compound-agent 1.6.5 → 1.7.3

package/CHANGELOG.md

@@ -9,6 +9,78 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 ## [Unreleased]
 
+## [1.7.3] - 2026-03-09
+
+### Added
+
+- **Update notification**: CLI checks npm registry for newer versions on startup (24h file-based cache, non-blocking). Notification displays after command output (TTY only) and in `ca prime` context.
+
+### Fixed
+
+- **Spec-dev epic type**: `bd create` in spec-dev Phase 4 now explicitly uses `--type=epic`, preventing epics from defaulting to task type. Plan phase also validates the epic type and corrects it if needed.
+- **Update-check hardening**: Added explicit `res.ok` check in `fetchLatestVersion`, removed dead `checkedAt` cache field, removed redundant type cast.
+
+## [1.7.2] - 2026-03-09
+
+### Added
+
+- **Loop review phase**: `ca loop` can now spawn independent AI reviewers (Claude Sonnet, Claude Opus, Gemini, Codex) in parallel after every N completed epics. Reviewers produce severity-tagged reports, an implementer session fixes findings, and reviewers are resumed (not fresh) to verify fixes. Iterates until all approve or max cycles reached. New CLI options: `--reviewers`, `--review-every`, `--max-review-cycles`, `--review-blocking`, `--review-model`. Gracefully skips unavailable CLIs.
+
+### Fixed
+
+- **Security: command injection in `ca test-summary --cmd`**: User-supplied test command is now validated against an allowlist of safe prefixes (`pnpm`, `npm`, `vitest`, etc.) and shell metacharacters are rejected.
+- **Security: shell injection in `ca doctor`**: Replaced `execSync(cmd, {shell})` with `execFileSync('bd', ['doctor'])` to avoid shell interpretation.
+- **Portable timeout for macOS**: Generated loop scripts now use a `portable_timeout()` wrapper that tries GNU `timeout`, then `gtimeout` (Homebrew coreutils), then a shell-based kill/watchdog fallback. Previously failed silently on macOS.
+- **Session ID python3 fallback**: Review phase session ID management now falls back to python3 when `jq` is unavailable, with a centralized `read_session_id()` helper.
+- **Git diff window stability**: Replaced fragile `HEAD~$N..HEAD` commit-count arithmetic with SHA-based `$REVIEW_BASE_SHA..HEAD` diff ranges, immune to rebases and cherry-picks.
+- **ID collision risk**: Memory item IDs now use 64-bit entropy (16 hex chars) instead of 32-bit (8 hex chars).
+- **JSONL resilience**: Malformed lines in JSONL files are now skipped with try/catch per line instead of crashing the entire read.
+- **Stdin timeout leak**: `clearTimeout` now called in `finally` block for stdin reads in retrieval and hooks.
+- **Double JSONL read eliminated**: `readMemoryItems()` now returns `deletedIds` set, removing the need for a separate `wasLessonDeleted()` file read.
+- **FTS5 trigger optimization**: SQLite update trigger now scoped to FTS-indexed columns only, reducing unnecessary FTS rebuilds.
+- **Clustering noise accuracy**: Single-item clusters now correctly returned as `noise` instead of an always-empty noise array.
+- **Embed-worker path validation**: `embed-worker` command now validates that `repoRoot` exists and is a directory before proceeding.
+- **Script check timeout**: Rule-based script checks now have a 30-second default timeout, configurable via `check.timeout`.
+
+### Changed
+
+- **Anchored approval detection**: Review loop now uses `^REVIEW_APPROVED` anchored grep to prevent false positives from partial-line matches.
+- **Numeric option validation**: `--review-every` and `--max-review-cycles` now reject NaN, negative, and non-integer values.
+- **`isModelUsable()` replaced**: `compound` command now uses lightweight `isModelAvailable()` (fs check) instead of loading the 278MB model just to probe.
+- **Dead code removed**: `addCompoundAgentHook()`, back-compat hook aliases (`CLAUDE_READ_TRACKER_HOOK_CONFIG`, `CLAUDE_STOP_AUDIT_HOOK_CONFIG`), and `wasLessonDeleted()` removed.
+- **Hardcoded model extracted**: Five occurrences of `'claude-opus-4-6'` in loop.ts extracted to `DEFAULT_MODEL` constant.
+- **EPIC_ID_PATTERN deduplicated**: `watch.ts` now imports `LOOP_EPIC_ID_PATTERN` from `loop.ts` instead of maintaining a duplicate.
+- **`warn()` output corrected**: `shared.ts` warn helper now writes to `stderr` instead of `stdout`.
+- **Templates import fixed**: `templates.ts` now imports `VERSION` from `../version.js` instead of barrel re-export.
+
+## [1.7.1] - 2026-03-09
+
+### Added
+
+- **Scenario testing integration**: Spec-dev Phase 3 now generates scenario tables from EARS requirements and Mermaid diagrams with five categories (happy, error, boundary, combinatorial, adversarial). Review phase verifies coverage via a new `scenario-coverage-reviewer` agent using heuristic AI-driven matching.
+- **Scenario coverage reviewer**: New medium-tier AgentTeam reviewer that matches test files against epic scenario tables and flags gaps (P1) or partial coverage (P2). Spawned for diffs >100 lines.
+
+### Fixed
+
+- **Stale reviewer count in tests**: Updated "5 reviewer perspectives" test to "6" with `scenario-coverage` assertion. Removed no-op `.replace('security-', 'security-')` in escalation wiring test.
+
+## [1.7.0] - 2026-03-08
+
+### Added
+
+- **Loop observability**: Generated loop script now writes `.loop-status.json` (real-time epic/attempt/status) and `loop-execution.jsonl` (append-only result log with per-epic duration and end-of-loop summary). Enables `ca watch --status` and post-mortem forensics.
+- **ESLint rule `no-solo-trivial-assertion`**: Custom rule that warns when a test's only assertion is `toBeDefined()`, `toBeTruthy()`, `toBeFalsy()`, or `toBeNull()`. Registered but not yet enabled (requires cleanup of ~40 existing violations).
+
+### Fixed
+
+- **Loop 0-byte log resilience**: `extract_text` pipeline could produce 0-byte log files while the trace JSONL had valid content, causing the loop to falsely detect failure. New `detect_marker()` function checks the macro log first (anchored grep), then falls back to the trace JSONL (unanchored grep). Includes health check warning on extraction failure.
+- **Search fallback when embeddings unavailable**: `retrieveForPlan()` no longer throws when the embedding model is missing or broken. Falls back to keyword-only search with a console warning.
+
+### Changed
+
+- **Anti-cargo-cult reviewer strengthened**: Added three new subtle anti-patterns to the reviewer agent: solo trivial assertions, substring-only `toContain()` checks, and keyword-presence tests. Each with bad/good examples.
+- **Loop template extraction**: Bash script templates moved to `loop-templates.ts` to stay within lint limits.
+
 ## [1.6.5] - 2026-03-07
 
