token-pilot 0.19.1 → 0.22.2

package/CHANGELOG.md

@@ -1,580 +1,736 @@
-# Changelog
-
-All notable changes to Token Pilot will be documented in this file.
-
-The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.1.0/),
-and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
-
-## [0.19.1] - 2026-04-15
-
-### Added
-- **`decisions` field in `session_snapshot`** — stores key decisions with reasoning (e.g., "removed sysfee step — caused double counting"). Prevents the model from revisiting rejected approaches after context compaction.
-
-## [0.19.0] - 2026-04-15
-
-### Added
-- **`session_snapshot` tool** — capture current session state (goal, confirmed facts, files, blockers, next step) as a compact markdown block (<200 tokens). Call before context compaction or when switching direction in long sessions.
-- **`max_tokens` parameter** on `smart_read` and `smart_read_many` — token budget per read. Output auto-downgrades through three levels: full content → structural outline → compact (symbol names + line ranges only). Enables context-constrained sessions.
-- **Session compaction advisory** — policy engine now tracks total tool calls and tokens returned. Advises calling `session_snapshot()` when thresholds are reached (default: every 15 calls or after 8,000 tokens). Configurable via `compactionCallThreshold` and `compactionTokenThreshold`.
-- **"Why This Approach Works"** section in README explaining the 3-level optimization strategy.
-
-### Changed
-- **21 tools** (was 20) — added `session_snapshot`.
-- **MCP instructions** updated with `session_snapshot` workflow and `max_tokens` guidance.
-- Benchmark numbers updated: 55 files, 102K raw → 9K outline tokens (91% savings).
-
-## [0.18.0] - 2026-04-05
-
-### Added
-- **`read_section` tool** — read a specific section from Markdown, YAML, JSON, or CSV files. Markdown: by heading name. YAML/JSON: by top-level key. CSV: by row range (`rows:1-50`). Much cheaper than reading the whole file.
-- **`read_for_edit` section parameter** — prepare edit context for non-code file sections. Works with all 4 formats.
-- **Markdown outline with line ranges** — `smart_read` on `.md` files now shows `[L5-20]` ranges and hints for `read_section`.
-- **YAML/JSON section ranges** — `smart_read` on `.yaml`/`.json` shows top-level key ranges.
-- **CSV smart_read** — shows columns, row count, sample rows, and hints for row-range reading.
-- **4 section parsers** — `markdown-sections.ts`, `yaml-sections.ts`, `json-sections.ts`, `csv-sections.ts`.
-
-### Changed
-- **20 tools** (was 19) — added `read_section`.
-- **492 tests** (was 441).
-
-### Fixed
-- `npm audit` — resolved brace-expansion, path-to-regexp, picomatch vulnerabilities.
-
-## [0.17.0] - 2026-04-02
-
-### Added
-- **`smart_read` scope parameter** — `scope="nav"` returns names + line ranges only (2-3x smaller), `scope="exports"` shows only public API. Default `scope="full"` unchanged.
-- **`smart_read` auto-delta** — when a file changed since last load (within 120s), shows ADDED/REMOVED/UNCHANGED symbols instead of full re-read. Config: `smartRead.autoDelta.enabled`.
-- **`read_symbol` include_edit_context** — optional `include_edit_context=true` appends raw code block (max 60 lines) to save a separate `read_for_edit` call. Large symbols fall back to `read_for_edit`.
-- **`find_usages` mode=list** — compact `file:line` output for initial discovery, 5-10x smaller than full mode.
-- **`smart_read_many` per-file dedup** — skips files already in context and unchanged, returns compact reminder instead.
-- **Actionable hints** — `read_for_edit` suggests `read_diff` after editing. Config: `display.actionableHints`.
-- **`symbol-display-constants.ts`** — shared display constants for symbol rendering.
-
-### Changed
-- **441 tests** (was 427) — new tests for scope, list mode, include_edit_context, dedup.
-- **MCP instructions** updated with scope/mode/include_edit_context guidance.
-- **find_usages context rendering** — sequential instead of concurrent to prevent shared cache race condition.
-
-## [0.16.1] - 2026-03-21
-
-### Added
-- **Hook interception tracking** — PreToolUse hook now records denied Read calls (file path, line count, estimated tokens) to `.token-pilot/hook-denied.jsonl`. Session analytics shows how many tokens the hook saved by intercepting unbounded reads on large code files.
-- **`session_analytics` hook savings** — compact report adds "Hook: intercepted N reads, saved ~X tokens" line. Verbose mode shows per-file breakdown of intercepted reads.
-
-## [0.16.0] - 2026-03-21
-
-### Added
-- **`read_symbols` tool** — batch read multiple symbols from one file in a single call (max 10). File is read once, AST resolved once. Saves N-1 round-trips vs calling `read_symbol` N times.
-- **`read_for_edit` batch mode** — new `symbols` array parameter reads multiple symbol edit contexts in one call. Each symbol returns raw code ready for Edit tool's `old_string`.
-- **`find_usages` context_lines** — new `context_lines` parameter (0-10) shows surrounding source code for each match. Eliminates follow-up `read_symbol` calls after finding usages.
-- **`smart_diff` affected symbols summary** — consolidated "AFFECTED SYMBOLS" section at the top of diff output, grouped by MODIFIED/ADDED/REMOVED. See all changed functions/classes at a glance.
-
-### Changed
-- **19 tools** (was 18) — added `read_symbols`.
-- **MCP instructions** — added batch read_symbols to decision rules and refactor workflow.
-- **427 tests** (unchanged — all pass with new features).
-
-## [0.15.0] - 2026-03-19
-
-### Added
-- **Regex fallback parser (TS/JS)** — `smart_read` now works for TypeScript/JavaScript files even without ast-index binary. Parses classes, functions, interfaces, types, enums, and class methods via regex. Zero dependencies, 130 lines. Covers ~80% of new users who fail to download ast-index.
-- **Regex fallback parser (Python)** — `smart_read` now works for Python files without ast-index. Parses classes, functions, async functions, decorators (`@dataclass`, `@app.route`), module constants (`UPPER_CASE`), methods with visibility detection (`_private`, `__dunder__`). 150 lines.
-- **Benchmark script** — `scripts/benchmark.ts` measures real token savings on public repos (express, fastify, flask). 92% average savings across 97 files ≥50 lines. Run: `npx tsx scripts/benchmark.ts`.
-- **Guide skill** — `/guide` command shows a quick-reference table of all Token Pilot tools with usage examples and recommended workflow.
-- **`hooks.denyThreshold` config** — hook deny threshold is now configurable in `.token-pilot.json` (default: 300, was hardcoded 500). Intercepts ~2x more native Read calls.
-
-### Changed
-- **Compact session analytics** — `session_analytics` report reduced from ~30 lines to ~5 lines. Shows calls, tokens saved, top 5 tools, top 3 files, cache hit rate on a single screen. Verbose mode (`verbose=true`) restores full breakdown.
-- **`server.ts` refactor** — extracted tool definitions to `server/tool-definitions.ts` and token estimate helpers to `server/token-estimates.ts` (−500 lines from server.ts).
-- **`find_usages` output** — results grouped by file with compact rendering. Single match per file on one line, multiple matches indented under file header.
-- **Stale references** — all `grep_search` hints updated to `Grep` (code-audit, find-unused, find-usages).
-- **README** — benchmark table with real data from 4 public repos. Updated savings claims from 80% to 90% (backed by benchmark).
-- **427 tests** (was 393).
-
-### Fixed
-- **`npx token-pilot` CLI** — symlink path resolution in `isDirectRun` check. All CLI commands now work correctly via npx.
-- **Regex fallback was dead code** — parsers existed but weren't wired into `client.ts` `outline()` method. Now properly called as fallback when ast-index unavailable.
-
-## [0.14.1] - 2026-03-14
-
-### Fixed
-- **CI: Node.js 24 runtime** — opted into `FORCE_JAVASCRIPT_ACTIONS_TO_NODE24` for GitHub Actions, resolving deprecation warnings for `actions/checkout@v4` and `actions/setup-node@v4`.
-- **CI: test matrix** — updated from Node 18+22 to Node 20+22 (Node 18 is EOL).
-- **Test: git commit in CI** — `read-for-edit` tests now pass `-c user.name` / `-c user.email` to `git commit`, fixing failures in environments without global git config.
-
-## [0.14.0] - 2026-03-14
-
-### Added
-- **R&D Track 0: Instrumentation** — per-call decision trace capturing file size, context state, estimated vs actual cost, and cheaper alternative suggestions. Integrated into all 18 tool handlers via `recordWithTrace()`.
-- **R&D Track 1: Budget Planner** — advisory layer suggesting cheaper tool alternatives (e.g. `smart_read` → `read_diff` when file already in context, → `read_symbol` when symbol known). Analytics-only, no blocking.
-- **R&D Track 2: Intent Router** — classifies tool calls into 7 intents (edit/debug/explore/review/analyze/search/read). Per-intent breakdown in session analytics.
-- **R&D Track 3: Edit Prep Mode** — `read_for_edit` with `include_callers`, `include_tests`, `include_changes` enrichment options.
-- **R&D Track 4: Session Cache** — tool-result-level caching with file/AST/git invalidation.
-- **R&D Track 5: Confidence-Based Escalation** — confidence metadata (high/medium/low) appended to `smart_read`, `read_symbol`, `read_for_edit`, `find_usages` responses. Shows known unknowns and suggested next steps.
-- **R&D Track 6: Working Set / Dedup** — compact reminders for already-loaded files and symbols.
-- **R&D Track 7: Related Files Ranking** — scored ranking with 6 signals (test +5, import +4, importer +3, same-dir +2, recently-changed +2, multi-ref +1). HIGH VALUE / MEDIUM / LOW buckets.
-- **R&D Track 8: Architecture Fingerprint** — caches architecture in `.token-pilot-fingerprint.json` (24h TTL). Amortizes `project_overview` cost across sessions.
-- **R&D Track 9: Verified Savings Dashboard** — savings breakdown by category (compression/cache/dedup), session cache hit rate, dedup stats.
-- **R&D Track 10: Team Policy Mode** — configurable policies: `preferCheapReads`, `maxFullFileReads`, `warnOnLargeReads`, `requireReadForEditBeforeEdit`.
-- **7 new core modules** — `confidence.ts`, `intent-classifier.ts`, `budget-planner.ts`, `decision-trace.ts`, `session-cache.ts`, `architecture-fingerprint.ts`, `policy-engine.ts`.
-- **35 new tests** — confidence (11), architecture-fingerprint (11), policy-engine (13). Total: 393 tests.
-
-### Changed
-- **`session_analytics`** — per-intent breakdown, decision insights, savings by category.
-- **`project_overview`** — saves/loads architecture fingerprint for cross-session caching.
-- **Config** — added `policies` section to `TokenPilotConfig`.
-
-## [0.13.0] - 2026-03-14
-
-### Added
-- **Version check for all components** — on startup, checks token-pilot (npm), ast-index (GitHub releases), and context-mode (npm) in parallel. Non-blocking, fire-and-forget. Shows update notifications in stderr.
-- **`autoUpdate` config flag** — `updates.autoUpdate: true` in `.token-pilot.json` auto-downloads new ast-index binary on startup. Default: `false` (notify only). token-pilot and context-mode only notify (separate processes).
-- **`checkBinaryUpdate()`** — compares installed ast-index version vs latest GitHub release.
-- **`isNewerVersion()` utility** — semver comparison: strip `v` prefix, compare segments. Handles different lengths (`1.0` vs `1.0.1`).
-- **Common Lisp extensions** — `.lisp`, `.lsp`, `.cl`, `.asd` added to `CODE_EXTENSIONS` for ast-index v3.28+ compatibility.
-- **9 new tests** — `isNewerVersion()` covering major/minor/patch, same version, older, `v` prefix, different segment lengths, large numbers, real-world versions. Total: 217 tests.
-
-### Changed
-- **`doctor` command** — now shows 3 sections: token-pilot (installed/latest), ast-index (installed/latest/auto-update status), context-mode (detected/latest npm). Previously only showed ast-index binary status.
-- **`install-ast-index` command** — now also updates existing binary if newer version available on GitHub.
-- **`printHelp()`** — fixed tool count: 18 (was incorrectly showing 12 since v0.8.0).
-- **Startup update check** — replaced single `checkLatestVersion()` with `checkAllUpdates()` covering all 3 components via `Promise.allSettled`.
-
-### Fixed
-- **`test_summary` PHPUnit parser** — now counts both `Failures:` and `Errors:` (was only counting failures).
-- **`test_summary` cargo parser** — correctly identifies failure name-list section (no `----` markers) vs detail section.
-- **`test_summary` token estimation** — uses shared `estimateTokens()` instead of local duplicate.
-- **`smart_log` category detection** — `documentation` now matches docs pattern, `tests` (plural) matches test pattern, `optimize`/`optimization` match perf pattern.
-- **`explore_area` path boundary** — `startsWith(path + '/')` prevents `src/auth` matching `src/authorize/`.
