Runs Phase 4 review/publish/verify for a cli-web-* CLI: implementation review by 3 parallel agents, the tiered quality checklist (Tier 1 critical fail-fast, then comprehensive), pip install + smoke test, and per-CLI skill generation. Use when a CLI's tests pass and it is ready to be validated and published.
Instalar com Codex ou Claude Copie este prompt, cole no Codex, Claude ou outro assistente e deixe que ele revise a página da skill e instale para você.
Um comando direto ignora o prompt de revisão. Verifique a origem antes de executá-lo.
Instruções da origem · Visualização somente leitura
name
standards
version
0.5.0
description
Runs Phase 4 review/publish/verify for a cli-web-* CLI: implementation review by 3 parallel agents, the tiered quality checklist (Tier 1 critical fail-fast, then comprehensive), pip install + smoke test, and per-CLI skill generation. Use when a CLI's tests pass and it is ready to be validated and published.
when_to_use
Trigger phrases: "validate CLI", "publish CLI", "review CLI", "smoke test", "quality check", "start Phase 4", "quality checklist", "generate Claude skill", "verify implementation quality", or after the testing skill completes. Not for capture, implementation, or test writing.
Quality gate for cli-web-* CLIs. This skill owns the complete Phase 4:
independent implementation review, structural quality checklist, publishing,
and end-user smoke testing. Nothing ships until this phase passes.
Copy this checklist and check off items as you complete them:
TEST.md has both Part 1 (plan) and Part 2 (results)
All core modules are implemented and functional
<APP>.md (API map) exists and documents all endpoints
If tests are not passing, invoke the testing skill first. If this gate or
the phase state is in a failed/inconsistent state, follow
skills/shared/RECOVERY.md §phase-state Check Failures.
Optional pre-review coverage scan: before dispatching the review agents,
you MAY run the gap-analyzer skill
(${CLAUDE_PLUGIN_ROOT}/skills/gap-analyzer/SKILL.md, pass
APP_PATH=<app>/agent-harness) to diff captured endpoints
(<APP>.md + traffic-capture/traffic-analysis.json) against implemented
commands. It is optional here because the traffic-fidelity-reviewer agent
covers endpoint coverage during Step 1; gap-analyzer is the mandatory first
step of /refine, where no reviewer pass exists. Run it here when coverage
looks doubtful and you want the structured report before the agents start.
Site Profile Exceptions
Not all checks apply to every CLI. When evaluating, consider the site profile:
No-auth sites (public APIs): Skip auth-related checks (auth.py required,
auth commands, auth smoke test). Mark as N/A.
Read-only sites (no write operations): Skip write operation smoke test.
Verify reads return real data instead.
API-key auth sites: auth login takes a key argument, not playwright-cli.
auth refresh is not applicable — use auth logout instead.
Mark inapplicable checks as "N/A — [reason]" rather than creating dead-code stubs.
Step 1: Implementation Review (3 Parallel Agents)
Before checking structure or publishing, verify the code actually does the
right thing. Tests prove it runs; this step proves it's correct.
Dispatch 3 plugin agents in the same message using the Agent tool:
traffic-fidelity-reviewer — API coverage (reads .md + client.py + commands/)
output-ux-reviewer — User experience (runs --help, checks REPL, validates JSON)
Pass each agent: APP_PATH={app}/agent-harness, APP_NAME={app}, and site
profile (auth_type, is_read_only). The agents are defined in the plugin's
agents/ directory.
Agent
Focus
What it reads
What it catches
Traffic Fidelity
API coverage
<APP>.md + client.py + commands/
Missing endpoints, wrong params, broken response parsing, dead client methods, stale API map
Protocol leaks, stale REPL help, dead command files, broken entry points
Each agent scores findings on a 0-100 confidence scale. When all 3 return:
Filter out findings with confidence < 75 (noise)
Categorize remaining findings:
Critical (90-100): Bugs, missing endpoints, data loss, auth broken
Important (75-89): Wrong fields, incomplete parsing, missing options
Minor (75, edge cases): Help text gaps, cosmetic issues
Present the review report
Fix all Critical issues before proceeding — re-run only the affected
agent to verify the fix
Fix Important issues (not strictly blocking but strongly recommended)
Gate: Do not proceed to Step 2 until Critical count = 0.
Step 2: Structural Quality Checklist (tiered)
The checklist is tiered (see references/quality-checklist.md "Tiers"):
Tier 1 (critical) failures block publish; Tier 2 (comprehensive)
failures are warnings that should still be fixed.
