メニュー
Sign an approval decision with the owner's ed25519 delegate key over canonical JSON, so every call Claudia makes is attributable and verifiable. Use before posting or logging any approval decision.
Codex または Claude でインストール この Prompt をコピーして Codex、Claude、または他のアシスタントに貼り付けると、Skill ページを確認してインストールできます。
SOC 職業分類に基づく
Guides and best practices for working with Neon Serverless Postgres. Covers setup, connection methods, branching, autoscaling, scale-to-zero, read replicas, connection pooling, Neon Auth, and the Neon CLI, MCP server, REST API, TypeScript SDK, and Python SDK. Use when users ask about "Neon setup", "connect to Neon", "Neon project", "DATABASE_URL", "serverless Postgres", "Neon CLI", "neonctl", "Neon MCP", "Neon Auth", "@neondatabase/serverless", "@neondatabase/neon-js", "scale to zero", "Neon autoscaling", "Neon read replica", or "Neon connection pooling".
Make a Better Auth app an OIDC/OAuth issuer (it signs tokens other apps verify) and give an AI agent its own bounded, on-behalf-of credential. Use when configuring @better-auth/oauth-provider, the jwt/JWKS plugin, OAuth client registration, the consent flow, resource-server token verification, or @better-auth/agent-auth. The official better-auth-* skills do NOT cover token issuance or agent identity — this one does, with config verified against Better Auth 1.6.x.
Expose the local connector gateway over a free Cloudflare quick-tunnel (auth OFF) so a claude.ai or ChatGPT custom connector can reach it, and hand the human the public URL. Use this for Part 5 whenever someone wants to run their connector live, get the connector URL, (re)start or refresh the tunnel, or take it down. This is the auth-OFF demo path (AUTH_DISABLED=1); real OAuth + a real sign-in service is the AI Identity course.
Teaches loop engineering — designing the perceive-decide-act-verify cycle that turns a single model call into a working agent. Use when the learner asks about agent loops, the run-it/verify rhythm, retries, stopping conditions, or why their agent does nothing or never stops.
Claudia's autonomous act loop and hard backstop, run on each heartbeat. Reads the pending Paperclip approval queue, classifies each item against the delegated envelope in code, then clears the routine band (signed and logged), surfaces the consequential, and refuses any gate failure. Use on every wake.
Append and query the owner's local append-only governance ledger, the parallel reasoning stream that records every decision Claudia makes (posted, surfaced, or refused) with its signature and rationale. Use after every decision and when the owner reviews the week.
| name | sign-decision |
| description | Sign an approval decision with the owner's ed25519 delegate key over canonical JSON, so every call Claudia makes is attributable and verifiable. Use before posting or logging any approval decision. |
Every decision you make carries a signature, so it can be told apart from a call the owner made themselves and can be verified later. This skill signs the decision payload with your ed25519 private key. Public key only is ever shared; the private key never leaves the machine, is never printed, logged, or committed.
Before you write a governance_ledger row or post a paperclipApprovalDecision. The signature goes in the ledger row's attestation field (and optionally the approval decisionNote). A decision without a verifying signature fails gate 2 and is not posted.
This skill ships executable. All four scripts are dependency-free Node ESM (crypto, no npm). Run them with node.
scripts/canonical.mjs: RFC-8785 (JCS) canonical JSON over the decision payload. Recursive key sort by UTF-16 code unit, no insignificant whitespace, so signer and verifier byte-match. node canonical.mjs runs a self-test proving key order is stable. Both sign and verify import it, so they canonicalize identically by construction.scripts/keygen.mjs: generate the ed25519 keypair, once. Private key to ~/.openclaw/keys/identic-ai.pem (chmod 600), public to identic-ai.pub. The private key is NEVER printed; it prints only the public SHA-256 fingerprint and runs a sign/verify self-test. It refuses to overwrite an existing private key (--force only for deliberate rotation).scripts/sign.mjs: read a decision payload as JSON on stdin, canonicalize it, sign the canonical bytes with the private key, print ONLY the base64 signature to stdout. The public fingerprint goes to stderr (public_fingerprint=...) for the ledger to record.scripts/verify.mjs: verify a signature against the PUBLIC key only (--sig <base64>, payload via --payload <file> or stdin). Exit 0 = valid, exit 2 = invalid, so callers gate on it.A plain object: at minimum the approval id, the action (approve / reject / request_revision / surface_to_owner), the principal (owner_identic_ai), who you act for, and a timestamp. Keep the field set stable; signer and verifier must agree on it exactly. govern-queue's decide.mjs builds this payload and calls sign.mjs + verify.mjs for you on every cleared decision; the signature lands in the ledger row's attestation_b64.
.gitignore already ignores *.pem and keys/; do not undo that.crypto, is the whole design.