-- **Validation consistency** — `validateSmartLogArgs` and `validateTestSummaryArgs` now use `optionalString`/`optionalNumber` helpers, reject empty strings, check integers.
-
-## [0.11.0] - 2026-03-14
-
-### Added
-- **`smart_log` tool** — structured git log with commit category detection (feat/fix/refactor/docs/test/chore/style/perf). Shows author breakdown, file stats (+/-), per-commit file list. Filters by path and ref. Raw git log → compact summary.
-- **`test_summary` tool** — runs test command and returns structured summary: total/passed/failed/skipped + failure details. Parsers for vitest, jest, pytest, phpunit, go test, cargo test, rspec, mocha + generic fallback. 200 lines of raw output → 10-15 lines.
-- **38 new tests** — smart_log parser (5), categorizer (4), test_summary parsers (17), runner detection (8), validation (4). Total: 208 tests (was 170).
-
-### Changed
-- **18 tools** (was 16) — added `smart_log`, `test_summary`
-- **MCP instructions** — added smart_log and test_summary to workflow guidance
-
-## [0.10.0] - 2026-03-14
-
-### Added
-- **`smart_diff` tool** — structural git diff with AST symbol mapping. Shows which functions/classes were modified/added/removed instead of raw patch output. Supports scopes: `unstaged`, `staged`, `commit` (ref required), `branch` (ref required). Small diffs (<=30 lines) include actual hunks, large diffs show summary. Returns `rawTokens` for precise savings analytics.
-- **`explore_area` tool** — one-call directory exploration combining outline + imports + tests + git changes. Replaces 3-5 separate tool calls when starting work on an area. Sections: `outline` (recursive depth 2), `imports` (external deps + who imports this area), `tests` (matching test/spec files), `changes` (recent git log). All sections run in parallel via `Promise.allSettled`.
-- **26 new tests** — smart_diff parser (10), symbol mapping (5), validation (11). Total: 170 tests (was 144).
-
-### Changed
-- **16 tools** (was 14) — added `smart_diff`, `explore_area`
-- **MCP instructions** — updated workflow: `project_overview → explore_area → smart_read → read_symbol → read_for_edit → edit → smart_diff`
-- **`outlineDir` and `CODE_EXTENSIONS` exported** from outline.ts for reuse by explore_area
-
-## [0.9.0] - 2026-03-08
-
-### Added
-- **`module_info` tool** — analyze module dependencies, dependents, public API, and unused deps. Uses ast-index v3.27.0 module commands (`modules`, `module-deps`, `module-dependents`, `module-api`, `unused-deps`). Includes degradation check when ast-index is unavailable.
-- **`project_overview` dual-detection** — shows BOTH ast-index type detection AND config-file detection (package.json, composer.json, Cargo.toml, pyproject.toml, go.mod) with CONFIDENCE scoring (high/medium/low/unknown). Detects frameworks, quality tools (PHPStan, ESLint, Vitest, Jest, Biome, etc.), CI pipelines (GitHub Actions, GitLab CI, Jenkins), and Docker.
-- **`project_overview` `include` parameter** — filter sections: `["stack"]` for quick type check, `["quality","ci"]` for tooling overview. Default: all sections.
-- **`find_usages` post-filters** — `scope` (path prefix), `kind` (definitions/imports/usages), `lang` (14 languages by extension), `limit` (per category, 1-500). All filters optional, backward compatible.
-- **`outline` recursive mode** — `recursive=true` with `max_depth` (default 2, max 5) recurses into subdirectories. At max depth shows file counts only.
-- **`src/core/project-detector.ts`** — extracted config-based detection logic into reusable module. Framework detection maps for PHP (7), JS (10), Python (5). Quality tools scanner (13 tools). CI pipeline detector (6 platforms).
-- **ast-index client: 5 module methods** — `modules()`, `moduleDeps()`, `moduleDependents()`, `unusedDeps()`, `moduleApi()` with JSON-first + text fallback parsing.
-- **ast-index types: 4 module interfaces** — `AstIndexModuleEntry`, `AstIndexModuleDep`, `AstIndexUnusedDep`, `AstIndexModuleApi`.
-
-### Fixed
-- **`module_info` token savings** — `tokensWouldBe` was equal to `tokensReturned` (0% savings). Now estimates manual analysis cost correctly.
-- **`outline` recursive overflow** — added `MAX_OUTLINE_LINES=500` guard to prevent runaway output on large projects with `recursive=true`.
-- **`project_overview` "frontend" label** — removed hardcoded "frontend" suffix for secondary stacks (Node.js is not always frontend).
-- **Ruff detection** — no longer double-reads `pyproject.toml`. Checks `ruff.toml`/`.ruff.toml` first, falls back to `pyproject.toml [tool.ruff]` only if needed.
-- **44 new tests** — validators (23) + project-detector (21). Total: 144 tests (was 100).
-
-### Changed
-- **14 tools** (was 13) — added `module_info`
-- **Tool descriptions** — updated with `(v1.1: ...)` version hints for enhanced tools
-- **MCP instructions** — added module_info to "COMBINE BOTH" workflow section
-- **Version sync** — package.json, plugin.json, marketplace.json all at 0.9.0
-
-## [0.8.3] - 2026-03-08
-
-### Fixed
-- **code_audit pattern search — root cause fix** — `ast-index agrep` does not support `--limit` flag. Token Pilot was passing `--limit 50` which caused the command to fail silently, returning 0 results across v0.8.0–v0.8.2. Removed the flag; results are now limited via `.slice()` after parsing.
-
-## [0.8.2] - 2026-03-08
-
-### Fixed
-- **code_audit pattern search** — inject `node_modules/.bin` into PATH so `ast-index agrep` can find `sg` (ast-grep) when it's installed as optional dependency but not in system PATH.
-- **code_audit annotations** — strip `@` prefix from annotation names (`@Injectable` → `Injectable`). ast-index expects names without `@`.
-
-## [0.8.1] - 2026-03-08
-
-### Added
-- **ast-grep auto-install** — `@ast-grep/cli` added as optional dependency. `code_audit(check="pattern")` now works out-of-the-box without manual `brew install ast-grep`.
-- **MCP instructions: security audit guidance** — instructions now recommend Grep for security patterns (password, token, secret, credential) and `find_unused` for dead code detection.
-
-### Changed
-- **ast-index stats → JSON parsing** — `--format json` for reliable file count extraction instead of regex on text output.
-
-## [0.8.0] - 2026-03-07
-
-### Added
-- **`code_audit` tool** — find code quality issues in one call: TODO/FIXME comments (`check="todo"`), deprecated symbols (`check="deprecated"`), structural code patterns via ast-grep (`check="pattern"`), decorator search (`check="annotations"`), or combined audit (`check="all"`).
-- **Incremental index update on file changes** — file watcher now triggers `ast-index update` (debounced 2s) after edits. Keeps index fresh for find_usages, find_unused, code_audit.
-- **ast-index client methods** — `agrep()`, `todo()`, `deprecated()`, `annotations()`, `incrementalUpdate()`.
-
-### Fixed
-- **smart_read on directories** — now returns helpful message instead of EISDIR crash.
-- **MCP instructions** — added "COMBINE BOTH" section for audit tasks (Token Pilot + Grep).
-
-## [0.7.6] - 2026-03-07
-
-### Added
-- **`npx token-pilot init`** — one command creates `.mcp.json` with both token-pilot and context-mode configured. Idempotent — safely updates existing configs without overwriting.
-- **MCP Server Instructions** — protocol-level `instructions` field tells AI agents WHEN to use Token Pilot tools instead of built-in defaults. Works universally on all MCP clients.
-- **Improved tool descriptions** — each tool explicitly states what built-in tool it replaces (e.g. "Use INSTEAD OF Read/cat").
-
-### Fixed
-- **3 high severity vulnerabilities** — updated hono and express-rate-limit.
-- **npm package size** — excluded source maps from package. 505 kB → 286 kB (−43%).
-- **Accurate thresholds** — README and instructions now correctly state smallFileThreshold=200 (was 80).
-- **read_diff documentation** — clarified that smart_read must be called BEFORE editing to create baseline snapshot.
-
-### Changed
-- **README** — honest metrics (60-80%), Quick Start with `init` command, MCP instructions section, Codex/Antigravity support.
-- **npm keywords** — added `codex`, `cline`, `model-context-protocol`, `token-savings`.
-
-## [0.7.4] - 2026-03-07
-
-### Added
-- **MCP Server Instructions** — protocol-level `instructions` field tells AI agents WHEN to use Token Pilot tools instead of built-in Read/cat/Grep. Works universally on Claude Code, Cursor, Codex, Antigravity, and any MCP-compatible client. Includes rules for when NOT to use Token Pilot (regex search, raw content copy-paste).
-- **Improved tool descriptions** — each tool now explicitly states what built-in tool it replaces (e.g. "Use INSTEAD OF Read/cat", "Use INSTEAD OF Grep/ripgrep"). Agents can make informed decisions from description alone, without needing project-level rules files.
-
-## [0.7.3] - 2026-03-07
-
-### Fixed
-- **read_diff diagnostic** — when cache miss occurs, now shows resolved absolute path and all cached file paths. This reveals path mismatches between smart_read and read_diff calls (e.g. different relative paths resolving to different absolute paths).
-
-## [0.7.2] - 2026-03-07
-
-### Fixed
-- **read_diff on small files** — `smart_read` small-file pass-through (≤150 lines) returned content without caching in fileCache. `read_diff` always showed "No previous read" for small files because the baseline was never stored. Now all files are cached regardless of size.
-
-## [0.7.1] - 2026-03-07
-
-### Fixed
-- **read_diff after read_for_edit** — `read_for_edit` now caches the full file content, so `read_diff` can use it as baseline after edits. Previously returned "No previous read" because read_for_edit didn't populate the file cache.
-- **outline on intermediate directories** — directories with only subdirectories (no direct code files) now show subdirectory listing with recursive code file counts instead of "No code files found". Enables progressive drill-down: `outline("module/") → outline("module/infrastructure/")`.
-
-## [0.7.0] - 2026-03-07
-
-### Fixed
-- **Project root detection** — complete rewrite of how token-pilot discovers the working project:
-  1. **MCP roots** (new, primary) — uses MCP protocol `listRoots()` to get workspace root from Claude Code. Works for all tools including `find_usages`, `find_unused`, `project_overview` (no file path needed).
-  2. **INIT_CWD/PWD env vars** (new) — when started via `npx`, npm sets `INIT_CWD` to the invoking directory. Catches cases where `process.cwd()` is `/` but the real project root is available in env.
-  3. **Git detect from file path** (improved) — now triggers from any tool call args (`path`, `paths`, `file`, `module`), not just `smart_read`.
-- **ast-index tools always disabled** — `find_usages`, `find_unused`, `project_overview` never triggered auto-detect because they have no `path` argument. Now all tools trigger detection via MCP roots.
-- **Error messages** — changed "project root is too broad" to actionable "call smart_read() on any project file first" when MCP roots unavailable.
-- **`isDangerousRoot`** — moved to shared `core/validation.ts` (was duplicated in `index.ts`).
-
-## [0.6.5] - 2026-03-07
-
-### Fixed
-- **AST index rebuild race condition** — concurrent tool calls no longer trigger multiple simultaneous rebuilds. `ensureIndex()` now deduplicates via shared promise. If rebuild fails due to lock file (another process running), falls back to existing index if available instead of throwing.
-- **Rebuild timeout** — increased from 60s to 120s for large projects where indexing takes longer.
-
-## [0.6.4] - 2026-03-07
-
-### Fixed
-- **CRITICAL: Hook installer** — malformed `settings.json` no longer silently destroyed. Distinguishes ENOENT (create fresh) from JSON parse error (abort with message). Uninstall also reports specific errors.
-- **CRITICAL: Server startup** — `startServer()` now has `.catch()` handler. Unhandled promise rejections no longer crash the process silently.
-- **Non-code handler** — removed `.xml` and `.csv` from `isNonCodeStructured` (no handler existed for them, fell through to null).
-- **Symbol resolver** — removed dangerous basename-only fallback in `pathMatches` (`index.ts` no longer matches any `index.ts`). Fixed hardcoded `endLine = start_line + 10` → uses `end_line` from ast-index or 50-line fallback.
-- **Config loader** — added prototype pollution guard (`__proto__`, `constructor`, `prototype` keys skipped in deepMerge). Parse errors now logged instead of silently swallowed.
-- **File cache** — size tracking now uses `Buffer.byteLength()` instead of `string.length` (chars ≠ bytes for non-ASCII). Removed dead `isSmallFile()` method.
-- **Validation** — `optionalNumber` now rejects `NaN` and `Infinity`.
-- **Token estimation** — `smart_read_many` now uses `estimateTokens()` instead of `length/4`.
-- **Analytics** — `project_overview` calls now tracked in session analytics.
-- **read_for_edit** — raised `MAX_EDIT_LINES` from 20 to 60 (20 was too aggressive, truncated most functions).
-- **related_files** — raised symbol search limit from 5 to 10 for reverse import detection.
