Reusable agent skills · v0.4.3

A small library, deeply composed.

Each skill is a single, well-scoped capability — the workflow layer above the protocol. Drop-in for Claude Code and Codex CLI. Apache 2.0.

Install + wire ↓ Browse source ↗

skills

Foundation

Core protocol awareness. Loaded once per session — teaches the runtime how to use the CLI and message protocol. Every other skill builds on it.

1 skill

oacp

foundation

The protocol primer — teaches a runtime how to use the CLI and message protocol. Foundation for every other skill.

$ /oacp

claudecodex oacp ≥ 0.3.0 source ↗

02 ·

Coordination

Move work between agents safely — general-purpose inbox handling and the review-loop pair that closes the agent-to-agent code review cycle.

3 skills

check-inbox

General-purpose inbox processor. Scans for pending OACP messages of any type, processes each, then exits. Pair with a loop or oacp watch for continuous monitoring.

$ /check-inbox

claudecodex oacp ≥ 0.1.0 source ↗

review-loop-reviewer

Reads a PR diff, produces structured findings, renders a verdict, and sends feedback to the author’s inbox. Supports multi-round reviews.

$ /review-loop-reviewer 42 --author claude

claudecodex oacp ≥ 0.1.0 source ↗

review-loop-author

Picks up review findings, applies fixes, pushes updates, and sends a review_addressed message back. Closes the loop.

$ /review-loop-author 42 --reviewer codex

claudecodex oacp ≥ 0.1.0 source ↗

03 ·

Workspace

Keep the workspace healthy — environment and inbox diagnostics, plus knowledge-layer audits across skills, memory files, and config.

2 skills

doctor

Wraps oacp doctor for self-diagnostics. Checks environment, workspace, inbox health, schemas, agent status. Auto-fixes safe issues.

$ /doctor

claudecodex oacp ≥ 0.1.2 source ↗

self-improve

Audits the agent’s knowledge layer — skills, memory files, config — for staleness, contradictions, gaps, and bloat. Proposes surgical fixes.

$ /self-improve

claudecodex oacp ≥ 0.1.0 source ↗

04 ·

Memory

Persist what was learned. Capture session summaries to the cortex inbox, and consolidate across projects into a single source of truth.

2 skills

debrief

Captures a structured session summary at end of session and sends it to the cortex inbox. Wrap-up calls this automatically; you can also run it standalone.

$ /debrief

claudecodex oacp ≥ 0.1.0 source ↗

sync

Consolidates session debriefs across projects into a single source of truth (SSOT). Run once daily — typically each morning.

$ /sync

claudecodex oacp ≥ 0.1.0 source ↗

05 ·

Lifecycle

Bookend a session well. Lifecycle skills orchestrate the end of every run — cleanup, debrief, self-improve, commit, push, sync — in strict order.

1 skill

wrap-up

lifecycle

End-of-session orchestrator. Cleans up merged branches and stale worktrees, runs debrief and self-improve, commits, pushes, and syncs OACP memory — in strict order.

composes

↳debrief

↳self-improve

↳commit + push

↳memory sync

$ /wrap-up

claudecodex oacp ≥ 0.3.0 source ↗

Install + wire

Three steps. Three minutes.

The CLI does the wiring. Skills land in the right place for each runtime — no manual symlinks, no per-agent config drift.

Read setup docs →

install.sh
bash

    
        01
        # 1. install the cli
      

        02
        $ pip install oacp-cli
      

        03
        
      

        04
        # 2. set up your runtime — wires the foundation skill automatically
      

        05
        $ oacp setup claude --project my-app
      

        06
        $ oacp setup codex  --project my-app
      

        07
        
      

        08
        # 3. invoke skills from inside the agent
      

        09
          claude  →  /oacp           # foundation
      

        10
          claude  →  /check-inbox    # coordination
      

        11
          claude  →  /wrap-up        # lifecycle

Up next

Read the shared-memory spec →