 ### Fixed

package/README.md

@@ -190,6 +190,11 @@ The CLI binary is `ca` (alias: `compound-agent`).
 | `ca loop -o <path>` | Custom output path (default: `./infinity-loop.sh`) |
 | `ca loop --max-retries <n>` | Max retries per epic on failure (default: 1) |
 | `ca loop --force` | Overwrite existing script |
+| `ca loop --reviewers <names...>` | Enable review phase with specified reviewers (claude-sonnet, claude-opus, gemini, codex) |
+| `ca loop --review-every <n>` | Review every N completed epics (0 = end-only, default: 0) |
+| `ca loop --max-review-cycles <n>` | Max review/fix iterations (default: 3) |
+| `ca loop --review-blocking` | Fail loop if review not approved after max cycles |
+| `ca loop --review-model <model>` | Model for implementer fix sessions (default: claude-opus-4-6) |
 
 ### Knowledge
 
@@ -263,7 +268,7 @@ pnpm lint             # Type check + ESLint
 
 | Script | Duration | Use Case |
 |--------|----------|----------|
-| `pnpm test:fast` | ~6s | Rapid feedback during development |
+| `pnpm test:fast` | ~12s | Rapid feedback during development |
 | `pnpm test` | ~60s | Full suite before committing |
 | `pnpm test:changed` | varies | Only tests affected by recent changes |
 | `pnpm test:watch` | - | Watch mode for TDD workflow |