Weave 5 Whys into harness failure response paths

[strix-tkellogg]

Summary

• System prompt: Routes introspection → 5 Whys when patterns emerge, and routes prediction misses to structured reflection
• Post-turn failure context: When a turn ends with error or circuit breaker, the next turn's prompt includes reflection guidance pointing to 5 Whys
• Prediction-review → 5 Whys bridge: SKILL.md now guides agents to decompose surprising misses instead of just logging prediction_true: false
• Chat history scan: Prediction-review scheduled job now also scans recent chat for corrections, error reactions, and repeated attempts — catches failures that predictions miss

PR #80 already covers cycle detection (integration point #2). This PR covers the remaining four.

The core problem: the 5 Whys skill existed as a callable tool but nothing in the harness actually routed agents toward it when failure happened. Like a fire extinguisher in a locked cabinet with no sign. Now four paths converge on it.

Files changed

• prompts.py — Two new skill routing bullets in system prompt (introspection→5Whys, prediction miss→5Whys)
• app.py — _last_turn_failure field captures error/circuit-breaker context, injected into next turn via render_turn_prompt
• config.py — Default scheduler job expanded: prediction misses trigger 5 Whys, plus chat history failure scan
• prediction-review/SKILL.md — New "When a Prediction Misses" section with 5 Whys routing criteria

Test plan

• 197 tests pass, 1 skipped
• render_turn_prompt accepts optional last_turn_failure param (backward compatible, defaults to None)
• Builtin skills tests pass (8/8)
• Manual verification: trigger a circuit breaker loop and confirm next turn gets failure context

🤖 Generated with Claude Code