Bug Description

Extended-thinking Claude models (4.6+, e.g. Opus 4.8) crash-loop the gateway with a non-retryable HTTP 400 when a parallel tool batch is interrupted before every tool result returns.

These models emit a signed thinking block on assistant turns that also fire tool_use blocks. Anthropic signs that block against the full, original turn content, and on replay it must be passed back byte-for-byte. agent/anthropic_adapter.py::_strip_orphaned_tool_blocks() legitimately strips a tool_use whose matching tool_result never arrived (parallel batch interrupted, context compression, or session truncation) — but that mutates the latest assistant turn, and _manage_thinking_signatures() then replays the now-stale signed thinking block verbatim. Anthropic rejects it:

messages.N.content.M: `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 400 is classified non-retryable, so the gateway reloads the same poisoned transcript from the persisted store every turn → infinite crash-loop with no self-recovery (a soft session reset does not clear it, because history is rebuilt from the store). The drifting content.M index in the error is just the changing count of stripped tool_use blocks across rebuilds.

Steps to Reproduce

1. Run the gateway with an extended-thinking model (e.g. claude-opus-4-8 on the native Anthropic endpoint, extended thinking ON).
2. Send a message whose first turn fires a large parallel tool batch with thinking enabled.
3. Have the batch interrupted before every tool_result comes back (e.g. /stop, an error on one tool, or compression mid-flight).
4. On the next turn the orphaned tool_use is stripped, the signed thinking block's signature no longer matches, and Anthropic 400s.
5. The gateway loops on the same error indefinitely.

Minimal unit reproduction (no network):

from agent.anthropic_adapter import convert_messages_to_anthropic
messages = [
    {"role": "assistant", "content": "",
     "tool_calls": [
         {"id": "tc_kept",   "function": {"name": "a", "arguments": "{}"}},
         {"id": "tc_orphan", "function": {"name": "b", "arguments": "{}"}},
     ],
     "reasoning_details": [
         {"type": "thinking", "thinking": "plan", "signature": "sig"},
     ]},
    {"role": "tool", "tool_call_id": "tc_kept", "content": "result A"},
]
_, result = convert_messages_to_anthropic(messages)
# Before the fix: the latest assistant turn still carries the signed `thinking`
# block whose signature was computed over the original (un-stripped) 3-block turn
# → Anthropic 400 on replay.

Expected Behavior

When a structural mutation (orphan-strip / merge / truncation) invalidates a thinking-block signature on the latest assistant turn, Hermes should either preserve the block verbatim (impossible — the block was mutated) or demote it to plain text so the turn replays cleanly and the model can re-plan. The gateway must not enter a non-retryable crash-loop.

Actual Behavior

The stale signed thinking block is replayed verbatim, Anthropic returns a non-retryable HTTP 400 ("blocks in the latest assistant message cannot be modified"), and the gateway crash-loops because the poisoned transcript is rebuilt from the store on every turn. No self-recovery; a soft session reset does not clear it.

Affected Component

Agent Core (conversation loop, context compression, memory)

Messaging Platform (if gateway-related)

N/A (CLI only)

Operating System

macOS 26.4 (Apple Silicon)

Python Version

3.11.15

Hermes Version

v0.15.1 (2026.5.29)

Root Cause Analysis

agent/anthropic_adapter.py:

• _strip_orphaned_tool_blocks() removes orphaned tool_use blocks from assistant turns (correct), but does not account for the fact that a co-located thinking/redacted_thinking block's Anthropic signature was computed over the original turn content and is now dead.
• _manage_thinking_signatures() latest-assistant branch then replays any block with a signature verbatim — including the now-invalid one — producing the 400.

The two functions are individually correct but compose into the bug: strip mutates, signature-management trusts the mutated turn's signature.

Proposed Fix

Flag the turn in _strip_orphaned_tool_blocks() when stripping a tool_use from a turn that also holds a thinking block; propagate that flag through _merge_consecutive_roles(); and in _manage_thinking_signatures() demote all thinking blocks on a flagged latest turn to text (preserving the reasoning) instead of replaying a dead signature. Intact turns are unaffected.

PR ready: #35846 (with two regression tests).

Are you willing to submit a PR for this?

• I'd like to fix this myself and submit a PR