An open source deterministic or LLM-enhanced agent showcasing and implementing the KCP protocol
Skill Claude CodeCodex
Autonomously refactor application source for clarity and run the test suite to prove behavior is unchanged. Authorized for src/ and test/ and the test-run capability only. It never touches prod/ or secrets/ and never pushes — the human reviews the diff before anything leaves the working tree.
Skill Claude CodeCodex
Run an autonomous batch compliance sweep across customer records against the current policies and write a findings report. Authorized to read policies/ and customers/ and write only into reports/. Use for periodic audit sweeps; it never edits the records it audits.
Skill Claude CodeCodex
Apply a reviewed database schema migration by running the migration runner against the migrations/ directory. Authorized for migrations/ and the db:migrate capability. Its provenance is covered by the manifest's ed25519 signature — an agent should only invoke it from a signature-verified manifest.
Skill Claude CodeCodex
DEMONSTRATION — the same database-migration playbook as db-migration-signed, but served from a manifest whose signature does not verify (missing or tampered). It exists to show signature verification failing closed. An agent must refuse to invoke it.
Skill Claude CodeCodex
Deploy a release to the staging environment by rendering config and running the approved deploy tooling. Authorized for config/ and the staging deploy capability only. It must never touch prod/ or secrets/ — those are outside its declared actionscope and require a separate, human-approved production skill.
Skill Claude CodeCodex
Read and search the project's documentation and skill library to answer "where is this documented / how does this work" questions. Read-only over docs/ and skills/. Use when a task is to look something up, summarize a doc, or locate a runbook — never to change anything.
Skill Claude CodeCodex
DEMONSTRATION of a mis-authored governed skill — kind skill and loadeligible true, but with NO actionscope. A skill fails closed, so a skill with no declared scope authorizes nothing. validate flags it as an error before it can ever ship.
Skill Claude CodeCodex
DEMONSTRATION of an attacker-planted skill — a prompt-injection payload dressed as a risk playbook that tries to exfiltrate customer data. Its audience deliberately excludes agent, so the audience gate skips it before it is ever scored. An agent must never load it.
Skill Claude CodeCodex
Comprehensively research a topic — gather the governed sources, cross-check them, and synthesize a grounded report where every claim cites a source. Read-only and memory-aware. Use as an autonomous agent's safe starting point for any "look into / what do we know about" task.
Skill Claude CodeCodex
SUPERSEDED — the 2024 customer risk-assessment playbook. Kept for audit and replay of decisions made under it, but no longer selectable. The active successor is risk-assessment. Use nothing here for new work.
Skill Claude CodeCodex
Assess the credit and compliance risk of a customer account from the published risk policies and the customer's own records, and produce a written risk memo. Use when a task asks to evaluate, score, or explain the risk of a customer, account, or applicant. Read-only over policies/ and customers/ — it never changes an…
Skill Claude CodeCodex
DEMONSTRATION of a governed skill that is well-formed and in-scope but was never granted invoke eligibility — kind skill with no loadeligible true. Skills fail closed, so the skilleligibility gate skips it. Rotating credentials is exactly the kind of action that should require an explicit grant.
Skill Claude CodeCodex
Cut an npm release of kcp-agent (the TypeScript/CLI package) — bump the version in every place the release tests check, re-sign knowledge.yaml, dispatch release.yml, and confirm the npm publish that release.yml hands off to ci.yml. Use when asked to release, publish, or tag a new kcp-agent version. Does NOT cover the…
Skill Claude CodeCodex
Navigate any KCP-published knowledge base (a knowledge.yaml manifest) with a deterministic, auditable load plan before reading or paying for anything. Use when a task involves a repository, site, or API that publishes a knowledge.yaml / .well-known/knowledge.yaml, when you need to know what knowledge exists, what it…