-        Asleep
-      </Badge>
-    )
-  }
-  if (effective === 'error') {
-    return (
-      <Badge
-        variant="destructive"
-        className={cn(base, 'border-destructive/30')}
-      >
-        <span className="size-1.5 rounded-full bg-destructive-foreground" />
-        Attention
-      </Badge>
-    )
-  }
-  if (effective === 'idle') {
-    return (
-      <Badge
-        variant="outline"
-        className={cn(
-          base,
-          'border-emerald-200 bg-emerald-50 text-emerald-900 hover:bg-emerald-50',
-        )}
-      >
-        <span className="size-1.5 rounded-full bg-emerald-500" />
-        Ready
-      </Badge>
-    )
-  }
-  return (
-    <Badge variant="outline" className={cn(base, 'text-muted-foreground')}>
-      <span className="size-1.5 rounded-full bg-muted-foreground/30" />
-      Setup
-    </Badge>
-  )
-}

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

@@ -54,40 +54,25 @@ 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
 }
 
 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 +81,6 @@ 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 +299,6 @@ export const ConversationInput: FC<ConversationInputProps> = ({
   placeholder,
   attachmentsEnabled = true,
   variant = 'conversation',
-  onStop,
 }) => {
   const [input, setInput] = useState('')
   const [selectedTabs, setSelectedTabs] = useState<chrome.tabs.Tab[]>([])
@@ -411,17 +379,10 @@ 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 handleSend = () => {
     const text = input.trim()
-    if (disabled || isStaging) return
-    if (streaming && !queueAware) return
+    if (disabled || isStaging || streaming) return
     if (!text && attachments.length === 0) return
     onSend({ text, attachments })
     setInput('')
@@ -551,7 +512,6 @@ export const ConversationInput: FC<ConversationInputProps> = ({
               )}
             />
           </div>
-          {streaming && onStop ? <StopButton onStop={onStop} /> : null}
           <VoiceButton
             isRecording={voice.isRecording}
             isTranscribing={voice.isTranscribing}
@@ -569,13 +529,12 @@ export const ConversationInput: FC<ConversationInputProps> = ({
               !!disabled ||
               voice.isRecording ||
               voice.isTranscribing ||
-              (streaming && !queueAware)
+              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 ? (

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/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/pending-initial-message.test.ts

@@ -1,109 +0,0 @@
-import { afterEach, describe, expect, it } from 'bun:test'
-import type { StagedAttachment } from '@/lib/attachments'
-import {
-  consumePendingInitialMessage,
-  peekPendingInitialMessage,
-  setPendingInitialMessage,
-} from './pending-initial-message'
-
-function makeAttachment(id: string): StagedAttachment {
-  return {
-    id,
-    kind: 'image',
-    mediaType: 'image/png',
-    name: `${id}.png`,
-    dataUrl: `data:image/png;base64,${id}`,
-    payload: {
-      kind: 'image',
-      mediaType: 'image/png',
-      name: `${id}.png`,
-      dataUrl: `data:image/png;base64,${id}`,
-    },
-  }
-}
-
-afterEach(() => {
-  // Drain any leftover pending entry so tests don't leak into each
-  // other (the module-scope state survives across `it` blocks).
-  consumePendingInitialMessage('drain')
-  // If still set, clear by consuming with the matching id.
-  const leftover = peekPendingInitialMessage()
-  if (leftover) consumePendingInitialMessage(leftover.agentId)
-})
-
-describe('pending-initial-message', () => {
-  it('consume returns the payload set for the same agentId', () => {
-    setPendingInitialMessage({
-      agentId: 'agent-a',
-      text: 'hello',
-      attachments: [makeAttachment('one')],
-      createdAt: Date.now(),
-    })
-    const result = consumePendingInitialMessage('agent-a')
-    expect(result?.text).toBe('hello')
-    expect(result?.attachments).toHaveLength(1)
-    expect(result?.attachments[0]?.id).toBe('one')
-  })
-
-  it('consume is destructive — second call returns null', () => {
-    setPendingInitialMessage({
-      agentId: 'agent-a',
-      text: 'hello',
-      attachments: [],
-      createdAt: Date.now(),
-    })
-    expect(consumePendingInitialMessage('agent-a')).not.toBeNull()
-    expect(consumePendingInitialMessage('agent-a')).toBeNull()
-  })
-
-  it('consume returns null and preserves entry when agentId differs', () => {
-    setPendingInitialMessage({
-      agentId: 'agent-a',
-      text: 'hello',
-      attachments: [],
-      createdAt: Date.now(),
-    })
-    expect(consumePendingInitialMessage('agent-b')).toBeNull()
-    expect(peekPendingInitialMessage()?.agentId).toBe('agent-a')
-    expect(consumePendingInitialMessage('agent-a')).not.toBeNull()
-  })
-
-  it('returns null for entries older than the TTL', () => {
-    setPendingInitialMessage({
-      agentId: 'agent-a',
-      text: 'old',
-      attachments: [],
-      createdAt: Date.now() - 11_000, // older than 10 s TTL
-    })
-    expect(consumePendingInitialMessage('agent-a')).toBeNull()
-  })
-
-  it('replaces a previous pending entry when set is called again', () => {
-    setPendingInitialMessage({
-      agentId: 'agent-a',
-      text: 'first',
-      attachments: [],
-      createdAt: Date.now(),
-    })
-    setPendingInitialMessage({
-      agentId: 'agent-b',
-      text: 'second',
-      attachments: [makeAttachment('two')],
-      createdAt: Date.now(),
-    })
-    expect(consumePendingInitialMessage('agent-a')).toBeNull()
-    const result = consumePendingInitialMessage('agent-b')
-    expect(result?.text).toBe('second')
-    expect(result?.attachments[0]?.id).toBe('two')
-  })
-
-  it('no-ops when set is called with empty agentId', () => {
-    setPendingInitialMessage({
-      agentId: '',
-      text: 'oops',
-      attachments: [],
-      createdAt: Date.now(),
-    })
-    expect(peekPendingInitialMessage()).toBeNull()
-  })
-})

packages/browseros-agent/apps/agent/entrypoints/app/agent-command/pending-initial-message.ts

@@ -1,81 +0,0 @@
-import type { StagedAttachment } from '@/lib/attachments'
-
-/**
- * Same-tab in-memory handoff between the `/home` composer and the
- * chat screen at `/home/agents/:agentId`. URL search params (`?q=`)
- * carry the text fine, but cannot carry binary attachments — a multi-
- * megabyte image dataUrl would explode URL length limits and round-
- * trip badly. This module is the rich-data side channel for the same
- * navigation: the composer writes here, the chat screen reads here on
- * mount.
- *
- * Intentionally module-scope. Same render tree, same tab — no need
- * for sessionStorage (which would force JSON-serialising the dataUrls
- * and re-parsing on the read side). Cross-tab handoff is out of
- * scope: the user typing at home in tab A and switching to tab B's
- * chat would surface an empty registry there, which is the correct
- * behaviour.
- */
-
-export interface PendingInitialMessage {
-  agentId: string
-  text: string
-  attachments: StagedAttachment[]
-  createdAt: number
-}
-
-/**
- * 10s TTL on the entry. A stale entry from a back-button journey
- * shouldn't fire on a future visit; if real-world latency makes 10s
- * too tight under slow harness boot, bump but never make it
- * indefinite.
- */
-const PENDING_TTL_MS = 10_000
-
-let pending: PendingInitialMessage | null = null
-let pendingTimer: ReturnType<typeof setTimeout> | null = null
-
-function clearPending(): void {
-  pending = null
-  if (pendingTimer !== null) {
-    clearTimeout(pendingTimer)
-    pendingTimer = null
-  }
-}
-
-export function setPendingInitialMessage(payload: PendingInitialMessage): void {
-  // Defensive: the home composer should never call this without an
-  // agent selected. If it somehow does, no-op rather than holding a
-  // payload we can't route.
-  if (!payload.agentId) return
-  clearPending()
-  pending = payload
-  pendingTimer = setTimeout(clearPending, PENDING_TTL_MS)
-}
-
-/**
- * Destructive read. Returns the entry only if `agentId` matches and
- * the entry is fresh; clears the entry on success so Strict-Mode
- * double-invokes can't double-send.
- */
-export function consumePendingInitialMessage(
-  agentId: string,
-): PendingInitialMessage | null {
-  if (!pending) return null
-  if (pending.agentId !== agentId) return null
-  if (Date.now() - pending.createdAt >= PENDING_TTL_MS) {
-    clearPending()
-    return null
-  }
-  const entry = pending
-  clearPending()
-  return entry
-}
-
-/**
- * Non-mutating read for tests. Production code should never need this
- * — use `consume` and own the lifecycle.
- */
-export function peekPendingInitialMessage(): PendingInitialMessage | null {
-  return pending
-}

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

@@ -36,15 +36,6 @@ interface UseAgentConversationOptions {
   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(
@@ -220,46 +211,31 @@ export function useAgentConversation(
   }
   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).
+  // On mount (and whenever the agent changes), check whether the
+  // server has an in-flight turn for this agent and reattach to it.
+  // This is what makes the chat resilient across tab close/reopen,
+  // refresh, and navigation: the runtime call kept running on the
+  // server while we were away. Effect only depends on `agentId` —
+  // the event handler is read off a ref so this doesn't re-subscribe
+  // every render.
   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
+        if (streamAbortRef.current) return // a fresh send already in flight
 
         // 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.
+        // to render into. We don't have the user message text on
+        // resume; the assistant turn is what we're catching up on.
         setTurns((prev) => [
           ...prev,
           {
             id: crypto.randomUUID(),
-            userText: active.prompt ?? '',
+            userText: '',
             parts: [],
             done: false,
             timestamp: active.startedAt,
@@ -271,7 +247,6 @@ export function useAgentConversation(
         lastSeqRef.current = null
         streamAbortRef.current = abortController
         setStreaming(true)
-        weStartedStream = true
 
         const response = await attachToHarnessTurn(agentId, {
           turnId: active.turnId,
@@ -290,20 +265,10 @@ export function useAgentConversation(
         // 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) {
+        if (!cancelled) {
+          if (streamAbortRef.current === abortController) {
+            streamAbortRef.current = null
+          }
           turnIdRef.current = null
           lastSeqRef.current = null
           setStreaming(false)
@@ -316,7 +281,7 @@ export function useAgentConversation(
       cancelled = true
       abortController.abort()
     }
-  }, [agentId, activeTurnIdDep])
+  }, [agentId])
 
   const send = async (input: string | SendInput) => {
     const normalized: SendInput =

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/agents/AgentList.tsx

@@ -2,75 +2,67 @@ 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 { compareAgentsByPinThenRecency } from './agents-list-order'
+import type { HarnessAgent, HarnessAgentAdapter } from './agent-harness-types'
 import type { AgentListItem } from './agents-page-types'
 import type { AgentLiveness } from './LivenessDot'
 
 interface AgentListProps {
   agents: AgentListItem[]
-  /** Optional per-agent activity metadata, keyed by `agentId`. */
+  /**
+   * Optional per-agent activity metadata. Keyed by `agentId`. Missing
+   * entries fall back to status='unknown' / lastUsedAt=null and the
+   * row renders an "unknown" dot. The server will populate this once
+   * the activity tracker ships; the page works without it.
+   */
   activity?: Record<
     string,
     { status: AgentLiveness; lastUsedAt: number | null }
   >
-  /** Lookup table from harness id → enriched agent record. */
+  /**
+   * Lookup table from harness agent id → adapter + reasoning effort,
+   * sourced from `useHarnessAgents`. Lets the row card render the
+   * correct adapter icon and chips for harness agents (legacy
+   * /claw/agents entries fall back to inferring from `runtimeLabel`).
+   */
   harnessAgentLookup?: Map<string, HarnessAgent>
-  /** Adapter catalog (carries per-adapter health). */
-  adapters: HarnessAdapterDescriptor[]
   loading: boolean
   deletingAgentKey: string | null
   onCreateAgent: () => void
   onDeleteAgent: (agent: AgentListItem) => void
-  onPinToggle: (agent: AgentListItem, next: boolean) => void
 }
 
 export const AgentList: FC<AgentListProps> = ({
   agents,
   activity,
   harnessAgentLookup,
-  adapters,
   loading,
   deletingAgentKey,
   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 by recency: most recently used first; never-used agents drop
+  // to the bottom in id-stable order so the list doesn't reshuffle on
+  // every refresh. The pinned exception is the gateway's `main` agent
+  // when it's never been touched — keep it at the top 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,
-        id: agent.agentId,
-        pinned: harness?.pinned ?? false,
-        lastUsedAt: activity?.[agent.agentId]?.lastUsedAt ?? null,
-      }
+    const withScore = agents.map((agent) => {
+      const lastUsedAt = activity?.[agent.agentId]?.lastUsedAt ?? null
+      return { agent, lastUsedAt }
     })
-    return withMeta
-      .sort(compareAgentsByPinThenRecency)
+    return withScore
+      .sort((a, b) => {
+        const aPinned = a.agent.agentId === 'main' && a.lastUsedAt === null
+        const bPinned = b.agent.agentId === 'main' && b.lastUsedAt === null
+        if (aPinned && !bPinned) return -1
+        if (!aPinned && bPinned) 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])
+  }, [activity, agents])
 
   if (loading && agents.length === 0) {
     return (
@@ -88,23 +80,18 @@ export const AgentList: FC<AgentListProps> = ({
     <div className="grid gap-3">
       {ordered.map((agent) => {
         const harness = harnessAgentLookup?.get(agent.agentId)
-        const adapter: HarnessAgentAdapter | 'unknown' =
+        const adapter: HarnessAgentAdapter | undefined =
           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}
+            agent={agent}
+            status={activity?.[agent.agentId]?.status}
+            lastUsedAt={activity?.[agent.agentId]?.lastUsedAt}
+            adapter={adapter}
+            reasoningEffort={harness?.reasoningEffort ?? null}
             onDelete={onDeleteAgent}
-            onPinToggle={onPinToggle}
+            deleting={deletingAgentKey === agent.key}
           />
         )
       })}
@@ -112,53 +99,10 @@ export const AgentList: FC<AgentListProps> = ({
   )
 }
 
