Cluster 13 Extended-Thinking Wedge Diagnostic

5 questions · ~2 minutes · browser-only, no signup, no telemetry · CC0 · v1.0 (2026-05-30)
The wedge: When extended thinking is enabled, a session can hit API Error: 400 messages.N.content.M: thinking or redacted_thinking blocks in the latest assistant message cannot be modified and never recover within that session. The corrupted message persists in the transcript; every subsequent turn re-sends and re-fails identically. Documented across 15+ open issues (central case #63147, 33 reactions) — reproduces on v2.1.143 through v2.1.156. The cluster has four sub-patterns with different triggers and different recoveries. Misapplied /clear burns context on 13B/C/D when a transcript edit could have preserved it.
New as of 2026-05-29: Under /loop or other autonomous-resume harnesses, a one-time 400 becomes an unrecoverable infinite loop — the loop queues another continuation, the latest assistant message is unchanged on disk, the next request hits the same 400, the loop replays. @LMS927369 reported this on the #63147 thread. Question 5 below detects this amplification.
Disclosure: The prevention section links to two SessionStart hooks in this repository (PR #445 advisory + PR #469 opt-in BLOCK for autonomous run) and the published field guide. The hooks detect the 13A precursor before resume actually fires the 400. Free, MIT-licensed. No signup, no telemetry, no analytics beacons — this page is a single self-contained HTML file you can save and run offline.

5 Questions

1. When did the 400 first fire?
The triggering action narrows which sub-pattern's recovery surface applies.
2. What does the on-disk transcript look like?
Run: jq -rc 'select(.type=="assistant") | .message.content[]? | select(.type=="thinking") | [(.thinking|length),(.signature|length)] | @tsv' <your-session>.jsonl | head -20 against ~/.claude/projects/<slug>/<id>.jsonl.
3. Did you switch models mid-session (/model) while extended thinking was active?
Mid-session model switches persist thinking signatures from the previous model in transcript history.
4. How many thinking blocks does the wedged session contain (rough order of magnitude)?
Scale is the precondition: 13A often fires at the high end, while 13B/13C can wedge a session with few thinking blocks.
5. Is the session running under /loop, headless agent fleet, or scheduled autonomous resume?
Critical: under autonomous resume, the one-time 400 becomes an infinite loop because nothing in the loop layer reads stderr to bail.

Diagnosis

Top likely sub-pattern

Recovery (apply matching the top sub-pattern)

Prevention (going forward)

Reference materials

If you keep hitting this category of failure

This page covers Cluster 13 specifically. If extended thinking failures are a recurring tax on your operation — burning quota, killing sessions, derailing long-horizon runs — the published material extends to: