Confidence Gate

Principle

When generating tests, scaffolding fixtures, classifying risk, or proposing any non-trivial test artifact, emit a confidence assessment before writing code. If confidence is below the threshold, stop and ask the user instead of generating plausible-looking output built on guesses.

Rationale

The failure mode of LLM-generated tests is rarely “refused to try” — it is “generated something plausible that passes locally and breaks silently in CI.” Hallucinated selectors, invented endpoint paths, fabricated risk scores, and reverse-engineered schemas all produce code that looks correct and tests nothing real. A confidence gate makes that failure mode loud by forcing the agent to declare its evidence and its unknowns before any artifact is committed.

Required output shape

Every non-trivial test artifact proposal must include:

Confidence: <1-10>
Rationale: <one or two sentences citing concrete evidence from the repo or contract>
Unknowns: <bulleted list of things the agent does not know>

The Rationale must cite a file path, a contract document, an existing pattern, or a captured observation. Vague rationale (“based on standard patterns”, “looks similar to other tests”) is not evidence and forces the score down.

Threshold rule

Confidence ≥ 7 — proceed with generation.
Confidence 5–6 — proceed but surface the assumptions to the user in the output so they can correct mid-flight.
Confidence < 5 — STOP. Do not generate. Ask the user to resolve the most-blocking Unknown first.

When to apply

Apply the gate when generating or proposing:

Selectors and page objects. Must have explored the live application via playwright-cli or read existing page object patterns. Confidence < 5 if neither.
Endpoint paths and request shapes. Must have read the OpenAPI / Swagger contract or existing endpoint enums. Confidence < 5 if the endpoint is being invented.
Risk classification (test-design, NFR). Must cite probability and impact evidence. Confidence < 5 if scoring is vibes-based.
Fixture composition. Must understand existing mergeTests patterns and fixture boundaries in the repo. Confidence < 5 if composing blindly.
Schema authoring (Zod, Ajv, JSON Schema). Must have a documented contract source (OpenAPI, JSON schema, existing schema file). Confidence < 5 if reverse-engineering from a single sample response.
Data factories. Must understand the production data shape and constraints. Confidence < 5 if guessing field validity rules.

When NOT to apply

Mechanical refactors with clear scope (rename a variable, add a tag, update an import).
Reading or summarizing existing artifacts.
Producing reports from already-gathered data.
Trivial test additions that copy an existing pattern exactly.

The gate exists to prevent fabrication, not to bureaucratize obvious work.

Anti-patterns

❌ Vanity scores. Confidence: 9 with no Rationale, or Rationale that does not cite evidence. Score the evidence, not the optimism.

❌ Listing then ignoring Unknowns. Listing unknowns and then proceeding anyway when Confidence is below threshold. If the gate is below threshold, the only valid next action is to ask the user.

❌ Asking generically. Asking “should I proceed?” instead of resolving the most-blocking Unknown with a concrete one-sentence question.

❌ Inflating to clear the bar. Adjusting Confidence upward to avoid the stop rule. If the evidence is weak, the score is weak; resolve the evidence, not the number.

Patterns that work

✅ Cite the source. “Confidence: 8 — Rationale: read src/openapi/users.yaml line 142-167 and existing schema at tests/api/users.schema.ts.”

✅ One concrete Unknown. When below threshold, ask one specific question: “Is POST /users/{id}/role documented anywhere? I can’t find it in the OpenAPI spec and there are no existing tests for it.”

✅ Promote evidence. When the user answers the Unknown, the Rationale gets stronger and Confidence rises legitimately. The gate is a feedback loop, not a checkpoint.

test-quality.md — Definition of Done for tests; the gate protects DoD compliance.
risk-governance.md — risk scoring discipline that informs Rationale for risk-related gates.
probability-impact.md — scoring scales used in risk-related Rationale.
selector-resilience.md — selector confidence specifically.
playwright-cli.md — the sanctioned exploration tool that promotes selector Confidence.

confidence-gate.md 4.5KB 原始文件 永久連結 Blame 文件歷史