-function inferAdapterFromLabel(label: string): HarnessAgentAdapter | 'unknown' {
+function inferAdapterFromLabel(label: string): HarnessAgentAdapter | undefined {
   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
+  return undefined
 }

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

@@ -1,99 +1,270 @@
+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 { Badge } from '@/components/ui/badge'
+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 { 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'
+import { AdapterIcon, adapterLabel } from './AdapterIcon'
+import {
+  canDelete as canDeleteAgent,
+  canRename as canRenameAgent,
+  displayName,
+  formatRelativeTime,
+  workspaceLabel,
+} from './agent-display.helpers'
+import type { HarnessAgentAdapter } from './agent-harness-types'
+import type { AgentListItem } from './agents-page-types'
+import { type AgentLiveness, LivenessDot } from './LivenessDot'
 
-interface AgentRowCardProps extends AgentRowCallbacks {
-  data: AgentRowData
-  /** Whether THIS agent is mid-delete; renders a spinner in the menu. */
+interface AgentRowCardProps {
+  agent: AgentListItem
+  /**
+   * Per-agent extras the listing surface provides on top of the
+   * minimal `AgentListItem` shape. `lastUsedAt` survives server
+   * restart (sourced from acpx session record); `status` is in-memory
+   * server-side.
+   */
+  status?: AgentLiveness
+  lastUsedAt?: number | null
+  /** Adapter the agent belongs to. Drives icon + label. */
+  adapter?: HarnessAgentAdapter
+  /** Reasoning effort chip (claude/codex/openclaw catalog). */
+  reasoningEffort?: string | null
+  /** Modeled directly off the inbound delete handler so the parent owns the dialog. */
+  onDelete: (agent: AgentListItem) => void
+  /** Whether THIS agent is mid-delete; renders a spinner in place of the trash icon. */
   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,
+  agent,
+  status = 'unknown',
+  lastUsedAt,
+  adapter,
+  reasoningEffort,
   onDelete,
-  onPinToggle,
+  deleting,
 }) => {
+  const navigate = useNavigate()
+  const adapterId = adapter ?? inferAdapterFromListItem(agent)
+  const workspace = workspaceLabel(agent)
+  const lastUsedLabel = formatRelativeTime(lastUsedAt ?? null)
+  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={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',
+        'group rounded-xl border border-border bg-card p-4 shadow-sm transition-all',
+        'hover:border-[var(--accent-orange)]/50 hover:shadow-sm',
       )}
     >
       <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)}
+        {/* Adapter tile + liveness dot in the corner. */}
+        <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={adapterId} className="h-6 w-6" />
+          </div>
+          <LivenessDot
+            status={status}
+            detail={livenessDetail(status, lastUsedAt)}
+            className="absolute -right-0.5 -bottom-0.5"
           />
-
-          <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 className="min-w-0 flex-1">
+          <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>
+            )}
+          </div>
+
+          <div className="mb-2 flex flex-wrap items-center gap-1.5 text-xs">
+            <Badge variant="secondary" className="font-normal">
+              {adapterLabel(adapterId)}
+            </Badge>
+            {agent.modelLabel && agent.modelLabel !== 'default' && (
+              <Badge variant="outline" className="font-normal">
+                {agent.modelLabel}
+              </Badge>
+            )}
+            {reasoningEffort && reasoningEffort !== 'medium' && (
+              <Badge variant="outline" className="font-normal">
+                {reasoningEffort}
+              </Badge>
+            )}
+          </div>
+
+          <div className="flex flex-wrap items-center gap-2 text-muted-foreground text-xs">
+            <span>Last used {lastUsedLabel}</span>
+            {workspace && (
+              <>
+                <span aria-hidden>•</span>
+                <span className="truncate font-mono" title={workspace}>
+                  {workspace}
+                </span>
+              </>
+            )}
+          </div>
+        </div>
+
+        <div className="flex shrink-0 items-center gap-2">
+          <Button variant="outline" size="sm" onClick={handleChat}>
+            <MessageSquare className="mr-1.5 h-3 w-3" />
+            Chat
+          </Button>
+          <DropdownMenu>
+            <DropdownMenuTrigger asChild>
+              <Button
+                variant="ghost"
+                size="icon"
+                aria-label={`More actions for ${displayName(agent)}`}
+                className="h-8 w-8"
+              >
+                <MoreHorizontal className="h-4 w-4" />
+              </Button>
+            </DropdownMenuTrigger>
+            <DropdownMenuContent align="end" className="w-44">
+              <DropdownMenuItem onSelect={() => void handleCopyId()}>
+                <Copy className="mr-2 h-3.5 w-3.5" />
+                Copy id
+              </DropdownMenuItem>
+              <RenameMenuItem disabled={!allowRename} />
+              <ResetHistoryMenuItem />
+              <DropdownMenuSeparator />
+              <DropdownMenuItem
+                onSelect={() => onDelete(agent)}
+                disabled={!allowDelete || deleting}
+                className="text-destructive focus:text-destructive"
+              >
+                {deleting ? (
+                  <Loader2 className="mr-2 h-3.5 w-3.5 animate-spin" />
+                ) : (
+                  <Trash2 className="mr-2 h-3.5 w-3.5" />
+                )}
+                Delete
+              </DropdownMenuItem>
+            </DropdownMenuContent>
+          </DropdownMenu>
+        </div>
       </div>
     </div>
   )
 }
+
+const RenameMenuItem: FC<{ disabled: boolean }> = ({ disabled }) => {
+  const item = (
+    <DropdownMenuItem disabled className="text-muted-foreground">
+      <Pencil className="mr-2 h-3.5 w-3.5" />
+      Rename
+    </DropdownMenuItem>
+  )
+  if (!disabled) return item
+  // Disabled but with a hint so users know it's coming, not broken.
+  return (
+    <TooltipProvider delayDuration={300}>
+      <Tooltip>
+        <TooltipTrigger asChild>
+          <span className="block w-full">{item}</span>
+        </TooltipTrigger>
+        <TooltipContent side="left" className="text-xs">
+          Rename coming soon
+        </TooltipContent>
+      </Tooltip>
+    </TooltipProvider>
+  )
+}
+
+const ResetHistoryMenuItem: FC = () => {
+  const item = (
+    <DropdownMenuItem disabled className="text-muted-foreground">
+      <RotateCcw className="mr-2 h-3.5 w-3.5" />
+      Reset history
+    </DropdownMenuItem>
+  )
+  return (
+    <TooltipProvider delayDuration={300}>
+      <Tooltip>
+        <TooltipTrigger asChild>
+          <span className="block w-full">{item}</span>
+        </TooltipTrigger>
+        <TooltipContent side="left" className="text-xs">
+          Reset history coming soon
+        </TooltipContent>
+      </Tooltip>
+    </TooltipProvider>
+  )
+}
+
+function inferAdapterFromListItem(
+  agent: AgentListItem,
+): HarnessAgentAdapter | 'unknown' {
+  const label = agent.runtimeLabel?.toLowerCase()
+  if (label?.includes('claude')) return 'claude'
+  if (label?.includes('codex')) return 'codex'
+  if (label?.includes('openclaw')) return 'openclaw'
+  return 'unknown'
+}
+
+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/AgentsPage.tsx

@@ -44,7 +44,6 @@ import {
   useCreateHarnessAgent,
   useDeleteHarnessAgent,
   useHarnessAgents,
-  useUpdateHarnessAgent,
 } from './useAgents'
 import { useOpenClawAgents, useOpenClawMutations } from './useOpenClaw'
 
@@ -77,7 +76,6 @@ export const AgentsPage: FC = () => {
   } = useOpenClawAgents(openClawAgentsEnabled)
   const createHarnessAgent = useCreateHarnessAgent()
   const deleteHarnessAgent = useDeleteHarnessAgent()
-  const updateHarnessAgent = useUpdateHarnessAgent()
   const {
     setupOpenClaw,
     createAgent: createOpenClawAgent,
@@ -344,24 +342,12 @@ export const AgentsPage: FC = () => {
           agents={agentListItems}
           activity={agentActivity}
           harnessAgentLookup={harnessAgentLookup}
-          adapters={adapters}
           loading={agentsLoading}
           deletingAgentKey={deletingAgent ? deletingAgentKey : null}
           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/agent-display.helpers.ts

@@ -1,5 +1,4 @@
 import type { AgentListItem } from './agents-page-types'
-import type { AgentLiveness } from './LivenessDot'
 
 /**
  * Display rules for the redesigned agent rows. Pure helpers — no React,
@@ -83,25 +82,3 @@ export function formatRelativeTime(epochMs: number | null): string {
   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

@@ -56,43 +56,6 @@ export interface HarnessAgent {
    * 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 +66,6 @@ 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 {

