Bug

During context truncation in long sessions (566+ messages, 4.6MB), limitHistoryTurns() can slice at user message boundaries, dropping an assistant(tool_use) while keeping its matching toolResult. This creates an invalid message sequence that the Anthropic API rejects on every retry.

Root Cause

1. repairToolUseResultPairing() runs via sanitizeSessionHistory() before limitHistoryTurns() — so the repair is done, then limitHistoryTurns re-introduces orphans by naive slicing
2. dist/agents/pi-extensions/transcript-sanitize.js exists and would fix this on every context event (every API call), but it's never loaded in buildEmbeddedExtensionPaths() — dead code

Fix (applied locally, works)

1. Load transcript-sanitize extension in buildEmbeddedExtensionPaths() — safety net on every API call
2. Call sanitizeToolUseResultPairing() after limitHistoryTurns() in both attempt.js and compact.js

Files

• dist/agents/pi-embedded-runner/extensions.js — add resolvePiExtensionPath('transcript-sanitize')
• dist/agents/pi-embedded-runner/run/attempt.js — repair after limitHistoryTurns
• dist/agents/pi-embedded-runner/compact.js — same

Repro

Run a session past ~500 messages with heavy tool use until auto-compaction triggers. The session becomes unrecoverable with roles must alternate / incorrect role information errors.