fix(compaction): strip thinking/redacted_thinking blocks to prevent Anthropic API rejection by coletoncodes · Pull Request #43783 · openclaw/openclaw

coletoncodes · 2026-03-12T07:11:56Z

Summary

Problem: Compaction corrupts thinking/redacted_thinking blocks through the serialization pipeline (sanitizeSurrogates, field reordering, empty-text dropping). Anthropic then rejects all subsequent API calls with "thinking or redacted_thinking blocks in the latest assistant message cannot be modified", permanently breaking the session.
Why it matters: Any user with an Anthropic model + extended thinking + safeguard compaction hits this after enough messages. The session becomes unrecoverable.
What changed: (1) dropThinkingBlocks now handles redacted_thinking, (2) new stripThinkingFromNonLatestAssistant strips thinking from historical messages during sanitization, (3) compaction path strips all thinking blocks before its summary LLM call.
What did NOT change: The regular API call path still preserves thinking blocks in the latest assistant message (Anthropic requires them). No changes to compaction logic itself, session serialization, or provider capabilities.

Change Type (select all)

Scope (select all touched areas)

Linked Issue/PR

User-visible / Behavior Changes

Sessions using Anthropic models with extended thinking (e.g. claude-opus-4-6, claude-sonnet-4-6) and safeguard compaction will no longer break after compaction triggers. Previously, compaction would corrupt thinking blocks and permanently break the session.

Security Impact (required)

New permissions/capabilities? No
Secrets/tokens handling changed? No
New/changed network calls? No
Command/tool execution surface changed? No
Data access scope changed? No

Repro + Verification

Environment

OS: macOS 15.5 (Darwin 24.6.0)
Runtime/container: Node.js, OpenClaw 2026.3.9
Model/provider: anthropic/claude-sonnet-4-6
Integration/channel: Telegram
Relevant config: compaction.mode: "safeguard", thinking: adaptive

Steps

Configure agent with anthropic/claude-sonnet-4-6 and compaction.mode: "safeguard"
Chat via Telegram until context window triggers compaction
Before fix: session breaks with "thinking blocks cannot be modified" error
After fix: compaction completes, conversation continues normally

Expected

Compaction succeeds without corrupting thinking blocks
Session continues normally after compaction

Actual (before fix)

Compaction LLM call rejected by Anthropic API
Error leaks to Telegram chat
Session becomes permanently unresponsive

Evidence

Failing test/log before + passing after
Trace/log snippets
Screenshot/recording
Perf numbers (if relevant)

13 tests covering dropThinkingBlocks (including redacted_thinking) and stripThinkingFromNonLatestAssistant (latest preservation, non-latest stripping, interleaved messages, empty content fallback).

Human Verification (required)

Verified scenarios: Reproduced on Telegram with anthropic/claude-sonnet-4-6 and safeguard compaction. After fix, compaction triggers successfully and conversation continues.
Edge cases checked: redacted_thinking blocks, messages with only thinking content (empty text fallback), interleaved user/toolResult messages, conversations where only the latest assistant has thinking blocks.
What you did not verify: Bedrock provider path (no Bedrock credentials available), GitHub Copilot Claude path.

Review Conversations

I replied to or resolved every bot review conversation I addressed in this PR.
I left unresolved only the conversations that still need reviewer or maintainer judgment.

Compatibility / Migration

Backward compatible? Yes
Config/env changes? No
Migration needed? No

Failure Recovery (if this breaks)

How to disable/revert this change quickly: Revert this commit. The stripThinkingFromNonLatestAssistant call is gated behind policy.preserveSignatures and the compaction stripping is gated behind the same flag, so only Anthropic providers are affected.
Files/config to restore: thinking.ts, google.ts, compact.ts
Known bad symptoms reviewers should watch for: If thinking blocks are being stripped when they shouldn't be, you'd see Anthropic responses without extended thinking in sessions that have it enabled.

Risks and Mitigations

Risk: Stripping thinking blocks from non-latest messages means compaction summaries won't include reasoning content from earlier messages.
- Mitigation: Thinking blocks are internal reasoning, not user-facing content. Compaction summaries are based on text/tool_use blocks which contain the actual conversation substance. Anthropic explicitly allows omitting thinking blocks from non-latest messages.

