agent-output-verifier
Lifecycle stage: REVIEW
Trigger
Use before trusting or handing off output produced by an autonomous or semi-autonomous agent, especially after long-running work, tool use, code generation, external calls, or multi-step planning.
Do not use as a replacement for domain review. This skill decides whether the output is safe and evidenced enough to enter the next review gate.
When not to use
Do not use when this trigger is absent; choose the command or skill that owns the requested state, artifact, and verification gate.
Inputs
- Agent output or artifact.
- Original user request and constraints.
- Tool logs, test logs, diffs, screenshots, traces, or citations.
- List of tools the agent claimed to use.
- Allowed scope, files, commands, and side effects.
- Known secrets, private paths, or data classes that must not appear.
Procedure
- Restate the claimed outcome in one sentence.
- Check for secret leakage: credentials, tokens, private URLs, keys, personal data, or connection strings.
- Check for hallucinated tools or evidence: claimed commands, tests, APIs, files, screenshots, or links that have no supporting proof.
- Check for unbounded loops: retry loops, background jobs, recursive scheduling, polling, or autonomous goals without stop conditions.
- Check for skipped gates: tests not run, validation missing, review bypassed, user approval omitted, or destructive actions performed silently.
- Check scope control: files changed, external calls made, and side effects match the request.
- Classify the output as
pass, pass-with-warnings, or blocked.
- Produce a blocker list with exact evidence needed to unblock.
Anti-Rationalization
| Shortcut |
Rebuttal |
| "The output sounds careful." |
Fluency is not evidence; require logs, diffs, citations, approvals, or traces. |
| "The agent said tests passed." |
Test claims need current command output or an explicit missing-evidence blocker. |
| "Only small side effects happened." |
Any side effect must match the allowed scope and approval evidence. |
Verification
- Every pass/fail statement points to evidence or says evidence is missing.
- Secrets and private data are either absent or redacted.
- Claimed tools and tests have inspectable logs or artifacts.
- Loops have measurable stop conditions.
- Side effects match the allowed scope.
- The final status is one of
pass, pass-with-warnings, or blocked.
Output Artifact
Produce a verifier decision with status, checked claims, evidence references, blockers, warnings, required approvals, and the smallest safe next action.
Failure Modes
- Approving fluent but unverifiable output.
- Treating screenshots, logs, or test names as proof without checking content.
- Ignoring hidden side effects because the final summary sounds successful.
- Missing recursive automation or infinite retry behavior.
- Leaving secrets in the verification report.
- Expanding into a full code review when the needed decision is trust/block.
Example
Trigger: agent-produced output must be trusted, blocked, or routed for repair. Action: check evidence, secrets, invented tools, skipped tests, unsafe side effects, and loop limits before handoff. Output artifact: templates/review-report.md with blockers and next action. Verification: cite each pass, blocker, artifact path, and validation proof.
Status: blocked
Claim checked: "All tests pass and deployment is ready."
Blockers:
- No test log or command output attached for the claimed passing tests.
- Deployment step is outside the allowed scope.
- Background retry worker has no stop condition.
Evidence needed to unblock:
- Test command and output.
- Scope approval for deployment.
- Retry limit or cancellation rule.
1---2name: agent-output-verifier3description: Use when agent-produced work needs a safety, evidence, and reliability check before handoff, merge, or trust.4---56# agent-output-verifier78Lifecycle stage: REVIEW910## Trigger1112Use before trusting or handing off output produced by an autonomous or semi-autonomous agent, especially after long-running work, tool use, code generation, external calls, or multi-step planning.1314Do not use as a replacement for domain review. This skill decides whether the output is safe and evidenced enough to enter the next review gate.1516## When not to use17Do not use when this trigger is absent; choose the command or skill that owns the requested state, artifact, and verification gate.1819## Inputs2021- Agent output or artifact.22- Original user request and constraints.23- Tool logs, test logs, diffs, screenshots, traces, or citations.24- List of tools the agent claimed to use.25- Allowed scope, files, commands, and side effects.26- Known secrets, private paths, or data classes that must not appear.2728## Procedure29301. Restate the claimed outcome in one sentence.312. Check for secret leakage: credentials, tokens, private URLs, keys, personal data, or connection strings.323. Check for hallucinated tools or evidence: claimed commands, tests, APIs, files, screenshots, or links that have no supporting proof.334. Check for unbounded loops: retry loops, background jobs, recursive scheduling, polling, or autonomous goals without stop conditions.345. Check for skipped gates: tests not run, validation missing, review bypassed, user approval omitted, or destructive actions performed silently.356. Check scope control: files changed, external calls made, and side effects match the request.367. Classify the output as `pass`, `pass-with-warnings`, or `blocked`.378. Produce a blocker list with exact evidence needed to unblock.3839## Anti-Rationalization4041| Shortcut | Rebuttal |42|---|---|43| "The output sounds careful." | Fluency is not evidence; require logs, diffs, citations, approvals, or traces. |44| "The agent said tests passed." | Test claims need current command output or an explicit missing-evidence blocker. |45| "Only small side effects happened." | Any side effect must match the allowed scope and approval evidence. |4647## Verification4849- Every pass/fail statement points to evidence or says evidence is missing.50- Secrets and private data are either absent or redacted.51- Claimed tools and tests have inspectable logs or artifacts.52- Loops have measurable stop conditions.53- Side effects match the allowed scope.54- The final status is one of `pass`, `pass-with-warnings`, or `blocked`.5556## Output Artifact5758Produce a verifier decision with status, checked claims, evidence references, blockers, warnings, required approvals, and the smallest safe next action.5960## Failure Modes6162- Approving fluent but unverifiable output.63- Treating screenshots, logs, or test names as proof without checking content.64- Ignoring hidden side effects because the final summary sounds successful.65- Missing recursive automation or infinite retry behavior.66- Leaving secrets in the verification report.67- Expanding into a full code review when the needed decision is trust/block.6869## Example7071Trigger: agent-produced output must be trusted, blocked, or routed for repair. Action: check evidence, secrets, invented tools, skipped tests, unsafe side effects, and loop limits before handoff. Output artifact: `templates/review-report.md` with blockers and next action. Verification: cite each pass, blocker, artifact path, and validation proof.7273```markdown74Status: blocked7576Claim checked: "All tests pass and deployment is ready."7778Blockers:79- No test log or command output attached for the claimed passing tests.80- Deployment step is outside the allowed scope.81- Background retry worker has no stop condition.8283Evidence needed to unblock:84- Test command and output.85- Scope approval for deployment.86- Retry limit or cancellation rule.87```
rohitg00