Task: Scoring engine

Description

Implement the core scoring engine that computes a composite 0-100 quality score for each AI response. The engine calculates three sub-scores (code correctness, user feedback, static analysis) and combines them using configurable weights. Scoring runs asynchronously to avoid blocking response delivery.

Acceptance Criteria

• packages/opencode/src/quality/scorer.ts created
• Correctness sub-score (0-100): based on tool invocation success/failure, error presence in response, LSP diagnostic delta after changes
• User feedback sub-score (0-100): starts at neutral (70), adjusted retroactively when signals arrive (accept: +, revert: -30, redo: -15, edit-after-apply: -5)
• Static analysis sub-score (0-100): based on LSP diagnostic count delta (new errors: heavy penalty, new warnings: moderate, resolved issues: bonus)
• Composite score = weighted sum using config weights, clamped to 0-100
• Scoring is async — uses queueMicrotask() or setTimeout(0) after response delivery
• Partial scoring supported — if LSP data unavailable, score with available signals only (weight redistributed)
• Score results persisted to SQLite via store.ts
• Scoring errors caught and logged, never propagated to main workflow

Technical Details

• Import config weights from quality config
• Import store functions from store.ts for persistence
• Correctness signals extracted from message metadata: tool results (success/error), response errors, token usage patterns
• Static analysis signals from LSP diagnostics snapshot before/after response application
• User feedback is deferred — initial score computed without it, updated retroactively
• Expose Quality.scoreResponse(messageMetadata, lspBefore, lspAfter) as main API

Dependencies

• Task 001: SQLite storage layer (for persistence)
• Task 002: Quality configuration (for weights and settings)

Effort Estimate

• Size: L
• Hours: 8-10
• Parallel: false (depends on 001, 002)