一键导入
evaluate
Run the application and verify sprint contract criteria via API tests, Playwright interaction, and schema validation.
用 Codex 或 Claude 帮你安装 复制这段 Prompt,粘贴到 Codex、Claude 或其他助手里,让它检查 Skill 页面并帮你完成安装。
菜单
Run the application and verify sprint contract criteria via API tests, Playwright interaction, and schema validation.
用 Codex 或 Claude 帮你安装 复制这段 Prompt,粘贴到 Codex、Claude 或其他助手里,让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
Autonomous build loop with Karpathy ratcheting, GAN evaluator, and session chaining. Iterates story groups until all features pass or stopping criteria met.
Socratic interview to create a Business Requirements Document. First step in the SDLC pipeline.
Generate system architecture, machine-readable schemas, and UI mockups. Spawns planner + ui-designer concurrently.
Evaluation patterns — sprint contract format, three-layer verification, scoring rubric references.
Standard GitHub issue workflow. Branch, reproduce, fix, test, PR.
Generate production code and tests for a story group using agent teams for parallel execution.
| name | evaluate |
| description | Run the application and verify sprint contract criteria via API tests, Playwright interaction, and schema validation. |
| argument-hint | [group-id] |
| context | fork |
| agent | evaluator |
Verify that the implemented group meets all sprint contract criteria by running live checks against the application: API calls, Playwright browser interaction, and schema validation.
/evaluate C
Evaluates group C's sprint contract. The group ID matches a node in specs/stories/dependency-graph.md and a file at sprint-contracts/{group}.json.
Before running /evaluate, verify:
sprint-contracts/{group}.json exists and is valid JSON.project-manifest.json exists with api_base_url, ui_base_url, and health_check fields.Read .claude/skills/evaluation/SKILL.md for project-specific evaluation patterns, custom assertion helpers, and any environment-specific overrides.
Read sprint-contracts/{group}.json. The contract contains:
api_checks: list of HTTP endpoint checks.playwright_checks: list of browser interaction sequences.design_checks: list of visual and component checks (evaluated in Full mode only).architecture_checks.files_must_exist: list of file paths that must be present on disk.features: list of feature IDs this group satisfies.Read project-manifest.json. Extract:
api_base_url — base URL for all API checks (e.g., http://localhost:8000).ui_base_url — base URL for Playwright navigation (e.g., http://localhost:3000).health_check — path to the health endpoint (e.g., /health).Run a health check to confirm the application is live:
curl --retry 5 --retry-delay 3 -sf {api_base_url}{health_check}
If the health check fails after 5 retries, immediately record a FAIL with failure_layer: "docker" and stop. Do not proceed to API or Playwright checks. A broken stack is not a partial pass.
For each entry in api_checks:
Execute the request via Bash:
curl -s -w '\n%{http_code}' -X {method} {api_base_url}{path}
Include -H headers and -d body as specified in the check entry.
Parse the response: the last line is the HTTP status code; everything before it is the response body.
Verify status code matches expect.status. A mismatch is a FAIL for this check.
Verify the response body contains every string listed in expect.body_contains. A missing string is a FAIL for this check.
If the check entry contains a schema_ref field, validate the response body against the schema:
python3 -c "
import json, jsonschema, sys
body = json.loads(sys.stdin.read())
schema = json.load(open('specs/design/api-contracts.schema.json'))
ref = schema['{schema_ref}']
jsonschema.validate(body, ref)
print('schema valid')
" <<< '{response_body}'
A schema validation error is a FAIL for this check.
Record each check as PASS or FAIL with the actual vs. expected values.
Before reporting an API check as FAILED, read the server logs:
docker compose logs backend --tail=50 2>&1
Include the relevant error from the logs in the failure report. This gives the generator the actual stack trace, not just "got 500 instead of 200."
For each performance_checks entry in the contract:
# Measure response time
time_ms=$(curl -s -o /dev/null -w "%{time_total}" -X {method} {api_base_url}{endpoint} | awk '{printf "%.0f", $1 * 1000}')
If time_ms > max_response_time_ms, report as WARN (not BLOCK — performance is advisory unless critical).
For each entry in playwright_checks:
Use Playwright MCP tools to execute the interaction sequence:
browser_navigate — navigate to a URL.browser_click — click an element (use getByRole, getByText, or getByLabel; never CSS selectors).browser_fill_form — fill form fields.browser_snapshot — capture the DOM snapshot for assertion.Execute each step in the order specified. Do not reorder or skip steps.
Verify each assertion listed in the check entry:
browser_navigate landed on the expected path.Use expect().toBeVisible() for visibility assertions. Never use waitForTimeout() — if an element is not immediately visible, the check fails.
Record each check as PASS or FAIL with a description of what was asserted and what was found.
Skip this layer entirely in Lean or Solo mode.
In Full mode, delegate to the design-critic agent:
design_checks entries from the sprint contract.ui_base_url.Record the design-critic's verdicts as-is. Do not override them.
For each path listed in architecture_checks.files_must_exist:
This check does not require Docker to be running.
After all checks complete, update features.json for every feature ID listed in the sprint contract's features array:
passes: true if all checks for that feature passed, false otherwise.last_evaluated: current timestamp in ISO 8601 format.failure_reason: null if passing; otherwise a human-readable description of the first failure (e.g., "GET /users/1 returned 404, expected 200").failure_layer: null if passing; otherwise one of "api", "playwright", "design", "unit_test", "docker".Do not remove existing fields from features.json. Merge the updates into the existing structure.
Write the full evaluation report to specs/reviews/evaluator-report.md:
# Evaluator Report — Group {group}
Date: {ISO 8601 timestamp}
VERDICT: PASS | FAIL
## API Checks
- [PASS] POST /users → 201 ✓
- [FAIL] GET /users/1 → expected 200, got 404
- [PASS] DELETE /users/1 → 204 ✓
## Playwright Checks
- [PASS] Upload page renders ✓
- [FAIL] Submit button not clickable
- [PASS] Success message visible after form submit ✓
## Design Checks
- [PASS] Button uses primary color token ✓
- [SKIP] Design checks skipped (Lean mode)
## Architecture Checks
- [PASS] All expected files exist ✓
- [FAIL] Missing: src/repository/user-repository.ts
## Features Updated
- F001: PASS
- F002: FAIL (api: GET /users/1 expected 200, got 404)
- F003: PASS
The overall VERDICT is PASS only if every check across all layers passes. A single FAIL in any layer produces a FAIL verdict.
| Mode | Layer 1 (API) | Layer 2 (Playwright) | Layer 3 (Design) |
|---|---|---|---|
| Full | Run | Run | Run |
| Lean | Run | Run | Skip |
| Solo | No-op — print "Solo mode: skipping evaluator" and exit |
Determine the current mode from project-manifest.json field mode. Default to Full if the field is absent.
getByRole, getByText, getByLabel). CSS selectors break with minor UI changes and are not permitted.failure_layer: "docker" and stop. Do not attempt workarounds or partial evaluations.