fix(agents): preserve thinking/redacted_thinking blocks in stripThoughtSignatures by iamhitarth · Pull Request #29620 · openclaw/openclaw

iamhitarth · 2026-02-28T09:11:53Z

Summary

stripThoughtSignatures removes thought_signature fields from content blocks for cross-provider compatibility (e.g. Gemini)
However, it also modifies thinking and redacted_thinking blocks via object spread + delete
Anthropic's API requires these blocks to be identical to the original response in the latest assistant message
On multi-turn conversations, the modified thinking blocks cause the API to reject with: "thinking or redacted_thinking blocks in the latest assistant message cannot be modified. These blocks must remain as they were in the original response."

Fix

Skip type: "thinking" and type: "redacted_thinking" blocks in stripThoughtSignatures — these blocks should never have their signatures stripped since they must be preserved verbatim for Anthropic.

Fixes #29618

Reproduction

Use Anthropic provider with extended thinking enabled
Have a multi-turn conversation in a Discord thread where thinking blocks are stored in session
On a subsequent turn, the session history replay modifies the thinking blocks via sanitizeSessionHistory → sanitizeSessionMessagesImages → stripThoughtSignatures
API rejects with the error above

Test plan

Verify multi-turn conversations with extended thinking no longer produce "cannot be modified" errors
Verify cross-provider signature stripping still works for non-thinking blocks (e.g. text blocks with thought_signature)

🤖 Generated with Claude Code

…htSignatures stripThoughtSignatures removes thought_signature fields from content blocks for cross-provider compatibility. However, it also modifies thinking and redacted_thinking blocks, which Anthropic's API requires to be identical to the original response. When these modified blocks appear in the latest assistant message on a subsequent turn, the API rejects the request with: "thinking or redacted_thinking blocks in the latest assistant message cannot be modified" Skip thinking/redacted_thinking block types in stripThoughtSignatures to preserve them unchanged. Co-Authored-By: Claude Opus 4.6 <[email protected]>

chatgpt-codex-connector

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 004e91311b

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

Open a pull request for review
Mark a draft as ready
Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

chatgpt-codex-connector · 2026-02-28T09:15:21Z

src/agents/pi-embedded-helpers/bootstrap.ts

+    if (blockType === "thinking" || blockType === "redacted_thinking") {
+      return block;


Gate thinking-block bypass to Anthropic requests

Returning early for every thinking/redacted_thinking block disables the allowBase64Only signature sanitization that Gemini paths rely on (resolveTranscriptPolicy sets this for Google/OpenRouter in src/agents/transcript-policy.ts:115-124). In provider-switch sessions (for example, Claude history containing thought_signature: "msg_*"), those non-base64 signatures are now forwarded unchanged, undoing the cross-provider compatibility behavior stripThoughtSignatures is intended to provide and potentially causing Gemini follow-up requests to fail. This bypass should be scoped to Anthropic replay requirements rather than applied unconditionally.

Useful? React with 👍 / 👎.

greptile-apps · 2026-02-28T09:16:29Z

Greptile Summary

Adds an early return in stripThoughtSignatures to preserve thinking/redacted_thinking blocks unchanged, fixing API rejections from Anthropic which requires these blocks to remain identical to the original response.

Changes:

Checks block type before processing
Returns thinking/redacted_thinking blocks immediately without modification
Preserves object identity and all fields (including thought_signature) for these block types

Impact:

Fixes multi-turn conversation errors with extended thinking enabled
Changes existing behavior: thinking blocks with msg_* signatures are now preserved instead of stripped (correct per Anthropic requirements)
Test file needs updating to reflect new behavior

Confidence Score: 4/5

Safe to merge after updating test expectations
The fix correctly addresses a real API compatibility issue with clear, efficient logic. The only concern is that an existing test expects the old (incorrect) behavior and needs updating. The change is well-justified by Anthropic's API requirements.
Verify test file src/agents/pi-embedded-helpers.sanitizeuserfacingtext.test.ts passes with updated expectations for thinking block preservation

_{Last reviewed commit: 004e913}

greptile-apps

_{1 file reviewed, 1 comment}

_{Edit Code Review Agent Settings | Greptile}

greptile-apps · 2026-02-28T09:16:32Z

src/agents/pi-embedded-helpers/bootstrap.ts

+    // Never modify thinking/redacted_thinking blocks — Anthropic requires these
+    // to be identical to the original API response.
+    const blockType = (block as { type?: unknown }).type;
+    if (blockType === "thinking" || blockType === "redacted_thinking") {
+      return block;
+    }


test file needs updating - the test at lines 148-160 in src/agents/pi-embedded-helpers.sanitizeuserfacingtext.test.ts expects thinking blocks with msg_* signatures to be stripped, but this fix preserves them unchanged (which is correct per Anthropic's requirements)

Prompt To Fix With AI

This is a comment left during a code review. Path: src/agents/pi-embedded-helpers/bootstrap.ts Line: 68-73 Comment: test file needs updating - the test at lines 148-160 in `src/agents/pi-embedded-helpers.sanitizeuserfacingtext.test.ts` expects `thinking` blocks with `msg_*` signatures to be stripped, but this fix preserves them unchanged (which is correct per Anthropic's requirements) How can I resolve this? If you propose a fix, please make it concise.

openclaw-barnacle · 2026-03-06T04:05:28Z

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

AuburyEssentian · 2026-03-06T06:02:14Z

Still relevant — this fix prevents Anthropic API rejections when thinking blocks are passed back in a multi-turn conversation. The stripThoughtSignatures function was stripping/transforming content blocks that must remain byte-identical to the original API response. Happy to rebase or address any feedback.

