Builds out the rest of the LLM enrichment roadmap on top of the semantic-search/RAG foundation. All features remain fully optional; codegraph behaves identically when no LLM is reachable. Tier 1 #3 — directory/module summaries: - Schema v6 adds directory_summaries(dir_path PK, summary, content_hash, model, generated_at) - src/llm/dir-summarizer.ts aggregates symbol summaries per dir into one paragraph (≤600 chars), keyed by stable content_hash; min 3 summarised symbols per dir as the threshold for "worth a paragraph" - Wired into background pass after summaries+embeddings - New MCP tool codegraph_module (per-dir lookup, or list-all when dirPath omitted) Tier 2 #4 — change-intent helper: - src/llm/change-intent.ts: standalone summarizeChange(client, name, kind, beforeBody, afterBody). Three modes (added | removed | modified) with shape-appropriate prompts. - CodeGraph.summarizeChange wrapper; no caching at this layer (PR-review tooling owns lifecycle), no MCP/CLI surface yet. Tier 2 #5 — role classification: - Schema v7 adds role + role_model columns on symbol_summaries plus idx_summaries_role - Closed label set: api_endpoint | business_logic | data_model | util | framework_glue | test_helper | unknown - parseRole tolerates punctuation, fences, and multi-word "Business Logic" variants by snake_casing as a fallback - Wired into background pass after dir summaries - New MCP tool codegraph_role (filter by label); CodeGraph methods findNodesByRole, getRoleCounts Tier 3 #7 — dead-code judge: - SQL pre-filter findOrphanedSymbols: kind in {function|method| class|component} + is_exported=0 + zero incoming `calls` edges, with path-pattern exemptions for tests/scripts/bench - LLM judge returns JSON {verdict: dead|live|uncertain, confidence, reason}; tolerant fence-stripper (incl. stray backticks) and confidence clamp [0..1]; results sorted dead-first then by confidence - New MCP tool codegraph_dead_code with verdict filter Tier 3 colbymchenry#8 — naming-drift checker: - sampleSiblingNames pulls ≤30 same-kind names from other files - LLM judge returns {consistent, suggestion, reason}; defaults to "consistent" on parse failure or <5 siblings (advisory, not enforcement) - CodeGraph.checkNamingDrift wrapper; designed for sync-hook callers, no MCP surface yet Agent-as-LLM bridge (NEW — works without any local LLM): - src/llm/agent-bridge.ts exposes a contract for the agent in the user's session (Claude Code, etc.) to fill the summary cache directly when no Ollama is available. - codegraph_pending_summaries returns un-summarised symbols with bodies + content_hash for the agent to summarise. - codegraph_save_summaries persists agent results, re-validating content_hash against current disk to reject stale answers. - Cache shape is identical to the local-LLM path; the two coexist cleanly. - Reuses validatePathWithinRoot for body reads; same SUMMARIZABLE_KINDS / MIN_BODY_LINES filter as local pass for cache-key parity. Reviewer-flagged fixes (from independent review of the tier work): - dir-summarizer cache-hit branch no longer double-increments `done` (the `finally` always runs on `continue`) - classifier parseRole gains snake_case fallback so "Business Logic" and similar multi-word responses no longer silently degrade to "unknown" - dead-code parseJudgeResponse strips stray backticks before JSON.parse so fenced multi-line responses don't fall back to uncertain - Post-chat abort guards added to summarizer, dir-summarizer, classifier, dead-code: writes after a close()-cancelled chat are now skipped instead of persisted as stale entries - dir-summarizer module doc acknowledges prompt-injection accepted risk (data originates from user's own codebase) - dead-code test replaces tautological assertion with meaningful judged ≤ candidates / errors == 0 checks - parseRole exported for direct testing; new test covers canonical, fenced, multi-word, and trailing-punct inputs Tests: 7 new in __tests__/llm-tiers.test.ts - Background pass writes both directory summaries and role labels - summarizeChange added/removed/modified - Dead-code judge returns parsed verdicts in valid range, no errors - parseRole handles all four input shapes - Agent bridge round-trip without LLM (pending → save → coverage reflects writes; second batch short-circuits on cache) - Agent bridge stale-content-hash rejection Schema migrations cover v4 → v5 → v6 → v7 cleanly. Test schema version expectations bumped. Full suite 409/409 passing including the watcher tests that were previously flaky. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Svelte 5 #3

Description

Metadata

Metadata

Assignees

Labels

Projects

Milestone

Relationships

Development

Issue actions