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-list-order.test.ts

@@ -1,104 +0,0 @@
-import { describe, expect, it } from 'bun:test'
-import type { HarnessAgent } from './agent-harness-types'
-import {
-  compareAgentsByPinThenRecency,
-  orderAgentsByPinThenRecency,
-} from './agents-list-order'
-
-function makeAgent(input: {
-  id: string
-  pinned?: boolean
-  lastUsedAt?: number | null
-}): HarnessAgent {
-  return {
-    id: input.id,
-    name: input.id,
-    adapter: 'codex',
-    permissionMode: 'approve-all',
-    sessionKey: 'session',
-    createdAt: 0,
-    updatedAt: 0,
-    pinned: input.pinned,
-    lastUsedAt: input.lastUsedAt,
-  }
-}
-
-describe('orderAgentsByPinThenRecency', () => {
-  it('floats pinned agents to the top regardless of recency', () => {
-    const result = orderAgentsByPinThenRecency([
-      makeAgent({ id: 'a', pinned: false, lastUsedAt: 1_000 }),
-      makeAgent({ id: 'b', pinned: true, lastUsedAt: 100 }),
-      makeAgent({ id: 'c', pinned: false, lastUsedAt: 500 }),
-    ])
-    expect(result.map((entry) => entry.id)).toEqual(['b', 'a', 'c'])
-  })
-
-  it('sorts by lastUsedAt desc within each pin group', () => {
-    const result = orderAgentsByPinThenRecency([
-      makeAgent({ id: 'older-pin', pinned: true, lastUsedAt: 100 }),
-      makeAgent({ id: 'newer-pin', pinned: true, lastUsedAt: 200 }),
-      makeAgent({ id: 'older', pinned: false, lastUsedAt: 50 }),
-      makeAgent({ id: 'newer', pinned: false, lastUsedAt: 80 }),
-    ])
-    expect(result.map((entry) => entry.id)).toEqual([
-      'newer-pin',
-      'older-pin',
-      'newer',
-      'older',
-    ])
-  })
-
-  it('seed-pins the gateway main agent above other never-used agents', () => {
-    const result = orderAgentsByPinThenRecency([
-      makeAgent({ id: 'aaa', pinned: false, lastUsedAt: null }),
-      makeAgent({ id: 'main', pinned: false, lastUsedAt: null }),
-      makeAgent({ id: 'zzz', pinned: false, lastUsedAt: null }),
-    ])
-    expect(result.map((entry) => entry.id)).toEqual(['main', 'aaa', 'zzz'])
-  })
-
-  it('drops the main seed-pin once the agent has been used', () => {
-    const result = orderAgentsByPinThenRecency([
-      makeAgent({ id: 'aaa', pinned: false, lastUsedAt: 999 }),
-      makeAgent({ id: 'main', pinned: false, lastUsedAt: 1 }),
-    ])
-    expect(result.map((entry) => entry.id)).toEqual(['aaa', 'main'])
-  })
-
-  it('puts never-used agents below recently-used ones', () => {
-    const result = orderAgentsByPinThenRecency([
-      makeAgent({ id: 'fresh', pinned: false, lastUsedAt: null }),
-      makeAgent({ id: 'used', pinned: false, lastUsedAt: 100 }),
-    ])
-    expect(result.map((entry) => entry.id)).toEqual(['used', 'fresh'])
-  })
-
-  it('id-stable tiebreaks two agents with identical lastUsedAt', () => {
-    const result = orderAgentsByPinThenRecency([
-      makeAgent({ id: 'b', pinned: false, lastUsedAt: 100 }),
-      makeAgent({ id: 'a', pinned: false, lastUsedAt: 100 }),
-    ])
-    expect(result.map((entry) => entry.id)).toEqual(['a', 'b'])
-  })
-})
-
-describe('compareAgentsByPinThenRecency', () => {
-  it('produces the same order as the harness-shape helper', () => {
-    const items = [
-      { id: 'older', pinned: false, lastUsedAt: 50 },
-      { id: 'newer', pinned: false, lastUsedAt: 80 },
-      { id: 'pinned', pinned: true, lastUsedAt: 1 },
-    ]
-    const sorted = [...items].sort(compareAgentsByPinThenRecency)
-    expect(sorted.map((item) => item.id)).toEqual(['pinned', 'newer', 'older'])
-  })
-
-  it('seeds the main agent above other never-used rows', () => {
-    const items = [
-      { id: 'zzz', pinned: false, lastUsedAt: null },
-      { id: 'main', pinned: false, lastUsedAt: null },
-    ]
-    const sorted = [...items].sort(compareAgentsByPinThenRecency)
-    expect(sorted.map((item) => item.id)).toEqual(['main', 'zzz'])
-  })
-})

packages/browseros-agent/apps/agent/entrypoints/app/agents/agents-list-order.ts

@@ -1,59 +0,0 @@
-import type { HarnessAgent } from './agent-harness-types'
-
-/**
- * Stable ordering for index-shaped agent surfaces (the `/agents` rail
- * and the chat-screen rail at `/agents/:agentId`). Pinned rows float
- * to the top, then recency desc, with never-used agents falling to
- * the bottom in id-stable order. The gateway's `main` agent gets
- * seed-pinned to the top of the never-used group so a fresh install
- * has an obvious starting point even before the user has used it.
- *
- * NOT the same rule as the home grid (`orderHomeAgents`): home is
- * action-shaped — active-turn floats to the top — so users can
- * resume what's running. The chat rail keeps recency stable so it
- * doesn't reshuffle as turns transition every 5s.
- */
-export function orderAgentsByPinThenRecency(
-  agents: HarnessAgent[],
-): HarnessAgent[] {
-  return [...agents].sort((a, b) => {
-    const aPinned = a.pinned ?? false
-    const bPinned = b.pinned ?? false
-    if (aPinned !== bPinned) return aPinned ? -1 : 1
-
-    const aSeed = a.id === 'main' && (a.lastUsedAt ?? null) === null
-    const bSeed = b.id === 'main' && (b.lastUsedAt ?? null) === null
-    if (aSeed && !bSeed) return -1
-    if (!aSeed && bSeed) return 1
-
-    const aValue = a.lastUsedAt ?? Number.NEGATIVE_INFINITY
-    const bValue = b.lastUsedAt ?? Number.NEGATIVE_INFINITY
-    if (aValue !== bValue) return bValue - aValue
-
-    return a.id.localeCompare(b.id)
-  })
-}
-
-/**
- * Same comparator, but operates over arbitrary records that carry
- * `pinned`, `lastUsedAt`, and an `id`-equivalent key. Used by the
- * `/agents` `AgentList` which pivots `AgentListItem` + harness
- * lookup into a sortable shape; both surfaces stay on identical
- * sort semantics through this adapter.
- */
-export function compareAgentsByPinThenRecency<
-  T extends { pinned: boolean; lastUsedAt: number | null; id: string },
->(a: T, b: T): number {
-  if (a.pinned !== b.pinned) return a.pinned ? -1 : 1
-
-  const aSeed = a.id === 'main' && a.lastUsedAt === null
-  const bSeed = b.id === 'main' && b.lastUsedAt === null
-  if (aSeed && !bSeed) return -1
-  if (!aSeed && bSeed) return 1
-
-  const aValue = a.lastUsedAt ?? Number.NEGATIVE_INFINITY
-  const bValue = b.lastUsedAt ?? Number.NEGATIVE_INFINITY
-  if (aValue !== bValue) return bValue - aValue
-
-  return a.id.localeCompare(b.id)
-}

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

@@ -8,7 +8,6 @@ import {
   type HarnessAdapterDescriptor,
   type HarnessAgent,
   type HarnessAgentHistoryPage,
-  type HarnessQueuedMessage,
   mapHarnessAgentToEntry,
 } from './agent-harness-types'
 import type { OpenClawStatus } from './useOpenClaw'
@@ -136,63 +135,6 @@ 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()
@@ -264,8 +206,6 @@ export interface HarnessActiveTurnInfo {
   lastSeq: number
   startedAt: number
   endedAt?: number
-  /** User message that kicked off the turn; null when not captured. */
-  prompt: string | null
 }
 
 /**
@@ -320,145 +260,3 @@ 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/sidepanel/index/sidepanel-chat-targets.test.ts

@@ -1,8 +1,5 @@
 import { describe, expect, it } from 'bun:test'
-import type {
-  HarnessAdapterDescriptor,
-  HarnessAgent,
-} from '@/entrypoints/app/agents/agent-harness-types'
+import type { HarnessAdapterDescriptor } from '@/entrypoints/app/agents/agent-harness-types'
 import type { LlmProviderConfig } from '@/lib/llm-providers/types'
 import {
   buildSidepanelChatTargets,
@@ -80,96 +77,58 @@ const adapters: HarnessAdapterDescriptor[] = [
   },
 ]
 
-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 })
+  it('returns LLM targets plus one ACP target per adapter model', () => {
+    const targets = buildSidepanelChatTargets({ providers, adapters })
 
     expect(targets.map((target) => target.id)).toEqual([
       'browseros',
       'anthropic-sonnet',
-      'agent-codex',
-      'agent-openclaw',
+      'acp:claude:sonnet:medium',
+      'acp:claude:haiku:medium',
+      'acp:codex:gpt-5.5:medium',
+      'acp:openclaw:default:medium',
     ])
   })
 
-  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')
+  it('emits a single default ACP target for adapters with no per-session model picker', () => {
+    const targets = buildSidepanelChatTargets({ providers, adapters })
+    const openclaw = targets.find(
+      (target) => target.id === 'acp:openclaw:default:medium',
+    )
 
     expect(openclaw).toMatchObject({
       kind: 'acp',
-      id: 'agent-openclaw',
-      agentId: 'agent-openclaw',
       adapter: 'openclaw',
       adapterName: 'OpenClaw',
       modelId: 'default',
       modelLabel: 'default',
-      name: 'Research Claw',
+      // Without a model picker, the target name is just the adapter
+      // name — the user picks the adapter, not a model under it.
+      name: 'OpenClaw',
       modelControl: 'best-effort',
-      reasoningEffort: 'high',
+      reasoningEffort: 'medium',
     })
   })
 
-  it('preserves adapter metadata for created agent targets', () => {
-    const targets = buildSidepanelChatTargets({ providers, adapters, agents })
-    const codex = targets.find((target) => target.id === 'agent-codex')
+  it('preserves ACP model-control and recommendation metadata', () => {
+    const targets = buildSidepanelChatTargets({ providers, adapters })
+    const haiku = targets.find(
+      (target) => target.id === 'acp:claude:haiku:medium',
+    )
 
-    expect(codex).toMatchObject({
+    expect(haiku).toMatchObject({
       kind: 'acp',
-      agentId: 'agent-codex',
-      adapter: 'codex',
-      adapterName: 'Codex',
-      modelId: 'gpt-5.5',
-      modelLabel: 'GPT-5.5',
-      modelControl: 'runtime-supported',
+      adapter: 'claude',
+      modelId: 'haiku',
+      modelControl: 'best-effort',
       recommended: true,
       reasoningEffort: 'medium',
-      reasoningEffortLabel: 'Medium',
     })
   })
 
-  it('still returns LLM targets when agents and adapters are unavailable', () => {
-    expect(
-      buildSidepanelChatTargets({ providers, adapters: [], agents: [] }),
-    ).toEqual([
+  it('still returns LLM targets when ACP adapters are unavailable', () => {
+    expect(buildSidepanelChatTargets({ providers, adapters: [] })).toEqual([
       {
         kind: 'llm',
         id: 'browseros',
@@ -190,7 +149,7 @@ describe('buildSidepanelChatTargets', () => {
 
 describe('resolveSidepanelChatTarget', () => {
   it('resolves selected LLM targets back to their provider config', () => {
-    const targets = buildSidepanelChatTargets({ providers, adapters, agents })
+    const targets = buildSidepanelChatTargets({ providers, adapters })
     const resolved = resolveSidepanelChatTarget({
       targets,
       defaultProviderId: 'browseros',
@@ -202,32 +161,13 @@ describe('resolveSidepanelChatTarget', () => {
   })
 
   it('falls back to the current default LLM provider when a persisted ACP target is stale', () => {
-    const targets = buildSidepanelChatTargets({
-      providers,
-      adapters,
-      agents: [],
-    })
+    const targets = buildSidepanelChatTargets({ providers, adapters: [] })
 
     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' },
+        selection: { kind: 'acp', id: 'acp:claude:haiku:medium' },
       }),
     ).toMatchObject({
       kind: 'llm',
@@ -240,8 +180,10 @@ 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')
+    const targets = buildSidepanelChatTargets({ providers, adapters })
+    const target = targets.find(
+      (candidate) => candidate.id === 'acp:codex:gpt-5.5:medium',
+    )
 
     await persistSidepanelChatTargetSelection(target, {
       setValue: async (value) => {
@@ -251,7 +193,7 @@ describe('persistSidepanelChatTargetSelection', () => {
 
     expect(savedSelection as SidepanelChatTargetSelection | null).toEqual({
       kind: 'acp',
-      id: 'agent-codex',
+      id: 'acp:codex:gpt-5.5:medium',
     })
     expect(providers).toEqual(originalProviders)
   })

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

@@ -1,6 +1,5 @@
 import type {
   HarnessAdapterDescriptor,
-  HarnessAgent,
   HarnessAgentAdapter,
 } from '@/entrypoints/app/agents/agent-harness-types'
 import type { LlmProviderConfig, ProviderType } from '@/lib/llm-providers/types'
@@ -20,7 +19,6 @@ export type SidepanelChatTarget =
       id: string
       name: string
       type: 'acp'
-      agentId: string
       adapter: HarnessAgentAdapter
       adapterName: string
       modelId: string
@@ -39,7 +37,6 @@ export type SidepanelChatTargetSelection = Pick<
 interface BuildSidepanelChatTargetsInput {
   providers: LlmProviderConfig[]
   adapters: HarnessAdapterDescriptor[]
-  agents?: HarnessAgent[]
 }
 
 interface ResolveSidepanelChatTargetInput {
@@ -66,49 +63,61 @@ let sidepanelChatTargetSelectionStorage:
 export function buildSidepanelChatTargets({
   providers,
   adapters,
-  agents = [],
 }: BuildSidepanelChatTargetsInput): SidepanelChatTarget[] {
   return [
     ...providers.map(toLlmTarget),
-    ...agents.map((agent) => toAcpTargetForAgent(agent, adapters)),
+    ...adapters.flatMap(toAcpTargetsForAdapter),
   ]
 }
 
