* fix: compaction config for small context windows (≤32K)
Raise COMPACTION_SMALL_CONTEXT_WINDOW from 16K to 32K so models like
Haiku 4.5 (30K context) use proportional 50% reserve instead of the
fixed 20K reserve. Also scale fixedOverhead for small contexts (capped
at 40% of context window) to prevent the doom loop where overhead alone
triggers compaction on every step.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* docs: add compaction tuning guidance to limits constants
Explain the relationship between SMALL_CONTEXT_WINDOW and
FIXED_OVERHEAD so devs know the 24K minimum constraint when
tweaking these values.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
* feat: add 2-stage pruning to compaction pipeline before LLM summarization
Add two new lightweight stages to the compaction prepareStep pipeline that
recover context tokens cheaply before falling back to expensive LLM
summarization:
- Stage 2: Use AI SDK's pruneMessages to remove old tool call/result
pairs beyond the last 6 messages entirely
- Stage 3: Replace remaining tool output values with short placeholders
("[Cleared — N chars]") while preserving tool call structure and IDs
Both stages re-estimate tokens from message content (not stale step
usage) after modifying messages. The existing LLM summarization and
sliding window fallback remain as Stage 4.
Also adds estimateTokensForThreshold() helper, clearToolOutputs()
function, and COMPACTION_PRUNE_KEEP_RECENT_MESSAGES /
COMPACTION_CLEAR_OUTPUT_MIN_CHARS constants.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* fix: reorder compaction pipeline — truncate before clear, protect recent tools
- Stage 0: Check threshold, return untouched when under (no data loss)
- Stage 1: Prune old tool call/result pairs beyond last 6 messages
- Stage 2: Truncate large tool outputs to 15K chars (keeps partial content)
- Stage 3: Clear old tool outputs with placeholders, protect last 2
- Stage 4: LLM-based compaction with sliding window fallback
clearToolOutputs now accepts keepRecentCount parameter (default 2) to
skip the N most recent tool messages from clearing.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* fix: limits fixes
* fix: address review — preserve toKeep context, derive test values from constants
- When Stage 3 (clearToolOutputs) doesn't resolve overflow, pass
truncated (not cleared) messages to Stage 4 so toKeep retains
meaningful tool outputs for the agent's immediate context
- Add comment explaining intentional conservatism in post-prune
token estimation (step usage is stale, must re-estimate safely)
- Refactor computeConfig tests to derive expected values from
AGENT_LIMITS constants instead of hardcoding magic numbers
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
- truncateToolOutputs: handle all output.type variants (text, json,
content) by checking output.value directly instead of branching on
type. The old code missed type 'content' (array of content parts),
causing 1M+ char tool results to pass through untouched.
- estimateTokens: change chars/4 to chars/3 — HTML/Markdown content
tokenizes at ~3.14 chars/token empirically, not 4.
- COMPACTION_FIXED_OVERHEAD: 5K → 12K to account for system prompt
(~2.5K tokens) + tool definitions as JSON Schema (~8-9K tokens).
- Apply truncateToolOutputs in prepareStep (Stage 0) before token
estimation, not just during summarization.
* fix: robust compaction with Pi-style token counting + overflow middleware
Root cause: getCurrentTokenCount() returned stale inputTokens from the
previous step, ignoring new tool results added to messages since that
step. A large tool output (DOM snapshot, page content) caused a token
jump that bypassed the compaction threshold check, leading to
context_length_exceeded errors (322K tokens sent, model max 262K).
Layer 1 — Accurate token counting (proactive):
- Adopt Pi coding agent's additive approach: base(inputTokens) +
outputTokens + estimate(trailing tool results)
- Trailing tool results are estimated by walking backwards from end of
messages array until a non-tool message is found
- Falls back to full estimation with safety multiplier when no real
usage data is available (first step of a turn)
Layer 2 — Context overflow middleware (reactive):
- LanguageModelV3Middleware that wraps doGenerate/doStream
- Catches context_length_exceeded errors at the model call level
- Truncates prompt (keeps system messages + most recent non-system
messages targeting 60% of context window)
- Retries the model call once
Verified end-to-end with real model (Gemini Flash Lite via OpenRouter)
on 16K context window: 4 compactions triggered correctly across 8
steps, no context_length_exceeded errors.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* fix: adopt Pi-style overflow detection patterns + fix truncation edge case
- Replace 6 generic substring matches with 17 provider-specific regex
patterns from Pi coding agent (Anthropic, OpenAI, Google, xAI, Groq,
OpenRouter, Bedrock, Copilot, llama.cpp, LM Studio, MiniMax, Kimi,
Mistral, z.ai)
- Fix truncatePrompt edge case: when the last message alone exceeds the
target, keepFrom was never updated → empty non-system messages. Now
always keeps at least the most recent non-system message.
- Add runtime guard for LanguageModelV3 cast in ai-sdk-agent.ts
- Add tests for false-positive rejection and truncation edge case
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
* feat: generalized compaction prompts with split turn handling
Replace browser-specific XML prompts with domain-agnostic markdown format.
Add split turn detection and parallel summarization for large single-turn
conversations. Switch compaction from generateText to streamText for
Fireworks API compatibility. Add comprehensive unit and E2E tests (84 total).
* fix: address code review issues for compaction (PR #391)
Enforce COMPACTION_MAX_SUMMARIZATION_INPUT cap, extract shared
callSummarizer helper, add runtime type guard for experimental_context,
move magic constants to AGENT_LIMITS, and remove dead constants.
* fix: cap truncatedTurnPrefix input to maxSummarizationInput
Apply the same sliding window cap to turn prefix messages that was
already applied to toSummarize, preventing unbounded LLM input for
long single-turn conversations with many tool calls.
* fix: reduce browseros-auto default context window to 200K
The 400K setting caused compaction to trigger at ~383K, but the actual
model limit is 262K. Conversations hit the hard limit before compaction
could kick in.
* feat: import all the missing tests before refactor
* fix: biome errors for tests
* fix: few type errors and add exceptiosn
* fix: few more type errors
* fix: remove agent port from tests
* fix: exclude tests from tsconfig, bun run tests natively
* fix: mcpServer test now waits for extension connected
