verify

name	verify
description	Skeptically validate that implementation is complete and works. Heavier than the verifier agent — exercises the feature, checks integration, and confirms regression protection.
allowed-tools	["Read","Grep","Glob","Bash(uv run poe agent-check)","Bash(uv run poe test)","Bash(uv run poe test-coverage)","Bash(git diff *)","Bash(git status)"]

/verify — Verify Implementation

PURPOSE

Skeptical end-to-end check: not just "compiles green" but "actually works."

CRITICAL

• Be skeptical — never accept "should work"; prove it.
• Run real commands, read real files — claims without evidence are not verified.
• Don't fix anything — surface failures; don't quietly patch.

STANDARD PATH

0. Exercise the feature

Before any infrastructure check, use what was implemented:

• New MCP tool? Invoke it via the test harness or MCP inspector.
• New API method? Make a test call (or run the matching integration test).
• Bug fix? Reproduce the original failure scenario; confirm it no longer fails.
• Behavior change? Demonstrate before-vs-after.

This catches "compiles but doesn't work" failures the test suite missed.

1. Code exists

• All claimed files present (not just planned)
• No stubs, TODOs, or "will do later" gaps
• All imports resolve

2. Code works

uv run poe agent-check    # format + lint + type check
uv run poe test           # tests

ALL green, no new warnings.

3. Code is complete

• All requirements addressed
• Edge cases handled (empty, error, None/UNSET)
• Error handling uses appropriate exception types

4. Code is integrated

For each new public function/class, run LSP findReferences. Zero references means something is unfinished — flag it.

5. Generated files intact

• api/**/*.py, models/**/*.py, client.py not manually edited
• If client was regenerated, pydantic models also regenerated

6. Coverage maintained

uv run poe test-coverage

Core logic ≥87%. New code has both success and error path coverage.

7. Regression check

• Confirm step 2 ran the FULL test suite (not a subset)
• No tests deleted or skipped accidentally
• If coverage dropped, identify what lost coverage and why

OUTPUT

## Verification Report

### Status: [PASS | FAIL]

### Verified
- [item]: [evidence]

### Failed
- [item]: [what's wrong + fix]

### Recommendations
- [improvements noticed]

EDGE CASES

• Instructions are wrong — if CLAUDE.md or another doc misled the implementation, fix the doc as part of verification. The goal is that future work doesn't repeat this failure.

RELATED

• verifier agent — fast mechanical gate (this skill is the heavier audit)
• pr-preparer agent — process readiness (commits, ADRs, help resource sync)
• code-reviewer agent — design review (this skill is correctness, not design)