-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,
+function toAcpTargetsForAdapter(
+  adapter: HarnessAdapterDescriptor,
+): SidepanelChatTarget[] {
+  const reasoning = adapter.reasoningEfforts.find(
+    (effort) => effort.id === adapter.defaultReasoningEffort,
   )
+  const reasoningEffort =
+    reasoning?.id ?? adapter.defaultReasoningEffort ?? 'medium'
 
-  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,
+  // Adapters with no per-session model picker (e.g. OpenClaw, whose
+  // model lives on the gateway-side agent record) still need exactly
+  // one sidepanel target so the user can pick the adapter at all.
+  if (adapter.models.length === 0) {
+    return [
+      {
+        kind: 'acp',
+        id: buildAcpTargetId(
+          adapter.id,
+          adapter.defaultModelId,
+          reasoningEffort,
+        ),
+        name: adapter.name,
+        type: 'acp',
+        adapter: adapter.id,
+        adapterName: adapter.name,
+        modelId: adapter.defaultModelId,
+        modelLabel: 'default',
+        modelControl: adapter.modelControl,
+        reasoningEffort,
+        reasoningEffortLabel: reasoning?.label,
+      },
+    ]
+  }
+
+  return adapter.models.map((model) => ({
+    kind: 'acp' as const,
+    id: buildAcpTargetId(adapter.id, model.id, reasoningEffort),
+    name: `${adapter.name} ${model.label}`,
+    type: 'acp' as const,
+    adapter: adapter.id,
+    adapterName: adapter.name,
+    modelId: model.id,
+    modelLabel: model.label,
+    modelControl: adapter.modelControl,
+    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({
@@ -163,6 +172,14 @@ function toLlmTarget(provider: LlmProviderConfig): SidepanelChatTarget {
   }
 }
 
+export function buildAcpTargetId(
+  adapter: HarnessAgentAdapter,
+  modelId: string,
+  reasoningEffort: string,
+): string {
+  return `acp:${adapter}:${modelId}:${reasoningEffort}`
+}
+
 async function getSidepanelChatTargetSelectionStorage(): Promise<SidepanelChatTargetSelectionStore> {
   if (sidepanelChatTargetSelectionStorage) {
     return sidepanelChatTargetSelectionStorage

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

@@ -1,9 +1,6 @@
 import { useCallback, useEffect, useMemo, useRef, useState } from 'react'
 import useDeepCompareEffect from 'use-deep-compare-effect'
-import {
-  useAgentAdapters,
-  useHarnessAgents,
-} from '@/entrypoints/app/agents/useAgents'
+import { useAgentAdapters } 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'
@@ -41,7 +38,6 @@ export const useChatRefs = () => {
     isLoading: isLoadingProviders,
   } = useLlmProviders()
   const { adapters, loading: isLoadingAdapters } = useAgentAdapters()
-  const { harnessAgents, loading: isLoadingAgents } = useHarnessAgents()
   const { personalization } = usePersonalization()
   const [targetSelection, setTargetSelection] =
     useState<SidepanelChatTargetSelection | null>(null)
@@ -61,9 +57,8 @@ export const useChatRefs = () => {
       buildSidepanelChatTargets({
         providers: llmProviders,
         adapters,
-        agents: harnessAgents,
       }),
-    [llmProviders, adapters, harnessAgents],
+    [llmProviders, adapters],
   )
 
   const selectedChatTarget = useMemo(
@@ -121,7 +116,6 @@ export const useChatRefs = () => {
     selectedChatTarget,
     selectChatTarget,
     selectedLlmProvider,
-    isLoadingProviders:
-      isLoadingProviders || isLoadingAdapters || isLoadingAgents,
+    isLoadingProviders: isLoadingProviders || isLoadingAdapters,
   }
 }

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

@@ -40,7 +40,7 @@ describe('buildSidepanelPreparedSendMessagesRequest', () => {
     })
   })
 
-  it('sends created-agent targets to the agent-id sidepanel route', () => {
+  it('sends ACP targets to the sidepanel ACP route with explicit target fields', () => {
     const request = buildSidepanelPreparedSendMessagesRequest({
       agentServerUrl: 'http://127.0.0.1:5151',
       target: acpTarget,
@@ -52,11 +52,12 @@ describe('buildSidepanelPreparedSendMessagesRequest', () => {
       ...commonRequestInput(),
     })
 
-    expect(request.api).toBe(
-      'http://127.0.0.1:5151/agents/agent-codex/sidepanel/chat',
-    )
+    expect(request.api).toBe('http://127.0.0.1:5151/agents/sidepanel/chat')
     expect(request.body).toEqual({
       conversationId,
+      adapter: 'codex',
+      modelId: 'gpt-5.5',
+      reasoningEffort: 'medium',
       message: 'Inspect the current tab',
       browserContext: {
         activeTab: { id: 10, url: 'https://example.com', title: 'Example' },
@@ -139,10 +140,9 @@ const llmTarget: SidepanelChatTarget = {
 
 const acpTarget: SidepanelChatTarget = {
   kind: 'acp',
-  id: 'agent-codex',
-  name: 'Review bot',
+  id: 'acp:codex:gpt-5.5:medium',
+  name: 'Codex GPT-5.5',
   type: 'acp',
-  agentId: 'agent-codex',
   adapter: 'codex',
   adapterName: 'Codex',
   modelId: 'gpt-5.5',

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

@@ -680,20 +680,13 @@ export const useChatSession = (options?: ChatSessionOptions) => {
   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,
+      provider_type: target?.kind === 'acp' ? 'acp' : llmTargetProvider?.type,
       model:
-        agentTarget?.modelId ??
-        llmTargetProvider?.modelId ??
-        selectedLlmProvider?.modelId,
+        target?.kind === 'acp'
+          ? target.modelId
+          : llmTargetProvider?.modelId || selectedLlmProvider?.modelId,
     })
 
     if (!isIntegrationsSyncedRef.current) {
@@ -770,8 +763,6 @@ export const useChatSession = (options?: ChatSessionOptions) => {
       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,
     })
 
     void selectChatTarget(target).catch((error) => {

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

@@ -34,10 +34,15 @@ export function buildSidepanelPreparedSendMessagesRequest({
   ...common
 }: BuildSidepanelPreparedSendMessagesRequestInput) {
   if (target?.kind === 'acp') {
+    // ACP session history is owned by AcpxRuntime through sessionKey, so LLM-only
+    // resume and approval fields are intentionally not forwarded.
     return {
-      api: `${agentServerUrl}/agents/${encodeURIComponent(target.agentId)}/sidepanel/chat`,
+      api: `${agentServerUrl}/agents/sidepanel/chat`,
       body: {
         conversationId: common.conversationId,
+        adapter: target.adapter,
+        modelId: target.modelId,
+        reasoningEffort: target.reasoningEffort,
         message: message ?? '',
         browserContext: common.browserContext,
         userSystemPrompt: common.userSystemPrompt,
@@ -66,9 +71,6 @@ export function toProviderOption(target: SidepanelChatTarget): Provider {
     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/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/cli/README.md

@@ -38,8 +38,8 @@ browseros-cli install                # downloads BrowserOS for your platform
 # If BrowserOS is installed but not running
 browseros-cli launch                 # opens BrowserOS, waits for server
 
-# Configure the CLI with the Server URL from BrowserOS settings
-browseros-cli init http://127.0.0.1:9000/mcp
+# Configure the CLI (auto-discovers running BrowserOS)
+browseros-cli init --auto            # detects server URL and saves config
 
