fix(agent): route 'thinking blocks cannot be modified' 400 to recovery

Anthropic returns a 400 when the thinking/redacted_thinking blocks in the
latest assistant message are mutated upstream: '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 classifier's thinking_signature branch only matched on the substring
'signature', so this variant fell through to a non-retryable client error
and hard-aborted the turn -- even though the existing strip-reasoning_details
-and-retry recovery would have healed it.

Broaden the 400 match to also catch 'cannot be modified' / 'must remain as
they were' (still gated on 'thinking'), routing it to the same recovery.
Adds a negative-case test so unrelated 'cannot be modified' 400s are not
swept in.

Defense-in-depth, orthogonal to the root-cause work in #35975 / #17861
(which prevent the block mutation in the first place). Only changes a
terminal-failure into a one-shot recovery.

Signed-off-by: Ian Culling <ian@culling.ca>
This commit is contained in:
Ian Culling 2026-05-31 18:12:42 -04:00 committed by Teknium
parent 6110aed9be
commit 86e10dd874
2 changed files with 57 additions and 3 deletions

View file

@ -549,14 +549,32 @@ def classify_api_error(
should_fallback=True,
)
# Anthropic thinking block signature invalid (400).
# Anthropic thinking block recovery (400). Two distinct failure modes,
# same recovery (strip all reasoning_details and retry without thinking
# blocks — see the thinking_signature handler in conversation_loop.py):
# 1. Signature mismatch: a thinking block is signed against the full
# turn content; any upstream mutation (context compression, session
# truncation, message merging) invalidates the signature.
# Pattern: "signature" + "thinking".
# 2. Frozen-block mutation: Anthropic rejects any change to the
# thinking/redacted_thinking blocks in the *latest* assistant
# message — "`thinking` or `redacted_thinking` blocks in the latest
# assistant message cannot be modified. These blocks must remain as
# they were in the original response." This carries no "signature"
# token, so the original pattern missed it and the turn hard-aborted
# as a non-retryable client error instead of self-healing.
# Pattern: "thinking" + ("cannot be modified" | "must remain as they were").
# Don't gate on provider — OpenRouter proxies Anthropic errors, so the
# provider may be "openrouter" even though the error is Anthropic-specific.
# The message pattern ("signature" + "thinking") is unique enough.
# The combined patterns are unique enough.
if (
status_code == 400
and "signature" in error_msg
and "thinking" in error_msg
and (
"signature" in error_msg
or "cannot be modified" in error_msg
or "must remain as they were" in error_msg
)
):
return _result(
FailoverReason.thinking_signature,