Skip to content

fix: drop thinking blocks for Anthropic to prevent session corruption#27142

Closed
slatem wants to merge 1 commit intoopenclaw:mainfrom
slatem:fix/drop-thinking-blocks-anthropic
Closed

fix: drop thinking blocks for Anthropic to prevent session corruption#27142
slatem wants to merge 1 commit intoopenclaw:mainfrom
slatem:fix/drop-thinking-blocks-anthropic

Conversation

@slatem
Copy link
Copy Markdown

@slatem slatem commented Feb 26, 2026

Problem

When extended thinking is enabled (thinking=low/high) on Anthropic models, session compaction or gateway restart can modify thinking/redacted_thinking blocks in the stored transcript. Anthropic's API requires these blocks to be byte-for-byte identical to the original response — any modification causes a permanent 400 invalid_request_error:

messages.N.content.1: `thinking` or `redacted_thinking` blocks in the latest
assistant message cannot be modified. These blocks must remain as they were
in the original response.

The session is bricked until /reset, losing all conversation context.

Root Cause

resolveTranscriptPolicy in src/agents/transcript-policy.ts only enables dropThinkingBlocks for GitHub Copilot Claude endpoints. For Anthropic's own API (anthropic-messages) and Bedrock (bedrock-converse-stream), thinking blocks are preserved in the transcript but can get corrupted during compaction or session reconstruction.

Fix

Enable dropThinkingBlocks for all Anthropic-compatible providers (isAnthropic), not just GitHub Copilot Claude. Anthropic's API accepts omitted thinking blocks — it only rejects modified ones. This is the minimal, safe fix:

- const dropThinkingBlocks = isCopilotClaude;
+ const dropThinkingBlocks = isAnthropic || isCopilotClaude;

Trade-off

The model loses thinking context from prior assistant turns. This is minor — thinking blocks are internal reasoning and don't significantly affect multi-turn coherence. The alternative (sessions permanently bricking) is far worse.

Tests

Added 5 new test cases covering:

  • Anthropic provider → dropThinkingBlocks: true
  • Bedrock Anthropic → dropThinkingBlocks: true
  • GitHub Copilot Claude → dropThinkingBlocks: true (existing behavior preserved)
  • OpenAI → dropThinkingBlocks: false
  • Google → dropThinkingBlocks: false

All 12 tests pass.

Fixes #19524
Fixes #25194

@openclaw-barnacle openclaw-barnacle bot added agents Agent runtime and tooling size: S labels Feb 26, 2026
@greptile-apps
Copy link
Copy Markdown
Contributor

greptile-apps bot commented Feb 26, 2026

Greptile Summary

Fixes critical session corruption bug in Anthropic models by extending dropThinkingBlocks to all Anthropic-compatible providers.

  • Root cause: Anthropic's API requires thinking/redacted_thinking blocks to be byte-for-byte identical to original response; session compaction or gateway restart can modify these blocks, causing permanent 400 errors that brick the session
  • Fix: Changed dropThinkingBlocks = isCopilotClaude to dropThinkingBlocks = isAnthropic || isCopilotClaude in src/agents/transcript-policy.ts:108
  • Coverage: Applies to anthropic-messages API, bedrock-converse-stream API, and GitHub Copilot Claude
  • Trade-off: Model loses thinking context from prior assistant turns, but this is acceptable since Anthropic accepts omitted thinking blocks (only rejects modified ones)
  • Tests: Added 5 new test cases covering Anthropic, Bedrock, GitHub Copilot Claude, and negative tests for OpenAI/Google

Confidence Score: 5/5

  • This PR is safe to merge with minimal risk
  • The fix is minimal (one line change), addresses a critical bug (permanent session corruption), reuses existing proven functionality (dropThinkingBlocks), includes comprehensive test coverage (5 new tests for positive and negative cases), and is well-documented with clear comments explaining the issue and solution. The change is defensive and does not affect other providers (OpenAI, Google, etc.).
  • No files require special attention

Last reviewed commit: d3035de

@slatem
fix: drop thinking/redacted_thinking blocks for Anthropic provider
804fa2f 
The Anthropic API rejects `thinking` and `redacted_thinking` blocks in
historical assistant messages when those blocks have been modified from
the original response (e.g. after session serialization round-trips).

Changes:
- Enable dropThinkingBlocks for Anthropic (was only Copilot Claude)
- Also strip `redacted_thinking` blocks (was only `thinking`)
- Add test coverage for redacted_thinking block handling
@slatem slatem force-pushed the fix/drop-thinking-blocks-anthropic branch from 9b76595 to 804fa2f Compare March 4, 2026 00:52
xiaoyaner0201
xiaoyaner0201 reviewed Mar 13, 2026
View reviewed changes
Copy link
Copy Markdown
Contributor

@xiaoyaner0201 xiaoyaner0201 left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

🔍 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:

  1. Creates a standalone stripThinkingFromNonLatestAssistant() without changing dropThinkingBlocks semantics (preserves Copilot behavior)
  2. Integrates via existing policy.preserveSignatures flag (proper transcript policy mechanism)
  3. Only PR that also fixes the compaction path (compact.ts) — others only fix sanitizeSessionHistory
  4. Handles both thinking and redacted_thinking block types
  5. 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.

@openclaw-barnacle
Copy link
Copy Markdown

openclaw-barnacle bot commented Mar 30, 2026

This pull request has been automatically marked as stale due to inactivity.
Please add updates or it will be closed.

@openclaw-barnacle openclaw-barnacle bot added the stale Marked as stale due to inactivity label Mar 30, 2026
@openclaw-barnacle
Copy link
Copy Markdown

openclaw-barnacle bot commented Apr 3, 2026

Closing due to inactivity.
If you believe this PR should be revived, post in #pr-thunderdome-dangerzone on Discord to talk to a maintainer.
That channel is the escape hatch for high-quality PRs that get auto-closed.

@openclaw-barnacle openclaw-barnacle bot closed this Apr 3, 2026
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Reviewers

1 more reviewer

@xiaoyaner0201 xiaoyaner0201 xiaoyaner0201 left review comments

Reviewers whose approvals may not affect merge requirements

Assignees

No one assigned

Labels

agents Agent runtime and tooling size: S stale Marked as stale due to inactivity

Projects

None yet

Milestone

No milestone

2 participants

@slatem @xiaoyaner0201