We Used HACP to Build HACP

What was proven

• Packet and report custody can be traced through explicit lifecycle states.
• Import verification supports evidence integrity checks.
• Verification does not prove completion.
• A human decision remains required after report verification.
• Stop outcomes are first-class and explicit.

How this supports delegated-work review

This dogfood evidence supports HACP's broader delegated-work review loop. PR Review Control Loop is one concrete example: the same packetized readback and owner gate discipline can reduce bad merges, duplicate bot noise, context juggling, and unclear human decisions.

Read the PR Review Control Loop workflow

Current clean-baseline bridge proof point

1. A human-approved packet leaves the app.
2. An owner-controlled local runner passes fail-closed preflight.
3. The runner executes exactly one pre-approved command.
4. The runner produces a report artifact for the trial.
5. The product accepts the report as evidence routed to human decision.

HACP does not run the command. In this trial, an owner-controlled local runner executed exactly one pre-approved command, produced a report, and the product accepted that report as evidence for a human decision. No further execution authority was granted.

The proof point is deliberately conservative: it records the control boundary around owner-controlled tool work, report custody, evidence import, and the next human decision. It does not mean HACP executes work or approves outcomes.

See the human-control boundaries

Boundary reminders

• HACP is a control plane for human-approved agent coordination.
• Human approval is an authority decision, not a checkbox.
• File and CLI transport move artifacts but do not grant authority.
• Report verification confirms integrity and custody checks, not completion.

What this does not claim

• Not autonomous orchestration.
• Not a hosted shell.
• Not model routing.
• Not a RabbitMQ replacement.
• Not proof of task completion.