OACP Skills

Reusable skills for AI coding agents coordinating over the Open Agent Coordination Protocol.

Skills

oacp

The OACP protocol primer — teaches a runtime how to use the CLI and message protocol: version contract, workspace orientation, oacp send recipe with type taxonomy, oacp inbox vs oacp watch, review-loop semantics, and conventions. Auto-triggers when an agent works with OACP. Foundation for the workflow skills below.

Runtimes: Claude Code, Codex

# Claude Code — auto-triggers based on context, or invoke explicitly
/oacp

---

check-inbox

Single-pass inbox processor. Scans for pending OACP messages, processes each one (read, act, reply, delete), then exits. In Claude Code, pair with Monitor + oacp watch (event-driven, preferred) or fall back to /loop polling for continuous monitoring. In Codex, run the command once or wrap it with /loop for recurring checks. Honors OACP Phase 1 receiver autonomy (always_pause / auto_review) with a 4-gate evaluator, audit events, and a threshold-exceeded checkpoint.

Runtimes: Claude Code, Codex

# Claude Code / Codex
/check-inbox                        # single-pass scan
/check-inbox --project myproject    # explicit project
/loop 2m /check-inbox               # polling fallback (every 2 min)

---

review-loop-reviewer

Reads a PR diff, produces structured findings (blocking / non-blocking), renders a verdict (LGTM or REQUEST_CHANGES), and sends feedback to the author's OACP inbox. Supports multi-round reviews with --poll.

Runtimes: Claude Code, Codex

# Claude Code (triggered via check-inbox when a review_request arrives)
/review-loop-reviewer 42 --author claude
/review-loop-reviewer 42 --author claude --poll   # wait for author's next push

---

review-loop-author

Picks up review findings from inbox or PR comments, applies fixes, pushes updates, and sends a review_addressed message back to the reviewer. Closes the review loop.

Runtimes: Claude Code, Codex

# Claude Code (triggered via check-inbox when review_feedback arrives)
/review-loop-author 42 --reviewer codex
/review-loop-author 42 --reviewer codex --poll     # wait for reviewer's next round

---

self-improve

Audits the agent's operating system — skills, memory files, runtime config, and AGENTS.md or CLAUDE.md instructions — for staleness, contradictions, gaps, and bloat. Proposes and applies surgical fixes with approval.

Runtimes: Claude Code, Codex

# Claude Code / Codex
/self-improve                  # full review (skills + memory + runtime config)
/self-improve skills           # review only skills that ran this session
/self-improve memory           # review memory files only
/self-improve agents-md        # Codex: review AGENTS.md files + settings only
/self-improve claude-md        # Claude Code: review CLAUDE.md files + settings only
/self-improve friction         # Claude Code: review only session-friction patterns
/self-improve <skill-name>     # review a single specific skill

---

doctor

Wraps oacp doctor for agent self-diagnostics. Runs checks across environment, workspace, inbox health, schemas, and agent status. Auto-fixes safe issues (missing inboxes, stale timestamps) and reports remaining blockers.

Runtimes: Claude Code, Codex

# Claude Code
/doctor                        # auto-detect project, full diagnostics
/doctor --project myproject    # explicit project name

# CLI directly
oacp doctor --project myproject --fix        # auto-fix safe issues
oacp doctor --project myproject --fix --json # structured output

---

wrap-up

End-of-session cleanup in one command. 8-step sequence: cleanup stale artifacts → optional debrief → org-memory event capture → /self-improve → commit current repo → optional OACP memory sync → pull-rebase + push → summary. Hard dependency on /self-improve; optional integrations with /debrief and the oacp CLI (≥0.3.0).

Runtimes: Claude Code, Codex

# Command
/wrap-up            # full sequence
/wrap-up --dry-run  # cleanup + debrief + self-improve only; skip commit, push, memory sync

---

org-memory-synthesis

Synthesize the cross-project org-memory layer that the OACP session-init hook auto-loads. Folds new events from $OACP_HOME/org-memory/events/ into the three curated SSOT files (recent.md, decisions.md, rules.md) via a marker-based incremental scan, then runs a 6-check audit (cross-file consistency, supersession asymmetry, stale-status, mechanical chronological order, category drift, migration leftovers). Pairs naturally with /wrap-up's org-memory step.

Runtimes: Claude Code, Codex

/org-memory-synthesis    # full synthesis + audit pass

How These Skills Work Together

The oacp skill teaches a runtime how to use the OACP CLI and protocol — it's the foundation the workflow skills build on. The next three form a complete agent-to-agent code review loop:

1. check-inbox monitors each agent's inbox for incoming messages (review requests, feedback, task assignments).
2. When a review request arrives, the reviewer agent runs review-loop-reviewer to analyze the PR diff and send structured findings back.
3. The author agent picks up the findings via review-loop-author, applies fixes, and sends a review_addressed message — closing the loop.

The remaining skills are complementary maintenance tooling: self-improve audits skill instructions, memory files, and config drift; doctor verifies the OACP environment and workspace are healthy, especially useful at session start; and org-memory-synthesis keeps the cross-project SSOT layer (the one auto-loaded at session init) trustworthy by folding events and auditing for drift.

Prerequisites

• OACP CLI >= 0.3.0 — install via pip install 'oacp-cli>=0.3.0'
• A supported runtime: Claude Code or Codex CLI
• An OACP workspace initialized with oacp init <project>

Installation

Option 1: skills.sh (Claude Code)

Install via skills.sh for Claude Code:

# Install all skills
npx skills add kiloloop/oacp-skills

# Install a single skill
npx skills add kiloloop/oacp-skills -s check-inbox

# Install globally (available across all projects)
npx skills add kiloloop/oacp-skills -g

Note: skills.sh installs the Claude Code variant of each skill. For Codex or other runtimes, use Option 2 with the runtime-specific subdirectory.

Option 2: Copy

Copy skill files directly into your project:

# Claude Code
mkdir -p .claude/skills/<skill-name>
cp skills/<skill-name>/SKILL.md .claude/skills/<skill-name>/SKILL.md

# Codex
mkdir -p .agents/skills/<skill-name>
cp skills/<skill-name>/codex/SKILL.md .agents/skills/<skill-name>/SKILL.md

Option 3: Symlink (for skill developers)

Symlink skills from a local clone so updates pull through automatically:

git clone https://github.com/kiloloop/oacp-skills.git ~/oacp-skills

# Claude Code — symlink a single skill
mkdir -p .claude/skills/check-inbox
ln -s ~/oacp-skills/skills/check-inbox/SKILL.md \
      .claude/skills/check-inbox/SKILL.md

# Symlink all skills at once
for skill in check-inbox doctor oacp review-loop-author review-loop-reviewer self-improve; do
  mkdir -p .claude/skills/$skill
  ln -sf ~/oacp-skills/skills/$skill/SKILL.md \
         .claude/skills/$skill/SKILL.md
done

Directory Structure

skills/
  <skill-name>/
    SKILL.md            # skill instructions (skills.sh compatible)
    skill.yaml          # machine-readable metadata (runtimes, category, OACP version)
    README.md           # human-readable docs
    shared/             # common intent, protocol refs
    claude/
      SKILL.md          # → symlink to ../SKILL.md
    codex/
      SKILL.md          # Codex-specific instructions
template/               # starter template for new skills

Contributing

See CONTRIBUTING.md for the skill structure, schema reference, and submission guidelines.

Related

• kiloloop/oacp — Open Agent Coordination Protocol CLI and spec

License

Apache-2.0