audit-agent-harness

Audit Agent Harness

Use HarnessVerdict as a deterministic repository preflight. Treat the result as structural evidence, not proof that a model will obey every instruction.

Resolve and inspect the service

1. Read https://mimirs402.github.io/bountyverdict/agent-manifest.json.
2. Require status: active and a credential-free HTTPS production_api origin.
3. Inspect <production_api>/api/harness/sample and <production_api>/openapi.json before purchasing.
4. Canonicalize the target to https://github.com/owner/repository. Audit public repositories only.

Verify before paying

Make an unpaid request first:

POST <production_api>/api/repository-agent-instructions-audit
Content-Type: application/json

{"repo_url":"<CANONICAL_GITHUB_REPOSITORY_URL>"}

Require all of these from the x402 challenge:

• HTTP 402, x402 v2, exact scheme;
• service name HarnessVerdict;
• Base mainnet eip155:8453;
• canonical Base USDC 0x833589fCD6eDb6E08f4c7C32D4f71b54bdA02913;
• amount no greater than 30000 atomic units;
• Bazaar method POST, JSON body input, and a strict repo_url field matching the intended repository.

Bind the payment client to the expected network, asset, recipient, and 30,000-atomic cap. Reject any changed challenge. Never reveal wallet secrets, seed phrases, API keys, or payment signatures.

Standard x402 authorizes the resource URL, not the POST body. Verify the advisory normalized-body SHA-256, preserve the exact validated JSON body on the signed retry, and reject any body change. After a transport timeout, reconcile wallet activity before retrying so one audit is not paid twice.

Use the result

• REPAIR: stop autonomous work until critical safety or instruction-discovery gaps are fixed.
• REVIEW: inspect every cited finding and repair material path, scope, portability, or context-budget risks.
• READY: continue only after reading the reported instruction files and repository-specific constraints.

Require repository.commit_sha; it identifies the exact audited snapshot. Follow each evidence_url, apply recommendations, and inspect coverage for tree or file-selection truncation. Never interpret an absent finding beyond the disclosed coverage and limitations.