AuburyEssentian · 2026-03-07T06:06:04Z

Closing as superseded — upstream merged equivalent fixes for thinking block preservation (strip only msg_* signatures, preserve thinking/redacted_thinking blocks verbatim) in the interim. The bug this addressed is no longer present in current main.

AuburyEssentian · 2026-03-07T06:06:11Z

Noting that upstream has since merged equivalent fixes for thinking block preservation (stripping only msg_* signatures, preserving thinking/redacted_thinking blocks verbatim). This PR may now be superseded — maintainers, feel free to close if the issue is fully addressed in current main.

…story sanitization The Anthropic API requires thinking and redacted_thinking blocks to be returned byte-for-byte identical in conversation history. Multiple sanitization functions were mutating these blocks via object spread and field deletion, causing API rejections with: "thinking or redacted_thinking blocks in the latest assistant message cannot be modified." Add isImmutableThinkingBlock() guard and apply it in all four mutation sites: - stripThoughtSignatures (bootstrap.ts) — was deleting signature fields - sanitizeToolResult (pi-embedded-subscribe.tools.ts) — was deleting .data - pruneProcessedHistoryImages (history-image-prune.ts) — was replacing blocks - dropThinkingBlocks (thinking.ts) — was missing redacted_thinking type Fixes openclaw#29618 Supersedes openclaw#29620 Co-Authored-By: Claude Opus 4.6 <[email protected]>

obviyus · 2026-04-01T11:05:46Z

Closing as duplicate; this was superseded by #58916.

openclaw-barnacle bot added agents Agent runtime and tooling size: XS labels Feb 28, 2026

chatgpt-codex-connector bot reviewed Feb 28, 2026

View reviewed changes

greptile-apps bot reviewed Feb 28, 2026

View reviewed changes

steipete mentioned this pull request Mar 2, 2026

fix(agents): skip thinking/redacted_thinking blocks in stripThoughtSignatures #29648

Closed

openclaw-barnacle bot added the stale Marked as stale due to inactivity label Mar 6, 2026

a-alphayed mentioned this pull request Mar 13, 2026

fix(agents): preserve thinking/redacted_thinking blocks in session history #44827

Open

5 tasks

openclaw-barnacle bot removed the stale Marked as stale due to inactivity label Mar 28, 2026

obviyus mentioned this pull request Apr 1, 2026

fix: preserve anthropic thinking blocks on replay #58916

Merged

obviyus closed this Apr 1, 2026

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

fix(agents): preserve thinking/redacted_thinking blocks in stripThoughtSignatures#29620

fix(agents): preserve thinking/redacted_thinking blocks in stripThoughtSignatures#29620
iamhitarth wants to merge 1 commit intoopenclaw:mainfrom
iamhitarth:fix/preserve-anthropic-thinking-blocks

iamhitarth commented Feb 28, 2026

Uh oh!

chatgpt-codex-connector bot left a comment

Uh oh!

chatgpt-codex-connector bot Feb 28, 2026

Uh oh!

greptile-apps bot commented Feb 28, 2026

Uh oh!

greptile-apps bot left a comment

Uh oh!

greptile-apps bot Feb 28, 2026

Uh oh!

openclaw-barnacle bot commented Mar 6, 2026

Uh oh!

AuburyEssentian commented Mar 6, 2026

Uh oh!

AuburyEssentian commented Mar 7, 2026

Uh oh!

AuburyEssentian commented Mar 7, 2026

Uh oh!

obviyus commented Apr 1, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

		if (blockType === "thinking" \|\| blockType === "redacted_thinking") {
		return block;

Uh oh!

Conversation

iamhitarth commented Feb 28, 2026

Summary

Fix

Reproduction

Test plan

Uh oh!

chatgpt-codex-connector bot left a comment

Choose a reason for hiding this comment

💡 Codex Review

Uh oh!

chatgpt-codex-connector bot Feb 28, 2026

Choose a reason for hiding this comment

Uh oh!

greptile-apps bot commented Feb 28, 2026

Greptile Summary

Confidence Score: 4/5

Uh oh!

greptile-apps bot left a comment

Choose a reason for hiding this comment

Uh oh!

greptile-apps bot Feb 28, 2026

Choose a reason for hiding this comment

Uh oh!

openclaw-barnacle bot commented Mar 6, 2026

Uh oh!

AuburyEssentian commented Mar 6, 2026

Uh oh!

AuburyEssentian commented Mar 7, 2026

Uh oh!

AuburyEssentian commented Mar 7, 2026

Uh oh!

obviyus commented Apr 1, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants