agent-receipt

Offline claim-to-evidence receipts for agent handoffs. A receipt records what was observed; it is not proof that an agent's claim is true.

Install

uv tool install agent-receipt

Or with pipx:

pipx install agent-receipt

For contributors working from this repository:

uv tool install ./packages/agent-receipt

Safety model

• Evidence paths are POSIX paths relative to a caller-supplied, trusted workspace root. Absolute paths, . / .., NULs, and symlink escapes are rejected.
• Rechecks require a verifier-supplied --recheck-root; the receipt never chooses a filesystem root or command working directory.
• Text evidence is literal substring matching only; v1 has no regex mode.
• File reads, receipt strings, claim/evidence counts, command arguments, captured command output, and command timeouts are bounded. A command that exceeds the output limit is terminated and recorded as failed.
• The optional context is either null or exactly binds packet_digest, input_commit, and output_manifest_digest. There is no arbitrary metadata.

content_digest detects inconsistent or non-rehashed changes to the unsigned receipt body. Because it is unkeyed, an active attacker can replace both the body and digest. An Ed25519 signature authenticates signer attribution against a verifier-provided public key; it still does not establish claim truth.

Build and verify

agent-receipt build \
  --workspace-root . \
  --agent worker --task "write report" --claim report="report exists" \
  --file-exists report=./report.md \
  --text-contains 'report=./report.md::PASS' \
  --out receipt.json

agent-receipt verify receipt.json --recheck --recheck-root .

Rechecks compare a current file hash to the hash observed while building the receipt. They therefore catch later file changes.

Command evidence

Commands run without a shell, with a constrained environment and bounded output. They always run from the trusted build root; no cwd is stored in the receipt. They can still execute repository code, so they are an explicit execution decision.

Command rechecks require both a verifier-selected root and an exact tuple allowlist:

agent-receipt verify receipt.json --recheck-commands --recheck-root . \
  --allow-command '/absolute/path/to/python -m pytest -q'

Inspect the executable, every argument, and the repository state before adding an allowlist entry. The executable path must be absolute, so PATH cannot silently select a different program. An unlisted receipt command fails closed. The verifier, not the worker receipt, defines success as exit code 0, uses a fixed 20-second timeout, and ignores worker-selected output expectations. At most five distinct commands may occur in one receipt.

Verification reports assurance and coverage. reported means no evidence was rerun, partially_rechecked means only some evidence was independently rerun, and fully_rechecked means every stored evidence item was rerun. Blocked command evidence is counted separately and never presented as rechecked. By default, verification exits successfully only for fully_rechecked. A lower threshold requires an explicit decision such as --minimum-assurance reported; that mode validates structure and internal consistency, not claim truth.

Canonical output manifests

The controller can inventory a completed output directory without executing any of its contents:

agent-receipt manifest create --workspace-root ./worker-output --json
agent-receipt manifest verify --workspace-root ./worker-output \
  --expected-digest '<independently trusted digest>' --json

Creation writes OUTPUT_MANIFEST.json under the selected root by default. The manifest uses the strict agent-output-manifest/v1 schema: relative POSIX paths are unique and sorted, and every regular file has a SHA-256 and byte count. Its digest is SHA-256 over canonical UTF-8 JSON with sorted object keys and compact separators, excluding the storage newline. Verification rebuilds the inventory and requires the exact file set, sizes, and hashes—not just the files named in the manifest. JSON with duplicate object keys is rejected.

Traversal is descriptor-relative and bounded by file count, entry count, depth, per-file size, total bytes, path length, and manifest size. Symbolic links, hard links, special files, path traversal, duplicate paths, and detectable path/content changes during scanning fail closed. receipt.json and OUTPUT_MANIFEST.json at the root are control files and are excluded by default; a custom CLI manifest path inside the root is also excluded exactly. No broad ignore patterns are applied. Creation and verification each require two complete, identical scans. This catches late changes missed by an earlier per-entry check, but it is not a transactional filesystem snapshot and cannot prevent a write after the final system call. Stop the worker before scanning and keep the output root quiescent until acceptance.

To bind a receipt to a handoff, supply all three strict context values at build and verify time: --packet-digest, --input-commit, and --output-manifest-digest. Verification compares the receipt context exactly to the verifier-provided values. Use the digest printed by manifest create, transfer it through a trusted channel, and independently run manifest verify before accepting the receipt. The digest is unkeyed; use the receipt's Ed25519 signature when signer attribution is needed.

Signatures

agent-receipt build ... --sign-private-key worker.pem --key-id worker-2026
agent-receipt verify receipt.json --trusted-key worker-2026=worker.pub \
  --minimum-assurance reported

Keep private keys out of receipts, repositories, command lines, and logs. Key IDs are labels; the verifier's supplied public key is the trust anchor. The explicit reported threshold above checks attribution and internal consistency only. Add the appropriate independent rechecks before accepting any claim.