Flaker Add Adapter

flaker — add a test-result adapter

flaker import <file> --adapter <type> parses external test reports (vitest, playwright, junit, …) into TestCaseResult[] and inserts them into the DuckDB store. New adapters plug into a single switch.

When this skill applies

• "flaker に の adapter を足したい"
• "flaker import で を読みたい"
• "'s test report を flaker に取り込みたい"

Files involved (3 to add, 1 to edit)

The bundler (scripts/build-package.mjs) discovers files transitively from the entry point — no manual export to add anywhere else.

Adapter shape

import type { TestCaseResult, TestResultAdapter } from "./types.js";
import { resolveTestIdentity } from "../identity.js";

interface RawShape { /* mirror only the fields actually consumed */ }

export const <name>Adapter: TestResultAdapter = {
  name: "<name>",
  parse(input: string): TestCaseResult[] {
    const report = JSON.parse(input) as RawShape;
    // … walk the report …
    // Use resolveTestIdentity(...) to build each result so the testId
    // hashing is consistent with the rest of flaker.
    return out;
  },
};

TestCaseResult (see src/cli/adapters/types.ts) requires:

• suite, testName, status ("passed" | "failed" | "skipped" | "flaky"), durationMs, retryCount
• taskId (optional but strongly recommended for stable identity across runs)
• errorMessage, failureLocation, artifacts, variant (optional)

Status mapping conventions

These are conventions across the existing adapters — keep them:

• recovered / retry succeeded → flaky (with retryCount: 1).
• error / timeout → failed.
• An entry with no failure but non-empty errors[] is still failed (page-level rollup; chaosbringer adapter does this for chaos crawls).
• Skipped / pending → skipped.

errorMessage hygiene

Long unprocessed messages bloat the DuckDB row. Adapters should:

1. Dedupe (the same network error firing 50 times shouldn't show 50× in storage).
2. Cap message count + per-message length (chaosbringer adapter caps at 4 unique × 200 chars; pick what fits the report's verbosity).
3. (+N more) suffix counts UNIQUE messages dropped, NOT duplicates removed by dedupe — this distinction was the #79 review fix. Conflating them over-reports.

Variant for run identity

When the source report has a per-run identifier (seed, build number, run id), put it in variant:

variant: { source: "<adapter-name>", seed: String(report.seed), runId: report.runId }

This lets flaker status / KPI queries split rows by run when needed.

Off-shape input handling

JSON.parse failures should propagate (caller passes a bad file → caller's problem). Inside the adapter, defend only against:

• Missing top-level array fields (if (!Array.isArray(report?.pages)) return []).
• Unexpected status strings (treat as "failed" with the raw status in errorMessage).
• Off-origin / malformed URLs in URL-style testNames (keep the raw string; don't throw).

Register

// src/cli/adapters/index.ts
import { <name>Adapter } from "./<name>.js";

// inside createTestResultAdapter switch:
case "<name>": return <name>Adapter;

The CLI's --help text for flaker import is hard-coded and goes stale; do NOT chase it in this PR (separate cleanup). The actual switch is the source of truth.

Tests

tests/adapters/<name>.test.ts — vitest, inline JSON fixture, one assertion per behaviour:

• Status mapping (passed / failed / skipped / flaky / recovered).
• retryCount and durationMs round-trip.
• errorMessage dedup + length cap + (+N more) semantics.
• Empty / off-shape input returns [] instead of throwing.
• Same-shape inputs produce the same testId (resolveTestIdentity stability).

Aim for 8-12 tests. Less is undertesting; more is usually testing JSON.parse.

End-to-end smoke (recommended, not required)

Build the CLI (pnpm run build), then:

node dist/cli/main.js import <real-report.json> \
  --adapter <name> \
  --commit smoke-001 --branch smoke --source local

Should print Imported N test results. Silent + exit 0 means the adapter returned [] — usually a shape-mismatch bug in parse.

CHANGELOG

### Added

- New `<name>` test-result adapter: `flaker import <file> --adapter <name>`. <one-sentence what it parses, what it produces>. <Why this is useful>.

What NOT to do

• Don't ship per-row dynamic identity — testId should be deterministic from (suite, testName, taskId, variant), never include timestamps or run-local IDs. Otherwise flaky detection compares rows that look distinct.
• Don't emit zero-row results when the report has data — that's the #1 bug class. End-to-end smoke catches it.
• Don't use console.log for diagnostics in the parse function — the import command's caller might be parsing JSON output.