workflow-verify

name	workflow-verify
description	Run non-mutating Smelter verification with fresh command evidence before any pass or fail claims.
version	0.55
type	workflow
consumes	current codebase (no RED requirement)
produces	verify_report.md
default_pattern	A
default_agent	qa-tester
supports_patterns	["A","C"]
team_template	{"C":{"parallel_split_by":"phase","agents":["qa-tester","code-reviewer","qa-tester"],"aggregator":"arbitrator","sync_point":"verify_report.md"}}
can_delegate_to	["e2e-testing-patterns","tdd-guide"]
gate	{"postcondition":[{"file_exists":"verify_report.md"},{"all_phases_reported":true}]}

workflow-verify

Overview

Non-modifying verification of the current codebase. Runs test and static verification, records E2E scope requirements, and produces verify_report.md. Used as the entry skill of /verify mode before the pipeline proceeds to workflow-e2e, workflow-e2e-review, and workflow-human-check.

Core principle: Evidence before claims. A pass/fail statement is only valid after the proving command or artifact has been freshly produced and read.

Violating the letter of this rule is violating the spirit of this rule.

Announce at start: "I'm using workflow-verify to run tests + static checks, record E2E scope, and produce verify_report.md before real-interface verification."

The Iron Law

NO VERIFY CLAIM WITHOUT FRESH EVIDENCE

The skill completes when verify_report.md exists with fresh Phase 1 and Phase 2 evidence plus a Phase 3 E2E scope section that tells workflow-e2e what real interface must be exercised. Skipping evidence = gate fail.

Verification concerns

Phase 1 — Test code execution

Run the project's existing test suites, scoped to the changed surface when possible.

# scoped unit / integration (example — adapt per project)
pnpm vitest run <changed_test_files>

# or pytest, jest, go test — whatever the repo ships

Record pass/fail counts, failing case names, error snippets.

Phase 2 — Code logic inspection (static)

Run static analyzers to catch logic issues without executing the code.

pnpm typecheck            # or: tsc --noEmit
pnpm lint                 # eslint / prettier
node scripts/feature-version-check.mjs report   # state schema integrity
node scripts/session-end.mjs --dry              # forbidden-reference scan

Record: typecheck errors, lint violations, state-schema drift, doc-sync issues.

Phase 3 — E2E interface scope

Phase 3 is NOT satisfied by a test-runner command. Phase 1 already ran test code. This skill records which real interfaces must be exercised. The next pipeline stage, workflow-e2e, drives those interfaces and captures artifacts.

Record the actual interface that users / callers / clients touch. Every required surface must later produce an artifact file under .smt/features/<slug>/artifacts/. Zero-artifact E2E runs fail in workflow-e2e / workflow-e2e-review.

Surface	Real interface	Required artifact
UI / Frontend	Real browser, Playwright launched	video (.webm) + screenshots (.png)
CLI / Script	Real subprocess (stdin, argv, exit code)	command transcript (argv + stdin + stdout + stderr + exit)
HTTP API	Real server on a real port + real fetch/curl	request/response log (method, path, status, body, timing)
Database	Real DB engine	query log + before/after state samples
Hook script	Real cat payload | node hook.mjs	input/output JSON samples + exit code

Disallowed in Phase 3:

• Calling an internal handler function directly instead of through the interface.
• Mocking the interface under test (MSW for the HTTP surface, stubs for the DB driver, etc.).
• Running only pnpm test / vitest run — that's Phase 1.
• Recording only an acknowledgement signal (toast, banner, modal close, spinner idle, HTTP 2xx, exit code 0) without separately asserting the target effect that was supposed to materialize.

Phase 3 hands off to the full workflow-e2e real-interface contract, including the Effect-vs-Ack rule and state.json.scenarios[].effect_evidence requirement. A run with artifacts but no effect evidence still fails downstream E2E review.

Scope by the changed surface; full regression only on explicit user request (workflow-human-check).

