- Home
- Skills
- Code Quality & Review
- Agent Foreman Verify & Repair - Independent AI Code QA
Works with the AI Agents you already use
Agent Foreman Verify & Repair - Independent AI Code QA
The agent that wrote the code is the worst judge of it: it remembers what it meant to do, not what the diff does.
$9
Agent Foreman Verify & Repair - Independent AI Code QA
Example session with this skill installed
Use foreman-verify-repair on the last commit. Acceptance criteria: 1) cart totals round half-up to 2 decimals (2.345 -> 2.35); 2) a regression test proves it; 3) no other test changes behaviour.
- Read your context and instructions
- Compiled the agent foreman verify
Summary
The implementer reported: fixed, test added, all 212 tests pass. I did not take that at its word. A fresh reviewer that never saw the conversation checked the commit.
Round 1: criterion 2 FAILED - the new test passes at the parent commit too, so it never exercised the bug. One repair round later,
all three criteria PASS with re-runnable evidence.
Round 1 - independent review
| # | Criterion | Verdict | Evidence |
|---|---|---|---|
| 1 | Totals round half-up to 2 decimals | PASS | git show HEAD - rounding change in the total calculation; targeted test run: 6 passed |
| 2 | Regression test proves it | FAIL | Fail-first check: new test dropped into the parent commit (temporary worktree) also passes. It asserts 2.34, while the requirement is 2.345 -> 2.35 (test file and line cited) |
| 3 | No other behaviour change | PASS | Full suite at parent and at commit: same test count, no new failures |
Repair task (sent to a fresh worker)
- Failed criterion: 2 - regression test proves the fix
- Evidence: test passes at the parent commit; asserted value 2.34 does not exercise half-up rounding
- Property to preserve: the test must FAIL at the parent and PASS at the commit; no other test may change
Round 2 - second fresh reviewer
| # | Criterion | Verdict | Evidence |
|---|---|---|---|
| 1 | Totals round half-up to 2 decimals | PASS | unchanged, re-run: passed |
| 2 | Regression test proves it | PASS | parent commit: new test FAILS (2.34 != 2.35); fix commit: PASSES |
| 3 | No other behaviour change | PASS | suite re-run at both commits; no new failures, no test-count drop |
Every row lists the exact commands the reviewer ran, so anyone can re-run them.
Next steps
- Paste the verdict table into the pull request as the definition of done.
- The reviewer changed no code and removed its temporary worktree.
- If a criterion had failed twice, the loop would have stopped here and asked you with the evidence.
Connects securely to your tools. The creator never sees your data.
About this skill
The agent that wrote the code is the worst judge of it: it remembers what it meant to do, not what the diff does. Foreman Verify & Repair separates building from checking.
After an implementation, your agent spawns a fresh reviewer subagent that receives only the original acceptance criteria, the repository, the commit and its parent - never the implementer's conversation or summary. The reviewer reads git show, runs the tests itself, and checks:
- every criterion: PASS / FAIL / NOT-VERIFIABLE-HERE, each with evidence a reader can re-run;
- fail-first: new tests fail at the parent commit and pass at the fix (so the test actually tests it);
- regression as a property: no failure at the commit that the parent does not have, and no drop in the test count (skipped or deleted tests look green);
- commit hygiene: foreign files, author, message;
- invariants in full when shared contracts, permissions, persistence or deletion are touched.
Every FAIL becomes a repair task - the failed criterion, the evidence and the property the fix must preserve - for a fresh worker, followed by a fresh reviewer. After two failed repairs on the same criterion it stops and shows you the evidence. Agents without subagents get a clearly labelled self-review checklist.
Key capabilities
- Fresh, independent reviewer per change; never sees the implementer's summary
- Per-criterion verdicts with re-runnable evidence; prose counts as FAIL
- Fail-first verification at the parent commit in a temporary worktree
- Regression judged as a property, including test-count drops
- Commit hygiene and invariant checks
- Automatic repair tasks with the property each fix must preserve
- Clear stop conditions; partially met criteria are never accepted
- Ready-made reviewer prompt, verdict format, repair template and self-review fallback
Who it is for
Developers and teams using AI agents for real changes, reviewers who want evidence before merging, and orchestrated multi-agent setups that need a definition of done.
Install / usage
Add the skill and say: "Use foreman-verify-repair on the last commit. Acceptance criteria: 1) ... 2) ..." The agent spawns the reviewer, reports the verdict and runs the repair loop on failures. Works with foreman-orchestrator automatically.
Limitations
Independent mode needs an agent that can spawn subagents; otherwise the self-review fallback is used and labelled as weaker than independent QA. Criteria that can only be checked in production are marked NOT-VERIFIABLE-HERE, never faked.
Changelog
- 1.0.0 - First release: reviewer prompt, verdict format, fail-first procedure, repair loop, self-review fallback.
Find more skills like this on the Claude Skills Marketplace, the Codex Skills Marketplace and the Gemini Skills Marketplace, or browse Claude Skills for Code Quality & Review.
Independent mode needs an agent that can spawn subagents; without them a self-review checklist is used and labelled as weaker than independent QA. QA quality depends on how specific the acceptance criteria are. Criteria that can only be checked in production are marked NOT-VERIFIABLE-HERE, never faked. Requires git and the project's own test commands.
How to use
- 1
Connect your agent to the Agensi MCP
Copy the prompt and paste it into Claude, ChatGPT, Cursor or any agent that supports MCP. It sets up the connection for you. Manual setup
- 2
Sign in to Agensi
Your agent shows a sign-in link. Sign in and click Allow.
- 3
Ask for what you need
Your agent loads this skill and uses it on your task. It always asks before buying a skill.
Reviews
No reviews yet
Be one of the first to try it. Every listed skill goes through the trust checks below.
Security scanned
Checked by our 8-point scan before listing
Fresh listing
Recently published to Agensi
Money-back guarantee
Doesn't do what its page promises? You get your money back.
Trust & safety
Security scan
Security scan passed 4 days ago
- Software
Git