-
-### Removed
-- Dead config options `cache.ttlMinutes` and `context.autoForgetMinutes` (declared but never used).
-
-## [0.6.3] - 2026-03-03
-
-### Changed
-- **Hook deny threshold** — raised from 200 to 500 lines. Files ≤500 lines pass through Read without denial roundtrip. Eliminates token overhead on medium files where hook denial costs more than outline saves.
-- **Adaptive fallback** — lowered from 90% to 70%. If outline ≥70% of raw file size, returns raw content. More aggressive at avoiding outlines that barely save tokens.
-- **Tool descriptions** — trimmed marketing language, percentages, and cross-references. ~250 fewer tokens in tool list per session.
-- **Outline cap** — top-level symbols capped at 40, class members at 30. Prevents outline explosion on files with 100+ methods.
-
-## [0.6.2] - 2026-03-02
-
-### Removed
-- **Dead handler files** — deleted `changed-symbols.ts` (removed in v0.5.0) and `find-callers.ts` (removed in v0.4.0). Were never registered in server but lingered as dead code.
-
-## [0.6.1] - 2026-03-02
-
-### Changed
-- **`smallFileThreshold`** — raised from 80 to 200 lines. Benchmark showed medium files (100-300 lines) had negative savings (-25%) because AST outline was larger than the raw file. Files ≤200 lines now pass through as raw content.
-- **`smart_read` adaptive fallback** — after generating outline, compares token count vs raw file. If outline ≥ 90% of raw size, returns raw content instead. Eliminates negative savings on any file size, regardless of language or threshold.
-- **`session_analytics` honest metrics** — replaced all hardcoded multipliers (`*5`, `*3`) with real full-file token counts from file cache. `tokensWouldBe` now reflects actual file size, not fabricated numbers. Non-file tools (related_files, outline, find_usages) report 1:1 (no savings claim).
-
-## [0.6.0] - 2026-03-02
-
-### Changed
-- **Read hook** — upgraded from advisory (`decision: "suggest"`) to blocking (`permissionDecision: "deny"`) for unbounded Read calls on large code files (>200 lines). Bounded Read (with offset/limit) is still allowed. Uses official `hookSpecificOutput` format per Claude Code docs.
-- **`read_for_edit` output** — already includes exact `Read(path, offset, limit)` command that passes through the hook, giving AI a clear path: `read_for_edit` → bounded `Read` → `Edit`.
-
-### Added
-- **Edit hook** — new PreToolUse hook matching Edit tool. Adds `additionalContext` suggesting `read_for_edit` for minimal code context. Doesn't block Edit — just provides a hint.
-- **Hook installer** — now installs and manages both Read and Edit hooks. Uninstall removes all Token Pilot hooks.
-
-## [0.5.3] - 2026-03-02
-
-### Changed
-- **`find_unused`** — completely rewritten with universal approach. Removed 60+ hardcoded framework-specific names. Now uses ast-index data: constructors filtered by name (`constructor`/`__init__`), Python dunder methods by `__*__` pattern, decorated symbols detected via `outline()` and shown separately with their decorators. No framework-specific knowledge.
-- **`formatFrameworkInfo`** (smart_read display) — removed hardcoded TypeORM (`Column`, `PrimaryGeneratedColumn`) and class-validator (`IsEmail`, `MinLength`) parsing. Now only detects standard HTTP verbs (GET/POST/PUT/DELETE/PATCH) which are protocol-level, not framework-specific. All other decorators shown as-is (`@DecoratorName`).
-- **`outline`** — route detection now universal. Instead of hardcoding `@Controller`, detects any class decorator with a path argument as route prefix. HTTP verb detection uses same universal pattern. Non-HTTP decorators shown as-is.
-
-## [0.5.2] - 2026-03-02
-
-### Fixed
-- **`project_overview`** — HINT no longer references deleted `search_code()`, now suggests `find_usages()` and `outline()`
-- **`related_files` imported_by** — now searches both `imports` AND `usages` from refs (not just imports), with increased limit (30). Cross-language filtering preserves same-family matches while removing false positives.
-- **`find_unused`** — excludes framework-implicit symbols (replaced by universal approach in 0.5.3)
-- **README** — updated handler file list (removed deleted handlers, added new ones)
-
-## [0.5.1] - 2026-03-02
-
-### Fixed
-- **`read_for_edit` symbol mode** — large symbols (>20 lines) now return only the first 20 lines instead of the entire method. Prevents returning 300+ lines when only a signature is needed for editing.
-- **`related_files` imported_by** — filter cross-language false positives. A TypeScript file no longer shows Python/Go/Rust files as importers. Refs are filtered by language family (JS/TS, Python, Go, JVM, etc.).
-- **`session_analytics`** — honest savings metric for `read_for_edit`. Reduced multiplier from 30x to 3x (realistic comparison vs `Read` with offset/limit, not vs full file).
-
-## [0.5.0] - 2026-03-02
-
-### Added
-- **`read_for_edit`** — killer feature for edit workflow. Returns RAW code (no line numbers) around a symbol or line, ready to copy as `old_string` for Edit. 97% fewer tokens than reading full file before editing.
-- **`related_files`** — import graph for any file: what it imports, what imports it, test files. Saves 3-5 Read calls per task.
-- **`outline`** — compact overview of all code files in a directory. One call instead of 5-6 smart_read calls. Framework-aware: shows HTTP routes for NestJS controllers.
-- **`read_symbol` show parameter** — `show: "full"|"head"|"tail"|"outline"` controls truncation. Default: auto (full ≤300 lines, outline >300).
-- **Framework-aware decorators** — smart_read/outline parse NestJS (`@Controller`+`@Get` → HTTP routes), TypeORM (`@Column` → types), class-validator (`@IsEmail` → constraints).
-
-### Removed
-- **`search_code`** — worse than Grep in practice, find_usages + Grep cover all use cases
-- **`export_ast_index`** — never used in real work, infrastructure tool only
-- **`context_status`** — debugging tool, not user-facing
-- **`forget`** — manual context management = poor design, should be automatic
-- **`changed_symbols`** — git diff + smart_read covers this use case
-
-### Changed
-- **12 focused tools** instead of 14 — removed 5 low-value, added 3 high-impact
-- Edit-heavy sessions: 5-10% → 40-50% token savings (via read_for_edit)
-- Average sessions: 20-25% → 45-55% token savings
-
-## [0.4.1] - 2026-03-02
-
-### Added
-- **Auto-install PreToolUse hook**: hook installs automatically on server start (Claude Code), no manual `install-hook` needed
-- **AI instructions template**: README includes ready-to-copy block for `.cursorrules` / `CLAUDE.md`
-
-### Changed
-- **Tool descriptions rewritten** — explicit "ALWAYS use instead of Read/cat", "use instead of Grep" for AI prioritization
-- README updated: PreToolUse hook section, MCP Tools table with "Instead of" column
-
-## [0.4.0] - 2026-03-02
-
-### Added
-- **Python class method parser**: smart_read/read_symbol shows all methods inside Python classes with visibility, decorators, async detection
-- **PHP class method parser**: same for PHP classes with public/private/protected, static
-- **Version display**: `project_overview` and `session_analytics` show `TOKEN PILOT v{version}`
-
-### Changed
-- **Removed find_callers** — did not save tokens vs grep, ast-index limitation with `this.method()` calls
-- **Removed find_implementations** — did not save tokens vs grep, ast-index limitation with decorators
-- **Removed class_hierarchy** — did not save tokens vs grep, poor results from ast-index
-- **14 focused tools** instead of 17 — only tools that actually save tokens or provide unique value
-
-### Fixed
-- **Mega-symbol truncation**: symbols >300 lines show head (50) + tail (30) + method outline instead of 71KB overflow
-- **Recursive findFlat**: unqualified method names (`run`, `_build_summary`) found inside class children
-
-## [0.3.2] - 2026-03-01
-
-### Fixed
-- **Python class methods**: smart_read now shows all methods inside Python classes (ast-index only returns class-level, token-pilot parses `def` methods with visibility, decorators, async detection)
-- **read_symbol Python**: `Orchestrator.run`, `Orchestrator._build_summary` — qualified and unqualified method access works (was returning entire 829-line class)
-- **Mega-symbol truncation**: symbols >300 lines show head (50) + tail (30) + method outline instead of 71KB overflow
-- **findFlat recursive**: unqualified method names (`run`, `_build_summary`) now found inside class children
-
-## [0.3.1] - 2026-03-01
-
-### Fixed
-- **find_usages**: combine `refs` + `search` with exact word boundary filtering — 0% result loss vs grep (was 40% loss with refs-only)
-- **read_symbol**: fix `Class.method` qualified names for flat outlines (ast-index lists methods as siblings, not children)
-- **read_symbol**: filter ast-index leaf name fallback by requested file (was returning symbols from wrong files)
-- **YAML smart_read**: 3-level nested parser with scalar values, array counts (was only showing top-level keys)
-- Removed all "Use Grep as fallback" hints — token-pilot gives complete results on its own
-
-## [0.3.0] - 2026-03-01
-
-### Added
-- **find_callers** tool — find all callers of a function, with optional call hierarchy tree (depth parameter)
-- **changed_symbols** tool — show symbol-level git changes (added/modified/removed) vs a base branch
-- **find_unused** tool — detect potentially unused/dead symbols in the project
-- 8 new ast-index client methods: `refs`, `map`, `conventions`, `callers`, `callTree`, `changed`, `unusedSymbols`, `fileImports`
-- Incremental index updates via `ast-index update` (fast) instead of full rebuild
-
-### Fixed
-- **find_usages**: rewritten to use `ast-index refs` — returns definitions + imports + usages in one call (was losing ~66% of results)
-- **project_overview**: rewritten to use `ast-index map` + `conventions` — shows architecture, frameworks, naming patterns, directory map with symbol kinds
-- **search_code**: deduplication of results (removes duplicate file:line entries)
-- **read_symbol**: structure-first lookup for `Class.method` qualified names with ast-index leaf fallback
-- **export_ast_index**: `all_indexed=true` option exports all files from ast-index, not just cached ones
-- **YAML smart_read**: expand one level of nesting (shows nested keys under top-level sections)
-
-### Changed
-- Total MCP tools: 14 → 17
-- ast-index commands used: 8 → 16
-- Index updates are now incremental by default (falls back to full rebuild only when needed)
-
-## [0.2.4] - 2026-03-01
-
-### Fixed
-- **search_code**: filter out garbage entries with empty file paths or `:undefined` lines
-- **read_symbol**: support `Class.method` and `Class::method` qualified names (structure-first lookup, ast-index leaf fallback)
-- **export_ast_index**: `all_indexed=true` option exports all files from ast-index, not just cached ones
-- **YAML smart_read**: expand one level of nesting (shows service names, nested keys under top-level sections)
-- Improved empty-cache message in export_ast_index with hint about `all_indexed`
-
-## [0.2.3] - 2026-03-01
-
-### Fixed
-- **ensureIndex**: plain rebuild first (indexes full monorepo), fallback to `--sub-projects` only if <5 files
-- **smart_read**: non-code files (YAML, JSON, Markdown, TOML) use structural summary instead of raw content dump
-- **smart_read**: unsupported files return truncated 60-line preview instead of full raw content
-- **class_hierarchy**: proper parser for ast-index text output (Parents/Children sections)
-- **project_overview**: uses directory name when package.json has no `name` field
-
-## [0.2.2] - 2026-03-01
-
-### Fixed
-- Published 0.2.1 contained stale Haiku files in dist/ (tsc doesn't clean old outputs)
-- Added `prebuild` script (`rm -rf dist`) to prevent stale artifacts
-- Added `chmod +x` in prepublishOnly to ensure bin is executable
-
-## [0.2.1] - 2026-03-01
-
-### Fixed
-- **RC3**: `search_code` now merges all ast-index result types (content_matches + symbols + files + references) — previously only used content_matches which was often empty
-- **RC4**: `class_hierarchy` and `implementations` parse text format as fallback when JSON parse fails
-- **RC6**: `read_symbol` auto-fetches outline from ast-index if no cached structure — no longer requires prior smart_read
-- `ensureIndex` uses `--sub-projects` flag for monorepo indexing
-
-### Removed
-- Reverted Haiku v0.2.1 — removed broken PersistentFileCache, DiffEngine, RealTokenEstimator, ContextWindowTracker, smart-read-xml, context-markup
-- Removed 3 heavy native dependencies: `better-sqlite3`, `js-tiktoken`, `diff`
-
-### Added
-- `start.sh` — bootstrap script for Claude Code plugin system
-
-## [0.2.0] - 2026-03-01
-
-### Fixed
-- **P0**: ast-index errors no longer silently swallowed — all search/usages/implementations/hierarchy/outline/symbol log errors to stderr
-- **P0**: `exec()` now captures and logs ast-index subprocess stderr
-- **P0**: `projectRoot` detected via `git rev-parse --show-toplevel` instead of `process.cwd()` (fixes wrong index root)
-- **P1**: `forget(all=true)` now clears both ContextRegistry and FileCache (fixes stale export_ast_index/read_diff after forget)
-- **P1**: `forget(path=X)` also invalidates FileCache for that path
-- **P2**: `read_symbol` supports PHP `::` separator (e.g. `RefundProcessor::refund`)
-- **P2**: `findInStructure` recursion fixed — supports 3+ level nesting (Namespace::Class::method)
-- `ensureIndex()`: verify index has content after `stats` — force rebuild if 0 files indexed
-
-### Changed
-- `project_overview`: now shows directory listing + ast-index stats (files, symbols, references) instead of stub
-- `project_overview`: added PHP (`composer.json`) detection
-
-## [0.1.6] - 2026-03-01 (unpublished)
-
-### Fixed
-- `ensureIndex()`: verify index has content after `stats` — force rebuild if 0 files indexed (fixes empty search results on first run)
-
-## [0.1.5] - 2026-03-01
-
-### Fixed
-- PreToolUse hook: read file path from stdin (Claude Code hook format) instead of `$FILE_PATH` env var
-- Hook now auto-suggests `smart_read` for large code files when Claude tries to use `Read`
-- `session_analytics`: now tracks all tools (read_symbol, read_range, read_diff, smart_read_many, search_code, find_usages, find_implementations, class_hierarchy) — previously only tracked smart_read
-- Empty search/usages/implementations results now show diagnostic hints (ast-index status, fallback suggestions)
-- `ensureIndex()` now logs build progress and errors to stderr
-
-## [0.1.4] - 2026-03-01
-
-### Fixed
-- Lazy file watcher: watch only loaded files instead of entire project root (fixes crash on Docker volumes, home dir, WSL)
-
-## [0.1.3] - 2026-03-01
-
-### Fixed
-- Chokidar file watcher error handler (partial fix, superseded by 0.1.4)
-
-## [0.1.2] - 2026-03-01
-
-### Fixed
-- ast-index integration: `outline` now parses text output (JSON format not supported in v3.24.0)
-- ast-index `symbol` response: handle array format, normalize field names (`line`→`start_line`, `path`→`file`)
-- ast-index `search` response: handle `{content_matches: [...]}` wrapper
-- ast-index `usages` response: map `context`→`text`, `path`→`file`
-- Server version now read dynamically from package.json
-
-### Added
-- `token-pilot doctor` — diagnostics command (checks ast-index, Node.js, config, updates)
-- `token-pilot --version` — print current version
-- Update check on server startup (non-blocking, logs to stderr)
-- `/mcp add` installation method documented for Claude Code chat
-- Troubleshooting section in README
-
-## [0.1.1] - 2026-03-01
-
-### Added
-- `npx -y token-pilot` — zero-install for any MCP client (Cursor, Cline, Continue, etc.)