 # Verify connection
 browseros-cli health
@@ -52,7 +52,7 @@ browseros-cli init <url>             # non-interactive — pass URL directly
 browseros-cli init                   # interactive — prompts for URL
 ```
 
-Config is saved to `~/.config/browseros-cli/config.yaml`. If `browseros-cli health` cannot connect, copy the current Server URL from BrowserOS Settings > BrowserOS MCP and run `browseros-cli init <Server URL>` again.
+Config is saved to `~/.config/browseros-cli/config.yaml`. The CLI also auto-discovers the server from `~/.browseros/server.json` (written by BrowserOS on startup).
 
 ### CLI updates
 
@@ -126,9 +126,9 @@ To connect Claude Code, Gemini CLI, or any MCP client, see the [MCP setup guide]
 | `--debug` | `BOS_DEBUG=1` | Debug output |
 | `--timeout, -t` | | Request timeout (default: 2m) |
 
-Priority for server URL: `--server` flag > `BROWSEROS_URL` env > config file
+Priority for server URL: `--server` flag > `BROWSEROS_URL` env > `~/.browseros/server.json` > config file
 
-If no server URL is configured, the CLI exits with setup instructions pointing to `install`, `launch`, and `init <Server URL>`.
+If no server URL is configured, the CLI exits with setup instructions pointing to `install`, `launch`, and `init`.
 
 ## Testing
 
@@ -179,7 +179,7 @@ apps/cli/
 │   └── config.go       # Config file (~/.config/browseros-cli/config.yaml)
 ├── cmd/
 │   ├── root.go         # Root command, global flags
-│   ├── init.go         # Server URL configuration (URL arg or interactive)
+│   ├── init.go         # Server URL configuration (URL arg, --auto, interactive)
 │   ├── install.go      # install (download BrowserOS for current platform)
 │   ├── launch.go       # launch (find and start BrowserOS, wait for server)
 │   ├── open.go         # open (new_page / new_hidden_page)

packages/browseros-agent/apps/cli/cmd/init.go

@@ -17,6 +17,8 @@ import (
 )
 
 func init() {
+	var autoDiscover bool
+
 	cmd := &cobra.Command{
 		Use:   "init [url]",
 		Short: "Configure the BrowserOS server connection",
@@ -32,8 +34,9 @@ You can provide the full URL or just the port number:
   browseros-cli init http://127.0.0.1:9000/mcp
   browseros-cli init 9000
 
-Modes:
+Three modes:
   browseros-cli init <url>    Non-interactive (full URL or port number)
+  browseros-cli init --auto   Auto-discover from ~/.browseros/server.json
   browseros-cli init          Interactive prompt`,
 		Annotations: map[string]string{"group": "Setup:"},
 		Args:        cobra.MaximumNArgs(1),
@@ -46,9 +49,22 @@ Modes:
 
 			switch {
 			case len(args) == 1:
+				// Non-interactive: URL provided as argument
 				input = args[0]
 
+			case autoDiscover:
+				// Auto-discover: server.json → config → probe common ports
+				discovered := probeRunningServer()
+				if discovered == "" {
+					output.Error("auto-discovery failed: no running BrowserOS found.\n\n"+
+						"  If not running:    browseros-cli launch\n"+
+						"  If not installed:  browseros-cli install", 1)
+				}
+				input = discovered
+				fmt.Printf("Auto-discovered server at %s\n", input)
+
 			default:
+				// Interactive prompt (original behavior)
 				fmt.Println()
 				bold.Println("BrowserOS CLI Setup")
 				fmt.Println()
@@ -79,14 +95,12 @@ Modes:
 				output.Errorf(1, "invalid URL: %s", input)
 			}
 
+			// Verify connectivity
 			fmt.Printf("Checking connection to %s ...\n", baseURL)
 			client := &http.Client{Timeout: 5 * time.Second}
 			resp, err := client.Get(baseURL + "/health")
 			if err != nil {
-				output.Errorf(1, "cannot connect to %s: %v\n\n"+
-					"Open BrowserOS Settings > BrowserOS MCP and copy the Server URL.\n"+
-					"Then run: browseros-cli init <Server URL>\n"+
-					"Example:  browseros-cli init http://127.0.0.1:9000/mcp", baseURL, err)
+				output.Errorf(1, "cannot connect to %s: %v\nIs BrowserOS running?", baseURL, err)
 			}
 			resp.Body.Close()
 
@@ -107,5 +121,6 @@ Modes:
 		},
 	}
 
+	cmd.Flags().BoolVar(&autoDiscover, "auto", false, "Auto-discover server URL from ~/.browseros/server.json")
 	rootCmd.AddCommand(cmd)
 }

packages/browseros-agent/apps/cli/cmd/install.go

@@ -28,7 +28,7 @@ Linux:   Downloads AppImage (or .deb with --deb flag)
 
 After installation:
   browseros-cli launch        # start BrowserOS
-  browseros-cli init <url>    # configure the CLI with the Server URL`,
+  browseros-cli init --auto   # configure the CLI`,
 		Annotations: map[string]string{"group": "Setup:"},
 		Args:        cobra.NoArgs,
 		Run: func(cmd *cobra.Command, args []string) {
@@ -81,7 +81,7 @@ After installation:
 			fmt.Println()
 			bold.Println("Next steps:")
 			dim.Println("  browseros-cli launch        # start BrowserOS")
-			dim.Println("  browseros-cli init <url>    # use the Server URL from BrowserOS settings")
+			dim.Println("  browseros-cli init --auto   # configure the CLI")
 		},
 	}
 

packages/browseros-agent/apps/cli/cmd/launch.go

@@ -1,7 +1,6 @@
 package cmd
 
 import (
-	"encoding/json"
 	"fmt"
 	"net/http"
 	"os"
@@ -39,7 +38,6 @@ If BrowserOS is already running, reports the server URL.`,
 
 			if url := probeRunningServer(); url != "" {
 				green.Printf("BrowserOS is already running at %s\n", url)
-				dim.Printf("Next: browseros-cli init %s\n", mcpEndpointURL(url))
 				return
 			}
 
@@ -65,7 +63,7 @@ If BrowserOS is already running, reports the server URL.`,
 
 			green.Printf("BrowserOS is ready at %s\n", url)
 			fmt.Println()
-			dim.Printf("Next: browseros-cli init %s\n", mcpEndpointURL(url))
+			dim.Println("Next: browseros-cli init --auto")
 		},
 	}
 
@@ -77,77 +75,39 @@ If BrowserOS is already running, reports the server URL.`,
 // Server probing
 // ---------------------------------------------------------------------------
 
-var commonBrowserOSPorts = []int{9100, 9200, 9300}
-
-// probeRunningServer checks launch discovery, explicit config, and common ports for a running server.
+// probeRunningServer checks server.json, config, and common ports for a running server.
 func probeRunningServer() string {
-	client := &http.Client{Timeout: 2 * time.Second}
+	check := func(baseURL string) bool {
+		client := &http.Client{Timeout: 2 * time.Second}
+		resp, err := client.Get(baseURL + "/health")
+		if err != nil {
+			return false
+		}
+		resp.Body.Close()
+		return resp.StatusCode == 200
+	}
 
-	if url := loadBrowserosServerURL(); url != "" && checkServerHealth(client, url) {
+	// 1. server.json — written by BrowserOS on startup with the actual port
+	if url := loadBrowserosServerURL(); url != "" && check(url) {
 		return url
 	}
 
-	if url := defaultServerURL(); url != "" && checkServerHealth(client, url) {
+	// 2. Saved config / env var
+	if url := defaultServerURL(); url != "" && check(url) {
 		return url
 	}
 
-	return probeCommonServerPorts(client)
-}
-
-func checkServerHealth(client *http.Client, baseURL string) bool {
-	resp, err := client.Get(baseURL + "/health")
-	if err != nil {
-		return false
-	}
-	resp.Body.Close()
-	return resp.StatusCode == 200
-}
-
-func probeCommonServerPorts(client *http.Client) string {
-	for _, port := range commonBrowserOSPorts {
+	// 3. Probe common BrowserOS ports as last resort
+	for _, port := range []int{9100, 9200, 9300} {
 		url := fmt.Sprintf("http://127.0.0.1:%d", port)
-		if checkServerHealth(client, url) {
+		if check(url) {
 			return url
 		}
 	}
+
 	return ""
 }
 
-type serverDiscoveryConfig struct {
-	ServerPort       int    `json:"server_port"`
-	URL              string `json:"url"`
-	ServerVersion    string `json:"server_version"`
-	BrowserOSVersion string `json:"browseros_version,omitempty"`
-	ChromiumVersion  string `json:"chromium_version,omitempty"`
-}
-
-// loadBrowserosServerURL reads BrowserOS's runtime discovery file for launch readiness only.
-//
-// Normal command resolution must not call this because it can override a URL the
-// user explicitly saved with `browseros-cli init <Server URL>`.
-func loadBrowserosServerURL() string {
-	home, err := os.UserHomeDir()
-	if err != nil {
-		return ""
-	}
-
-	data, err := os.ReadFile(filepath.Join(home, ".browseros", "server.json"))
-	if err != nil {
-		return ""
-	}
-
-	var sc serverDiscoveryConfig
-	if err := json.Unmarshal(data, &sc); err != nil {
-		return ""
-	}
-
-	return normalizeServerURL(sc.URL)
-}
-
-func mcpEndpointURL(baseURL string) string {
-	return strings.TrimSuffix(baseURL, "/") + "/mcp"
-}
-
 // ---------------------------------------------------------------------------
 // Platform-native installation detection
 // ---------------------------------------------------------------------------
@@ -157,8 +117,7 @@ func mcpEndpointURL(baseURL string) string {
 // macOS:   `open -Ra "BrowserOS"` — queries Launch Services (finds apps anywhere)
 // Linux:   checks /usr/bin/browseros (.deb), browseros.desktop, or AppImage files
 // Windows: checks executable at %LOCALAPPDATA%\BrowserOS\Application\BrowserOS.exe
-//
-//	and registry uninstall key (per-user Chromium install pattern)
+//          and registry uninstall key (per-user Chromium install pattern)
 func isBrowserOSInstalled() bool {
 	switch runtime.GOOS {
 	case "darwin":
@@ -312,11 +271,14 @@ func waitForServer(maxWait time.Duration) (string, bool) {
 
 	for time.Now().Before(deadline) {
 		// server.json is written by BrowserOS on startup with the actual port
-		if url := loadBrowserosServerURL(); url != "" && checkServerHealth(client, url) {
-			return url, true
-		}
-		if url := probeCommonServerPorts(client); url != "" {
-			return url, true
+		if url := loadBrowserosServerURL(); url != "" {
+			resp, err := client.Get(url + "/health")
+			if err == nil {
+				resp.Body.Close()
+				if resp.StatusCode == 200 {
+					return url, true
+				}
+			}
 		}
 		fmt.Print(".")
 		time.Sleep(1 * time.Second)

packages/browseros-agent/apps/cli/cmd/launch_test.go

@@ -1,99 +0,0 @@
-package cmd
-
-import (
-	"fmt"
-	"net"
-	"net/http"
-	"net/http/httptest"
-	"net/url"
-	"os"
-	"path/filepath"
-	"strconv"
-	"testing"
-	"time"
-
-	"browseros-cli/config"
-)
-
-func TestProbeRunningServerUsesDiscoveryBeforeConfig(t *testing.T) {
-	home := t.TempDir()
-	t.Setenv("HOME", home)
-	t.Setenv("USERPROFILE", home)
-	t.Setenv("XDG_CONFIG_HOME", t.TempDir())
-	t.Setenv("BROWSEROS_URL", "")
-
-	discoveredServer := newHealthyServer(t)
-	configServer := newHealthyServer(t)
-
-	serverDir := filepath.Join(home, ".browseros")
-	if err := os.MkdirAll(serverDir, 0755); err != nil {
-		t.Fatalf("os.MkdirAll() error = %v", err)
-	}
-	data := []byte(fmt.Sprintf(`{"url":%q}`, discoveredServer.URL))
-	if err := os.WriteFile(filepath.Join(serverDir, "server.json"), data, 0644); err != nil {
-		t.Fatalf("os.WriteFile() error = %v", err)
-	}
-	if err := config.Save(&config.Config{ServerURL: configServer.URL}); err != nil {
-		t.Fatalf("config.Save() error = %v", err)
-	}
-
-	got := probeRunningServer()
-	if got != normalizeServerURL(discoveredServer.URL) {
-		t.Fatalf("probeRunningServer() = %q, want %q", got, normalizeServerURL(discoveredServer.URL))
-	}
-}
-
-func TestWaitForServerUsesCommonPortFallback(t *testing.T) {
-	home := t.TempDir()
-	t.Setenv("HOME", home)
-	t.Setenv("USERPROFILE", home)
-
-	server := newHealthyServer(t)
-	port := serverPort(t, server.URL)
-
-	originalPorts := commonBrowserOSPorts
-	commonBrowserOSPorts = []int{port}
-	t.Cleanup(func() {
-		commonBrowserOSPorts = originalPorts
-	})
-
-	got, ok := waitForServer(100 * time.Millisecond)
-	if !ok {
-		t.Fatal("waitForServer() ok = false, want true")
-	}
-	if got != normalizeServerURL(server.URL) {
-		t.Fatalf("waitForServer() = %q, want %q", got, normalizeServerURL(server.URL))
-	}
-}
-
-func newHealthyServer(t *testing.T) *httptest.Server {
-	t.Helper()
-
-	server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
-		if r.URL.Path != "/health" {
-			http.NotFound(w, r)
-			return
-		}
-		w.WriteHeader(http.StatusOK)
-	}))
-	t.Cleanup(server.Close)
-	return server
-}
-
-func serverPort(t *testing.T, rawURL string) int {
-	t.Helper()
-
-	parsed, err := url.Parse(rawURL)
-	if err != nil {
-		t.Fatalf("url.Parse() error = %v", err)
-	}
-	_, portText, err := net.SplitHostPort(parsed.Host)
-	if err != nil {
-		t.Fatalf("net.SplitHostPort() error = %v", err)
-	}
-	port, err := strconv.Atoi(portText)
-	if err != nil {
-		t.Fatalf("strconv.Atoi() error = %v", err)
-	}
-	return port
-}

packages/browseros-agent/apps/cli/cmd/root.go

@@ -2,8 +2,10 @@ package cmd
 
 import (
 	"context"
+	"encoding/json"
 	"fmt"
 	"os"
+	"path/filepath"
 	"strconv"
 	"strings"
 	"time"
@@ -287,15 +289,18 @@ func drainAutomaticUpdateCheckWithTimeout(done <-chan struct{}, timeout time.Dur
 	}
 }
 
-// defaultServerURL returns the implicit target from user-controlled settings only.
-//
-// BrowserOS writes a discovery file at runtime, but normal commands intentionally
-// ignore it so a saved URL is not silently overridden by another running server.
 func defaultServerURL() string {
+	// 1. Explicit env var always wins
 	if env := normalizeServerURL(os.Getenv("BROWSEROS_URL")); env != "" {
 		return env
 	}
 
+	// 2. Live discovery file from running BrowserOS (most current)
+	if url := loadBrowserosServerURL(); url != "" {
+		return url
+	}
+
+	// 3. Saved config (may be stale if port changed)
 	cfg, err := config.Load()
 	if err == nil {
 		if url := normalizeServerURL(cfg.ServerURL); url != "" {
@@ -306,6 +311,33 @@ func defaultServerURL() string {
 	return ""
 }
 
+type serverDiscoveryConfig struct {
+	ServerPort       int    `json:"server_port"`
+	URL              string `json:"url"`
+	ServerVersion    string `json:"server_version"`
+	BrowserOSVersion string `json:"browseros_version,omitempty"`
+	ChromiumVersion  string `json:"chromium_version,omitempty"`
+}
+
+func loadBrowserosServerURL() string {
+	home, err := os.UserHomeDir()
+	if err != nil {
+		return ""
+	}
+
+	data, err := os.ReadFile(filepath.Join(home, ".browseros", "server.json"))
+	if err != nil {
+		return ""
+	}
+
+	var sc serverDiscoveryConfig
+	if err := json.Unmarshal(data, &sc); err != nil {
+		return ""
+	}
+
+	return normalizeServerURL(sc.URL)
+}
+
 func normalizeServerURL(raw string) string {
 	normalized := strings.TrimSpace(raw)
 
@@ -337,10 +369,8 @@ func validateServerURL(raw string) (string, error) {
 
 	return "", fmt.Errorf(
 		"BrowserOS server URL is not configured.\n\n" +
-			"  Open BrowserOS Settings > BrowserOS MCP and copy the Server URL.\n" +
-			"  Save it with:       browseros-cli init <Server URL>\n" +
-			"  Example:            browseros-cli init http://127.0.0.1:9000/mcp\n" +
-			"  If BrowserOS is closed:  browseros-cli launch\n" +
-			"  If not installed:        browseros-cli install",
+			"  If BrowserOS is running:  browseros-cli init --auto\n" +
+			"  If BrowserOS is closed:   browseros-cli launch\n" +
+			"  If not installed:         browseros-cli install",
 	)
 }

packages/browseros-agent/apps/cli/cmd/root_test.go

@@ -1,13 +1,8 @@
 package cmd
 
 import (
-	"os"
-	"path/filepath"
-	"strings"
 	"testing"
 	"time"
-
-	"browseros-cli/config"
 )
 
 func TestSetVersionUpdatesRootCommand(t *testing.T) {
@@ -105,76 +100,6 @@ func TestShouldSkipAutomaticUpdates(t *testing.T) {
 	}
 }
 
-func TestDefaultServerURLUsesEnvBeforeConfig(t *testing.T) {
-	t.Setenv("XDG_CONFIG_HOME", t.TempDir())
-	t.Setenv("BROWSEROS_URL", "http://127.0.0.1:9115/mcp")
-
-	if err := config.Save(&config.Config{ServerURL: "http://127.0.0.1:9000/mcp"}); err != nil {
-		t.Fatalf("config.Save() error = %v", err)
-	}
-
-	got := defaultServerURL()
-	if got != "http://127.0.0.1:9115" {
-		t.Fatalf("defaultServerURL() = %q, want %q", got, "http://127.0.0.1:9115")
-	}
-}
-
-func TestDefaultServerURLUsesSavedConfig(t *testing.T) {
-	t.Setenv("XDG_CONFIG_HOME", t.TempDir())
-	t.Setenv("BROWSEROS_URL", "")
-
-	if err := config.Save(&config.Config{ServerURL: "http://127.0.0.1:9115/mcp"}); err != nil {
-		t.Fatalf("config.Save() error = %v", err)
-	}
-
-	got := defaultServerURL()
-	if got != "http://127.0.0.1:9115" {
-		t.Fatalf("defaultServerURL() = %q, want %q", got, "http://127.0.0.1:9115")
-	}
-}
-
-func TestDefaultServerURLIgnoresBrowserOSServerJSON(t *testing.T) {
-	home := t.TempDir()
-	t.Setenv("HOME", home)
-	t.Setenv("USERPROFILE", home)
-	t.Setenv("XDG_CONFIG_HOME", t.TempDir())
-	t.Setenv("BROWSEROS_URL", "")
-
-	serverDir := filepath.Join(home, ".browseros")
-	if err := os.MkdirAll(serverDir, 0755); err != nil {
-		t.Fatalf("os.MkdirAll() error = %v", err)
-	}
-	data := []byte(`{"url":"http://127.0.0.1:9999"}`)
-	if err := os.WriteFile(filepath.Join(serverDir, "server.json"), data, 0644); err != nil {
-		t.Fatalf("os.WriteFile() error = %v", err)
-	}
-
-	if got := defaultServerURL(); got != "" {
-		t.Fatalf("defaultServerURL() = %q, want empty", got)
-	}
-}
-
-func TestNormalizeServerURLAcceptsMCPEndpoint(t *testing.T) {
-	got := normalizeServerURL(" http://127.0.0.1:9115/mcp ")
-	if got != "http://127.0.0.1:9115" {
-		t.Fatalf("normalizeServerURL() = %q, want %q", got, "http://127.0.0.1:9115")
-	}
-}
-
-func TestValidateServerURLExplainsManualInit(t *testing.T) {
-	_, err := validateServerURL("")
-	if err == nil {
-		t.Fatal("validateServerURL() error = nil, want setup instructions")
-	}
-	msg := err.Error()
-	if !strings.Contains(msg, "browseros-cli init <Server URL>") {
-		t.Fatalf("validateServerURL() error = %q, want manual init instructions", msg)
-	}
-	if strings.Contains(msg, "init --auto") {
-		t.Fatalf("validateServerURL() error = %q, should not mention init --auto", msg)
-	}
-}
-
 func TestDrainAutomaticUpdateCheckWithTimeoutWaitsForCompletion(t *testing.T) {
 	done := make(chan struct{})
 	returned := make(chan struct{})

packages/browseros-agent/apps/cli/mcp/client.go

@@ -44,7 +44,10 @@ func (c *Client) connect(ctx context.Context) (*sdkmcp.ClientSession, error) {
 
 	session, err := sdkClient.Connect(ctx, transport, nil)
 	if err != nil {
-		return nil, fmt.Errorf("cannot connect to BrowserOS at %s: %w%s", c.BaseURL, err, connectionSetupInstructions())
+		return nil, fmt.Errorf("cannot connect to BrowserOS at %s: %w\n\n"+
+			"  If BrowserOS is running on a different port:  browseros-cli init --auto\n"+
+			"  If BrowserOS is not running:                  browseros-cli launch\n"+
+			"  If not installed:                             browseros-cli install", c.BaseURL, err)
 	}
 	return session, nil
 }
@@ -184,7 +187,10 @@ func (c *Client) Status() (map[string]any, error) {
 func (c *Client) restGET(path string) (map[string]any, error) {
 	resp, err := c.HTTPClient.Get(c.BaseURL + path)
 	if err != nil {
-		return nil, fmt.Errorf("cannot connect to BrowserOS at %s: %w%s", c.BaseURL, err, connectionSetupInstructions())
+		return nil, fmt.Errorf("cannot connect to BrowserOS at %s: %w\n\n"+
+			"  If BrowserOS is running on a different port:  browseros-cli init --auto\n"+
+			"  If BrowserOS is not running:                  browseros-cli launch\n"+
+			"  If not installed:                             browseros-cli install", c.BaseURL, err)
 	}
 	defer resp.Body.Close()
 
@@ -199,14 +205,3 @@ func (c *Client) restGET(path string) (map[string]any, error) {
 	}
 	return data, nil
 }
-
-// connectionSetupInstructions explains how to recover from a stale or missing server URL.
-func connectionSetupInstructions() string {
-	return "\n\n" +
-		"  Open BrowserOS Settings > BrowserOS MCP and copy the Server URL.\n" +
-		"  Save it with:       browseros-cli init <Server URL>\n" +
-		"  Example:            browseros-cli init http://127.0.0.1:9000/mcp\n" +
-		"  Run once with:      browseros-cli --server <Server URL> health\n" +
-		"  If BrowserOS is closed:  browseros-cli launch\n" +
-		"  If not installed:        browseros-cli install"
-}

packages/browseros-agent/apps/cli/npm/README.md

@@ -31,8 +31,8 @@ browseros-cli install
 # Start BrowserOS
 browseros-cli launch
 
-# Configure MCP settings with the Server URL from BrowserOS settings
-browseros-cli init http://127.0.0.1:9000/mcp
+# Auto-configure MCP settings for your AI tools
+browseros-cli init --auto
 
 # Verify everything is working
 browseros-cli health

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

@@ -9,13 +9,11 @@ Evaluation framework for BrowserOS browser automation agents. Runs tasks from st
 - **BrowserOS binary** at `/Applications/BrowserOS.app` (macOS) or `BROWSEROS_BINARY` pointing at it
 - **Bun** runtime
 - **API keys** for your LLM provider (and `CLAUDE_CODE_OAUTH_TOKEN` if you use `performance_grader`)
-- **Python 3.10+ with `agisdk`** for AGI SDK / REAL Bench grading. Set `BROWSEROS_EVAL_PYTHON` if your default `python3` is older.
 
 ## Quick Start
 
