compound-agent 2.2.0 → 2.3.0
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- package/CHANGELOG.md +35 -0
- package/package.json +5 -5
package/CHANGELOG.md
CHANGED
|
@@ -7,6 +7,41 @@ All notable changes to this project will be documented in this file.
|
|
|
7
7
|
The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.1.0/),
|
|
8
8
|
and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
|
|
9
9
|
|
|
10
|
+
## [2.3.0] - 2026-03-25
|
|
11
|
+
|
|
12
|
+
### Added
|
|
13
|
+
|
|
14
|
+
- **Smarter failure escalation** (Epic 1): When Claude Code tools fail repeatedly (≥3 same-target or ≥3 total), the post-tool-failure hook now searches the lesson database for relevant past solutions instead of showing a static tip. Includes injectable `LessonSearchFunc`, rune-safe token extraction from error output, BM25-scored FTS5 search with 500ms timeout, 4-level fallback chain, and confidence annotations for low-scoring matches.
|
|
15
|
+
- **Architect intelligence** (Epic 2): Architect skill gains a research sufficiency gate (search `ca search` + `docs/research/` before spec writing) and automatic integration verification epic creation during decomposition.
|
|
16
|
+
- **Acceptance criteria protocol** (Epic 3): AC table generation from EARS requirements in the plan phase, AC reading in the work phase, and an AC gate between plan and work in the cook-it pipeline.
|
|
17
|
+
- **Lesson-calibrated reviews** (Epic 4): Review skill now searches past lessons per-reviewer (3-5 cap, recency bias) and conditionally triggers runtime verification via the new `runtime-verifier` agent role skill for ephemeral Playwright testing.
|
|
18
|
+
- **Integration verification** (Epic 5): Template drift detection test that verifies reviewer names in skill templates have matching agent role skill directories.
|
|
19
|
+
- **Context-aware FTS5 search**: New `SearchKeywordScoredORContext()` method threads `context.Context` through to `QueryContext`, enforcing the 500ms timeout contract at the database layer.
|
|
20
|
+
|
|
21
|
+
### Fixed
|
|
22
|
+
|
|
23
|
+
- **Failure threshold off-by-one**: `sameTargetThreshold` changed from 2 to 3 to match EARS spec FE-1 (≥3 failures before escalation).
|
|
24
|
+
- **Phantom reviewer names**: Fixed `cct-reviewer` → `cct-subagent` and `docs-reviewer` → `doc-gardener` in review and compound skill templates.
|
|
25
|
+
- **Cyclop complexity violation**: Extracted `installAgentRoleSkillReferences()` helper from `InstallAgentRoleSkills()` in `primitives.go` to stay within the cyclomatic complexity limit.
|
|
26
|
+
- **Comma-separated epic IDs in loop generator**: `ca loop` now converts comma-separated epic IDs to space-separated for bash `for` loop compatibility.
|
|
27
|
+
|
|
28
|
+
## [2.2.1] - 2026-03-24
|
|
29
|
+
|
|
30
|
+
### Added
|
|
31
|
+
|
|
32
|
+
- **Loop generator production parity**: Brought the Go `ca loop` script generator to full parity with the hardened production `infinity-loop.sh`. Ported crash handler (EXIT trap with status file logging), 3-layer memory safety (repo-scoped orphan cleanup, memory gate, background watchdog with PID tracking), `parse_json()` with jq/python3 fallback, dependency-aware epic ordering (`check_deps_closed()`), dual-file anchored marker detection, and CLI prerequisite checks.
|
|
33
|
+
- **Loop review and improve phases**: Implemented 8 missing flags from the former TS CLI: `--reviewers`, `--review-every`, `--max-review-cycles`, `--review-blocking`, `--review-model`, `--improve`, `--improve-max-iters`, `--improve-time-budget`. Review phase supports multi-model spawning, session management, and review cycles with implementer fix loop. Improve phase supports topic discovery, iteration with rollback, and time budget.
|
|
34
|
+
- **Field-tested enhancements**: Git status check after epic completion (auto-commits if dirty), git push at loop end (with remote availability check), reviewer availability summary logging.
|
|
35
|
+
- **Infinity-loop reference docs**: Added troubleshooting section and real-world example from compound-agent's own 6-epic loop run.
|
|
36
|
+
|
|
37
|
+
### Fixed
|
|
38
|
+
|
|
39
|
+
- **Review triggers were dead code**: Review phase triggers were placed after an exit statement and never executed.
|
|
40
|
+
- **Undefined variable in periodic trigger**: `$RESULT` replaced with `$SUCCESS` in periodic review trigger.
|
|
41
|
+
- **Review base SHA reset in loop**: `REVIEW_BASE_SHA` was incorrectly reset inside the while loop on every iteration.
|
|
42
|
+
- **Improve phase stdin piping**: `extract_text` was using file args instead of stdin pipe.
|
|
43
|
+
- **Conditional exit**: Exit line now omitted when improve phase follows review phase.
|
|
44
|
+
|
|
10
45
|
## [2.2.0] - 2026-03-23
|
|
11
46
|
|
|
12
47
|
### Added
|
package/package.json
CHANGED
|
@@ -1,6 +1,6 @@
|
|
|
1
1
|
{
|
|
2
2
|
"name": "compound-agent",
|
|
3
|
-
"version": "2.
|
|
3
|
+
"version": "2.3.0",
|
|
4
4
|
"type": "module",
|
|
5
5
|
"description": "Learning system for Claude Code — avoids repeating mistakes across sessions",
|
|
6
6
|
"bin": {
|
|
@@ -50,10 +50,10 @@
|
|
|
50
50
|
"knowledge-management"
|
|
51
51
|
],
|
|
52
52
|
"optionalDependencies": {
|
|
53
|
-
"@syottos/darwin-arm64": "2.
|
|
54
|
-
"@syottos/darwin-x64": "2.
|
|
55
|
-
"@syottos/linux-arm64": "2.
|
|
56
|
-
"@syottos/linux-x64": "2.
|
|
53
|
+
"@syottos/darwin-arm64": "2.3.0",
|
|
54
|
+
"@syottos/darwin-x64": "2.3.0",
|
|
55
|
+
"@syottos/linux-arm64": "2.3.0",
|
|
56
|
+
"@syottos/linux-x64": "2.3.0"
|
|
57
57
|
},
|
|
58
58
|
"author": "Nathan Delacrétaz",
|
|
59
59
|
"license": "MIT",
|