2a. Tier 1 fail-fast first. Run only the critical checks and fix every
FAIL before doing anything else — there is no point reviewing a CLI whose
structure, packaging, or --json envelope is broken:
The summary shows per-tier counts. Exit is non-zero only on Tier 1 failures
(add --strict to make Tier 2 failures blocking too). Fix Tier 2 FAILs
before publishing unless explicitly deferred with a reason.
The validator automates the mechanical checks; the remaining judgment-based
items (documentation quality, error message guidance, fixture realism) are
reviewed manually per references/quality-checklist.md.
Step 3: Install and Verify
setup.py was generated by scaffold-cli.py in Phase 2 (namespace packages,
entry point, profile-correct dependencies) — verify it, don't rewrite it.
cd <app>/agent-harness
pip install -e .
which cli-web-<app> # entry point on PATH
cli-web-<app> --help# command tree renders
cli-web-<app> doctor # environment self-diagnosis (install, auth, deps)
If doctor reports fail-level findings, fix them before the smoke test —
they are exactly the problems an end user would hit first.
Step 4: End-User Smoke Test (MANDATORY)
Run the automated smoke test first for quick validation:
This checks CLI binary resolution, --help, --version, auth status, and --json
output for protocol leaks. Then proceed with manual verification below.
This is the most critical verification step. The agent MUST simulate what a real
end user would do after pip install cli-web-<app>. If this fails, the pipeline
is NOT complete -- go back and fix the issue.
If no-auth site: Skip items 1-2 (auth). Go directly to item 3 (READ).
If read-only site: Skip item 4 (WRITE). Verify reads return real data.
1. Authenticate as an end user would:
cli-web-<app> auth login
This uses Python sync_playwright() -- opens a browser, user logs in,
cookies saved. This is what end users will run. If this fails, the CLI is
broken for end users.
2. Verify auth status shows LIVE VALIDATION OK:
cli-web-<app> auth status
Must show: cookies present, tokens valid. If it shows "expired", "redirect",
or any auth failure -- STOP. Fix auth before proceeding.
3. Run a READ operation and verify real data:
cli-web-<app> --json <first-resource> list
This must return real data from the live API -- NOT an error, NOT empty,
NOT "auth not configured". Verify the JSON response contains expected fields.
4. Run a WRITE operation and verify it actually worked:
This is the step the agent most commonly skips. Reading data is easy -- the
real test is whether the CLI can CREATE, UPDATE, or GENERATE something.
# For CRUD apps (Monday, Notion, Jira):
cli-web-<app> --json <resource> create --name "smoke-test-$(date +%s)"
cli-web-<app> --json <resource> list # verify the created item appears
cli-web-<app> --json <resource> delete --id <id-from-create>
# For generation apps (Suno, Midjourney, NotebookLM audio):
cli-web-<app> --json <resource> generate --prompt "test" --wait# Verify: JSON response contains a real ID, status=complete, not an error# If the command has --output, verify the file was downloaded and size > 0# For search/query apps:
cli-web-<app> --json search "test query"# Verify: results array is non-empty
If ANY write/generate command fails, the pipeline is NOT complete.
Reading a list of existing items only proves auth works -- it does NOT prove
the CLI can actually do useful work. The whole point is to CREATE things,
not just read them.
5. Only after items 1-4 ALL pass, continue to Step 5.
Smoke Test Checklist
auth login works (Python playwright, API key, or N/A for no-auth)
auth status shows valid (or N/A for no-auth)
At least one READ returns real data
At least one WRITE/CREATE/GENERATE succeeds (or N/A for read-only)
The CLI works standalone -- no debug Chrome, no port 9222, no MCP
Output sanity: no raw protocol data leaks in --json output (see below)
Output Sanity
Run every command with --json and check for raw protocol leaks (wrb.fr, af.httprm,
empty [], null required fields). Full red-flags table:
skills/shared/CONVENTIONS.md §Protocol-Leak Smoke Check.
#1 gap to watch for: Agent runs list (GET with auth — easy), declares done, but
never tests create/generate (POST with CSRF, encoding). Always test at least one write.
Step 5: Generate the Per-CLI Skill
See "Generate Claude Skill" below. The skill must pass the quality gate
before Step 6.