 ```bash
 cd apps/eval
-cp .env.example .env.development
 # Edit .env.development with your keys, then:
 bun run eval
 ```
@@ -25,62 +23,17 @@ 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-opus
-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 |
 |------|-------------|
 | `single` | Single LLM agent driven by the BrowserOS tool loop (CDP) |
 | `orchestrator-executor` | High-level orchestrator + per-step executor (LLM or Clado visual model) |
-| `claude-code` | External Claude Code CLI driven through BrowserOS MCP |
 
 ### Single agent
 
@@ -121,24 +74,6 @@ The orchestrator works with any LLM provider. The executor can be another LLM, o
 }
 ```
 
-### Claude Code
-
-Claude Code runs as an external `claude -p` subprocess. The eval runner passes a task-scoped MCP config that points Claude Code at the active worker's BrowserOS MCP endpoint, while the eval capture layer still saves messages, screenshots, trajectory metadata, and grader outputs.
-
-```json
-{
-  "agent": {
-    "type": "claude-code",
-    "model": "opus"
-  }
-}
-```
-
-```bash
-BROWSEROS_EVAL_PYTHON=/path/to/python3 bun run eval run --config configs/legacy/claude-code-agisdk-real.json
-bun run eval suite --config configs/legacy/claude-code-agisdk-real.json --publish r2
-```
-
 ## Graders
 