-- Claude Code plugin marketplace support (`.claude-plugin/marketplace.json`)
-- `start.sh` bootstrap script — auto `npm install` + `npm run build` on first run
-- `npm publish` ready (`files` field, `prepublishOnly` script)
-- Universal install instructions in README for Claude Code, Cursor, Cline
-
-### Changed
-- `.mcp.json` now uses `start.sh` for reliable bootstrap
-- README reorganized: npx as primary install, from-source as fallback
-
-## [0.1.0] - 2026-03-01
-
-### Added
-
-- **Core Reading Tools**: `smart_read`, `read_symbol`, `read_range`, `read_diff`, `smart_read_many`
-  - AST-based structural overviews saving 80-95% tokens
-  - Small file pass-through (< 80 lines returned in full)
-  - O(n) diff algorithm for re-reads
-  - Advisory context registry with compact reminders
-- **Search & Navigation**: `search_code`, `find_usages`, `find_implementations`, `class_hierarchy`, `project_overview`
-  - Powered by ast-index (tree-sitter + SQLite FTS5)
-  - Cross-file symbol resolution
-- **Context Management**: `session_analytics`, `context_status`, `forget`
-  - Token savings tracking per tool and per file
-  - Advisory (non-blocking) context tracking
-- **Integration**: `export_ast_index`
-  - context-mode detection and complementary architecture
-  - AST data export for BM25 cross-indexing
-- **Infrastructure**
-  - Git HEAD watcher with selective cache invalidation on branch switch
-  - File watcher (chokidar) for automatic cache invalidation
-  - LRU file cache with configurable size limit
-  - Input validation for all tools (path traversal protection)
-  - Auto-download of ast-index binary from GitHub releases
-  - PreToolUse hook installer for Claude Code
-  - Claude Code plugin format (.claude-plugin/)
-  - Non-code structural summaries (JSON, YAML, Markdown, TOML)
-  - Configurable via `.token-pilot.json`
+# Changelog
+
+All notable changes to Token Pilot will be documented in this file.
+
+The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.1.0/),
+and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
+
+## [0.22.2] - 2026-04-18
+
+### Fixed
+
+- **`session_snapshot` silently dropped `decisions[]`** — the tool schema exposed the field and the renderer consumed it, but the server dispatch's inline cast type omitted it, so every snapshot lost its Decisions section. Fix: added `decisions?: string[]` to the cast. Regression-guarded by new `tests/handlers/session-snapshot.test.ts` covering every schema field.
+- **Help text tool count out of date** — `token-pilot --help` said `MCP Tools (20)` but the server registers 22. Corrected count + listed all 22 (including `read_section` and `read_symbols`).
+- **README doc drift** — hard-coded `(21)` in the MCP Tools heading and "six subagents" throughout. Replaced with count-free phrasing; added Tier 1 / Tier 2 tables covering all 11 subagents; added `session_budget` to the Session tools row.
+
+### Changed
+
+- **Session-registry flush on signal termination** — `SessionRegistryManager.flushAll()` is now wired to `SIGINT` and `SIGTERM` in addition to `beforeExit` (the latter doesn't fire on signal-based termination).
+- Clarified the `shutdownFlush` comment about `process.exit()` limitations.
+- Added a one-line intro to the README subagents section explaining the Tier 1 vs Tier 2 split.
+
+### Numbers
+- 881 tests green (+2 regression tests for `session_snapshot`), `tsc --noEmit` clean.
+
+## [0.22.1] - 2026-04-18
+
+### Added — TP-02l Tier 2 subagents (5 new)
+
+Five more `tp-*` specialists, installed alongside the existing six via `npx token-pilot install-agents`:
+
+- **`tp-debugger`** — bug diagnosis via call-tree traversal (`find_usages` + `read_symbol` + `smart_log`). Given a stack trace or error, finds the root-cause line without Reading whole files.
+- **`tp-migration-scout`** — pre-migration impact map. Given a target (API, symbol, dependency), emits a file-by-file checklist grouped by effort class (trivial / local / cross-file / needs-design).
+- **`tp-test-writer`** — writes tests for one specific symbol, mirroring the project's existing test style. Runs `test_summary` before declaring done — refuses to claim success on tests it didn't run.
+- **`tp-dead-code-finder`** — cross-checks `find_unused` with Grep, recent git history, and dynamic-lookup patterns before recommending deletion. Output only — never deletes.
+- **`tp-commit-writer`** — drafts a Conventional-Commit message from staged diff. Refuses to write when `test_summary` reports failures, when diff mixes types (asks to split), or when staged is empty.
+
+Total subagents now: **11** (6 Tier 1 + 5 Tier 2). Build pipeline auto-discovers `tp-*.md` files — no config changes required.
+
+### Numbers
+- 879 tests green, `tsc --noEmit` clean.
+
+## [0.22.0] - 2026-04-18
+
+### Added — TP-69m session-scoped dedup
+
+The `ContextRegistry` that remembers "this file / symbol / range is already in your context" used to live for the MCP server process lifetime — a restart or the way Claude Code spawns short-lived server instances threw the knowledge away. Now it is per-session and persisted to disk.
+
+Four mechanics shipped together:
+
+1. **`ContextRegistry` snapshot API** — new `toSnapshot()` / `loadSnapshot()` round-trip the state through plain JSON. Silent on malformed input (a broken snapshot file degrades to an empty registry, never crashes the server).
+2. **`SessionRegistryManager`** — owns a map of session_id → registry, LRU-caps the live set (default 8 sessions in memory), reads/writes `.token-pilot/context-registries/<id>.json`. Unsafe ids (empty, traversal, slashes) get an ephemeral in-memory registry that is never persisted.
+3. **Per-call `pickRegistry` in server.ts** — `smart_read`, `read_symbol`, `read_range`, `smart_read_many` now pick the right registry for each tool call based on args. No `session_id` → process-default (legacy behaviour). Flushes to disk after every successful dedup-aware call.
+4. **`force: true` escape hatch** — new optional arg on the four dedup tools. When compaction has evicted an earlier result from the agent's window, `force: true` returns the full content instead of a "you already loaded this" pointer. Critical: without it, a session-scoped dedup pointing to a compacted turn would be an impossible-to-escape pit.
+
+Schema additions on `smart_read` / `read_symbol` / `read_range` / `smart_read_many`: optional `session_id: string` and `force: boolean`. Backwards compatible — existing callers see no change.
+
+Shutdown: `SessionRegistryManager.flushAll()` is attached to `process.beforeExit` so any registries that missed their post-call flush still land on disk.
+
+### Numbers
+- 879 tests green, `tsc --noEmit` clean.
+
+## [0.21.2] - 2026-04-18
+
+### Added
+- **`session_snapshot` auto-persist + SessionStart resume pointer (TP-340)** — calling `session_snapshot` now writes the rendered block to `.token-pilot/snapshots/<iso>.md` and `latest.md` (opt-out via `persist: false`). SessionStart hook surfaces a one-line pointer when the latest snapshot is fresh (<24h), so a new window after `/clear`, compaction, or a fresh process can pick up the thread without re-hydrating context by hand. Retention keeps the last 10 archived snapshots.
+- **`session_budget` MCP tool (TP-hsz batch A)** — new tool reports the live session's saved tokens, configured budget, burn fraction (clamped 0..1), base threshold, and the effective threshold the adaptive curve would apply right now. Small payload (~80 tokens) — the agent can poll cheaply before a big read to decide whether to tighten up.
+- **Context-mode auto-suggest in Bash advisor (TP-hsz batch A)** — when `.mcp.json` advertises context-mode, the large-Bash-output advisory now mentions `mcp__context-mode__execute` as an option (sandbox keeps stdout out of the window). Sync detector — no async plumbing added to the hook.
+- **Time-to-compact projection in `session_budget` (TP-hsz batch B)** — payload now includes `eventCount`, `avgSavedPerEvent`, `eventsUntilExhaustion`, `firstEventMs`, `lastEventMs`. Agent can see how many more same-shape turns the adaptive budget will tolerate at the current burn rate.
+
+### Changed
+- **Snapshot resume pointer is tighter and more informative** — SessionStart "fresh snapshot" window narrowed from 24h to 2h (an unrelated next-day task shouldn't inherit yesterday's thread) and now surfaces the snapshot's `Goal:` extract inline so the agent can eyeball relevance before reading `latest.md`.
+- **Clarified adaptive-threshold / `session_budget` semantics** — `burnFraction` is Read-hook suppression pressure, NOT context-window occupancy. Token Pilot has no visibility into actual window state; the new docstrings and tool descriptions say so explicitly, and the `session_budget` payload carries a `semantics:` hint. No behaviour change; naming-only clarification before TP-69m builds on the same signal.
+
+### Numbers
+- 873 tests green, `tsc --noEmit` clean.
+
+## [0.21.1] - 2026-04-18
+
+### Added
+- **Adaptive Read-hook threshold (TP-bbo)** — opt-in `hooks.adaptiveThreshold` auto-lowers `denyThreshold` as the current session burns through `hooks.adaptiveBudgetTokens` (default 100k). Piecewise curve: unchanged below 30% burn, ×0.75 at 30–60%, ×0.5 at 60–80%, ×0.3 (floor 50 lines) beyond. Burn is read from `.token-pilot/hook-events.jsonl` `savedTokens` for the live `session_id`. Default off — zero behaviour change unless the user enables it. Env overrides: `TOKEN_PILOT_ADAPTIVE_THRESHOLD`, `TOKEN_PILOT_ADAPTIVE_BUDGET`.
+- **Save-doc CLI (TP-89n)** — `token-pilot save-doc <name>` persists any stdin text (curl, WebFetch, long research notes) to `.token-pilot/docs/<name>.md` so it survives compaction and can be re-read cheaply with `smart_read` / `read_range` instead of refetching the external source. `token-pilot list-docs` enumerates saved docs. Name validation refuses traversal / path separators; overwrite is explicit (`--overwrite`).
+
+### Numbers
+- 862 tests green, `tsc --noEmit` clean.
+
+## [0.21.0] - 2026-04-18
+
+### Added
+- **`doctor` Claude Code env-var advisor (TP-c08)** — surfaces the four knobs the community guide flags as giving 60-80% session savings with zero code change (`CLAUDE_CODE_SUBAGENT_MODEL=haiku`, `MAX_THINKING_TOKENS=10000`, `CLAUDE_AUTOCOMPACT_PCT_OVERRIDE=50`, `model=sonnet`). Pure advisory — never modifies user settings; reads both `process.env` and `~/.claude/settings.json` with fallback semantics.
