| name | live-proof-gate-ops |
| description | Define and enforce proof-before-done gates for agent workflows, automations, integrations, UI changes, and queued work. |
| memory_tags | ["domain:agent-systems","workflow:live-proof-gate","skill_role:evaluator","repo_boundary:tools","inputs:workflow","outputs:verification-plan","risk:medium"] |
Live Proof Gate Ops
Trigger
Use when:
- a task, automation, integration, or queue item needs a clear "done" verifier
- the user mentions proof gates, end-to-end testing, live verification, receipts, acceptance criteria, or "did it really work?"
- an agent workflow could pass mocks/tests while failing at the real external boundary
- a later-queue item is being promoted into active execution and needs a proof plan
Do not use when:
- the work is pure note-taking with no runtime, external, or user-facing boundary
- the user explicitly wants brainstorming only
- live testing would be unsafe, destructive, unauthorized, or privacy-invasive
Inputs
- Required: task or workflow objective, changed surface, expected user-visible or system-visible outcome
- Optional: repo docs, test commands, credentials/access status, environment constraints, screenshots, logs, prior failures, release/merge target
Workflow
- Identify the real boundary:
- file artifact, CLI command, web UI, mobile app, API/provider, email/social send path, scheduled automation, hardware/device, or human approval loop
- Define the minimum acceptable proof:
- command/test output for local code
- browser or screenshot proof for UI
- authenticated call/readback for external APIs
- sent/received/readback evidence for messaging or email
- generated artifact path plus validation for documents/media
- automation run result plus schedule/config readback for recurring jobs
- Separate proof levels:
unit: small deterministic testintegration: adjacent systems togetherlive: real account/provider/device/user pathreceipt: concise evidence recorded in a queue item, PR, journal, or handoff
- Decide the gate state:
pass: proof ran against the final candidate and matches the objectiveblocked: exact access, credential, account state, device, or approval is missingwaived: owner explicitly accepts missing live proof for this itemnot-applicable: pure docs/metadata/no live boundary; explain why
- Record the result:
- what was tested
- command/path/account boundary used, without secrets
- observed behavior
- artifact or link
- residual risk
- next action if blocked
- When durable or cross-run evidence is useful, optionally emit a
provider-neutral machine-readable receipt:
- Use any installed generic receipt runner; AI Task Receipt is one example,
not a required dependency.
- Capture task/run id, boundary, permissions, command or action type,
timestamps/duration, exit status, bounded/redacted output, verifier and
result, gate state, artifact reference, and residual risk.
- Keep a concise JSON representation suitable for automation and an
optional Markdown rendering for humans.
- Do not require paid multi-model execution. A receipt can wrap the single
authorized run that actually proves the boundary.
Outputs
Return:
- proof boundary
- required proof level
- gate state
- evidence/receipt
- missing access or exact owner question
- residual risk
- recommended next action
- optional machine-readable receipt path/id when durable evidence was requested
For queue items, add or update a proof field. For PRs or commits, include proof in the final report or landing comment.
Boundaries
- Do not fabricate live proof from confidence, mocks, docs, fixtures, or screenshots of unrelated states.
- Do not expose secrets, private data, internal identifiers, or credential values in receipts.
- Do not perform public, financial, account-mutating, or destructive actions unless the user explicitly approved that exact action.
- If live proof is blocked, finish safe prep work and ask for the exact missing access, waiver, or decision.
- A receipt is evidence packaging, not proof by itself; its verifier and
observed postcondition still determine the gate state.
