agent-work-receipt-designer

name	agent-work-receipt-designer
description	Design a normalized, machine-readable "work receipt" that answers what changed, why, by whom, and with what validation, for any AI-coding-agent task regardless of backend (Claude Code, Codex, Cursor, Aider, CI, or homegrown). Use when defining a receipt schema, normalizing heterogeneous agent logs into one evidence shape, gating PR merges on artifact-backed proof, or designing an OSS maintainer's fast-reject/fast-merge workflow. NOT for validating a DAG node's output against its own schema (use output-contract-enforcer), continuous runtime monitoring of a live coordination daemon (use runtime-verification-for-agents), or general Port Daddy claims/notes/salvage mechanics (use port-daddy-agent-skill).
license	Apache-2.0
allowed-tools	Read,Write,Edit,Bash,Grep,Glob
metadata	{"category":"Agent & Orchestration","tags":["work-receipt","provenance","validation-evidence","audit-trail","agent-coordination"],"provenance":{"kind":"first-party","owners":["port-daddy"]},"pairs-with":[{"skill":"output-contract-enforcer","reason":"Validates a single node's output shape; receipts validate the whole task's evidence, spend, and risk."},{"skill":"runtime-verification-for-agents","reason":"Runtime monitors catch live coordination violations; receipts are the durable record after the fact."},{"skill":"port-daddy-agent-skill","reason":"Supplies the claims/notes/salvage substrate a receipt's rollback pointer and provenance can anchor to."},{"skill":"agentic-coding-product-research","reason":"Names durable, review-reducing receipts as Port Daddy's wedge against tool-specific agent logs."}],"io-contract":{"kind":"deliverable","consumes":["[Truncated]","[Truncated]"],"produces":["[Truncated]","[Truncated]"]}}

Agent Work Receipt Designer

Design the receipt schema and discipline that lets any AI coding agent's work be reviewed, trusted, and reversed without re-reading the chat.

Use This For

• Defining a cross-tool receipt schema so Claude Code, Codex, Cursor, Aider, and CI runs all land in one evidence shape.
• Turning a raw agent transcript into a typed record: identity, intent, context used, actions, validation, spend, risks, rollback, provenance.
• Gating PR merges so a maintainer can reject a low-proof agent PR in seconds and merge a high-proof one with confidence.
• Auditing whether "tests pass" claims are backed by a captured exit code or artifact, or are just the agent's word.
• Designing durable, replayable handoffs that survive the originating session or worktree being deleted.

Do Not Use This For

• Validating one DAG node's output against its declared JSON schema mid-pipeline (output-contract-enforcer).
• Continuously auditing a live coordination daemon against formal invariants (runtime-verification-for-agents).
• General Port Daddy session/claims/salvage mechanics unrelated to producing a receipt (port-daddy-agent-skill).

Process

flowchart TD
  A[Collect raw backend log] --> B[Extract identity + intent]
  B --> C[Extract context used]
  C --> D[Extract actions: commands, tool calls, diff]
  D --> E[Extract validation: exit codes + artifacts only]
  E --> F[Rank risks worst-first]
  F --> G[Attach rollback pointer]
  G --> H[Hash/sign for provenance]
  H --> I[Lint receipt, publish durably]

1. Collect the raw log for the backend that ran (transcript JSONL, exec events, chat history, CI step logs). See references/backend-normalization.md for the per-backend extraction map.
2. Extract identity (agent/model/backend/sessionId/operator) and intent (goal/scope/stopCondition) — a vague stop condition is the earliest sign the resulting receipt will be weak.
3. Extract contextUsed: files actually read, governing rules consulted (CLAUDE.md/AGENTS.md/ADRs), attachments.
4. Extract actions: every command with its real exit code, tool-call tallies, and a diffSummary a reviewer can triage from before opening the diff.
5. Extract validation from tool-result payloads or captured logs only — never from the agent's own narration. A passed: true test with no exitCode and no artifactPath is a self-report, not proof; mark it instead.