v2.3.1

Changes

• Add root .claude-plugin/plugin.json manifest required for marketplace submission
• Sync plugin version to 2.3.1 across all manifests
• Fix old plugins/review-loop/ path references in design specs
• Remove external tooling artifacts (docs/superpowers/plans/)
• Update README with version history and marketplace info

v2.3.0: Autooptimize Skill & Test Harness

What's New

Autooptimize Skill (.claude/skills/autooptimize/)

Generalized version of Karpathy's autoresearch methodology for optimizing any project artifact — prompts, configs, orchestration logic, or code. Runs autonomous experiment loops: execute → eval → mutate → keep/discard.

Test Harness (autooptimize-harness/)

Comprehensive test infrastructure for measuring review-loop plugin effectiveness:

• Single-prompt testing — Test individual prompts (design-review, code-review, code-implement) via codex exec with fixtures containing known defects. ~1-5 min per run.
• End-to-end pipeline testing — Run the full review-loop via claude -p in isolated temp dirs, then execute produced code and tests automatically.
• LLM-as-judge eval — Binary pass/fail scoring using Claude as evaluator.
• 3 test scenarios — CLI calculator (simple), rate limiter (medium), KV store (complex).
• Fixtures — Design docs, code samples, and specs with known subtle defects for evaluating review quality.

Baseline Evaluation Results

Scenario	Duration	Design Rounds	Code Rounds	Tests	Status
CLI calculator	6 min	1	1	22/22	PASS
Rate limiter	95 min	6	2	32/32	PASS
KV store	~10 hr	6	2	88/88	PASS

Key findings:

• All review prompts score 100% on known-issue detection — prompt quality is already strong
• E2E pipeline produces correct, tested code across all complexity levels (142 total tests passing)
• Design stage convergence speed is the primary optimization target for complex tasks

review-loop v2.2.0: Independent Validation Round

What's New

Independent Validation Round

After the existing Claude-Codex review loop converges (or exhausts its 5 rounds), a fresh Codex instance with zero shared review history re-examines the output using blind-spot-focused prompts. If issues are found, they're fed back into the regular loop for fix + re-validation (max 2 cycles per stage).

This addresses the false convergence problem: Claude and Codex agree there are no remaining issues, but independent review still finds entirely new categories of problems — shared blind spots from operating under the same prompt framework.

New Modes

Mode	Role	Purpose
independent-design-review	Codex (READ-ONLY)	Blind-spot review of design docs
independent-code-review	Codex (READ-ONLY)	Blind-spot review of code changes
validation-design-fix	Codex (READ-ONLY)	Review design fixes after validation
validation-fix	Codex (implementer)	Fix code issues found by validation

Key Details

• Context isolation: prompt-assembly layer injects zero review history; prompt-instruction layer explicitly prohibits reading specs/reviews/ and .claude/
• Composite round tokens: c1/c2 for validation cycles, c1f1/c2f2 for fix rounds — no artifact naming collisions
• Failure handling: retry once on timeout/failure, skip + log on double failure, surface in final output (never falsely claim validation passed)
• Ordering: validation runs before verify round (verify remains the terminal pass)
• Role separation: AGENTS.md distinguishes independent reviewers, design fix reviewers, and code fixers with distinct constraints
• 20 new tests (37 total), all passing

File Changes

• 4 new prompt templates in review-loop/prompts/
• Modified: common.sh (4 new modes, composite round validation, append_diff_section with temp git index), review-loop.md (validation + verify + output flows), AGENTS.md (role separation), tests/review-loop.test.sh
• No changes to run-review-bg.sh, check-review.sh, kill-review.sh, hooks, or state file schema

See specs/design.md for the full implementation design.

v2.1.0

review-loop v2.1.0

Addresses two production issues discovered in v2:

Fresh Independent Reviews

• Every review round now mandates a full audit of the entire document/diff
• Verify rounds strip all prior review context and prepend a "FULL INDEPENDENT REVIEW" header
• Eliminates attention narrowing where successive rounds focused only on previously found issues

Optional Brainstorming Stage

• If superpowers:brainstorming is available, user is asked whether to brainstorm before designing
• Output saved to specs/brainstorm.md as supplementary context (task description stays authoritative)
• Session-scoped via brainstorm_done flag — stale artifacts from previous sessions are ignored
• Brainstorming suppressed in all subsequent Codex prompts

Improved Cancellation

• Records start_branch and start_sha in state file
• Cancel restores starting branch/commit and deletes session branch
• Handles detached HEAD correctly
• git reset -- . && git checkout -- . && git clean -fd clears staged + unstaged + untracked

Other Changes

• specs/brainstorm.md added to code-stage protected paths and staging exclusions
• Code-stage verify is Claude-only (no Codex invocation)
• Updated tests (18 total, all passing)
• Updated AGENTS.md for three-stage workflow

Design Review Audit Trail

The implementation spec went through 5 rounds of Codex design review + a verification pass. All review artifacts are preserved in specs/reviews/design/.

Full Changelog: 6630b26...v2.1.0