compound-agent 2.4.1 → 2.5.1
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 +26 -0
- package/package.json +5 -5
package/CHANGELOG.md
CHANGED
|
@@ -9,6 +9,32 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
|
|
|
9
9
|
|
|
10
10
|
## [Unreleased]
|
|
11
11
|
|
|
12
|
+
## [2.5.1] - 2026-03-28
|
|
13
|
+
|
|
14
|
+
### Added
|
|
15
|
+
|
|
16
|
+
- **Visual verification in polish audit**: Reviewers now auto-detect UI projects and take Playwright screenshots at 4 viewports (375px, 768px, 1024px, 1440px) as part of their audit, critiquing layout, spacing, contrast, hierarchy, and responsiveness alongside static code analysis. Graceful degradation when Playwright is unavailable or no UI detected (P3/INFO + [NEEDS_QA] fallback).
|
|
17
|
+
- **Visual verification in cook-it review**: Step 10b in the review skill takes Playwright screenshots when the Verification Contract requires `browser_evidence`, `design_craft_check`, or `responsive_check`. References QA Engineer detection logic for framework auto-detection.
|
|
18
|
+
|
|
19
|
+
## [2.5.0] - 2026-03-27
|
|
20
|
+
|
|
21
|
+
### Added
|
|
22
|
+
|
|
23
|
+
- **Polish Loop (Phase 6)**: New `ca polish` CLI command generates a standalone bash script for iterative quality refinement. Runs N cycles of: multi-model audit fleet evaluating full implementation against the build-great-things pre-ship checklist (34 quality items + 12 laziness anti-patterns), mini-architect decomposing findings into improvement epics, and inner infinity loop implementing them. Supports claude-opus, claude-sonnet, gemini, and codex as reviewers. Includes dry-run mode, crash handler with status JSON, reviewer CLI detection with graceful degradation, and per-cycle observability (individual reports, synthesized reports, architect logs).
|
|
24
|
+
- **Architect Phase 6**: Opt-in phase in architect SKILL.md that activates after the infinity loop completes. Gate 5 for user confirmation. Generates and launches the polish loop script with monitoring commands.
|
|
25
|
+
- **Polish loop reference docs**: `architect/references/polish-loop/README.md` (configuration reference) and `audit-prompt.md` (audit prompt design explaining how it differs from the review fleet).
|
|
26
|
+
|
|
27
|
+
### Fixed
|
|
28
|
+
|
|
29
|
+
- **Polish loop: dry-run crash** (P0): `POLISH_EPICS` variable initialized before the loop and at start of each cycle to prevent unbound variable error under `set -u`.
|
|
30
|
+
- **Polish loop: crash handler exit code** (P0): EXIT trap now preserves the original exit code with `exit $exit_code` so callers detect failures.
|
|
31
|
+
- **Polish loop: log() ordering** (P1): `log()` function defined before crash handler in script assembly to prevent `log: command not found` during early failures.
|
|
32
|
+
- **Polish loop: missing ca prerequisite** (P1): Added `command -v ca` check alongside claude and bd.
|
|
33
|
+
- **Polish loop: ARG_MAX risk** (P1): Reviewer prompts piped via stdin (`-p - < file`) instead of command substitution to handle large specs.
|
|
34
|
+
- **Polish loop: architect heredoc expansion** (P1): Mini-architect prompt uses quoted heredoc + file-based injection to prevent shell expansion of report content containing `$` or backticks.
|
|
35
|
+
- **Polish loop: spec file validation** (P1): Script fails fast if spec file doesn't exist at runtime.
|
|
36
|
+
- **Polish loop: git commit before push**: Post-loop commits synthesized reports and status artifacts before pushing.
|
|
37
|
+
|
|
12
38
|
## [2.4.1] - 2026-03-26
|
|
13
39
|
|
|
14
40
|
### Fixed
|
package/package.json
CHANGED
|
@@ -1,6 +1,6 @@
|
|
|
1
1
|
{
|
|
2
2
|
"name": "compound-agent",
|
|
3
|
-
"version": "2.
|
|
3
|
+
"version": "2.5.1",
|
|
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": "2.5.1",
|
|
54
|
+
"@syottos/darwin-x64": "2.5.1",
|
|
55
|
+
"@syottos/linux-arm64": "2.5.1",
|
|
56
|
+
"@syottos/linux-x64": "2.5.1"
|
|
57
57
|
},
|
|
58
58
|
"author": "Nathan Delacrétaz",
|
|
59
59
|
"license": "MIT",
|