 | Name | Description |
@@ -161,21 +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 |
-| `BROWSEROS_EVAL_PYTHON` | Optional Python interpreter for JSON graders such as `agisdk_state_diff` |
-| `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` |
@@ -190,22 +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
@@ -215,7 +119,7 @@ Published runs are available at `EVAL_R2_CDN_BASE_URL/viewer.html?run=<run-id>`.
   "base_server_port": 9110,
   "base_extension_port": 9310,
   "load_extensions": false,
-  "headless": false
+  "headless": true
 }
 ```
 
@@ -233,12 +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 |
@@ -265,47 +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

@@ -7,7 +7,7 @@
     "baseUrl": "https://api.fireworks.ai/inference/v1",
     "supportsImages": true
   },
-  "dataset": "../../data/agisdk-real.jsonl",
+  "dataset": "../data/agisdk-real.jsonl",
   "num_workers": 4,
   "restart_server_per_task": true,
   "browseros": {

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

@@ -7,8 +7,8 @@
     "baseUrl": "https://openrouter.ai/api/v1",
     "supportsImages": true
   },
-  "dataset": "../../data/agisdk-real.jsonl",
-  "num_workers": 3,
+  "dataset": "../data/webbench-2of4-50.jsonl",
+  "num_workers": 10,
   "restart_server_per_task": true,
   "browseros": {
     "server_url": "http://127.0.0.1:9110",
@@ -21,6 +21,6 @@
   "captcha": {
     "api_key_env": "NOPECHA_API_KEY"
   },
-  "graders": ["agisdk_state_diff"],
+  "graders": ["performance_grader"],
   "timeout_ms": 1800000
 }

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

@@ -14,7 +14,7 @@
       "baseUrl": "https://clado-ai--clado-browseros-action-000159-merged-actionmod-f4a6ef.modal.run"
     }
   },
-  "dataset": "../../data/agisdk-real.jsonl",
+  "dataset": "../data/agisdk-real.jsonl",
   "num_workers": 10,
   "restart_server_per_task": true,
   "browseros": {
@@ -23,7 +23,7 @@
     "base_server_port": 9110,
     "base_extension_port": 9310,
     "load_extensions": false,
-    "headless": false
+    "headless": true
   },
   "captcha": {
     "api_key_env": "NOPECHA_API_KEY"

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-smoke.json

@@ -1,26 +0,0 @@
-{
-  "agent": {
-    "type": "single",
-    "provider": "openai-compatible",
-    "model": "moonshotai/kimi-k2.5",
-    "apiKey": "OPENROUTER_API_KEY",
-    "baseUrl": "https://openrouter.ai/api/v1",
-    "supportsImages": true
-  },
-  "dataset": "../../data/agisdk-real-smoke.jsonl",
-  "num_workers": 1,
-  "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/legacy/browseros-agent-kimi-k2-5-agisdk-real.json

@@ -1,26 +0,0 @@
-{
-  "agent": {
-    "type": "single",
-    "provider": "openai-compatible",
-    "model": "moonshotai/kimi-k2.5",
-    "apiKey": "OPENROUTER_API_KEY",
-    "baseUrl": "https://openrouter.ai/api/v1",
-    "supportsImages": true
-  },
-  "dataset": "../../data/agisdk-real.jsonl",
-  "num_workers": 3,
-  "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/legacy/browseros-agent-opus-4-6-agisdk-real.json

@@ -1,27 +0,0 @@
-{
-  "agent": {
-    "type": "single",
-    "provider": "bedrock",
-    "model": "global.anthropic.claude-opus-4-6-v1",
-    "region": "AWS_REGION",
-    "accessKeyId": "AWS_ACCESS_KEY_ID",
-    "secretAccessKey": "AWS_SECRET_ACCESS_KEY",
-    "supportsImages": true
-  },
-  "dataset": "../../data/agisdk-real.jsonl",
-  "num_workers": 2,
-  "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/legacy/claude-code-agisdk-real.json

@@ -1,23 +0,0 @@
-{
-  "agent": {
-    "type": "claude-code",
-    "model": "opus",
-    "extraArgs": ["--permission-mode", "bypassPermissions"]
-  },
-  "dataset": "../../data/agisdk-real.jsonl",
-  "num_workers": 1,
-  "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": false
-  },
-  "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/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/agisdk-evaluate.py

@@ -81,13 +81,30 @@ def main():
 
         reward_val = float(reward_val) if reward_val is not None else 0.0
         results = info.get("results", [])
+        # `info["results"]` aligns 1:1 with `tc.task.evals` — zip them so we can
+        # surface the human-readable description and JMESPath query alongside
+        # the pass/fail. Without this the only feedback was a stringified dict.
+        evals = list(getattr(tc.task, "evals", []))
 
         per_criterion = []
         softened_count = 0
-        for r in results:
+        for idx, r in enumerate(results):
             passed = bool(r[0])
-            detail = r[1] if len(r) > 1 else ""
-            entry: dict = {"passed": passed, "detail": str(detail)}
+            detail = r[1] if len(r) > 1 else {}
+            ev = evals[idx] if idx < len(evals) else None
+
+            actual_value = expected_value = None
+            if isinstance(detail, dict):
+                actual_value = detail.get("actual_value")
+                expected_value = detail.get("expected_value")
+
+            entry: dict = {
+                "passed": passed,
+                "description": getattr(ev, "description", "") or "",
+                "query": getattr(ev, "query", "") or "",
+                "expected_value": expected_value,
+                "actual_value": actual_value,
+            }
             if not _STRICT and not passed and _soft_string_match(detail):
                 entry["passed"] = True
                 entry["softened"] = True
@@ -100,9 +117,43 @@ def main():
         if all_pass and reward_val != 1.0:
             reward_val = 1.0
 
-        out_message = str(message)
-        if softened_count and all_pass:
-            out_message = f"Task passed (with {softened_count} softened string criterion/criteria)."
+        # Build a useful message: list every criterion with a pass/fail icon
+        # so the viewer's grader pill shows the full check-list, not just
+        # failures. This becomes the `reasoning` shown in the viewer.
+        if not per_criterion:
+            # Defensive: agisdk returned no criteria — fall back to its message.
+            out_message = str(message)
+        else:
+            failures = [c for c in per_criterion if not c["passed"]]
+            if all_pass:
+                header = (
+                    f"All {len(per_criterion)} criteria passed"
+                    + (
+                        f" ({softened_count} softened)."
+                        if softened_count
+                        else "."
+                    )
+                )
+            else:
+                header = (
+                    f"{len(failures)} of {len(per_criterion)} criteria failed:"
+                )
+
+            lines = []
+            for c in per_criterion:
+                icon = "✓" if c["passed"] else "✗"
+                desc = c["description"] or c["query"] or "<unknown>"
+                soft = " (softened)" if c.get("softened") else ""
+                if c["passed"]:
+                    lines.append(f"{icon} {desc}{soft}")
+                else:
+                    exp_s = repr(c["expected_value"])
+                    act_s = repr(c["actual_value"])
+                    lines.append(
+                        f"{icon} {desc}: expected {exp_s}, got {act_s}"
+                    )
+
+            out_message = header + "\n" + "\n".join(lines)
 
         print(
             json.dumps(

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

@@ -1,191 +0,0 @@
-#!/usr/bin/env bun
-
-import { mkdir, stat } from 'node:fs/promises'
-import { dirname, resolve } from 'node:path'
-import { query as claudeQuery } from '@anthropic-ai/claude-agent-sdk'
-import { readRunMetricSummary } from '../src/reporting/task-metrics'
-
-export const DEFAULT_REPORT_MODEL = 'claude-opus-4-6'
-export const DEFAULT_REPORT_MAX_TURNS = 300
-
-type Env = Record<string, string | undefined>
-type ClaudeQuery = (input: unknown) => AsyncIterable<Record<string, unknown>>
-
-export interface ReportAgentInvocation {
-  inputDir: string
-  outputPath: string
-  prompt: string
-}
-
-export interface GenerateEvalReportOptions {
-  inputDir: string
-  outputPath: string
-  runAgent?: (invocation: ReportAgentInvocation) => Promise<void>
-}
-
-interface ClaudeReportAgentDeps {
-  query?: ClaudeQuery
-  env?: Env
-}
-
-function usage(): string {
-  return `Usage: bun scripts/generate-report.ts --input <run-dir> --output <report.html>`
-}
-
-function parseArgs(
-  argv: string[],
-): Pick<GenerateEvalReportOptions, 'inputDir' | 'outputPath'> {
-  let inputDir = ''
-  let outputPath = ''
-  for (let i = 0; i < argv.length; i++) {
-    const arg = argv[i]
-    if (arg === '--input' || arg === '--run') {
-      inputDir = argv[++i] ?? ''
-    } else if (arg === '--output' || arg === '--out') {
-      outputPath = argv[++i] ?? ''
-    } else if (arg === '--help' || arg === '-h') {
-      console.log(usage())
-      process.exit(0)
-    }
-  }
-  if (!inputDir || !outputPath) {
-    throw new Error(usage())
-  }
-  return { inputDir, outputPath }
-}
-
-function claudeCodeEnv(env: Env): Env {
-  return {
-    CLAUDE_CODE_OAUTH_TOKEN: env.CLAUDE_CODE_OAUTH_TOKEN,
-    ANTHROPIC_API_KEY: env.ANTHROPIC_API_KEY,
-    HOME: env.HOME,
-    PATH: env.PATH,
-    SHELL: env.SHELL,
-    TMPDIR: env.TMPDIR,
-    TMP: env.TMP,
-    TEMP: env.TEMP,
-    USER: env.USER,
-    CLAUDECODE: '',
-  }
-}
-
-async function buildReportPrompt(
-  inputDir: string,
-  outputPath: string,
-): Promise<string> {
-  const metrics = await readRunMetricSummary(inputDir)
-
-  return `Analyze this BrowserOS eval run and write a shareable HTML report.
-
-Run directory: ${inputDir}
-Output file to write: ${outputPath}
-
-You are running with the run directory as cwd. Inspect the local artifacts:
-- summary.json for run totals and pass rate
-- each task directory's metadata.json for query, final answer, timing, screenshots, and grader results
-- each task directory's messages.jsonl for tool calls, tool errors, and recent trajectory
-- screenshots/ for visual evidence
-- grader-artifacts/ when present for grader-specific context
-
-Write the final report directly to the output file path above. Do not print the
-report instead of writing it. Do not modify any input artifacts. The only file
-you should create or overwrite is the requested report.html.
-
-The report should follow the style and density of the Shadowfax AGI SDK report:
-- Title like "AGI SDK Random-10 Failure Report" or a run-specific equivalent
-- Run directory and note that screenshots are embedded as data URIs
-- Summary cards for total tasks, passed, failed, pass rate, average duration, average steps, and average tool calls
-- A Metrics section with compact charts for Duration by task, Steps by task, Tool calls by task, and Tool errors by task
-- Task Summary table with task id, status, score, duration, steps, and prompt
-- Include tool calls and tool errors in the Task Summary table
-- Failure sections with stable anchors using each task id, for example <section id="agisdk-networkin-10">
-- For each failed task: Diagnosis, Evidence, Next Check, final screenshot, AGI SDK / grader criteria, final answer, and recent trajectory events
-- Make failure links in the summary table point to the task anchors
-- Keep the HTML self-contained: inline CSS and embedded final screenshots as data:image/png;base64 URIs
-- Escape user/model text correctly so task outputs cannot break the page
-
-Analysis guidance:
-- Focus on why the model failed: task understanding, browser/tool usage, missing verification, tool errors, max-step/timeout, bad final answer, or grader ambiguity
-- Use messages.jsonl strategically. Do not paste huge DOM outputs into the report. Summarize only the relevant recent trajectory and evidence.
-- Limit trajectory analysis to the most relevant 200-300 events/calls across the run. Prefer failed tasks and the final/key actions for each failure.
-- If a grader criterion is boolean-only or ambiguous, say so and identify what additional artifact would make it debuggable.
-
-Deterministic run metrics computed from metadata.json and messages.jsonl:
-\`\`\`json
-${JSON.stringify(metrics, null, 2)}
-\`\`\`
-
-After writing the file, verify that ${outputPath} exists and is non-empty.`
-}
-
-async function assertRunDir(inputDir: string): Promise<void> {
-  const inputStat = await stat(inputDir).catch(() => null)
-  if (!inputStat?.isDirectory()) {
-    throw new Error(`Not a run directory: ${inputDir}`)
-  }
-}
-
-async function assertReportWritten(outputPath: string): Promise<void> {
-  const outputStat = await stat(outputPath).catch(() => null)
-  if (!outputStat?.isFile() || outputStat.size === 0) {
-    throw new Error(`Report was not written: ${outputPath}`)
-  }
-}
-
-export async function runClaudeCodeReportAgent(
-  invocation: ReportAgentInvocation,
-  deps: ClaudeReportAgentDeps = {},
-): Promise<void> {
-  const query = deps.query ?? (claudeQuery as unknown as ClaudeQuery)
-  let resultSubtype: string | undefined
-
-  for await (const message of query({
-    prompt: invocation.prompt,
-    options: {
-      cwd: invocation.inputDir,
-      model: DEFAULT_REPORT_MODEL,
-      systemPrompt:
-        'You are an eval failure analyst. Produce a concise, evidence-backed, self-contained HTML report from local run artifacts.',
-      permissionMode: 'bypassPermissions',
-      allowDangerouslySkipPermissions: true,
-      maxTurns: DEFAULT_REPORT_MAX_TURNS,
-      env: claudeCodeEnv(deps.env ?? process.env),
-    },
-  })) {
-    if (message.type === 'result') {
-      resultSubtype =
-        typeof message.subtype === 'string' ? message.subtype : undefined
-    }
-  }
-
-  if (resultSubtype && resultSubtype !== 'success') {
-    throw new Error(`Claude Code report agent failed: ${resultSubtype}`)
-  }
-}
-
-export async function generateEvalReport(
-  options: GenerateEvalReportOptions,
-): Promise<void> {
-  const inputDir = resolve(options.inputDir)
-  const outputPath = resolve(options.outputPath)
-
-  await assertRunDir(inputDir)
-  await mkdir(dirname(outputPath), { recursive: true })
-
-  const invocation = {
-    inputDir,
-    outputPath,
-    prompt: await buildReportPrompt(inputDir, outputPath),
