fix(agent): strip thinking blocks from content arrays on signature recovery

[liuhao1024]

What does this PR do?

The thinking-signature recovery path only stripped the top-level reasoning_details field from messages, but Anthropic thinking blocks live inside assistant message content arrays. After recovery, convert_messages_to_anthropic() on the retry kept the signed thinking blocks from the last assistant message, causing the same HTTP 400 "Invalid signature in thinking block" error on every retry attempt.

Root Cause

When a user switches from a provider that emits thinking blocks with its own signatures (e.g. MiniMax via Anthropic transport) to Anthropic proper, the message history contains thinking blocks with MiniMax-specific signatures. Anthropic rejects these with HTTP 400. The recovery code at run_agent.py (thinking_signature branch) only did:

_m.pop("reasoning_details", None)

But the thinking blocks are in _m["content"] as a list of content blocks, not in reasoning_details. After the continue, the retry goes through convert_messages_to_anthropic() which re-processes messages and keeps signed thinking blocks from the last assistant message — the same invalid signatures that caused the original error.

Related Issue

N/A

Type of Change

• 🐛 Bug fix (non-breaking change that fixes an issue)

Changes Made

• See commit messages for detailed changes

How to Test

1. Run pytest tests/ -q — all tests should pass
2. Verify the specific scenario described above is resolved

Checklist

Code

• I've read the Contributing Guide
• My commit messages follow Conventional Commits (fix(scope):, feat(scope):, etc.)
• I searched for existing PRs to make sure this isn't a duplicate
• My PR contains only changes related to this fix/feature (no unrelated commits)
• I've run pytest tests/ -q and all tests pass
• I've added tests for my changes (required for bug fixes, strongly encouraged for features)
• I've tested on my platform: macOS 26.4.1

Documentation & Housekeeping

• I've updated relevant documentation (README, docs/, docstrings) — or N/A
• I've updated cli-config.yaml.example if I added/changed config keys — or N/A
• I've updated CONTRIBUTING.md or AGENTS.md if I changed architecture and workflows — or N/A
• I've considered cross-platform impact (Windows, macOS) per the compatibility guide — or N/A

[teknium1]

Thanks for identifying the content-array variant of this recovery gap. The premise still holds on current main: agent/conversation_loop.py:2843-2846 strips only reasoning_details, while agent/anthropic_adapter.py:1982-1985 converts assistant content arrays and :2311-2344 preserves signed thinking on the latest direct-Anthropic assistant turn.

Problems

• The current retry wire list is api_messages, created as shallow copies at agent/conversation_loop.py:787-830. Current recovery intentionally leaves canonical messages untouched (:2819-2836). The PR's messages mutation needs to be moved to api_messages; otherwise it risks persisted-history corruption and does not match the current retry architecture.
• tests/agent/test_thinking_signature_recovery.py only reimplements the proposed loop. It does not verify the actual retry or convert_messages_to_anthropic() output.

Suggested changes

• Filter thinking and redacted_thinking blocks by rebinding api_messages[*]["content"] in the current recovery branch.
• Add a real-path retry test proving the retry payload is clean and canonical history is unchanged.

This is an automated hermes-sweeper review.

[GottZ]

This was generated by AI during triage.

Summary

Two PRs address the repeated HTTP 400 caused by replaying invalid signed Anthropic thinking state. #24407 filters thinking blocks from assistant content arrays but mutates canonical history, whereas #70107 removes both replay channels from API-time copies while preserving canonical messages.

Related pull requests

• #24407 related — (+145/-7) — superseded / needs rework: The diff removes thinking and redacted-thinking blocks from canonical messages, directly addressing the reported content-array signature failure but risking persisted-history corruption; its tests only reproduce the loop rather than exercising the actual retry and Anthropic conversion path. Despite the keep_open review on #24407, that contributor review itself identifies these blocking issues and requires moving the fix to api_messages.
• #70107 duplicate — (+48/-10) — preferred fix: The diff updates the current recovery path to clear both reasoning_details and the ordered anthropic_content_blocks replay channel from shallow-copied api_messages, preventing the invalid signature from being resent while leaving canonical history unchanged.

Duplicates

#24407 and #70107 target the same signature-recovery retry failure and substantially overlap in intent; #70107 implements the fix against the current API-time replay architecture.

Suggested consolidation

Merge #70107 as the narrower current-architecture fix; #24407 can be closed as superseded because its present diff mutates canonical messages and lacks a real-path retry/conversion test, notwithstanding its keep_open review.

Complex graph

flowchart LR
    classDef open fill:#dbeafe,stroke:#1d4ed8,color:#1e3a8a
    classDef merged fill:#dcfce7,stroke:#15803d,color:#14532d
    classDef closed fill:#e5e7eb,stroke:#6b7280,color:#1f2937
    classDef unverified fill:#f3f4f6,stroke:#9ca3af,color:#374151
    classDef best stroke-width:3px,stroke:#b45309
    classDef target stroke-width:3px,stroke:#4338ca
    subgraph Dup24407 ["PRs duplicating each other"]
        P24407["PR #24407 (open)"]
        P70107["PR #70107 (open)"]
    end
    class P24407 open
    class P70107 open
    class P24407 target
    click P24407 "https://github.com/NousResearch/hermes-agent/pull/24407"
    click P70107 "https://github.com/NousResearch/hermes-agent/pull/70107"

Loading

Graph: solid arrow = fixes / best fix, dashed arrow = partial or unverified (see edge label); boxed group = PRs duplicating each other; amber border = best fix; indigo border = target; gray node = closed or no verify verdict yet (state tag in the node label).

Cross-PR triage: Reviewed 2 pull requests and 0 issues in this complex. Each diff was read against this issue; Assessment working set: 15 kB of PR diffs, 5 kB of issue/PR text, 1 kB of discussion (1 comments), 1 verify verdict. verdicts reflect diff content, not PR titles. Part of an automated triage batch.