Cf Test

Report candidate test-quality findings against observable contracts and authoritative invariants. Do not edit repository files other than the batched pass's recap, run the test suite, or turn test style preferences into findings.

Scope

Resolve exactly one test scope:

• Pending tests, the default when no target is named and the working tree holds test changes: existing test files selected by staged, unstaged, and untracked work together.
• The existing suite, when no target is named and the working tree holds no test change: every existing test file. A clean checkout does not make a request about tests unanswerable; it makes the whole suite the subject.
• A named history range: existing test files selected by the requested commits, range, or branch.
• Explicit tests: the named test files or directories, whether changed or unchanged.

Treat every selected test file as a primary test and inspect it as a whole. Treat changed production files in the same change set, production files directly imported by explicit tests, and identified authoritative documents as contract surfaces. Requirements, product or domain documentation, and acceptance criteria are authoritative only when they name the tested behavior; path proximity alone is insufficient. Contract surfaces provide evidence but are not themselves review targets.

Exclude generated, vendored, ignored, fixture-only, snapshot-output, and deleted files unless an authoritative requirement makes one directly relevant. If the repository holds no test file at all, report that and stop. Never widen a test scope to production files.

Hard Gates

• No invariant or observable contract, no missing-coverage, absence, brittle-string, or over-specification finding. A preference is not a requirement.
• Report one finding per nameable test or test table. Do not convert one assertion smell into a repository-wide pattern inventory.
• Keep every finding a candidate. Confirming intended behavior or changing production design belongs to the user or the receiving skill.
• Complete both lens groups before routing or suggesting edits.

Flow

1. Resolve primary tests, contract surfaces, and authoritative sources by path without loading the corpus in full.
2. Read references/dynamic-agents.md. Run the installed-local context gate against every primary test and already-selected contract surface. Follow its policy exactly; when it is not local, read references/delegated-execution.md and follow it.
3. Read references/assertion-quality.md and run both declared lens groups. For local, run them sequentially. For subagent-1, subagent-2, or batched, also read references/test-agent-brief.md and fill every placeholder with bounded contract surfaces. Every assignment applies both lens groups. The shared reference owns assignment and completion.
4. De-risk every reported candidate against the cited test, contract surface, and authoritative source. Inspect only the cited evidence needed to accept, narrow, or exclude it.
5. Route once after the complete pass. Use cf-scenario when expected behavior is ambiguous, cf-mr-wolf when a domain-model redesign such as making invalid states unrepresentable needs a decision, and a direct bounded test edit only after the invariant is confirmed.

Artifacts

• Owns .cflow/cf-test-recap.md: the batched pass's recap, written under the Recap File contract in references/delegated-execution.md. It is the pass's one repository write.
• When it already exists and the request is about it, work through it under the same contract instead of running the Flow.
• Create or update no other repository file, including any other .cflow/* file and todo files. If the user asks to persist findings elsewhere, finish this pass and name cf-todo as a separate next action.

Output Format

Return only:

• Scope: resolved scope, primary tests, contract surfaces, authoritative sources, and exclusions.
• Context budget: measured files, LOC, estimated tokens, selected policy, consent source, and actual model or runtime default; use local when no agent ran.
• Lenses: every lens from references/assertion-quality.md, marked reporting, silent, or not applicable with the absent condition.
• Findings: candidates grouped by the action they ask for, in the order remove, rewrite, strengthen, and by severity within each group. A test that does not pay for itself is the answer to a different question than a test that covers too little, and burying the first under the second answers neither. Each candidate carries claim, evidence, invariant or contract, impact, severity, confidence with its basis, false-positive check, and status candidate.
• Handoff: destination mapped to candidate ids, recommended first action, or none.
• Result: clear or candidates found, opening with how many cases are proposed for removal; then state that confirmation is still needed, that no files were modified except the recap when written, and the next action.