compound-loop

Engineering lifecycle skills with identical contracts, gates, and safety invariants in Claude Code and Codex.

compound-loop merges three lineages into one curated skill set:

• superpowers — process discipline (brainstorming, plans, TDD, systematic debugging, verification)
• compound-engineering — review pipeline rigor and knowledge compounding (persona lanes, confidence gating, docs/solutions)
• release-loop — a six-phase orchestrated lifecycle (Design → Plan → Implement → Review → Ship → Retro) with file-based state

Read PRINCIPLES.md first — it is the charter every skill enforces. Accumulated learnings from past work live in docs/solutions/, and CONCEPTS.md defines this repo's shared vocabulary — both are worth a look before working in an area they cover.

Skills

Skill	Role
release-loop	Orchestrates the six phases end to end; each phase invokes the standalone skill below
release	Owns the local versioned-release ceremony: artifact-derived CHANGELOG, synchronized manifests, USER-gated annotated tag
designing	Requirements exploration with scope tiering, rigor-gap probes, and mandatory measurable success criteria
planning	Implementation plans with stable unit IDs, per-unit test scenarios, and a risk-scored deepening pass
implementing	Plan execution with file-based handoffs, parallel-safety checks, and graceful degradation
tdd	Red-green-refactor discipline (Iron Law: no production code without a failing test)
debugging	Root-cause investigation with causal-chain gate and prediction discipline
worktree-isolation	Isolated workspaces: native tool → git worktree → work in place
reviewing	Lane-contract multi-perspective review with confidence gating and a machine-readable JSON mode
shipping	Verification gate, commit/PR protocol, CI watch loop, review-feedback resolution, merge gate
retrospective	Measured-vs-declared success comparison, carry-forward tracking, lessons
compound	Capture solved problems into docs/solutions/ and maintain CONCEPTS.md vocabulary
compound-refresh	Periodic audit of accumulated knowledge docs (keep/update/consolidate/replace/delete)

Install

Claude Code

/plugin marketplace add <path-or-repo>
/plugin install compound-loop

Codex

User-level skill discovery root is ~/.agents/skills/ (current official docs; verify with your Codex version — legacy setups used ~/.codex/skills/):

for s in skills/*/; do ln -s "$(pwd)/$s" ~/.agents/skills/"$(basename "$s")"; done

Skills are invoked as /name in Claude Code and $name in Codex.

Cross-harness contracts

• Skills communicate through files, never through harness-specific channels: review findings as JSON artifacts (mode:agent), knowledge capture with terminal signal strings (mode:headless), loop state in .release-loop/progress.md.
• Every parallel dispatch degrades: native parallel subagents → sequential passes → single-call fallback.
• Plan documents follow a single schema (frontmatter + stable U-IDs + categorized test scenarios) shared by planning, implementing, and reviewing. See schemas/.

Python compatibility gate

The supported interpreter boundary is declared once in schemas/python-support.json. Maintainers can run bash scripts/test-python-compatibility.sh all for the operational gate or bash scripts/test-python-compatibility.sh fixtures for its disposable fixture suite. The gate fails closed when either endpoint is unavailable or does not match the declared implementation and minor version. When command discovery is not suitable, PYTHON_OLDEST and PYTHON_NEWEST accept absolute interpreter paths.

The registry in scripts/test-python-compatibility.sh has two explicit classes: committed Python entry points and Python generated for later execution. Add new artifacts to that registry rather than copying endpoint policy into another file. Direct heredocs that execute immediately remain outside this boundary.