Add reflection: async dissonance detection

[tkellogg]

Summary

• Adds async self-review after each send_message — the agent's own model evaluates outgoing messages against criteria in a user-editable markdown file
• When dissonance is detected (confidence >= 0.7), a 🪞 reaction appears on the sent message
• Off by default (reflection.enabled: false), trivially enabled via config.yaml
• New builtin dissonance skill explains what 🪞 means and how to tune it
• Default criteria pre-populated from Strix's proven dissonance patterns (service wrap-ups, hollow validation, stance avoidance, ball-handing)

Architecture

• reflection.py — ReflectionHook class, fired from send_message after successful delivery
• Fire-and-forget via asyncio.create_task — does not block message delivery
• Side-effect only — result is NOT injected into the agent's conversation context
• Questions file read fresh each evaluation, so edits take effect immediately
• Uses the agent's configured model (not a separate API key)

Config

reflection:
  enabled: false
  questions_file: state/is-dissonant-prompt.md

Visible in default config with explanatory comment. The questions file path lets the user know where to look.

Files changed

File	Change
open_strix/reflection.py	Core reflection logic (201 lines)
open_strix/config.py	ReflectionConfig dataclass + parsing
open_strix/app.py	Hook initialization on startup
open_strix/tools.py	6-line hook call after send_message
open_strix/builtin_skills/dissonance/SKILL.md	Skill docs (what 🪞 means, tuning)
open_strix/builtin_skills/dissonance/default-questions.md	Pre-populated criteria
SETUP.md	Config reference docs
tests/test_reflection.py	23 tests

Test plan

• 23 new tests covering config parsing, questions file reading, hook behavior (skip empty, skip missing file, dissonance detected, no dissonance, low confidence, null result), and builtin skill discovery
• Full suite: 183 passed, 0 regressions (1 pre-existing uv PATH failure in onboarding test)
• Manual test: enable on Motley or Verge, verify 🪞 appears on appropriate messages

🤖 Generated with Claude Code