+- **`.claudeignore` generator (TP-rtg)** — `token-pilot init` now offers to create a `.claudeignore` with sensible defaults (node_modules, dist, build, __pycache__, lockfiles, source maps, …). Non-destructive: carries a magic-comment marker so re-runs refresh our own file in place but never clobber user-owned `.claudeignore`. `doctor` reports current status.
+- **CLAUDE.md hygiene check in `doctor` (TP-rtg)** — warns when `CLAUDE.md` exceeds 60 non-empty lines (that file loads into every Claude Code message; long rules are per-turn tax). Read-only; counts ignore blank lines and markdown horizontal rules.
+- **Bash output advisor (TP-jzh)** — new `PostToolUse:Bash` hook. When Bash stdout exceeds ~8000 characters, the hook appends a single-line `additionalContext` tip pointing the agent at cheaper alternatives (`mcp__token-pilot__test_summary` for test runs, bounded commands, head/tail piping). Cannot truncate output in-flight — Claude Code's PostToolUse is observational for non-MCP tools — but steers the next turn.
+
+### Changed
+- `.claude-plugin/hooks/hooks.json` and the installer now register the new PostToolUse:Bash hook alongside Read/Edit/SessionStart. Idempotent install adds it without touching existing hooks; uninstall removes PostToolUse too.
+
+### Numbers
+- 843 tests green, `tsc --noEmit` clean.
+
+## [0.20.2] - 2026-04-18
+
+### Changed
+- **`token-pilot init` now offers to install tp-* subagents** — after writing `.mcp.json`, if a TTY is attached the command asks `Install 6 tp-* subagents now? [Y/n]`. If yes, delegates to the full `install-agents` flow (scope prompt, idempotence, persistence). In non-TTY the next-step hint is printed instead of asking. Closes the gap where first-time users left `init` thinking everything was ready and only learned about subagents from a later stderr reminder.
+- **Refreshed the init success message** — replaced the v0.13-era "AST-aware code reading (60-80% token savings)" line with a description of the v0.20 enforcement-layer scope.
+
+## [0.20.1] - 2026-04-18
+
+### Fixed
+- **hook-events.jsonl not written** — the writeEvent helper in the hook dispatcher was fire-and-forget (`void appendEvent(...)`). `process.exit(0)` raced with the async fs write, so every event was silently dropped. Now awaits the write before returning. `token-pilot stats` and `stats --by-agent` finally show real data.
+
+## [0.20.0] - 2026-04-18
+
+### Added
+- **Enforcement layer (TP-816)** — four-component architecture that makes token-pilot actually used, not just advertised.
+  - **`deny-enhanced` hook mode** (new default) — `PreToolUse:Read` on qualifying large code files returns a structural summary (imports, exports, declarations, head/tail fallback) **inside the denial reason**. Works for every agent, including subagents that lack MCP access. `advisory` and `off` modes remain available.
+  - **SessionStart hook** — emits a one-shot reminder after every `/clear` / `/compact` / new session, listing the mandatory MCP tools and the installed `tp-*` subagents. Respects `sessionStart.enabled` independently of `hooks.mode`.
+  - **`bless-agents` CLI** — scans installed agents, classifies by tool-allowlist shape (wildcard / exclusion / explicit), and writes project-level overrides adding `mcp__token-pilot__*` to category-C agents. `unbless-agents` + `doctor` upstream-drift detection close the loop.
+  - **Subagent family (`tp-*`)** — six Tier-1 agents with tight response budgets and verdict-first output contract: `tp-run` (800), `tp-onboard` (600), `tp-pr-reviewer` (600), `tp-impact-analyzer` (400), `tp-refactor-planner` (500), `tp-test-triage` (500). Installed via `npx token-pilot install-agents` (user or project scope, idempotent with body-hash).
+- **`install-agents` / `uninstall-agents` CLI** — scope resolution (flag > persisted > prompt > error), idempotence matrix (unchanged / template-upgraded / user-edited / no-hash), `--force` to overwrite user-edited (never touches files without our marker).
+- **MCP startup reminder** — one-time stderr nudge when no `tp-*` agents are installed; silenced by `agents.reminder: false` or `TOKEN_PILOT_NO_AGENT_REMINDER=1`; suppressed inside subagents via `TOKEN_PILOT_SUBAGENT=1`.
+- **`hook-events.jsonl` telemetry** — new schema `{ts, session_id, agent_type, agent_id, event, file, lines, estTokens, summaryTokens, savedTokens}`; rotates at 10 MB, retains 30 days / 100 MB.
+- **`stats` CLI** — `token-pilot stats` (default totals + top files), `--session[=<id>]` (filter to one session, most recent by default), `--by-agent` (group by `agent_type`, null rendered as "main").
+- **`bench:hook` script** — `npm run bench:hook` reports p50/p95/p99 hook latency against a 1000-line fake file; thresholds from TP-816 §11 available as opt-in `--check=true` gate.
+
+### Changed
+- **Config** — new fields: `hooks.mode` (`off` | `advisory` | `deny-enhanced`, replaces legacy boolean `hooks.enabled`), `sessionStart.*`, `agents.scope`, `agents.reminder`, `hooks.migratedFrom`.
+- **Legacy migration** — `hooks.mode: "deny"` (v0.19) is rewritten to `"advisory"` on next load with a one-time stderr notice and `hooks.migratedFrom: "deny"` marker. Old `hooks.enabled: false` is migrated to `mode: "off"`. Both are idempotent.
+- **Env vars** — `TOKEN_PILOT_DENY_THRESHOLD=<n>` overrides `hooks.denyThreshold`. Documented alongside `TOKEN_PILOT_MODE`, `TOKEN_PILOT_BYPASS`, `TOKEN_PILOT_DEBUG`, `TOKEN_PILOT_NO_AGENT_REMINDER`, `TOKEN_PILOT_SUBAGENT`.
+
+### Deferred
+- **Live-LLM behavioural assertions** — the agent-behaviour acceptance ("uses MCP before raw Read; response within budget; no narration") requires a live Anthropic or Claude Code runner. Deterministic coverage (structure, budget ceiling, fixture compat) is in place; live dispatch moves to a v0.20.x follow-up.
+- **Claude Code marketplace plugin** — planned for a future release; `install-agents` remains the supported path.
+
+### Numbers
+- 806 tests green, `tsc --noEmit` clean.
+
+## [0.19.2] - 2026-04-15
+
+### Added
+- **npm-first binary install** — `install-ast-index` now tries `npm install -g @ast-index/cli` before falling back to GitHub download. Works on all platforms including Windows (no more "ZIP extraction not supported" error).
+- **npm binary discovery** — `findBinary` now checks the npm global prefix (`npm config get prefix`) as a 3rd resolution strategy: config → system PATH → npm global → managed install.
+
+### Fixed
+- **Hook installer uses absolute paths** — hooks now write `<node> <script> hook-read` instead of bare `token-pilot hook-read`. Fixes `token-pilot: not found` in `/bin/sh` environments (nvm, npx, non-login shells).
+- **Skip auto-install when running as plugin** — when `CLAUDE_PLUGIN_ROOT` is set, the MCP server no longer writes duplicate hooks into `.claude/settings.json`.
+- **Auto-upgrade broken hooks** — old hooks with bare `token-pilot` commands are automatically replaced with absolute-path versions on next server start.
+
+### Changed
+- **`BinaryStatus.source`** now includes `'npm'` as a value (shown in `doctor` and `session_analytics`).
+- **`search()` supports `--type` filter** — filter results by symbol type (`class`, `function`, `interface`, etc.). Leverages ast-index ≥3.30.0.
+- **`hierarchy()` supports `--in-file` / `--module` filters** — scope class hierarchy queries by filename or module path. Leverages ast-index ≥3.30.0.
+- **498 tests** (was 492).
+
+## [0.19.1] - 2026-04-15
+
+### Added
+- **`decisions` field in `session_snapshot`** — stores key decisions with reasoning (e.g., "removed sysfee step — caused double counting"). Prevents the model from revisiting rejected approaches after context compaction.
+
+## [0.19.0] - 2026-04-15
+
+### Added
+- **`session_snapshot` tool** — capture current session state (goal, confirmed facts, files, blockers, next step) as a compact markdown block (<200 tokens). Call before context compaction or when switching direction in long sessions.
+- **`max_tokens` parameter** on `smart_read` and `smart_read_many` — token budget per read. Output auto-downgrades through three levels: full content → structural outline → compact (symbol names + line ranges only). Enables context-constrained sessions.
+- **Session compaction advisory** — policy engine now tracks total tool calls and tokens returned. Advises calling `session_snapshot()` when thresholds are reached (default: every 15 calls or after 8,000 tokens). Configurable via `compactionCallThreshold` and `compactionTokenThreshold`.
+- **"Why This Approach Works"** section in README explaining the 3-level optimization strategy.
+
+### Changed
+- **21 tools** (was 20) — added `session_snapshot`.
+- **MCP instructions** updated with `session_snapshot` workflow and `max_tokens` guidance.
+- Benchmark numbers updated: 55 files, 102K raw → 9K outline tokens (91% savings).
+
+## [0.18.1] - 2026-04-13
+
+### Fixed
+- **Hook installer uses absolute paths** — hooks now write `<node> <script> hook-read` instead of bare `token-pilot hook-read`. Fixes `token-pilot: not found` errors in `/bin/sh` environments (nvm, npx, non-login shells).
+- **Skip auto-install when running as plugin** — when `CLAUDE_PLUGIN_ROOT` is set, the MCP server no longer writes duplicate hooks into `.claude/settings.json` (the plugin system handles this via `hooks.json`).
+- **Auto-upgrade broken hooks** — old hooks with bare `token-pilot` commands are automatically replaced with absolute-path versions on next server start.
+
+### Changed
+- **495 tests** (was 492).
+
+## [0.18.0] - 2026-04-05
+
+### Added
+- **`read_section` tool** — read a specific section from Markdown, YAML, JSON, or CSV files. Markdown: by heading name. YAML/JSON: by top-level key. CSV: by row range (`rows:1-50`). Much cheaper than reading the whole file.
+- **`read_for_edit` section parameter** — prepare edit context for non-code file sections. Works with all 4 formats.
+- **Markdown outline with line ranges** — `smart_read` on `.md` files now shows `[L5-20]` ranges and hints for `read_section`.
+- **YAML/JSON section ranges** — `smart_read` on `.yaml`/`.json` shows top-level key ranges.
+- **CSV smart_read** — shows columns, row count, sample rows, and hints for row-range reading.
+- **4 section parsers** — `markdown-sections.ts`, `yaml-sections.ts`, `json-sections.ts`, `csv-sections.ts`.
+
+### Changed
+- **20 tools** (was 19) — added `read_section`.
+- **492 tests** (was 441).
+
+### Fixed
+- `npm audit` — resolved brace-expansion, path-to-regexp, picomatch vulnerabilities.
+
+## [0.17.0] - 2026-04-02
+
+### Added
+- **`smart_read` scope parameter** — `scope="nav"` returns names + line ranges only (2-3x smaller), `scope="exports"` shows only public API. Default `scope="full"` unchanged.
+- **`smart_read` auto-delta** — when a file changed since last load (within 120s), shows ADDED/REMOVED/UNCHANGED symbols instead of full re-read. Config: `smartRead.autoDelta.enabled`.
+- **`read_symbol` include_edit_context** — optional `include_edit_context=true` appends raw code block (max 60 lines) to save a separate `read_for_edit` call. Large symbols fall back to `read_for_edit`.
+- **`find_usages` mode=list** — compact `file:line` output for initial discovery, 5-10x smaller than full mode.
+- **`smart_read_many` per-file dedup** — skips files already in context and unchanged, returns compact reminder instead.
+- **Actionable hints** — `read_for_edit` suggests `read_diff` after editing. Config: `display.actionableHints`.
+- **`symbol-display-constants.ts`** — shared display constants for symbol rendering.
+
+### Changed
+- **441 tests** (was 427) — new tests for scope, list mode, include_edit_context, dedup.
+- **MCP instructions** updated with scope/mode/include_edit_context guidance.
+- **find_usages context rendering** — sequential instead of concurrent to prevent shared cache race condition.
+
+## [0.16.1] - 2026-03-21
+
+### Added
+- **Hook interception tracking** — PreToolUse hook now records denied Read calls (file path, line count, estimated tokens) to `.token-pilot/hook-denied.jsonl`. Session analytics shows how many tokens the hook saved by intercepting unbounded reads on large code files.
+- **`session_analytics` hook savings** — compact report adds "Hook: intercepted N reads, saved ~X tokens" line. Verbose mode shows per-file breakdown of intercepted reads.
+
+## [0.16.0] - 2026-03-21
+
+### Added
+- **`read_symbols` tool** — batch read multiple symbols from one file in a single call (max 10). File is read once, AST resolved once. Saves N-1 round-trips vs calling `read_symbol` N times.