Files Changed

src/agents/pi-embedded-runner/thinking.ts — Fixed dropThinkingBlocks to handle redacted_thinking + added stripThinkingFromNonLatestAssistant
src/agents/pi-embedded-runner/thinking.test.ts — 13 tests covering both functions
src/agents/pi-embedded-runner/google.ts — Apply stripThinkingFromNonLatestAssistant in sanitizeSessionHistory for Anthropic providers
src/agents/pi-embedded-runner/compact.ts — Strip all thinking blocks before compaction LLM call

greptile-apps · 2026-03-12T07:16:58Z

Greptile Summary

This PR fixes a session-breaking bug where thinking/redacted_thinking blocks in Anthropic assistant messages were corrupted by the serialization pipeline during safeguard compaction, causing Anthropic to permanently reject subsequent API calls with "thinking blocks cannot be modified." The fix has two complementary parts: (1) stripThinkingFromNonLatestAssistant is added to the regular sanitizeSessionHistory path to prevent corruption of historical messages while leaving the latest assistant message untouched (Anthropic requires it to be byte-identical), and (2) dropThinkingBlocks — now extended to also handle redacted_thinking blocks — is called in the compaction path to strip all thinking blocks before the summarization LLM call, since compaction is generating a fresh summary and not continuing the conversation.

Changes summary:

thinking.ts: new isThinkingBlock helper, dropThinkingBlocks extended to cover redacted_thinking, new stripThinkingFromNonLatestAssistant function
google.ts: stripThinkingFromNonLatestAssistant inserted into sanitizeSessionHistory pipeline gated on policy.preserveSignatures
compact.ts: dropThinkingBlocks called (all messages) before the compaction LLM call, gated on transcriptPolicy.preserveSignatures
thinking.test.ts: 13 tests covering both new/updated functions and all described edge cases

One subtle side-effect to be aware of: In compact.ts, session.agent.replaceMessages(withoutThinking) permanently mutates the session's messages before the compaction LLM call. If that LLM call later fails (e.g. network error, timeout), the session is left with thinking blocks already stripped but no new compaction summary. In practice this is safe — Anthropic only rejects messages where thinking blocks are present but corrupted, not absent — but it is a visible state change on the failure path that is not rolled back.

Confidence Score: 4/5

This PR is safe to merge; it correctly addresses a reproducible, session-breaking bug for Anthropic + extended thinking + compaction users with well-tested changes.
The logic in all three call sites is correct and consistent. The 13 new tests cover the key edge cases thoroughly. The one point of caution is that compact.ts permanently mutates the session (stripping thinking blocks) before the compaction LLM call, so a failed compaction leaves the session in a modified-but-valid state. Because Anthropic only rejects present-but-corrupted thinking blocks — not absent ones — this degraded state is still API-compatible; it just loses thinking context for that turn. All changes are gated behind preserveSignatures so non-Anthropic providers are unaffected.
Pay attention to compact.ts lines 796-801 regarding the pre-mutation of session state before the compaction LLM call.

_{Last reviewed commit: 7b134e1}

…nthropic API rejection Compaction corrupts thinking/redacted_thinking blocks through the serialization pipeline (sanitizeSurrogates, field reordering), causing Anthropic to reject subsequent API calls with "thinking blocks cannot be modified". This permanently breaks sessions using extended thinking. Three changes: - dropThinkingBlocks now strips both thinking and redacted_thinking - stripThinkingFromNonLatestAssistant removes thinking blocks from historical messages (Anthropic allows omission from non-latest) - Compaction path strips all thinking blocks before the summary LLM call since compaction doesn't need them Closes openclaw#39887 Co-Authored-By: Claude Opus 4.6 <[email protected]>

xiaoyaner0201

🔍 Duplicate PR Group: Anthropic Thinking Block Compaction

This PR is one of 7 open PRs addressing the same issue: Anthropic thinking/redacted_thinking blocks from older assistant messages cause API errors during context compaction.

