Portable agent skills distilled from a publicly mirrored CC coding-agent codebase, then rewritten so they can be installed in Claude Code, Codex, and OpenClaw without depending on private runtime internals.

This repo is not a source dump. It is a cleaned skill pack:

• prompts extracted into reusable templates
• host-agnostic helper scripts
• portable SKILL.md bundles
• release and smoke-test docs for public distribution

If you are building with coding agents, these are the parts that usually separate a fun demo from a stable toolchain:

• memory that stays useful instead of rotting
• compression that preserves user corrections
• verification that does not trust "done"
• multi-agent routing that does not pollute the main context
• proactive jobs with explicit limits

This repo packages those patterns into six installable skills.

Most agent repos share the same hard problems:

• how to keep memory useful without storing stale code facts
• how to compress long conversations without losing user corrections
• how to verify claimed completion instead of trusting it
• how to coordinate multiple agents without turning the main context into garbage
• how to add lightweight proactive behavior without building a dangerous daemon

The six skills in this repo package those patterns into download-ready bundles.

• builders who want better agent behavior without forking a full agent product
• teams using Claude Code, Codex, or OpenClaw and wanting a shared skill layer
• people studying CC-style harness design but needing something publishable and portable
• anyone who wants prompts plus scripts, not prompts alone

Each bundle lives in skills/ and includes:

• SKILL.md
• references/prompt-template.md
• references/source-notes.md
• scripts/...
• README.md

If you only try three things, try these first:

1. install dream-memory
2. install verification-gate
3. install structured-context-compressor

That gives you one memory workflow, one quality gate, and one continuation artifact. It is the fastest way to feel whether this repo is useful.

Run the basic bundle checks first:

bash ./skills/check_all.sh

Install one skill into a host runtime:

mkdir -p ~/.claude/skills
cp -R ./skills/dream-memory ~/.claude/skills/

Then invoke it in Claude Code:

Use /dream-memory to consolidate the memory directory before we continue.

mkdir -p ~/.codex/skills
cp -R ./skills/dream-memory ~/.codex/skills/

Then invoke it in Codex:

Use /dream-memory and consolidate recent memory files into one stable index.

mkdir -p ~/.openclaw/workspace/skills
cp -R ./skills/dream-memory ~/.openclaw/workspace/skills/

Then check it in OpenClaw:

openclaw skills info dream-memory

Publish to ClawHub after checks:

bash ./skills/publish_all.sh 0.1.0

• Publisher guide: skills/README.md
• Release order and copy: skills/RELEASE_PLAN.md
• Smoke-test report: skills/TEST_REPORT.md

Recommended public GitHub repo name: cc-harness-skills

Good alternatives:

• cc-agent-skills
• cc-harness-kit
• cc-portable-skills

The public naming uses CC on purpose: it signals the origin of the ideas without hard-coding a full vendor product name into the repo identity.

Latest smoke-test status:

• Claude Code: six skills load successfully
• OpenClaw: six skills resolve as Ready
• Codex: bundle structure is verified; runtime smoke test is pending local auth

Full report: skills/TEST_REPORT.md

skills/
  dream-memory/
  memory-extractor/
  verification-gate/
  swarm-coordinator/
  structured-context-compressor/
  kairos-lite/

Internal working notes and non-publishable local experiments still live under .codex/skills/, but the public release surface for this project is the root skills/ directory.