Step 6: Register the CLI (registry-driven — most repo docs are generated)
registry.json is the source of truth: the CI test matrix, the README fleet
table, and the install block are all generated from it. Add the registry
entry, then run the generators — never edit those outputs by hand:
# 1. Add the entry to registry.json (schema below), then:
cli-web-devkit registry validate # entry <-> fleet cross-check
cli-web-devkit docs # regenerates README regions + registry site data
cli-web-devkit about --apply # syncs the GitHub "About" CLI count (needs gh admin)
cli-web-devkit resync --app <app> # vendored files in sync + manifest updated
cli-web-devkit drift # must report 0 drifted/missing# 2. Offline fleet contract for the new CLI (help/version/REPL/MCP/doctor):
python -m pytest tests/contract -q -k <app>
Remaining hand-edited files (independent — update in parallel if you like):
CHANGELOG.md — entry under [Unreleased] -> Added
CLAUDE.md — row in the Generated CLIs table
README.md hero badge counts (outside the generated markers)
cli_web/<app>/README.md — fill in the scaffolded skeleton
Start from the scaffolded skeletons.scaffold-cli.py (v2) already
rendered README.md and the per-CLI SKILL.md skeletons from
templates/README.md.tpl and templates/SKILL.md.tpl during Phase 2 —
fill in the remaining placeholders with actual CLI data from <app> --help
and <APP>.md rather than writing from scratch.
Generate Claude Skill
Goal: Create a project-local Claude skill so that Claude can use this CLI
automatically in future conversations — no manual lookup required.
IMPORTANT: The skill must exist in TWO locations:
.claude/skills/<app>-cli/SKILL.md — for Claude Code discovery (project-level)
<app>/agent-harness/cli_web/<app>/skills/SKILL.md — portable with pip install
(included via package_data in setup.py)
Create the skill once, then copy it to both locations.
Read references/skill-authoring.md first — it defines the frontmatter
rules, description format, body limits, and the standard section structure.
The skeleton rendered from templates/SKILL.md.tpl during Phase 2 already
follows it; fill in the FILL_IN markers.
Run cli-web-<app> --help and each group's --help — every command you
document must be verified against the real surface (a stale example is
worse than no example).
Validate before publishing: the skill must pass
python -m pytest ${CLAUDE_PLUGIN_ROOT}/scripts/tests/test_skill_quality.py
(frontmatter fields, description ≤1024 chars third-person, body ≤500 lines,
reference links resolve).
Update Repository README
Add presentation metadata to registry.json, then run cli-web-devkit docs to
regenerate the README table and registry site. Add a hand-written quick-start
example in the README's "Try Them" section only when it adds distinct value.
Update registry.json and CLAUDE.md
Add the new CLI to registry.json at the repo root:
{"name":"cli-web-<app>","website":"<website>","protocol":"<detected protocol>","auth":"<auth type>","directory":"<app>/agent-harness","namespace":"cli_web.<app>","commands":["<cmd1>","<cmd2>", ...],"install":"pip install cli-web-<app>","description":"<one-line public description>","site_icon":"<short label>","site_category":"<display category>","site_tags":["<filter-tag>"]}
Also add to the Generated CLIs table in CLAUDE.md.
Pipeline Complete
The pipeline is NOT done until ALL of these are checked:
Smoke Tests
Auth works (login + status, or N/A for no-auth)
At least one READ returns real data
At least one WRITE succeeds (or N/A for read-only)
cli_web/<app>/skills/SKILL.md exists (portable with pip install)
Based on the scaffolded skeleton from templates/SKILL.md.tpl
Package
setup.py has package_data={"": ["skills/*.md", "*.md"]}
__main__.py exists for python -m cli_web.<app> support
Documentation
cli_web/<app>/README.md exists (filled in from the templates/README.md.tpl skeleton)
<APP>.md API map exists
tests/TEST.md has Part 1 (plan) + Part 2 (results)
Repo-Level Updates
registry.json — entry with name, website, protocol, auth, commands,
install, description, skill path (+ canary read-only invocations for
no-auth CLIs). The CI matrix, README table, and registry site derive from
this entry.
cli-web-devkit registry validate + docs + drift all green
CHANGELOG.md — entry added under [Unreleased] → Added
CLAUDE.md — new row in Generated CLIs table
README.md hero badge counts updated (outside the generated markers)
Branch protection: if required status checks are pinned by job name, add
the new matrix job names — the matrix itself updates automatically from
registry.json (never edit tests.yml)
All key rules (naming, auth, --json, REPL, rate limits) are defined in
skills/shared/CONVENTIONS.md — HARNESS.md and CLAUDE.md only index them.
When every box above is checked, mark the pipeline finished:
