Verify a Change
Turn a completion claim into evidence tied to the actual change.
When to use
Use after implementation or before a review/merge decision. For SPEC.md drift,
use check. For a live browser flow, use webapp-testing or an available host's
flow-verification skill; consume its evidence without repeating the same run.
Do not run a full build for a wording-only edit or introduce a test framework
solely to satisfy this checklist.
Workflow
- Read the request, repository instructions and acceptance criteria. Identify
the real scope: working tree, staged changes, commit range or PR base/head.
Include untracked files when the task includes them. Do not assume HEAD~1
is the review base; it misses uncommitted work and multi-commit changes.
- Discover commands from manifests, CI and project docs. Preserve the project's
package manager, versions, framework, coverage policy and test boundaries.
Missing commands are unavailable checks, not a reason to invent npm scripts.
- Select the smallest checks that cover the changed behavior: build/types/lint
as applicable, focused regression tests, then broader checks required by the
repo or justified by integration risk. A zero exit code with zero collected
tests is not test coverage. Inspect skips and unexpected empty discovery.
Tests assert what the code's users can observe, so a test fails when the
behavior breaks and keeps passing when only the internals change: drive
the code through its users' interface with a concrete input, and derive
the expected value from the requirement, not from a run of the code. A
test that still passes when the behavior breaks proves nothing; a test
that fails when only internals change is pinned to the implementation.
- Run each command with its working directory and capture its real exit code.
Keep full output locally when needed; summarize without masking failures
through head/tail pipelines. In PowerShell, capture
$LASTEXITCODE
immediately after a native command, before another native command runs.
- Use the repo's configured secret/security checks for relevant changes. Never
label a grep for key-shaped strings a security audit or print matched secrets.
State scan scope and missing prerequisites. Inspect Markdown/config changes
too when they can carry credentials or executable agent instructions.
- Review the actual diff for unintended changes, boundary cases, generated
files and changes made after testing. Fix within the authorized scope, then
rerun checks affected by the fix. Do not endlessly repeat successful checks.
- Report evidence and remaining limitations. Readiness is not authorization to
commit, push, merge or deploy. Use existing authorization for those actions.
Evidence contract
| Check |
Required evidence |
| Build / types / lint |
Command, working directory, exit code, relevant findings |
| Tests |
Command, collected/passed/failed/skipped counts and coverage if measured |
| Behavior |
Requirement exercised and observed result; mocks versus real integration |
| Security |
Scanner and scope, redacted finding count, unavailable dependencies |
| Diff |
Base/head or working-tree scope, changed files, untested later changes |
Use PASS, FAIL, SKIPPED (reason), UNAVAILABLE (prerequisite) or
NOT APPLICABLE (reason) per check. Required unavailable/skipped checks leave
readiness unverified. Never turn them into PASS. A test summary printed before
a hang or nonzero process exit is incomplete evidence.
Prove it works
Verify against the real artifact, not a proxy, a self-report, or "it
compiles." Read the actual value, not a cached or derived representation;
check process liveness directly, not through derived state. When
verification fails, suspect the observation method before suspecting the
system. The strongest proof is a deterministic script that re-runs the same
comparison: write it, run it, and keep its output as an artifact a reviewer
can re-run instead of trusting your word.
Prerequisites and companions
This is instructions-only: it supplies no /verify command, hooks or scanner.
Use the loaded project's tools. Route through skill-router to find
python-testing, react-testing, webapp-testing, requesting-code-review,
security-best-practices or session-handoff when relevant.
Provenance
Adapted from ECC verification-loop at the reviewed revision.
See MIT notice. Local changes replace fixed commands/thresholds,
timed repetition and superficial security checks with project-scoped evidence.