+- **`read_for_edit` batch mode** — new `symbols` array parameter reads multiple symbol edit contexts in one call. Each symbol returns raw code ready for Edit tool's `old_string`.
+- **`find_usages` context_lines** — new `context_lines` parameter (0-10) shows surrounding source code for each match. Eliminates follow-up `read_symbol` calls after finding usages.
+- **`smart_diff` affected symbols summary** — consolidated "AFFECTED SYMBOLS" section at the top of diff output, grouped by MODIFIED/ADDED/REMOVED. See all changed functions/classes at a glance.
+
+### Changed
+- **19 tools** (was 18) — added `read_symbols`.
+- **MCP instructions** — added batch read_symbols to decision rules and refactor workflow.
+- **427 tests** (unchanged — all pass with new features).
+
+## [0.15.0] - 2026-03-19
+
+### Added
+- **Regex fallback parser (TS/JS)** — `smart_read` now works for TypeScript/JavaScript files even without ast-index binary. Parses classes, functions, interfaces, types, enums, and class methods via regex. Zero dependencies, 130 lines. Covers ~80% of new users who fail to download ast-index.
+- **Regex fallback parser (Python)** — `smart_read` now works for Python files without ast-index. Parses classes, functions, async functions, decorators (`@dataclass`, `@app.route`), module constants (`UPPER_CASE`), methods with visibility detection (`_private`, `__dunder__`). 150 lines.
+- **Benchmark script** — `scripts/benchmark.ts` measures real token savings on public repos (express, fastify, flask). 92% average savings across 97 files ≥50 lines. Run: `npx tsx scripts/benchmark.ts`.
+- **Guide skill** — `/guide` command shows a quick-reference table of all Token Pilot tools with usage examples and recommended workflow.
+- **`hooks.denyThreshold` config** — hook deny threshold is now configurable in `.token-pilot.json` (default: 300, was hardcoded 500). Intercepts ~2x more native Read calls.
+
+### Changed
+- **Compact session analytics** — `session_analytics` report reduced from ~30 lines to ~5 lines. Shows calls, tokens saved, top 5 tools, top 3 files, cache hit rate on a single screen. Verbose mode (`verbose=true`) restores full breakdown.
+- **`server.ts` refactor** — extracted tool definitions to `server/tool-definitions.ts` and token estimate helpers to `server/token-estimates.ts` (−500 lines from server.ts).
+- **`find_usages` output** — results grouped by file with compact rendering. Single match per file on one line, multiple matches indented under file header.
+- **Stale references** — all `grep_search` hints updated to `Grep` (code-audit, find-unused, find-usages).
+- **README** — benchmark table with real data from 4 public repos. Updated savings claims from 80% to 90% (backed by benchmark).
+- **427 tests** (was 393).
+
+### Fixed
+- **`npx token-pilot` CLI** — symlink path resolution in `isDirectRun` check. All CLI commands now work correctly via npx.
+- **Regex fallback was dead code** — parsers existed but weren't wired into `client.ts` `outline()` method. Now properly called as fallback when ast-index unavailable.
+
+## [0.14.1] - 2026-03-14
+
+### Fixed
+- **CI: Node.js 24 runtime** — opted into `FORCE_JAVASCRIPT_ACTIONS_TO_NODE24` for GitHub Actions, resolving deprecation warnings for `actions/checkout@v4` and `actions/setup-node@v4`.
+- **CI: test matrix** — updated from Node 18+22 to Node 20+22 (Node 18 is EOL).
+- **Test: git commit in CI** — `read-for-edit` tests now pass `-c user.name` / `-c user.email` to `git commit`, fixing failures in environments without global git config.
+
+## [0.14.0] - 2026-03-14
+
+### Added
+- **R&D Track 0: Instrumentation** — per-call decision trace capturing file size, context state, estimated vs actual cost, and cheaper alternative suggestions. Integrated into all 18 tool handlers via `recordWithTrace()`.
+- **R&D Track 1: Budget Planner** — advisory layer suggesting cheaper tool alternatives (e.g. `smart_read` → `read_diff` when file already in context, → `read_symbol` when symbol known). Analytics-only, no blocking.
+- **R&D Track 2: Intent Router** — classifies tool calls into 7 intents (edit/debug/explore/review/analyze/search/read). Per-intent breakdown in session analytics.
+- **R&D Track 3: Edit Prep Mode** — `read_for_edit` with `include_callers`, `include_tests`, `include_changes` enrichment options.
+- **R&D Track 4: Session Cache** — tool-result-level caching with file/AST/git invalidation.
+- **R&D Track 5: Confidence-Based Escalation** — confidence metadata (high/medium/low) appended to `smart_read`, `read_symbol`, `read_for_edit`, `find_usages` responses. Shows known unknowns and suggested next steps.
+- **R&D Track 6: Working Set / Dedup** — compact reminders for already-loaded files and symbols.
+- **R&D Track 7: Related Files Ranking** — scored ranking with 6 signals (test +5, import +4, importer +3, same-dir +2, recently-changed +2, multi-ref +1). HIGH VALUE / MEDIUM / LOW buckets.
+- **R&D Track 8: Architecture Fingerprint** — caches architecture in `.token-pilot-fingerprint.json` (24h TTL). Amortizes `project_overview` cost across sessions.
+- **R&D Track 9: Verified Savings Dashboard** — savings breakdown by category (compression/cache/dedup), session cache hit rate, dedup stats.
+- **R&D Track 10: Team Policy Mode** — configurable policies: `preferCheapReads`, `maxFullFileReads`, `warnOnLargeReads`, `requireReadForEditBeforeEdit`.
+- **7 new core modules** — `confidence.ts`, `intent-classifier.ts`, `budget-planner.ts`, `decision-trace.ts`, `session-cache.ts`, `architecture-fingerprint.ts`, `policy-engine.ts`.
+- **35 new tests** — confidence (11), architecture-fingerprint (11), policy-engine (13). Total: 393 tests.
+
+### Changed
+- **`session_analytics`** — per-intent breakdown, decision insights, savings by category.
+- **`project_overview`** — saves/loads architecture fingerprint for cross-session caching.
+- **Config** — added `policies` section to `TokenPilotConfig`.
+
+## [0.13.0] - 2026-03-14
+
+### Added
+- **Version check for all components** — on startup, checks token-pilot (npm), ast-index (GitHub releases), and context-mode (npm) in parallel. Non-blocking, fire-and-forget. Shows update notifications in stderr.
+- **`autoUpdate` config flag** — `updates.autoUpdate: true` in `.token-pilot.json` auto-downloads new ast-index binary on startup. Default: `false` (notify only). token-pilot and context-mode only notify (separate processes).
+- **`checkBinaryUpdate()`** — compares installed ast-index version vs latest GitHub release.
+- **`isNewerVersion()` utility** — semver comparison: strip `v` prefix, compare segments. Handles different lengths (`1.0` vs `1.0.1`).
+- **Common Lisp extensions** — `.lisp`, `.lsp`, `.cl`, `.asd` added to `CODE_EXTENSIONS` for ast-index v3.28+ compatibility.
+- **9 new tests** — `isNewerVersion()` covering major/minor/patch, same version, older, `v` prefix, different segment lengths, large numbers, real-world versions. Total: 217 tests.
+
+### Changed
+- **`doctor` command** — now shows 3 sections: token-pilot (installed/latest), ast-index (installed/latest/auto-update status), context-mode (detected/latest npm). Previously only showed ast-index binary status.
+- **`install-ast-index` command** — now also updates existing binary if newer version available on GitHub.
+- **`printHelp()`** — fixed tool count: 18 (was incorrectly showing 12 since v0.8.0).
+- **Startup update check** — replaced single `checkLatestVersion()` with `checkAllUpdates()` covering all 3 components via `Promise.allSettled`.
+
+### Fixed
+- **`test_summary` PHPUnit parser** — now counts both `Failures:` and `Errors:` (was only counting failures).
+- **`test_summary` cargo parser** — correctly identifies failure name-list section (no `----` markers) vs detail section.
+- **`test_summary` token estimation** — uses shared `estimateTokens()` instead of local duplicate.
+- **`smart_log` category detection** — `documentation` now matches docs pattern, `tests` (plural) matches test pattern, `optimize`/`optimization` match perf pattern.
+- **`explore_area` path boundary** — `startsWith(path + '/')` prevents `src/auth` matching `src/authorize/`.
+- **Validation consistency** — `validateSmartLogArgs` and `validateTestSummaryArgs` now use `optionalString`/`optionalNumber` helpers, reject empty strings, check integers.
+
+## [0.11.0] - 2026-03-14
+
+### Added
+- **`smart_log` tool** — structured git log with commit category detection (feat/fix/refactor/docs/test/chore/style/perf). Shows author breakdown, file stats (+/-), per-commit file list. Filters by path and ref. Raw git log → compact summary.
+- **`test_summary` tool** — runs test command and returns structured summary: total/passed/failed/skipped + failure details. Parsers for vitest, jest, pytest, phpunit, go test, cargo test, rspec, mocha + generic fallback. 200 lines of raw output → 10-15 lines.
+- **38 new tests** — smart_log parser (5), categorizer (4), test_summary parsers (17), runner detection (8), validation (4). Total: 208 tests (was 170).
+
+### Changed
+- **18 tools** (was 16) — added `smart_log`, `test_summary`
+- **MCP instructions** — added smart_log and test_summary to workflow guidance
+
+## [0.10.0] - 2026-03-14
+
+### Added
+- **`smart_diff` tool** — structural git diff with AST symbol mapping. Shows which functions/classes were modified/added/removed instead of raw patch output. Supports scopes: `unstaged`, `staged`, `commit` (ref required), `branch` (ref required). Small diffs (<=30 lines) include actual hunks, large diffs show summary. Returns `rawTokens` for precise savings analytics.
+- **`explore_area` tool** — one-call directory exploration combining outline + imports + tests + git changes. Replaces 3-5 separate tool calls when starting work on an area. Sections: `outline` (recursive depth 2), `imports` (external deps + who imports this area), `tests` (matching test/spec files), `changes` (recent git log). All sections run in parallel via `Promise.allSettled`.
+- **26 new tests** — smart_diff parser (10), symbol mapping (5), validation (11). Total: 170 tests (was 144).
+
+### Changed
+- **16 tools** (was 14) — added `smart_diff`, `explore_area`
+- **MCP instructions** — updated workflow: `project_overview → explore_area → smart_read → read_symbol → read_for_edit → edit → smart_diff`
+- **`outlineDir` and `CODE_EXTENSIONS` exported** from outline.ts for reuse by explore_area
+
+## [0.9.0] - 2026-03-08
+
+### Added
+- **`module_info` tool** — analyze module dependencies, dependents, public API, and unused deps. Uses ast-index v3.27.0 module commands (`modules`, `module-deps`, `module-dependents`, `module-api`, `unused-deps`). Includes degradation check when ast-index is unavailable.
+- **`project_overview` dual-detection** — shows BOTH ast-index type detection AND config-file detection (package.json, composer.json, Cargo.toml, pyproject.toml, go.mod) with CONFIDENCE scoring (high/medium/low/unknown). Detects frameworks, quality tools (PHPStan, ESLint, Vitest, Jest, Biome, etc.), CI pipelines (GitHub Actions, GitLab CI, Jenkins), and Docker.
+- **`project_overview` `include` parameter** — filter sections: `["stack"]` for quick type check, `["quality","ci"]` for tooling overview. Default: all sections.
+- **`find_usages` post-filters** — `scope` (path prefix), `kind` (definitions/imports/usages), `lang` (14 languages by extension), `limit` (per category, 1-500). All filters optional, backward compatible.
+- **`outline` recursive mode** — `recursive=true` with `max_depth` (default 2, max 5) recurses into subdirectories. At max depth shows file counts only.
+- **`src/core/project-detector.ts`** — extracted config-based detection logic into reusable module. Framework detection maps for PHP (7), JS (10), Python (5). Quality tools scanner (13 tools). CI pipeline detector (6 platforms).
+- **ast-index client: 5 module methods** — `modules()`, `moduleDeps()`, `moduleDependents()`, `unusedDeps()`, `moduleApi()` with JSON-first + text fallback parsing.
+- **ast-index types: 4 module interfaces** — `AstIndexModuleEntry`, `AstIndexModuleDep`, `AstIndexUnusedDep`, `AstIndexModuleApi`.
+
+### Fixed
+- **`module_info` token savings** — `tokensWouldBe` was equal to `tokensReturned` (0% savings). Now estimates manual analysis cost correctly.
+- **`outline` recursive overflow** — added `MAX_OUTLINE_LINES=500` guard to prevent runaway output on large projects with `recursive=true`.
+- **`project_overview` "frontend" label** — removed hardcoded "frontend" suffix for secondary stacks (Node.js is not always frontend).
+- **Ruff detection** — no longer double-reads `pyproject.toml`. Checks `ruff.toml`/`.ruff.toml` first, falls back to `pyproject.toml [tool.ruff]` only if needed.
