Summary

v0.4.1 — Context Optimization. Intercepts Claude Code Read tool re-reads to reduce token consumption via AST skeleton caching.

• AST utils refactor — new src/ast-utils.ts with shared findImports(), findFunctions(), findClasses(), findExports(). 4 detectors refactored, 68 lines of duplication removed.
• Skeleton extraction — compresses source files to structural summaries (imports + function signatures + class outlines). 91% reduction on engine.ts (684 lines → 49).
• SQLite cache — bun:sqlite with WAL mode. Tracks skeletons, session reads, and token savings per session.
• Hook handlers — PreToolUse smart-limits unchanged re-reads to 30 lines + skeleton. PostToolUse caches skeleton after each read. PostCompact prunes old sessions.
• vibecop init --context — generates Claude Code hooks with conflict detection. Resolves absolute paths for both local dev and npm installs.
• vibecop context benchmark — scans project, shows per-file compression ratios and projected savings. Runs under node (no bun required).
• MCP tool: vibecop_context_benchmark — same benchmark accessible through MCP server.
• Documentation — README, CHANGELOG, CONTRIBUTING, Starlight docs (architecture, MCP, Claude Code, CLI reference) all updated.

Test Coverage

610 tests pass (42 new). Type check clean. Lint clean. Both builds succeed.

New tests:
  test/ast-utils.test.ts           — 28 tests (findImports, findFunctions, findClasses, findExports)
  test/context/skeleton.test.ts    — 12 tests (extraction quality, edge cases)
  test/context/cache.test.ts       — 15 tests (SQLite ops, WAL mode, pruning)
  test/context/integration.test.ts — 11 tests (end-to-end hook protocol)
  test/init.test.ts                — 4 new tests (--context flag, conflict detection)

Validation Results

22/22 validation checks pass (bun scripts/validate-context.ts):

• Skeleton quality: 91% reduction on engine.ts, ~4,446 tokens saved per re-read
• Latency: P90 pre-hook 21ms (budget: <200ms)
• Protocol: all edge cases pass through cleanly (invalid JSON, missing session_id, unsupported files, partial reads, changed files)

Benchmark (on vibecop itself)

Files:        147 supported
Total tokens: ~171,448

Projected savings per session:
  20% re-read rate:  ~56,117 tokens saved (33% of total Read usage)
  40% re-read rate:  ~85,810 tokens saved (50% of total Read usage)
  60% re-read rate:  ~106,543 tokens saved (62% of total Read usage)

Architecture

Test plan

• 610 tests pass (bun test)
• Type check clean (tsc --noEmit)
• Lint clean (biome check)
• CLI build succeeds (dist/cli.js)
• Context build succeeds (dist/context.js)
• Benchmark runs under pure node (node dist/cli.js context benchmark)
• Stats runs under bun (bun dist/context.js stats)
• MCP tool handler returns structured JSON
• Zero bun:sqlite references in dist/cli.js

🤖 Generated with Claude Code