loop-verifier

name

loop-verifier

description

Verify agent work against project gates before declaring it done. Use this skill whenever working in a project with a .loop/GATES.md file, whenever finishing any coding task inside a loop, whenever the user asks "is this actually done", "verify this", "check your work", or before emitting any VERDICT line. Also use it when a task involves tests, refactors, bug fixes, or dependency updates, when deciding whether work should be escalated to a human, and when the user asks to "audit my loop", "check loop readiness", or evaluate their loop setup.

Loop Verifier

The checker half of the maker/checker split. Turns "the agent says it's done" into "the gates say it's done." Built on one anti-cheating rule from the research: checks must be derived from the TASK, before or independently of the implementation — an agent that writes its checks after coding tends to write checks that mirror its own mistakes.

The protocol

1. Before implementing: write the checks

From the task description and .loop/GATES.md, produce the concrete check list for this task:

• Which universal gates apply (they all do) and what they mean here
• Which task-type gates apply (code / bug fix / refactor / docs / dependency)
• The project-specific commands (build, test, lint) that must exit clean
• 3–5 task-specific checks: edge cases, invariants, a worked example with a known expected result

Write this list down BEFORE writing any implementation. This ordering is the whole defense against self-confirming verification.

2. After implementing: run the gates

• Execute every runnable check for real (tests, build, lint). Paste actual output, not a summary of expected output.
• For non-runnable checks, walk each one explicitly: state the check, quote the evidence from the work, mark pass/fail.
• The Decompose audit is always included: every requirement in the task gets a quote from the output that satisfies it. No quote = gap = fail.

3. Verdict rules

• PASS requires: all applicable gates green, all checks walked, no check skipped silently.
• FAIL must name which gate failed and why — that line feeds the retry.
• A check you couldn't run (missing env, no test framework) is reported as UNRUN, and UNRUN on a required gate means the verdict is FAIL, not PASS. Unverifiable is not verified.
• Never soften a verdict because effort was high. The verdict describes the work, not the attempt.

4. Escalation triggers (hand to the human immediately)

• Any action on the GATES.md "Never" list would be required to proceed
• The same gate has failed 3 times for the same root cause
• The task requires changing the gates themselves
• Trust level is L1 and the fix requires applying changes

Verifying by task type

Code change — new behavior gets a new test; run suite before and after; Verify Loop on edge cases (empty input, boundary values, mutation of caller data, error paths).

Bug fix — reproduce first. A fix without a reproduction is a guess. The proof is one test that failed before the fix and passes after.

Refactor — the only gate that matters: identical behavior. Suite green before AND after; public contracts untouched; if no tests exist, write characterization tests first or escalate.

Docs/writing — every factual claim about the code checked against the code. Then the Stranger Review pass from the loop-engineering skill.

Dependency update — read the changelog for breaking changes before updating, not after the build breaks.

Loop setup audit ("audit my loop")

When the user asks to audit their loop setup or loop readiness, score the project against this checklist and report gaps with concrete fixes:

1. Brain present: .loop/GATES.md, LESSONS.md, STATE.md exist and aren't untouched templates (project-specific gates filled in?)
2. Gates executable: build/test/lint commands defined and actually run in this environment
3. Trust level sane: L2+ only with a populated LESSONS.md showing clean history; flag L3 without an allowlist as a finding
4. Memory alive: lessons being both written AND applied (standing rules present after 2+ weeks of use is the health signal); flag files past ~30 entries for compression
5. Stop rules: max attempts set; escalation triggers reachable; "Never" list intact
6. No sprawl: if multiple loops run, check they share STATE.md's "Do not touch" list and don't overlap targets

Output: a short scorecard (each item pass/gap/fix), worst gap first. This replaces a standalone audit CLI — the checklist is the tool.

Maker/checker note

This skill is strongest when the checker runs in a separate context from the maker (a fresh subagent, or at minimum a hard persona break) — models judge their own work leniently and other work accurately. See docs/maker-checker.md in the repo for the split recipes.