+- **44 new tests** — validators (23) + project-detector (21). Total: 144 tests (was 100).
+
+### Changed
+- **14 tools** (was 13) — added `module_info`
+- **Tool descriptions** — updated with `(v1.1: ...)` version hints for enhanced tools
+- **MCP instructions** — added module_info to "COMBINE BOTH" workflow section
+- **Version sync** — package.json, plugin.json, marketplace.json all at 0.9.0
+
+## [0.8.3] - 2026-03-08
+
+### Fixed
+- **code_audit pattern search — root cause fix** — `ast-index agrep` does not support `--limit` flag. Token Pilot was passing `--limit 50` which caused the command to fail silently, returning 0 results across v0.8.0–v0.8.2. Removed the flag; results are now limited via `.slice()` after parsing.
+
+## [0.8.2] - 2026-03-08
+
+### Fixed
+- **code_audit pattern search** — inject `node_modules/.bin` into PATH so `ast-index agrep` can find `sg` (ast-grep) when it's installed as optional dependency but not in system PATH.
+- **code_audit annotations** — strip `@` prefix from annotation names (`@Injectable` → `Injectable`). ast-index expects names without `@`.
+
+## [0.8.1] - 2026-03-08
+
+### Added
+- **ast-grep auto-install** — `@ast-grep/cli` added as optional dependency. `code_audit(check="pattern")` now works out-of-the-box without manual `brew install ast-grep`.
+- **MCP instructions: security audit guidance** — instructions now recommend Grep for security patterns (password, token, secret, credential) and `find_unused` for dead code detection.
+
+### Changed
+- **ast-index stats → JSON parsing** — `--format json` for reliable file count extraction instead of regex on text output.
+
+## [0.8.0] - 2026-03-07
+
+### Added
+- **`code_audit` tool** — find code quality issues in one call: TODO/FIXME comments (`check="todo"`), deprecated symbols (`check="deprecated"`), structural code patterns via ast-grep (`check="pattern"`), decorator search (`check="annotations"`), or combined audit (`check="all"`).
+- **Incremental index update on file changes** — file watcher now triggers `ast-index update` (debounced 2s) after edits. Keeps index fresh for find_usages, find_unused, code_audit.
+- **ast-index client methods** — `agrep()`, `todo()`, `deprecated()`, `annotations()`, `incrementalUpdate()`.
+
+### Fixed
+- **smart_read on directories** — now returns helpful message instead of EISDIR crash.
+- **MCP instructions** — added "COMBINE BOTH" section for audit tasks (Token Pilot + Grep).
+
+## [0.7.6] - 2026-03-07
+
+### Added
+- **`npx token-pilot init`** — one command creates `.mcp.json` with both token-pilot and context-mode configured. Idempotent — safely updates existing configs without overwriting.
+- **MCP Server Instructions** — protocol-level `instructions` field tells AI agents WHEN to use Token Pilot tools instead of built-in defaults. Works universally on all MCP clients.
+- **Improved tool descriptions** — each tool explicitly states what built-in tool it replaces (e.g. "Use INSTEAD OF Read/cat").
+
+### Fixed
+- **3 high severity vulnerabilities** — updated hono and express-rate-limit.
+- **npm package size** — excluded source maps from package. 505 kB → 286 kB (−43%).
+- **Accurate thresholds** — README and instructions now correctly state smallFileThreshold=200 (was 80).
+- **read_diff documentation** — clarified that smart_read must be called BEFORE editing to create baseline snapshot.
+
+### Changed
+- **README** — honest metrics (60-80%), Quick Start with `init` command, MCP instructions section, Codex/Antigravity support.
+- **npm keywords** — added `codex`, `cline`, `model-context-protocol`, `token-savings`.
+
+## [0.7.4] - 2026-03-07
+
+### Added
+- **MCP Server Instructions** — protocol-level `instructions` field tells AI agents WHEN to use Token Pilot tools instead of built-in Read/cat/Grep. Works universally on Claude Code, Cursor, Codex, Antigravity, and any MCP-compatible client. Includes rules for when NOT to use Token Pilot (regex search, raw content copy-paste).
+- **Improved tool descriptions** — each tool now explicitly states what built-in tool it replaces (e.g. "Use INSTEAD OF Read/cat", "Use INSTEAD OF Grep/ripgrep"). Agents can make informed decisions from description alone, without needing project-level rules files.
+
+## [0.7.3] - 2026-03-07
+
+### Fixed
+- **read_diff diagnostic** — when cache miss occurs, now shows resolved absolute path and all cached file paths. This reveals path mismatches between smart_read and read_diff calls (e.g. different relative paths resolving to different absolute paths).
+
+## [0.7.2] - 2026-03-07
+
+### Fixed
+- **read_diff on small files** — `smart_read` small-file pass-through (≤150 lines) returned content without caching in fileCache. `read_diff` always showed "No previous read" for small files because the baseline was never stored. Now all files are cached regardless of size.
+
+## [0.7.1] - 2026-03-07
+
+### Fixed
+- **read_diff after read_for_edit** — `read_for_edit` now caches the full file content, so `read_diff` can use it as baseline after edits. Previously returned "No previous read" because read_for_edit didn't populate the file cache.
+- **outline on intermediate directories** — directories with only subdirectories (no direct code files) now show subdirectory listing with recursive code file counts instead of "No code files found". Enables progressive drill-down: `outline("module/") → outline("module/infrastructure/")`.
+
+## [0.7.0] - 2026-03-07
+
+### Fixed
+- **Project root detection** — complete rewrite of how token-pilot discovers the working project:
+  1. **MCP roots** (new, primary) — uses MCP protocol `listRoots()` to get workspace root from Claude Code. Works for all tools including `find_usages`, `find_unused`, `project_overview` (no file path needed).
+  2. **INIT_CWD/PWD env vars** (new) — when started via `npx`, npm sets `INIT_CWD` to the invoking directory. Catches cases where `process.cwd()` is `/` but the real project root is available in env.
+  3. **Git detect from file path** (improved) — now triggers from any tool call args (`path`, `paths`, `file`, `module`), not just `smart_read`.
+- **ast-index tools always disabled** — `find_usages`, `find_unused`, `project_overview` never triggered auto-detect because they have no `path` argument. Now all tools trigger detection via MCP roots.
+- **Error messages** — changed "project root is too broad" to actionable "call smart_read() on any project file first" when MCP roots unavailable.
+- **`isDangerousRoot`** — moved to shared `core/validation.ts` (was duplicated in `index.ts`).
+
+## [0.6.5] - 2026-03-07
+
+### Fixed
+- **AST index rebuild race condition** — concurrent tool calls no longer trigger multiple simultaneous rebuilds. `ensureIndex()` now deduplicates via shared promise. If rebuild fails due to lock file (another process running), falls back to existing index if available instead of throwing.
+- **Rebuild timeout** — increased from 60s to 120s for large projects where indexing takes longer.
+
+## [0.6.4] - 2026-03-07
+
+### Fixed
+- **CRITICAL: Hook installer** — malformed `settings.json` no longer silently destroyed. Distinguishes ENOENT (create fresh) from JSON parse error (abort with message). Uninstall also reports specific errors.
+- **CRITICAL: Server startup** — `startServer()` now has `.catch()` handler. Unhandled promise rejections no longer crash the process silently.
+- **Non-code handler** — removed `.xml` and `.csv` from `isNonCodeStructured` (no handler existed for them, fell through to null).
+- **Symbol resolver** — removed dangerous basename-only fallback in `pathMatches` (`index.ts` no longer matches any `index.ts`). Fixed hardcoded `endLine = start_line + 10` → uses `end_line` from ast-index or 50-line fallback.
+- **Config loader** — added prototype pollution guard (`__proto__`, `constructor`, `prototype` keys skipped in deepMerge). Parse errors now logged instead of silently swallowed.
+- **File cache** — size tracking now uses `Buffer.byteLength()` instead of `string.length` (chars ≠ bytes for non-ASCII). Removed dead `isSmallFile()` method.
+- **Validation** — `optionalNumber` now rejects `NaN` and `Infinity`.
+- **Token estimation** — `smart_read_many` now uses `estimateTokens()` instead of `length/4`.
+- **Analytics** — `project_overview` calls now tracked in session analytics.
+- **read_for_edit** — raised `MAX_EDIT_LINES` from 20 to 60 (20 was too aggressive, truncated most functions).
+- **related_files** — raised symbol search limit from 5 to 10 for reverse import detection.
+
+### Removed
+- Dead config options `cache.ttlMinutes` and `context.autoForgetMinutes` (declared but never used).
+
+## [0.6.3] - 2026-03-03
+
+### Changed
+- **Hook deny threshold** — raised from 200 to 500 lines. Files ≤500 lines pass through Read without denial roundtrip. Eliminates token overhead on medium files where hook denial costs more than outline saves.
+- **Adaptive fallback** — lowered from 90% to 70%. If outline ≥70% of raw file size, returns raw content. More aggressive at avoiding outlines that barely save tokens.
+- **Tool descriptions** — trimmed marketing language, percentages, and cross-references. ~250 fewer tokens in tool list per session.
+- **Outline cap** — top-level symbols capped at 40, class members at 30. Prevents outline explosion on files with 100+ methods.
+
+## [0.6.2] - 2026-03-02
+
+### Removed
+- **Dead handler files** — deleted `changed-symbols.ts` (removed in v0.5.0) and `find-callers.ts` (removed in v0.4.0). Were never registered in server but lingered as dead code.
+
+## [0.6.1] - 2026-03-02
+
+### Changed
+- **`smallFileThreshold`** — raised from 80 to 200 lines. Benchmark showed medium files (100-300 lines) had negative savings (-25%) because AST outline was larger than the raw file. Files ≤200 lines now pass through as raw content.
+- **`smart_read` adaptive fallback** — after generating outline, compares token count vs raw file. If outline ≥ 90% of raw size, returns raw content instead. Eliminates negative savings on any file size, regardless of language or threshold.
+- **`session_analytics` honest metrics** — replaced all hardcoded multipliers (`*5`, `*3`) with real full-file token counts from file cache. `tokensWouldBe` now reflects actual file size, not fabricated numbers. Non-file tools (related_files, outline, find_usages) report 1:1 (no savings claim).
+
+## [0.6.0] - 2026-03-02
+
+### Changed
+- **Read hook** — upgraded from advisory (`decision: "suggest"`) to blocking (`permissionDecision: "deny"`) for unbounded Read calls on large code files (>200 lines). Bounded Read (with offset/limit) is still allowed. Uses official `hookSpecificOutput` format per Claude Code docs.
+- **`read_for_edit` output** — already includes exact `Read(path, offset, limit)` command that passes through the hook, giving AI a clear path: `read_for_edit` → bounded `Read` → `Edit`.
+
+### Added
+- **Edit hook** — new PreToolUse hook matching Edit tool. Adds `additionalContext` suggesting `read_for_edit` for minimal code context. Doesn't block Edit — just provides a hint.
+- **Hook installer** — now installs and manages both Read and Edit hooks. Uninstall removes all Token Pilot hooks.
+
+## [0.5.3] - 2026-03-02
+
+### Changed
+- **`find_unused`** — completely rewritten with universal approach. Removed 60+ hardcoded framework-specific names. Now uses ast-index data: constructors filtered by name (`constructor`/`__init__`), Python dunder methods by `__*__` pattern, decorated symbols detected via `outline()` and shown separately with their decorators. No framework-specific knowledge.
+- **`formatFrameworkInfo`** (smart_read display) — removed hardcoded TypeORM (`Column`, `PrimaryGeneratedColumn`) and class-validator (`IsEmail`, `MinLength`) parsing. Now only detects standard HTTP verbs (GET/POST/PUT/DELETE/PATCH) which are protocol-level, not framework-specific. All other decorators shown as-is (`@DecoratorName`).
+- **`outline`** — route detection now universal. Instead of hardcoding `@Controller`, detects any class decorator with a path argument as route prefix. HTTP verb detection uses same universal pattern. Non-HTTP decorators shown as-is.
+
+## [0.5.2] - 2026-03-02
+
+### Fixed
+- **`project_overview`** — HINT no longer references deleted `search_code()`, now suggests `find_usages()` and `outline()`
+- **`related_files` imported_by** — now searches both `imports` AND `usages` from refs (not just imports), with increased limit (30). Cross-language filtering preserves same-family matches while removing false positives.
+- **`find_unused`** — excludes framework-implicit symbols (replaced by universal approach in 0.5.3)
+- **README** — updated handler file list (removed deleted handlers, added new ones)
+
+## [0.5.1] - 2026-03-02
+
+### Fixed
+- **`read_for_edit` symbol mode** — large symbols (>20 lines) now return only the first 20 lines instead of the entire method. Prevents returning 300+ lines when only a signature is needed for editing.
+- **`related_files` imported_by** — filter cross-language false positives. A TypeScript file no longer shows Python/Go/Rust files as importers. Refs are filtered by language family (JS/TS, Python, Go, JVM, etc.).
