← Back to Daily Builds

Daily build note · May 25, 2026

Agent Receipts Desk

A productized service and lightweight tool that creates reviewable receipts for important AI-agent work: what the agent saw, which tools it used, what it changed, what was approved, what...

AI Tools Runnable local MVP built Public demo coming soon

What shipped

Built a runnable local MVP for Agent Receipts Desk: a Python CLI and deterministic report generator that creates reviewable receipts for AI-agent work.

Core files:

  • agent_receipts_desk/cli.py - command routing for init, sample, new, render, and index.
  • agent_receipts_desk/models.py - receipt schema defaults, config defaults, JSON helpers.
  • agent_receipts_desk/git_evidence.py - read-only git collector for branch, commit, changed files, and diff stat.
  • agent_receipts_desk/risk.py - deterministic risk classifier.
  • agent_receipts_desk/render.py - Markdown, JSON, client-safe, and PR comment output generation.
  • agent_receipts_desk/redact.py - local redaction for paths, internal links, and secret-like strings.
  • agent_receipts_desk/samples.py - three realistic sample runs.
  • scripts/smoke.py - end-to-end verification for sample generation and redaction.

Architecture

  • Used Python 3.11+ and the standard library only, so the tool runs locally without dependency setup.
  • Kept the receipt schema provider-neutral and deterministic instead of relying on LLM summarization.
  • Stored all outputs as plain Markdown and JSON under agent-receipts/.
  • Used read-only git commands and clear degradation outside a git repo.
  • Made client-safe export a separate Markdown artifact so internal and external handoffs can coexist.

Trimmed scope

Deliberately excluded:

  • hosted SaaS, auth, billing, and cloud storage
  • live GitHub PR posting
  • live MCP discovery
  • direct Codex session parsing
  • PDF export
  • browser dashboard
  • external LLM calls or any Claude, Anthropic, or OpenRouter integration

The MVP focuses on the evidence packet: intake, risk labels, receipt generation, redaction, and index.

Limitations

  • Guided intake is intentionally simple and line-oriented.
  • Verification evidence is manually summarized; it does not yet import raw log files.
  • Redaction is pattern-based and should be reviewed before sending client-facing receipts.
  • Risk classification is deterministic and conservative, but it can miss nuanced security or business risks.
  • The JSON schema is stable enough for the MVP but not formally versioned with a JSON Schema file yet.

Verification

Smoke test command:

``bash python scripts/smoke.py ``

The smoke test generates three sample receipt packs and verifies:

  • receipt.json
  • receipt.md
  • client-safe.md
  • pr-comment.md
  • index generation
  • expected risk flags on the auth sample
  • client-safe redaction of private paths and internal links

Suggested next steps

  1. Add --from-intake support to render a user-authored intake JSON file.
  2. Add direct attachment import for test, lint, build, secret scan, browser QA, and deploy logs.
  3. Publish a JSON Schema for receipt.json.
  4. Add optional GitHub PR comment posting.
  5. Add a small browser UI for browsing agent-receipts/index.json.
  6. Add monthly rollups over receipt indexes for consulting deliverables.