sigmap 6.14.0 → 6.15.0

package/CHANGELOG.md

@@ -10,6 +10,32 @@ Format: [Semantic Versioning](https://semver.org/)
 
 ---
 
+## [6.15.0] — 2026-06-09
+
+### Added
+- **`verify-ai-output` — Hallucination Guard Reliable MVP (Phase 1, #232):**
+  - Two new deterministic detectors — **`fake-test-file`** (a referenced `*.test`/`*.spec`/`__tests__`/`test_*.py` path absent on disk, reported separately from `fake-file`) and **`fake-npm-script`** (`npm run X` where `X` is not a `package.json` script).
+  - **Closest-match suggestions** (`src/verify/closest-match.js`) — Levenshtein + file-proximity over the symbol index attaches a heuristic hint to flagged names ("Did you mean `loadConfig()` in `src/config/loader.js:42`?"). Labeled as heuristic, with its own confidence bucketing.
+  - **Finalized JSON schema** — every issue now carries `{ type, value, line, location, message, confidence, suggestion }`; `summary` gains `withSuggestion`. Detection confidence is `high` for path/dep/script checks and `medium` for symbol checks.
+  - **Parser hardening** — multi-line `import { … } from '…'` statements and TypeScript `import X = require('…')` are now detected; `npm`/`pnpm`/`yarn run` script references are extracted.
+  - **HTML report view** (`src/format/verify-report.js`) — `sigmap verify-ai-output <answer> --report [out.html]` writes a standalone, self-contained red/amber/green report (no external assets/scripts); a Markdown renderer shares the same structure for CI/PR comments.
+  - **Proof harness** — `npm run benchmark:verify` scores each detector group against labeled cases and enforces precision targets (file ≥ 95%, import ≥ 85%, symbol ≥ 75%, script ≥ 95%), emitting a precision/recall CSV. Runs offline via a synthetic self-test; point it at real repos with `--manifest`.
+  - New guide: `docs-vp/guide/verify-ai-output.md`.
+- **Memory tools — `note`, `status`, and the `read_memory` MCP tool (Phase 1.5, #233):** closes the cold-start gap so an agent can recall *what we were doing and why* without re-scanning the repo.
+  - **`sigmap note "<text>"`** — append to a cross-session decision log stored as append-only NDJSON at `.context/notes.ndjson` (each entry records text, ISO timestamp, and git branch). `sigmap note` with no text lists recent notes (`--list <N>`, `--json`). New module `src/session/notes.js`.
+  - **`sigmap status`** — repo state at a glance: branch (with an unborn-branch fallback), dirty-file count, last index run (time, version, file count) with a **staleness** signal (tracked files modified since the last index), and notes summary. `--json` supported.
+  - **`read_memory` MCP tool (11th tool)** — returns recent notes (most recent first) plus the last ranking-session focus from `ask`, formatted for agent consumption. Registered in `src/mcp/tools.js`, `handlers.js`, and `server.js`; bundled into the standalone binary.
+  - New guide: `docs-vp/guide/memory.md`.
+- The `verify-ai-output` (29 cases) and new `memory-tools` (13 cases) integration suites are now part of `npm run test:integration`.
+
+### Changed
+- MCP server now exposes **11 tools** (was 10) with the addition of `read_memory`; `version.json` `mcp_tools`, the `mcp.md` guide, and all tool-count test gates updated accordingly.
+
+### Fixed
+- `npm run test:integration` referenced a non-existent `test/integration/mcp-server.test.js`; corrected to `test/integration/mcp/server.test.js`.
+
+---
+
 ## [6.14.0] — 2026-06-07
 
 ### Added
@@ -20,6 +46,10 @@ Format: [Semantic Versioning](https://semver.org/)
   - Markdown report by default, `--json` for CI (`{ file, issues, summary }`). Exits `1` when any issue is found, `0` when clean.
   - New modules `src/verify/parsers.js` (file/import/symbol/code-block extraction) and `src/verify/hallucination-guard.js` (`verify(answerText, cwd, opts)`); all external lookups are injectable so the core is unit-testable.
 
+### Fixed
+
+- **Standalone binary build:** registered the new `src/verify/parsers` and `src/verify/hallucination-guard` modules in the `gen-context.js` `__factories` bundle. Without them the Release Binaries (Node SEA) build failed its pre-flight check (`missing from __factories`); `requireSourceOrBundled` falls back to `__require` in the single-file binary where `src/` is not present.
+
 ---
 
 ## [6.13.0] — 2026-06-05

package/README.md

@@ -47,10 +47,10 @@ SigMap extracts function and class signatures from your codebase and feeds the r
 
 ## Why SigMap?
 
-- **81.1% hit@5** — right file found in top 5 results (vs 13.6% baseline)
-- **96.5% token reduction** — average across 21 real repos
-- **53.3% task success rate** — up from 10% without context
-- **1.66 prompts per task** — down from 2.84 (41.8% fewer retries)
+- **75.6% hit@5** — right file found in top 5 results (vs 13.6% baseline)
+- **97.1% token reduction** — average across 21 real repos
+- **51.1% task success rate** — up from 10% without context
+- **1.73 prompts per task** — down from 2.84 (39.0% fewer retries)
 - **31 languages supported** — TypeScript, Python, Go, Rust, Java, R, and 25 others
 - **No vendor lock-in** — works with any AI assistant or local LLM
 - **No API costs** — use local models (Ollama, llama.cpp, vLLM) with zero token fees
@@ -63,7 +63,7 @@ SigMap extracts function and class signatures from your codebase and feeds the r
 
 | Without SigMap | With SigMap |
 |---|---|
-| ❌ Guessing which files are relevant | ✅ Right file in context — 81% of the time |
+| ❌ Guessing which files are relevant | ✅ Right file in context — 76% of the time |
 | ❌ Sending the full repo to your AI | ✅ Minimal context — only what matters |
 | ❌ Embeddings / vector DB required | ✅ Grounded answers, no infra needed |
 
@@ -87,13 +87,13 @@ Ask → Rank → Context → Validate → Judge → Learn
 ## Benchmark
 
 ```
-Benchmark : sigmap-v6.13-main (21 repositories, including R language)
-Date      : 2026-06-05
+Benchmark : sigmap-v6.15-main (21 repositories, including R language)
+Date      : 2026-06-09
 
-Hit@5          : 81.1%   (baseline 13.6%  — 6.0× lift)
-Token reduction: 96.5%   (across 21 repos)
-Prompt reduction : 41.8% (2.84 → 1.66 prompts per task)
-Task success   : 53.3%   (baseline 10%)
+Hit@5          : 75.6%   (baseline 13.6%  — 5.6× lift)
+Token reduction: 97.1%   (across 21 repos)
+Prompt reduction : 39.0% (2.84 → 1.73 prompts per task)
+Task success   : 51.1%   (baseline 10%)
 Repos tested   : 21 (JavaScript, Python, Go, Rust, Java, R, C++, C#, Dart, Swift, Ruby, PHP, Scala, Kotlin, and more)
 ```