Verification Loop

Verification Loop

Verification Loop runs the strongest practical checks for the current project, fixes failures when in scope, and repeats until the result is known.

Trigger Conditions

Use this skill when:

• Code, config, scripts, generated assets, or project docs changed.
• A feature/refactor/repair phase is complete.
• The user asks to verify, test, lint, build, typecheck, or fix current failures.
• $project-commander or $handoff-writer needs a readiness statement.

Do Not Trigger

Do not use this skill when:

• No files changed and the user only asked for conceptual analysis.
• Running commands would be destructive or require credentials the user has not provided.
• The project has no executable checks and a lighter file-format review is sufficient; record that limitation instead.

Command Discovery

Before running checks, discover project conventions:

• JavaScript/TypeScript: inspect package.json scripts and lockfile (pnpm-lock.yaml, package-lock.json, yarn.lock, bun.lockb).
• Python: inspect pyproject.toml, requirements.txt, tox.ini, noxfile.py, or pytest.ini.
• Rust/Go/.NET/Java: inspect standard manifests and existing CI config.
• If AGENTS.md defines verification commands, prefer those.

Verification Phases

Run the applicable phases in this order:

1. Build or compile check.
2. Type check.
3. Lint/static analysis.
4. Test suite or focused tests.
5. Security/secrets scan appropriate to the repo.
6. Diff review for unintended changes.

Skip unavailable phases only with a recorded reason.

Repair Loop

1. Run the check.
2. If it fails, read the error carefully and identify the smallest in-scope fix.
3. Apply the fix without reverting unrelated user changes.
4. Re-run the failed check.
5. Continue until checks pass, the issue is out of scope, or a real blocker is reached.

Required .agent/ Updates

Always update:

• .agent/LOG.md: commands, results, failures, fixes, reruns.
• .agent/TASKS.md: mark verification tasks and unresolved failures.
• .agent/HANDOFF.md: final verification status and exact next command if not clean.

Update when relevant:

• .agent/DECISIONS.md: chosen verification command, skipped gate rationale, accepted residual risk.
• .agent/REVIEW_CHECKLIST.md: mark verified items.

Other Skills

• Use $self-reviewer after verification passes or reaches a documented blocker.
• Return to $project-commander if verification reveals new implementation work.

Subagents

Do not delegate the final verification decision. A tester subagent may run focused checks in parallel only when subagents are available and permitted, but the main agent must inspect the final result.

Output Format

VERIFICATION REPORT
Build:     PASS/FAIL/SKIPPED - details
Types:     PASS/FAIL/SKIPPED - details
Lint:      PASS/FAIL/SKIPPED - details
Tests:     PASS/FAIL/SKIPPED - details
Security:  PASS/FAIL/SKIPPED - details
Diff:      PASS/FAIL/SKIPPED - changed files reviewed
Overall:   READY/NOT READY/BLOCKED
Issues:
1. ...
Next action:

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: fengsihao23-eng
• Source: fengsihao23-eng/shen-hua
• License: MIT

Install and usage instructions live in the source repository linked above.