/cli-anything-web

name: cli-anything-web
description: Generate a complete agent-native CLI for any web app by recording and analyzing network traffic via playwright-cli. Runs the full pipeline with site assessment, capture, implementation, testing, and verification.
argument-hint: <url> [--mitmproxy]
allowed-tools: Bash(*), Read, Write, Edit, mcp__chrome-devtools__*

CLI-Anything-Web: Full Pipeline

Read the methodology overview: @${CLAUDE_PLUGIN_ROOT}/HARNESS.md

Target URL and flags: $ARGUMENTS

Capture Mode Detection

Check if `--mitmproxy` flag is present in the arguments:

• **With `--mitmproxy`**: Use mitmproxy-capture.py for traffic capture (no body truncation, real-time filtering, enhanced analysis). Requires `pip install mitmproxy` (Python 3.12+).
• **Without `--mitmproxy`** (default): Use standard playwright-cli tracing (the original method).

Prerequisites Check

Step 1: Check playwright-cli availability

!`npx @playwright/cli@latest --version 2>&1 && echo "PLAYWRIGHT_OK" || echo "PLAYWRIGHT_FAIL"`

**If PLAYWRIGHT_OK** → use playwright-cli for all operations (primary path).

**If PLAYWRIGHT_FAIL** → If playwright-cli is not available, see HARNESS.md for the MCP fallback path.

Step 1b: If `--mitmproxy` flag was passed, also verify mitmproxy

!`python -c "import mitmproxy; print('MITMPROXY_OK')" 2>&1 || echo "MITMPROXY_FAIL"`

**If MITMPROXY_FAIL** → Tell user to `pip install mitmproxy` or drop the `--mitmproxy` flag.

NEVER use `mcp__claude-in-chrome__*` tools -- blocked, cannot capture request bodies.

Execution Plan

Run the full pipeline by invoking skills in sequence. **Each phase checks for prior completion and skips if already done.**

1. Check playwright-cli availability (see Prerequisites above) 2. **Check pipeline state:** `python ${CLAUDE_PLUGIN_ROOT}/scripts/phase-state.py status <app>`

• If a phase is already `done` → skip it and proceed to the next
• If a phase `failed` with `retryable` → retry automatically (once)
• If a phase `failed` with `fatal` → report and ask user
• Full decision tree (including corrupt/inconsistent state):

`${CLAUDE_PLUGIN_ROOT}/skills/shared/RECOVERY.md` §phase-state Check Failures 3. For each phase, check before running:

   python ${CLAUDE_PLUGIN_ROOT}/scripts/phase-state.py check <app> --phase <phase>
   # Exit 0 = skip (already done), Exit 1 = run

4. Invoke `capture` skill -- Phase 1 site assessment + traffic recording

• Also check capture checkpoint: `python ${CLAUDE_PLUGIN_ROOT}/scripts/capture-checkpoint.py restore <app>`

5. Invoke `methodology` skill -- Phase 2 analyze/design/implement

• Agent MUST read a reference CLI first (same protocol) — see methodology skill

6. Invoke `testing` skill -- Phase 3 test writing/documentation 7. Invoke `standards` skill -- Phase 4 publish and verify + generate Claude skill

After each phase completes, mark it:

python ${CLAUDE_PLUGIN_ROOT}/scripts/phase-state.py complete <app> --phase <phase> --output <output-path>

Each skill handles its phases completely and invokes the next when done. See HARNESS.md for the pipeline overview and critical rules.

Extract the app name from the URL (e.g., `monday.com` → `monday`, `notion.so` → `notion`).

Success Criteria

The command succeeds when: 1. All core modules are implemented and functional (`client.py`, `auth.py`, `session.py`, `models.py`) 2. CLI supports both one-shot commands and REPL mode 3. `--json` output mode works for all commands 4. All tests pass (100% pass rate) 5. Subprocess tests use `_resolve_cli()` and pass with `CLI_WEB_FORCE_INSTALLED=1` 6. TEST.md contains both plan (Part 1) and results (Part 2) 7. README.md documents installation and usage 8. `setup.py` is created and local installation works 9. CLI is available in PATH as `cli-web-<app>` 10. `.claude/skills/<app>-cli/SKILL.md` exists at the project root

Output Structure

<app-name>/
└── agent-harness/
    ├── <APP>.md               # API map, data model, auth scheme
    ├── setup.py               # PyPI config (find_namespace_packages)
    └── cli_web/               # Namespace package (NO __init__.py)
        └── <app>/             # Sub-package
            ├── __init__.py    # Required — marks as sub-package
            ├── README.md
            ├── <app>_cli.py   # Main CLI entry point
            ├── __main__.py
            ├── core/
            │   ├── client.py
            │   ├── auth.py
            │   ├── session.py
            │   └── models.py
            ├── utils/
            │   ├── repl_skin.py
            │   ├── output.py
            │   └── config.py
            └── tests/
                ├── TEST.md
                ├── test_core.py
                └── test_e2e.py

**No-auth sites:** Remove `core/auth.py`, `core/session.py`, and auth commands from the structure above. Only create what the CLI actually needs.

Progress Tracking

After each phase, report status in this format:

┌─────────┬────────┬────────────────────────────────────────────┐
│ Phase   │ Status │ Description                                │
├─────────┼────────┼────────────────────────────────────────────┤
│ Phase 1 │ ...    │ Capture — Traffic Capture + Auth           │
│ Phase 2 │ ...    │ Methodology — Analyze + Design + Implement │
│ Phase 3 │ ...    │ Test — Write Tests + Document Results      │
│ Phase 4 │ ...    │ Standards — Publish + Smoke Test + Skill   │
└─────────┴────────┴────────────────────────────────────────────┘