Related PRs

PR	Author	Date	Approach	Preserves latest	redacted_thinking	Compaction path	Tests	Clean diff
#24261	@Vaibhavee89	02-23	Guard functions + logging	N/A (no actual strip)	Detect only	❌	Medium	✅
#25381	@Nipurn123	02-24	Modify `dropThinkingBlocks` skip latest	✅	❌	❌	Good	❌ (bundles sendDice)
#27142	@slatem	02-26	Extend `dropThinkingBlocks` to Anthropic	❌ (strips all)	✅	❌	Limited	✅
#39919	@taw0002	03-08	New strip function in sanitizeSessionHistory	✅	✅	❌	Good	✅
#39940	@liangruochong44-ui	03-08	Policy flag for Anthropic	❌ (strips all)	✅	❌	Medium	❌ (bundles cron/SQLite/UI)
#43783	@coletoncodes	03-12	New `stripThinkingFromNonLatestAssistant` + compact path	✅	✅	✅	Best	✅
#44650	@rikisann	03-13	Runtime monkey-patch of node_modules	✅	✅	❌	None	⚠️ Fragile approach

Analysis

After reading all 7 diffs:

#43783 is the strongest candidate. It:

Creates a standalone stripThinkingFromNonLatestAssistant() without changing dropThinkingBlocks semantics (preserves Copilot behavior)
Integrates via existing policy.preserveSignatures flag (proper transcript policy mechanism)
Only PR that also fixes the compaction path (compact.ts) — others only fix sanitizeSessionHistory
Handles both thinking and redacted_thinking block types
Clean diff with comprehensive tests (toolResult interleaving, empty blocks, reference equality)

#39919 is a close second (same core function, clean diff, good tests) but misses the compaction path.

Notable discovery from #44650: sanitizeSurrogates may corrupt thinking block signatures — worth investigating separately even though the monkey-patch approach is not viable.

Not recommended: #27142 and #39940 strip thinking from ALL messages including latest (breaks multi-turn quality). #25381 and #39940 bundle unrelated changes.

Similarities identified via embedding-based clustering (cosine > 0.82). Maintainers: #43783 appears ready for review.

coletoncodes · 2026-04-02T01:40:40Z

Closing -- the fixes in this PR have been superseded by upstream work:

dropThinkingBlocks was refactored to preserve only the latest assistant message's thinking blocks (fix(agents): drop Anthropic thinking blocks on replay #44843, fix: preserve anthropic thinking blocks on replay #58916)
Cache-TTL pruning aligned with thinking replay sanitization (f70ad92)

The compaction-specific stripping (all thinking blocks before the summarization LLM call) may still be a gap, but will revisit from main if needed.

openclaw-barnacle bot added agents Agent runtime and tooling size: M labels Mar 12, 2026

coletoncodes force-pushed the fix/compaction-thinking-blocks branch from 56cd457 to b1dcf3d Compare March 12, 2026 17:54

coletoncodes and others added 2 commits March 12, 2026 16:08

fix(compaction): document intentional pre-mutation of thinking blocks

d5acc08

coletoncodes force-pushed the fix/compaction-thinking-blocks branch from b1dcf3d to d5acc08 Compare March 12, 2026 21:09

xiaoyaner0201 reviewed Mar 13, 2026

View reviewed changes

coletoncodes closed this Apr 2, 2026

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

fix(compaction): strip thinking/redacted_thinking blocks to prevent Anthropic API rejection#43783

fix(compaction): strip thinking/redacted_thinking blocks to prevent Anthropic API rejection#43783
coletoncodes wants to merge 2 commits intoopenclaw:mainfrom
ColetonCodes-LLC:fix/compaction-thinking-blocks

coletoncodes commented Mar 12, 2026

Uh oh!

greptile-apps bot commented Mar 12, 2026

Uh oh!

xiaoyaner0201 left a comment

Uh oh!

coletoncodes commented Apr 2, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

Uh oh!

Conversation

coletoncodes commented Mar 12, 2026

Summary

Change Type (select all)

Scope (select all touched areas)

Linked Issue/PR

User-visible / Behavior Changes

Security Impact (required)

Repro + Verification

Environment

Steps

Expected

Actual (before fix)

Evidence

Human Verification (required)

Review Conversations

Compatibility / Migration

Failure Recovery (if this breaks)

Risks and Mitigations

Files Changed

Uh oh!

greptile-apps bot commented Mar 12, 2026

Greptile Summary

Confidence Score: 4/5

Uh oh!

xiaoyaner0201 left a comment

Choose a reason for hiding this comment

🔍 Duplicate PR Group: Anthropic Thinking Block Compaction

Related PRs

Analysis

Uh oh!

coletoncodes commented Apr 2, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants