Install
openclaw skills install @ayobamih/agentproofUse AgentProof for approval-bound repository patches with explicit authority, exactly-once execution, independent verification, deterministic recovery, and signed receipts.
openclaw skills install @ayobamih/agentproofUse this skill only when the user wants a consequential repository change to pass through AgentProof's authorization and evidence protocol, or when they want to inspect, reconcile, compensate, or verify an existing AgentProof transaction or receipt.
AgentProof is not a generic code-editing shortcut. It protects one current action contract: agentproof.repository_patch.v1.
agentproof-dev-authority to authorize real, consequential, or production work. Development authority is only for disposable demonstrations and tests.The AgentProof CLI is currently a developer-preview package. This ClawHub skill intentionally does not auto-install the product code. Install AgentProof from the project's documented immutable release or source-consumption path, then ensure agentproof is on PATH.
Confirm the available contract before doing anything consequential:
agentproof --help
Build or receive a valid repository-patch request document that binds the exact repository root, operations, intent, policy, correlation ID, and state directory. Then run:
agentproof prepare repository-patch --input request.json > prepared.json
Review the prepared document before continuing. If it does not match the user's intended bytes and policy, stop.
Use an explicit expiry and nonce:
agentproof approval-request \
--input prepared.json \
--expires-at <ISO_TIMESTAMP> \
--nonce <NONCE> \
> approval-request.json
At this point, stop unless a legitimate authority mechanism supplies the signed approval needed for execution. Do not self-approve.
Only after a valid execution-request document exists and the required receipt-signing key path is explicitly available:
agentproof execute \
--input execution-request.json \
--receipt-key <private-key.pem> \
> receipt.json
Do not invent the execution request or signer material.
Use the trusted signer fingerprint supplied by the operator or authority configuration:
agentproof verify-receipt \
--input receipt.json \
--trust-fingerprint <sha256:...>
For production-authority evidence, require the expected environment when appropriate:
agentproof verify-receipt \
--input receipt.json \
--trust-fingerprint <sha256:...> \
--required-authority-environment production
A successful mutation is not enough. Report separately whether execution completed, the receipt is cryptographically valid, and the signer is trusted.
Inspect an existing transaction before attempting recovery:
agentproof status --input status-query.json
If execution state is ambiguous, use reconciliation rather than starting over:
agentproof reconcile \
--input reconciliation-query.json \
--receipt-key <private-key.pem>
Use compensation only when explicitly required and when all trust inputs are available:
agentproof compensate \
--input status-query.json \
--receipt-key <private-key.pem> \
--trust-fingerprint <sha256:...> \
--authority-environment <development|production>
Always distinguish these states instead of collapsing them into "done":
If any required authority, key, fingerprint, repository state, or receipt evidence is missing, say that it is missing and stop at the appropriate boundary.