ACT Phase 2: Feedback tool for explicit finding ratings

[DSado88]

Context

Phase 5.5 of the unified review skill now captures Claude's critic judgments (elevated vs dismissed findings) and user reactions. But the signal is implicit — inferred from conversation flow. An explicit feedback tool would provide cleaner labels.

Related: 5-model review of ACT integration (2026-03-10) — all models converged on "explicit user feedback > git-diff correlation." Grok proposed a 200 LoC alternative that captures 80% of the value.

Depends on: Phase 5.5 skill capture (done, in SKILL.md)
Related to: #12 (session transcript layer)

Proposal

Add a feedback tool to Squall MCP:

squall feedback <review_id> <model> <score>

• review_id: matches the results file timestamp
• model: which model's output is being rated
• score: 0 (noise), 1 (okay), 2 (actionable)

Schema change

ALTER TABLE model_events ADD COLUMN feedback_score INTEGER DEFAULT NULL;

Recommendation update

Extend query_recommendations to weight by feedback:

bayesian_success = (successes + feedback_up + 2) / (quality_n + feedback_n + 5)

UX

After a review synthesis, Claude can prompt: "Rate any standout models? (e.g. feedback 20260310 codex 2)" — but this should be optional, never blocking.

Implementation (~200 LoC)

1. New tool handler in src/server.rs (~30 lines)
2. Schema migration in src/memory/schema.rs (~10 lines)
3. Feedback write path in src/memory/local.rs (~40 lines)
4. Update query_recommendations in src/memory/global.rs (~30 lines)
5. Tests (~80 lines)

Success criteria

• Feedback scores persist in DuckDB and local models.md
• memory recommend output reflects feedback-weighted rankings
• No workflow disruption — feedback is always optional

Labels

ACT, memory, UX