+- **`session_analytics`** — honest savings metric for `read_for_edit`. Reduced multiplier from 30x to 3x (realistic comparison vs `Read` with offset/limit, not vs full file).
+
+## [0.5.0] - 2026-03-02
+
+### Added
+- **`read_for_edit`** — killer feature for edit workflow. Returns RAW code (no line numbers) around a symbol or line, ready to copy as `old_string` for Edit. 97% fewer tokens than reading full file before editing.
+- **`related_files`** — import graph for any file: what it imports, what imports it, test files. Saves 3-5 Read calls per task.
+- **`outline`** — compact overview of all code files in a directory. One call instead of 5-6 smart_read calls. Framework-aware: shows HTTP routes for NestJS controllers.
+- **`read_symbol` show parameter** — `show: "full"|"head"|"tail"|"outline"` controls truncation. Default: auto (full ≤300 lines, outline >300).
+- **Framework-aware decorators** — smart_read/outline parse NestJS (`@Controller`+`@Get` → HTTP routes), TypeORM (`@Column` → types), class-validator (`@IsEmail` → constraints).
+
+### Removed
+- **`search_code`** — worse than Grep in practice, find_usages + Grep cover all use cases
+- **`export_ast_index`** — never used in real work, infrastructure tool only
+- **`context_status`** — debugging tool, not user-facing
+- **`forget`** — manual context management = poor design, should be automatic
+- **`changed_symbols`** — git diff + smart_read covers this use case
+
+### Changed
+- **12 focused tools** instead of 14 — removed 5 low-value, added 3 high-impact
+- Edit-heavy sessions: 5-10% → 40-50% token savings (via read_for_edit)
+- Average sessions: 20-25% → 45-55% token savings
+
+## [0.4.1] - 2026-03-02
+
+### Added
+- **Auto-install PreToolUse hook**: hook installs automatically on server start (Claude Code), no manual `install-hook` needed
+- **AI instructions template**: README includes ready-to-copy block for `.cursorrules` / `CLAUDE.md`
+
+### Changed
+- **Tool descriptions rewritten** — explicit "ALWAYS use instead of Read/cat", "use instead of Grep" for AI prioritization
+- README updated: PreToolUse hook section, MCP Tools table with "Instead of" column
+
+## [0.4.0] - 2026-03-02
+
+### Added
+- **Python class method parser**: smart_read/read_symbol shows all methods inside Python classes with visibility, decorators, async detection
+- **PHP class method parser**: same for PHP classes with public/private/protected, static
+- **Version display**: `project_overview` and `session_analytics` show `TOKEN PILOT v{version}`
+
+### Changed
+- **Removed find_callers** — did not save tokens vs grep, ast-index limitation with `this.method()` calls
+- **Removed find_implementations** — did not save tokens vs grep, ast-index limitation with decorators
+- **Removed class_hierarchy** — did not save tokens vs grep, poor results from ast-index
+- **14 focused tools** instead of 17 — only tools that actually save tokens or provide unique value
+
+### Fixed
+- **Mega-symbol truncation**: symbols >300 lines show head (50) + tail (30) + method outline instead of 71KB overflow
+- **Recursive findFlat**: unqualified method names (`run`, `_build_summary`) found inside class children
+
+## [0.3.2] - 2026-03-01
+
+### Fixed
+- **Python class methods**: smart_read now shows all methods inside Python classes (ast-index only returns class-level, token-pilot parses `def` methods with visibility, decorators, async detection)
+- **read_symbol Python**: `Orchestrator.run`, `Orchestrator._build_summary` — qualified and unqualified method access works (was returning entire 829-line class)
+- **Mega-symbol truncation**: symbols >300 lines show head (50) + tail (30) + method outline instead of 71KB overflow
+- **findFlat recursive**: unqualified method names (`run`, `_build_summary`) now found inside class children
+
+## [0.3.1] - 2026-03-01
+
+### Fixed
+- **find_usages**: combine `refs` + `search` with exact word boundary filtering — 0% result loss vs grep (was 40% loss with refs-only)
+- **read_symbol**: fix `Class.method` qualified names for flat outlines (ast-index lists methods as siblings, not children)
+- **read_symbol**: filter ast-index leaf name fallback by requested file (was returning symbols from wrong files)
+- **YAML smart_read**: 3-level nested parser with scalar values, array counts (was only showing top-level keys)
+- Removed all "Use Grep as fallback" hints — token-pilot gives complete results on its own
+
+## [0.3.0] - 2026-03-01
+
+### Added
+- **find_callers** tool — find all callers of a function, with optional call hierarchy tree (depth parameter)
+- **changed_symbols** tool — show symbol-level git changes (added/modified/removed) vs a base branch
+- **find_unused** tool — detect potentially unused/dead symbols in the project
+- 8 new ast-index client methods: `refs`, `map`, `conventions`, `callers`, `callTree`, `changed`, `unusedSymbols`, `fileImports`
+- Incremental index updates via `ast-index update` (fast) instead of full rebuild
+
+### Fixed
+- **find_usages**: rewritten to use `ast-index refs` — returns definitions + imports + usages in one call (was losing ~66% of results)
+- **project_overview**: rewritten to use `ast-index map` + `conventions` — shows architecture, frameworks, naming patterns, directory map with symbol kinds
+- **search_code**: deduplication of results (removes duplicate file:line entries)
+- **read_symbol**: structure-first lookup for `Class.method` qualified names with ast-index leaf fallback
+- **export_ast_index**: `all_indexed=true` option exports all files from ast-index, not just cached ones
+- **YAML smart_read**: expand one level of nesting (shows nested keys under top-level sections)
+
+### Changed
+- Total MCP tools: 14 → 17
+- ast-index commands used: 8 → 16
+- Index updates are now incremental by default (falls back to full rebuild only when needed)
+
+## [0.2.4] - 2026-03-01
+
+### Fixed
+- **search_code**: filter out garbage entries with empty file paths or `:undefined` lines
+- **read_symbol**: support `Class.method` and `Class::method` qualified names (structure-first lookup, ast-index leaf fallback)
+- **export_ast_index**: `all_indexed=true` option exports all files from ast-index, not just cached ones
+- **YAML smart_read**: expand one level of nesting (shows service names, nested keys under top-level sections)
+- Improved empty-cache message in export_ast_index with hint about `all_indexed`
+
+## [0.2.3] - 2026-03-01
+
+### Fixed
+- **ensureIndex**: plain rebuild first (indexes full monorepo), fallback to `--sub-projects` only if <5 files
+- **smart_read**: non-code files (YAML, JSON, Markdown, TOML) use structural summary instead of raw content dump
+- **smart_read**: unsupported files return truncated 60-line preview instead of full raw content
+- **class_hierarchy**: proper parser for ast-index text output (Parents/Children sections)
+- **project_overview**: uses directory name when package.json has no `name` field
+
+## [0.2.2] - 2026-03-01
+
+### Fixed
+- Published 0.2.1 contained stale Haiku files in dist/ (tsc doesn't clean old outputs)
+- Added `prebuild` script (`rm -rf dist`) to prevent stale artifacts
+- Added `chmod +x` in prepublishOnly to ensure bin is executable
+
+## [0.2.1] - 2026-03-01
+
+### Fixed
+- **RC3**: `search_code` now merges all ast-index result types (content_matches + symbols + files + references) — previously only used content_matches which was often empty
+- **RC4**: `class_hierarchy` and `implementations` parse text format as fallback when JSON parse fails
+- **RC6**: `read_symbol` auto-fetches outline from ast-index if no cached structure — no longer requires prior smart_read
+- `ensureIndex` uses `--sub-projects` flag for monorepo indexing
+
+### Removed
+- Reverted Haiku v0.2.1 — removed broken PersistentFileCache, DiffEngine, RealTokenEstimator, ContextWindowTracker, smart-read-xml, context-markup
+- Removed 3 heavy native dependencies: `better-sqlite3`, `js-tiktoken`, `diff`
+
+### Added
+- `start.sh` — bootstrap script for Claude Code plugin system
+
+## [0.2.0] - 2026-03-01
+
+### Fixed
+- **P0**: ast-index errors no longer silently swallowed — all search/usages/implementations/hierarchy/outline/symbol log errors to stderr
+- **P0**: `exec()` now captures and logs ast-index subprocess stderr
+- **P0**: `projectRoot` detected via `git rev-parse --show-toplevel` instead of `process.cwd()` (fixes wrong index root)
+- **P1**: `forget(all=true)` now clears both ContextRegistry and FileCache (fixes stale export_ast_index/read_diff after forget)
+- **P1**: `forget(path=X)` also invalidates FileCache for that path
+- **P2**: `read_symbol` supports PHP `::` separator (e.g. `RefundProcessor::refund`)
+- **P2**: `findInStructure` recursion fixed — supports 3+ level nesting (Namespace::Class::method)
+- `ensureIndex()`: verify index has content after `stats` — force rebuild if 0 files indexed
+
+### Changed
+- `project_overview`: now shows directory listing + ast-index stats (files, symbols, references) instead of stub
+- `project_overview`: added PHP (`composer.json`) detection
+
+## [0.1.6] - 2026-03-01 (unpublished)
+
+### Fixed
+- `ensureIndex()`: verify index has content after `stats` — force rebuild if 0 files indexed (fixes empty search results on first run)
+
+## [0.1.5] - 2026-03-01
+
+### Fixed
+- PreToolUse hook: read file path from stdin (Claude Code hook format) instead of `$FILE_PATH` env var
+- Hook now auto-suggests `smart_read` for large code files when Claude tries to use `Read`
+- `session_analytics`: now tracks all tools (read_symbol, read_range, read_diff, smart_read_many, search_code, find_usages, find_implementations, class_hierarchy) — previously only tracked smart_read
+- Empty search/usages/implementations results now show diagnostic hints (ast-index status, fallback suggestions)
+- `ensureIndex()` now logs build progress and errors to stderr
+
+## [0.1.4] - 2026-03-01
+
+### Fixed
+- Lazy file watcher: watch only loaded files instead of entire project root (fixes crash on Docker volumes, home dir, WSL)
+
+## [0.1.3] - 2026-03-01
+
+### Fixed
+- Chokidar file watcher error handler (partial fix, superseded by 0.1.4)
+
+## [0.1.2] - 2026-03-01
+
+### Fixed
+- ast-index integration: `outline` now parses text output (JSON format not supported in v3.24.0)
+- ast-index `symbol` response: handle array format, normalize field names (`line`→`start_line`, `path`→`file`)
+- ast-index `search` response: handle `{content_matches: [...]}` wrapper
+- ast-index `usages` response: map `context`→`text`, `path`→`file`
+- Server version now read dynamically from package.json
+
+### Added
+- `token-pilot doctor` — diagnostics command (checks ast-index, Node.js, config, updates)
+- `token-pilot --version` — print current version
+- Update check on server startup (non-blocking, logs to stderr)
+- `/mcp add` installation method documented for Claude Code chat
+- Troubleshooting section in README
+
+## [0.1.1] - 2026-03-01
+
+### Added
+- `npx -y token-pilot` — zero-install for any MCP client (Cursor, Cline, Continue, etc.)
+- Claude Code plugin marketplace support (`.claude-plugin/marketplace.json`)
+- `start.sh` bootstrap script — auto `npm install` + `npm run build` on first run
+- `npm publish` ready (`files` field, `prepublishOnly` script)
+- Universal install instructions in README for Claude Code, Cursor, Cline
+
+### Changed
+- `.mcp.json` now uses `start.sh` for reliable bootstrap
+- README reorganized: npx as primary install, from-source as fallback
+
+## [0.1.0] - 2026-03-01
+
+### Added
+
+- **Core Reading Tools**: `smart_read`, `read_symbol`, `read_range`, `read_diff`, `smart_read_many`
+  - AST-based structural overviews saving 80-95% tokens
+  - Small file pass-through (< 80 lines returned in full)
+  - O(n) diff algorithm for re-reads
+  - Advisory context registry with compact reminders
+- **Search & Navigation**: `search_code`, `find_usages`, `find_implementations`, `class_hierarchy`, `project_overview`
+  - Powered by ast-index (tree-sitter + SQLite FTS5)
+  - Cross-file symbol resolution
+- **Context Management**: `session_analytics`, `context_status`, `forget`
+  - Token savings tracking per tool and per file
+  - Advisory (non-blocking) context tracking
+- **Integration**: `export_ast_index`
+  - context-mode detection and complementary architecture
+  - AST data export for BM25 cross-indexing
+- **Infrastructure**
+  - Git HEAD watcher with selective cache invalidation on branch switch
+  - File watcher (chokidar) for automatic cache invalidation
+  - LRU file cache with configurable size limit
+  - Input validation for all tools (path traversal protection)
+  - Auto-download of ast-index binary from GitHub releases
+  - PreToolUse hook installer for Claude Code
+  - Claude Code plugin format (.claude-plugin/)
+  - Non-code structural summaries (JSON, YAML, Markdown, TOML)
+  - Configurable via `.token-pilot.json`