1---2name: verification-loop3description: >-4 Verify a completed change with the project's actual build, type, lint, test and5 diff checks before reporting readiness. Use for verify this change, ready for6 review, prove this works, or completion checks across any supported harness.7license: MIT8metadata:9 source: https://skillmd.com/skills/hongyime/verification-loop10 author: Local setup, adapted from ECC11 version: "1.0.0"12 platform: "Codex, Antigravity (agy), OpenCode, Muse on Windows, macOS, Linux (Debian)"13---1415# Verify a Change1617Turn a completion claim into evidence tied to the actual change.1819## When to use2021Use after implementation or before a review/merge decision. For SPEC.md drift,22use `check`. For a live browser flow, use `webapp-testing` or an available host's23flow-verification skill; consume its evidence without repeating the same run.24Do not run a full build for a wording-only edit or introduce a test framework25solely to satisfy this checklist.2627## Workflow28291. Read the request, repository instructions and acceptance criteria. Identify30 the real scope: working tree, staged changes, commit range or PR base/head.31 Include untracked files when the task includes them. Do not assume HEAD~132 is the review base; it misses uncommitted work and multi-commit changes.332. Discover commands from manifests, CI and project docs. Preserve the project's34 package manager, versions, framework, coverage policy and test boundaries.35 Missing commands are unavailable checks, not a reason to invent npm scripts.363. Select the smallest checks that cover the changed behavior: build/types/lint37 as applicable, focused regression tests, then broader checks required by the38 repo or justified by integration risk. A zero exit code with zero collected39 tests is not test coverage. Inspect skips and unexpected empty discovery.40 Tests assert what the code's users can observe, so a test fails when the41 behavior breaks and keeps passing when only the internals change: drive42 the code through its users' interface with a concrete input, and derive43 the expected value from the requirement, not from a run of the code. A44 test that still passes when the behavior breaks proves nothing; a test45 that fails when only internals change is pinned to the implementation.464. Run each command with its working directory and capture its real exit code.47 Keep full output locally when needed; summarize without masking failures48 through head/tail pipelines. In PowerShell, capture `$LASTEXITCODE`49 immediately after a native command, before another native command runs.505. Use the repo's configured secret/security checks for relevant changes. Never51 label a grep for key-shaped strings a security audit or print matched secrets.52 State scan scope and missing prerequisites. Inspect Markdown/config changes53 too when they can carry credentials or executable agent instructions.546. Review the actual diff for unintended changes, boundary cases, generated55 files and changes made after testing. Fix within the authorized scope, then56 rerun checks affected by the fix. Do not endlessly repeat successful checks.577. Report evidence and remaining limitations. Readiness is not authorization to58 commit, push, merge or deploy. Use existing authorization for those actions.5960## Evidence contract6162| Check | Required evidence |63|---|---|64| Build / types / lint | Command, working directory, exit code, relevant findings |65| Tests | Command, collected/passed/failed/skipped counts and coverage if measured |66| Behavior | Requirement exercised and observed result; mocks versus real integration |67| Security | Scanner and scope, redacted finding count, unavailable dependencies |68| Diff | Base/head or working-tree scope, changed files, untested later changes |6970Use **PASS**, **FAIL**, **SKIPPED (reason)**, **UNAVAILABLE (prerequisite)** or71**NOT APPLICABLE (reason)** per check. Required unavailable/skipped checks leave72readiness unverified. Never turn them into PASS. A test summary printed before73a hang or nonzero process exit is incomplete evidence.7475## Prove it works7677Verify against the real artifact, not a proxy, a self-report, or "it78compiles." Read the actual value, not a cached or derived representation;79check process liveness directly, not through derived state. When80verification fails, suspect the observation method before suspecting the81system. The strongest proof is a deterministic script that re-runs the same82comparison: write it, run it, and keep its output as an artifact a reviewer83can re-run instead of trusting your word.8485## Prerequisites and companions8687This is instructions-only: it supplies no `/verify` command, hooks or scanner.88Use the loaded project's tools. Route through `skill-router` to find89`python-testing`, `react-testing`, `webapp-testing`, `requesting-code-review`,90`security-best-practices` or `session-handoff` when relevant.9192## Provenance9394Adapted from [ECC verification-loop at the reviewed revision](https://github.com/affaan-m/ECC/blob/e482e579415fde18357cafce70f177ae19fd7f03/skills/verification-loop/SKILL.md).95See [MIT notice](LICENSE.txt). Local changes replace fixed commands/thresholds,96timed repetition and superficial security checks with project-scoped evidence.