Fable 5 safeguards false-positive [reasoning_extraction] terminates code-review subagents (mutation-testing vocabulary)

Summary

Fable 5 safeguards repeatedly reject legitimate subagent conversations with Details: [reasoning_extraction], terminating background agents mid-task. The affected conversations are internal code-review/QA subagents doing adversarial testing of our own codebase: briefs contain phrases like "remove the auth guard mutation and verify the test turns red", "mutate the SQL fence and confirm the suite fails", i.e. standard mutation-testing and security-gate vocabulary. Nothing in the sessions requests harmful content.

Pattern observed

• Started ~21:30 CEST 2026-08-20, immediately after a session usage limit reset.
• First hit only resumed subagents (continuation via transcript), then also freshly spawned subagents with similar review briefs.
• Same subagent briefs run fine when the subagent is spawned on Opus (model: opus), so the trigger appears specific to the Fable 5 safeguard layer.
• Error text: "Fable 5's safeguards flagged this message ... Claude Code can't respond to this message with Fable 5. Details: [reasoning_extraction]".

Request IDs

• req_011CeERWmqZuRXvTWH66SRry
• req_011CeERZeFzVaGEHMxHHfT3Y
• req_011CeERZwjxKFdKhs1UjYiNy
• req_011CeERbQTLH256z2VKFiYEH
• req_011CeERbVAr2t2bQvmPrMgCX
• req_011CeERczmrgDBU9DQYPvPDf
• req_011CeERfersZMNUsHTtwGYxk
• req_011CeERgnp8i9WY85faunHep
• req_011CeERhkSHghqJQ2dXxdUZr
• req_011CeERi39d2cYx5Bqr6QMYa

Environment

• Claude Code CLI on macOS (darwin 25.6.0), model claude-fable-5, subagents via the Agent tool (background, worktree isolation).
• Impact: multi-agent review pipelines stall; workaround is per-agent model: opus override.

Expected

Adversarial mutation-testing language in internal code-review agents should not trip the [reasoning_extraction] safeguard, or the safeguard should allow retry rather than terminating the agent.

---
Internal triage form (81 Studios house format, kept for our own tracking):

BEHOV: Fable 5 safeguard falsely terminates legitimate code-review subagents, stalling our review pipeline.
MÄTNING: 10 request IDs above, all [reasoning_extraction], 2026-08-20 ~21:30-21:55 CEST; identical briefs succeed with model: opus.
ÅTGÄRD: Report upstream (this issue); interim workaround is per-agent Opus override.
TAK: Not requesting any safeguard removal; only fixing the false positive on mutation-testing vocabulary. Measurement is falsified if the same briefs later fail on Opus too (then it is not Fable-specific).