| name | verify |
| description | Skeptically verify that an implementation actually works — exercise the feature, run validation tiers, check coverage, confirm new code is wired up. Delegates to the verifier agent. |
| allowed-tools | Read, Grep, Glob, Bash(git status), Bash(git diff*), Bash(git log*), Bash(uv run poe quick-check), Bash(uv run poe agent-check), Bash(uv run poe check), Bash(uv run poe full-check), Bash(uv run poe test), Bash(uv run poe test-coverage) |
/verify — Skeptical implementation check
PURPOSE
Prove the implementation works with evidence — not assumptions. Catches "compiles but doesn't actually run" failures the test suite can miss.
CRITICAL
- Trust evidence, not claims — run every command, read every file, check every output. Never accept "should work" without proof.
- Exercise the feature first — before checking infrastructure, actually invoke the new MCP tool / API method / fixed code path
- Use
LSP findReferences on new symbols — zero callers means it's not wired up, even if imported
- Generated files must be intact — if
api/, models/, or client.py changed, both regenerate-client and generate-pydantic must have run
STANDARD PATH
1. Pick the validation tier
| Tier | Cmd | When |
|---|
| 1 | uv run poe quick-check | During iterative development |
| 2 | uv run poe agent-check | Before committing |
| 3 | uv run poe check | Before opening PR |
| 4 | uv run poe full-check | Before requesting review |
Use Tier 3 by default for verification. If the change is small and isolated, Tier 2 is enough.
2. Exercise the feature
- New MCP tool → invoke via test harness or MCP inspector
- New API method → make a test call (or via integration test)
- Bug fix → reproduce the original scenario; confirm it no longer fails
- Behavior change → demonstrate old vs new
3. Run the verifier agent
Delegate the systematic checklist to the verifier agent. It walks: code exists, code works (validation tier passes), code is complete (no TODO stubs), code is integrated (LSP findReferences on new public symbols), generated-files intact, coverage maintained, regression check.
4. Confirm
- Validation tier passes clean (no warnings)
- New public functions/classes have callers (or are intentionally part of a public API surface)
- Coverage on core ≥ 87%
- No tests skipped or deleted unintentionally
5. Report
## Verification Report
### Status: PASS | FAIL
### Verified
- [item]: [evidence: command output, file:line, test count]
### Failed
- [item]: [what's wrong, what to fix]
### Recommendations
- [improvements noticed]
EDGE CASES
- [Coverage decreased] — identify which file lost coverage and why; document if the loss is acceptable (e.g. removed dead code) or fix.
- [Tests passed but feature doesn't actually work] — the feature exercise step (§2) caught it; otherwise verification is incomplete.
- [Verification reveals a doc gap] — fix CLAUDE.md or relevant guide as part of the verification work, not later.
RELATED
verifier agent — runs the full systematic checklist
pr-preparer agent — checks process readiness (commit format, coverage threshold)
code-reviewer agent — assesses code quality, not just whether it works