-  }
-  await (options.runAgent ?? runClaudeCodeReportAgent)(invocation)
-  await assertReportWritten(outputPath)
-}
-
-if (import.meta.main) {
-  try {
-    await generateEvalReport(parseArgs(Bun.argv.slice(2)))
-  } catch (error) {
-    console.error(error instanceof Error ? error.message : String(error))
-    process.exit(1)
-  }
-}

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/claude-code/index.ts

@@ -1,238 +0,0 @@
-import { writeFile } from 'node:fs/promises'
-import { join } from 'node:path'
-import { DEFAULT_TIMEOUT_MS } from '../../constants'
-import type { ClaudeCodeAgentConfig, UIMessageStreamEvent } from '../../types'
-import { withEvalTimeout } from '../../utils/with-eval-timeout'
-import type { AgentContext, AgentEvaluator, AgentResult } from '../types'
-import {
-  type ClaudeCodeProcessRunner,
-  createClaudeCodeProcessRunner,
-} from './process-runner'
-import {
-  ClaudeCodeStreamParser,
-  shouldCaptureScreenshotForTool,
-} from './stream-parser'
-
-export interface ClaudeCodeEvaluatorDeps {
-  processRunner?: ClaudeCodeProcessRunner
-}
-
-export class ClaudeCodeEvaluator implements AgentEvaluator {
-  private processRunner: ClaudeCodeProcessRunner
-
-  constructor(
-    private ctx: AgentContext,
-    deps: ClaudeCodeEvaluatorDeps = {},
-  ) {
-    this.processRunner = deps.processRunner ?? createClaudeCodeProcessRunner()
-  }
-
-  async execute(): Promise<AgentResult> {
-    const { config, task, capture, taskOutputDir } = this.ctx
-    const startTime = Date.now()
-    const timeoutMs = config.timeout_ms ?? DEFAULT_TIMEOUT_MS
-
-    await capture.messageLogger.logUser(task.query)
-
-    if (config.agent.type !== 'claude-code') {
-      throw new Error('ClaudeCodeEvaluator only supports claude-code config')
-    }
-    const agentConfig = config.agent
-
-    const mcpConfigPath = join(taskOutputDir, 'claude-code-mcp.json')
-    await writeFile(
-      mcpConfigPath,
-      JSON.stringify(
-        buildClaudeCodeMcpConfig(config.browseros.server_url),
-        null,
-        2,
-      ),
-    )
-
-    const parser = new ClaudeCodeStreamParser()
-    const toolNamesById = new Map<string, string>()
-    const prompt = buildClaudeCodePrompt(task.query)
-    const args = buildClaudeCodeArgs({
-      prompt,
-      mcpConfigPath,
-      config: agentConfig,
-    })
-
-    const { terminationReason } = await withEvalTimeout(
-      timeoutMs,
-      capture,
-      async (signal) => {
-        const runResult = await this.processRunner.run({
-          executable: agentConfig.claudePath,
-          args,
-          cwd: taskOutputDir,
-          signal,
-          onStdoutLine: async (line) => {
-            const events = parser.pushLine(line)
-            for (const event of events) {
-              await this.handleStreamEvent(event, toolNamesById)
-            }
-          },
-        })
-
-        if (runResult.exitCode !== 0) {
-          const message =
-            runResult.stderr.trim() ||
-            `Claude Code exited with status ${runResult.exitCode}`
-          capture.addError('agent_execution', message, {
-            exitCode: runResult.exitCode,
-          })
-          if (!parser.getLastText()) {
-            throw new Error(message)
-          }
-        }
-
-        for (const error of runResult.streamErrors ?? []) {
-          capture.addWarning(
-            'message_logging',
-            `Claude Code stream event processing failed: ${error}`,
-          )
-        }
-
-        return runResult
-      },
-    )
-
-    const endTime = Date.now()
-    const finalAnswer = parser.getLastText() ?? capture.getLastAssistantText()
-    const metadata = {
-      query_id: task.query_id,
-      dataset: task.dataset,
-      query: task.query,
-      started_at: new Date(startTime).toISOString(),
-      completed_at: new Date(endTime).toISOString(),
-      total_duration_ms: endTime - startTime,
-      total_steps: parser.getToolCallCount() || capture.getScreenshotCount(),
-      termination_reason: terminationReason,
-      final_answer: finalAnswer,
-      errors: capture.getErrors(),
-      warnings: capture.getWarnings(),
-      device_pixel_ratio: capture.screenshot.getDevicePixelRatio(),
-      agent_config: {
-        type: 'claude-code' as const,
-        model: agentConfig.model,
-      },
-      grader_results: {},
-    }
-
-    await capture.trajectorySaver.saveMetadata(metadata)
-
-    return {
-      metadata,
-      messages: capture.getMessages(),
-      finalAnswer,
-    }
-  }
-
-  private async handleStreamEvent(
-    event: UIMessageStreamEvent,
-    toolNamesById: Map<string, string>,
-  ): Promise<void> {
-    const { capture, task } = this.ctx
-    let screenshot: number | undefined
-
-    if (event.type === 'tool-input-available') {
-      toolNamesById.set(event.toolCallId, event.toolName)
-      if (isPageInput(event.input)) {
-        capture.setActivePageId(event.input.page)
-      }
-    }
-
-    if (
-      event.type === 'tool-output-available' ||
-      event.type === 'tool-output-error'
-    ) {
-      const toolName = toolNamesById.get(event.toolCallId)
-      if (toolName && shouldCaptureScreenshotForTool(toolName)) {
-        screenshot = await this.captureScreenshot()
-      }
-    }
-
-    await capture.messageLogger.logStreamEvent(event, screenshot)
-    capture.emitEvent(task.query_id, {
-      ...event,
-      ...(screenshot !== undefined && { screenshot }),
-    })
-  }
-
-  private async captureScreenshot(): Promise<number | undefined> {
-    const { capture, task } = this.ctx
-    try {
-      const screenshot = await capture.screenshot.capture(
-        capture.getActivePageId(),
-      )
-      capture.emitEvent(task.query_id, {
-        type: 'screenshot-captured',
-        screenshot,
-      })
-      return screenshot
-    } catch {
-      return undefined
-    }
-  }
-}
-
-function isPageInput(input: unknown): input is { page: number } {
-  return (
-    typeof input === 'object' &&
-    input !== null &&
-    'page' in input &&
-    typeof input.page === 'number'
-  )
-}
-
-function buildClaudeCodePrompt(taskQuery: string): string {
-  return [
-    'You are running inside BrowserOS eval.',
-    'Use the BrowserOS MCP tools to interact with the already-open browser and complete the user task.',
-    'When the task is complete, respond with the final answer only.',
-    'If blocked, explain the blocker clearly.',
-    '',
-    `Task: ${taskQuery}`,
-  ].join('\n')
-}
-
-function buildClaudeCodeArgs({
-  prompt,
-  mcpConfigPath,
-  config,
-}: {
-  prompt: string
-  mcpConfigPath: string
-  config: ClaudeCodeAgentConfig
-}): string[] {
-  const args = [
-    '-p',
-    prompt,
-    '--mcp-config',
-    mcpConfigPath,
-    '--strict-mcp-config',
-    '--output-format',
-    'stream-json',
-    '--verbose',
-  ]
-
-  if (config.model) args.push('--model', config.model)
-  args.push(...config.extraArgs)
-
-  return args
-}
-
-function buildClaudeCodeMcpConfig(serverUrl: string) {
-  const trimmed = serverUrl.replace(/\/$/, '')
-  const url = trimmed.endsWith('/mcp') ? trimmed : `${trimmed}/mcp`
-  return {
-    mcpServers: {
-      browseros: {
-        type: 'http',
-        url,
-        headers: { 'X-BrowserOS-Source': 'sdk-internal' },
-      },
-    },
-  }
-}

packages/browseros-agent/apps/eval/src/agents/claude-code/process-runner.ts

@@ -1,114 +0,0 @@
-export interface ClaudeCodeRunOptions {
-  executable: string
-  args: string[]
-  cwd: string
-  signal?: AbortSignal
-  onStdoutLine: (line: string) => Promise<void>
-}
-
-export interface ClaudeCodeRunResult {
-  exitCode: number
-  stderr: string
-  streamErrors?: string[]
-}
-
-export interface ClaudeCodeProcessRunner {
-  run(options: ClaudeCodeRunOptions): Promise<ClaudeCodeRunResult>
-}
-
-export interface SpawnOptions {
-  cwd: string
-  signal?: AbortSignal
-  onStdoutLine: (line: string) => Promise<void>
-}
-
-export interface CreateClaudeCodeProcessRunnerDeps {
-  spawn?: (cmd: string[], options: SpawnOptions) => Promise<ClaudeCodeRunResult>
-}
-
-export function createClaudeCodeProcessRunner(
-  deps: CreateClaudeCodeProcessRunnerDeps = {},
-): ClaudeCodeProcessRunner {
-  const spawn = deps.spawn ?? spawnClaudeCode
-  return {
-    run: async ({ executable, args, cwd, signal, onStdoutLine }) =>
-      spawn([executable, ...args], { cwd, signal, onStdoutLine }),
-  }
-}
-
-async function spawnClaudeCode(
-  cmd: string[],
-  options: SpawnOptions,
-): Promise<ClaudeCodeRunResult> {
-  const proc = Bun.spawn({
-    cmd,
-    cwd: options.cwd,
-    stdin: 'ignore',
-    stdout: 'pipe',
-    stderr: 'pipe',
-  })
-
-  const abort = () => {
-    try {
-      proc.kill('SIGTERM')
-    } catch {
-      // Process may already have exited.
-    }
-  }
-  options.signal?.addEventListener('abort', abort, { once: true })
-
-  try {
-    const streamErrors: string[] = []
-    const stdoutPromise = readLines(
-      proc.stdout,
-      options.onStdoutLine,
-      streamErrors,
-    )
-    const stderrPromise = new Response(proc.stderr).text()
-    const exitCode = await proc.exited
-    await stdoutPromise
-    const stderr = await stderrPromise
-    return { exitCode, stderr, streamErrors }
-  } finally {
-    options.signal?.removeEventListener('abort', abort)
-  }
-}
-
-async function readLines(
-  stream: ReadableStream<Uint8Array>,
-  onLine: (line: string) => Promise<void>,
-  streamErrors: string[],
-): Promise<void> {
-  const reader = stream.getReader()
-  const decoder = new TextDecoder()
-  let buffer = ''
-
-  while (true) {
-    const { done, value } = await reader.read()
-    if (done) break
-
-    buffer += decoder.decode(value, { stream: true })
-    const lines = buffer.split('\n')
-    buffer = lines.pop() ?? ''
-    for (const line of lines) {
-      await emitLine(line, onLine, streamErrors)
-    }
-  }
-
-  buffer += decoder.decode()
-  if (buffer.length > 0) {
-    await emitLine(buffer, onLine, streamErrors)
-  }
-}
-
-async function emitLine(
-  line: string,
-  onLine: (line: string) => Promise<void>,
-  streamErrors: string[],
-): Promise<void> {
-  try {
-    await onLine(line)
-  } catch (error) {
-    streamErrors.push(error instanceof Error ? error.message : String(error))
-  }
-}

packages/browseros-agent/apps/eval/src/agents/claude-code/stream-parser.ts

@@ -1,142 +0,0 @@
-import { randomUUID } from 'node:crypto'
-import type { UIMessageStreamEvent } from '../../types'
-
-type JsonObject = Record<string, unknown>
-
-export class ClaudeCodeStreamParser {
-  private lastText: string | null = null
-  private toolCallCount = 0
-
-  pushLine(line: string): UIMessageStreamEvent[] {
-    const trimmed = line.trim()
-    if (!trimmed) return []
-
-    let parsed: unknown
-    try {
-      parsed = JSON.parse(trimmed)
-    } catch {
-      return []
-    }
-
-    if (!isObject(parsed)) return []
-
-    if (parsed.type === 'assistant') {
-      return this.parseAssistantMessage(parsed)
-    }
-    if (parsed.type === 'user') {
-      return this.parseUserMessage(parsed)
-    }
-    if (parsed.type === 'result' && typeof parsed.result === 'string') {
-      this.lastText = parsed.result
-    }
-
-    return []
-  }
-
-  getLastText(): string | null {
-    return this.lastText
-  }
-
-  getToolCallCount(): number {
-    return this.toolCallCount
-  }
-
-  private parseAssistantMessage(message: JsonObject): UIMessageStreamEvent[] {
-    const content = contentBlocks(message)
-    const events: UIMessageStreamEvent[] = []
-
-    for (const block of content) {
-      if (block.type === 'text' && typeof block.text === 'string') {
-        const id = randomUUID()
-        this.lastText = block.text
-        events.push(
-          { type: 'text-start', id },
-          { type: 'text-delta', id, delta: block.text },
-          { type: 'text-end', id },
-        )
-      } else if (
-        block.type === 'tool_use' &&
-        typeof block.id === 'string' &&
-        typeof block.name === 'string'
-      ) {
-        this.toolCallCount++
-        events.push({
-          type: 'tool-input-available',
-          toolCallId: block.id,
-          toolName: block.name,
-          input: block.input,
-        })
-      }
-    }
-
-    return events
-  }
-
-  private parseUserMessage(message: JsonObject): UIMessageStreamEvent[] {
-    const content = contentBlocks(message)
-    const events: UIMessageStreamEvent[] = []
-
-    for (const block of content) {
-      if (
-        block.type !== 'tool_result' ||
-        typeof block.tool_use_id !== 'string'
-      ) {
-        continue
-      }
-
-      if (block.is_error === true) {
-        events.push({
-          type: 'tool-output-error',
-          toolCallId: block.tool_use_id,
-          errorText: stringifyToolContent(block.content),
-        })
-      } else {
-        events.push({
-          type: 'tool-output-available',
-          toolCallId: block.tool_use_id,
-          output: normalizeToolContent(block.content),
-        })
-      }
-    }
-
-    return events
-  }
-}
-
-export function shouldCaptureScreenshotForTool(toolName: string): boolean {
-  if (!toolName.startsWith('mcp__browseros__')) return false
-  return !toolName.endsWith('__take_screenshot')
-}
-
-function contentBlocks(message: JsonObject): JsonObject[] {
-  const inner = isObject(message.message) ? message.message : message
-  return Array.isArray(inner.content) ? inner.content.filter(isObject) : []
-}
-
-function isObject(value: unknown): value is JsonObject {
-  return typeof value === 'object' && value !== null
-}
-
-function normalizeToolContent(content: unknown): unknown {
-  if (!Array.isArray(content)) return content
-  return content.map((item) => {
-    if (
-      isObject(item) &&
-      item.type === 'text' &&
-      typeof item.text === 'string'
-    ) {
-      return item.text
-    }
-    return item
-  })
-}
-
-function stringifyToolContent(content: unknown): string {
-  const normalized = normalizeToolContent(content)
-  if (typeof normalized === 'string') return normalized
-  try {
-    return JSON.stringify(normalized)
-  } catch {
-    return String(normalized)
-  }
-}