Output — verify_report.md

## Phase 1 — Tests
- runner: <vitest | jest | pytest | ...>
- scope: <paths>
- total: <N>  pass: <n>  fail: <m>
- failures:
  - <case_name> — <file:line> — <error excerpt>

## Phase 2 — Static Inspection
- typecheck: <pass | N errors>
- lint: <pass | N violations>
- state-schema: <pass | orphan features>
- doc-sync: <pass | drift items>

## Phase 3 — E2E Interface
- required surfaces: <ui | cli | api | db | hook | none>
- required scenarios: <...>
- artifact status: pending workflow-e2e | exempt with reason

## Summary verdict
- Result: pass | fail
- If fail → suggested next mode: /fix (for bugs or trivial fixes; surface exemptions cover CSS/copy/typo)

Pass / fail semantics

The skill itself completes as soon as tests/static checks have run, E2E scope has been recorded, and the report file exists. "Pass" vs "fail" in the report is an evaluation of the codebase evidence so far, not permission to skip downstream E2E or human-check.

• Skill pass (hook gate): verify_report.md exists with Phase 1, Phase 2, and Phase 3 E2E-scope sections.
• Report verdict = pass_so_far: zero Phase 1 failures and zero Phase 2 errors; E2E still proceeds when a surface is required.
• Report verdict = fail: any Phase 1 or Phase 2 failure → the report's suggested next mode drives routing after the user decision.

Fail routing (producer chain)

• test_run cause → workflow-coding (tests broke; fix code) → likely whitelist_violation in /verify mode → mode upgrade to /fix (or chain auto-advance)
• typecheck / lint / build cause → same: mode upgrade to /fix
• assertion cause (e2e) → same
• file_absent (verify_report.md missing) → self-rerun

Iron Law compliance

• Never modifies source code or tests. Strictly read + execute + report.
• Never halts — emits a report even if phases fail.
• No retries — one pass per invocation; routing on fail.

Scope exemption

exempt.tdd == true (mode default) → skips writing new tests, but still runs existing ones in Phase 1. exempt.e2e == false by default → E2E Phase 3 is mandatory unless tasker explicitly flags the change as having no interface surface.

Multi-Pass Verification

This is not a review skill. It does not run Multi-Pass rounds. One invocation = one report.

Red Flags - STOP

Thought	Reality
"Phase 1 passed, report clean"	Report needs tests, static checks, and E2E scope. Downstream E2E still runs when required.
"Phase 3 is just like pnpm test"	Phase 3 is real-interface E2E, not a test runner. pnpm test is Phase 1. Phase 3 requires artifacts.
"Ack signal is enough for Phase 3"	Phase 3 inherits workflow-e2e's Effect-vs-Ack rule. Ack alone fails.
"I'll skip Phase 2 if Phase 1 passed"	Typecheck / lint / state-schema / doc-sync catch different bugs than tests.
"Looks good"	Looks good is not evidence. Run the command, read the output, record the result.
"One failure means I should modify tests to make them pass"	workflow-verify NEVER modifies source or tests. It reads + runs + reports only.

Rationalization Prevention

Excuse	Reality
"No interface surface, skip Phase 3"	Tasker decides exempt.e2e. If not declared exempt, run Phase 3.
"Full regression would take too long"	Scoped by changed surface. Full regression is only on workflow-human-check request.
"I already ran this last session"	Re-run. Session-local state is irrelevant; verify_report.md must be fresh.

Terminal State — Required Next Skill

Skill completes when verify_report.md exists with tests, static inspection, and E2E scope sections. The required next skill is determined by the /verify pipeline.

Report verdict	Next action
pass_so_far with E2E required	workflow-e2e
pass_so_far with no interface surface	workflow-human-check may summarize the no-E2E rationale
fail with bug / regression	Continue to human-check with suggested mode upgrade to /fix unless the active chain explicitly routes recovery
fail with trivial diff (typo / CSS / copy)	Suggest /fix with surface exemption
file_absent (report missing)	Self-rerun

/verify is read-only by design. It does not chain into implementation skills. Mode upgrade to /fix is an explicit user decision, not an automatic source-edit transition.

Do NOT:

• Modify source code or tests to "fix" phase failures within this skill
• Invoke workflow-coding directly — that is a mode upgrade decision
• Skip producing verify_report.md because "all passed locally"
• Stop after verify_report.md when workflow-e2e or workflow-human-check remains in the active pipeline