Playwright CLI

Use playwright-cli directly (already installed) to execute browser automation from natural-language requests.

Quick Start

1. Verify the CLI is available:

   command -v playwright-cli >/dev/null 2>&1

1. Open a browser session:

   playwright-cli open https://example.com

1. Capture refs before ref-based commands:

   playwright-cli snapshot

Natural-Language Execution Workflow

1. Parse the request into atomic actions.
2. Convert each action to command templates from references/command-map.md.
3. Run playwright-cli snapshot before commands that require element refs (click, fill, check, select, hover, drag, etc.) unless a fresh snapshot is already available.
4. Re-run playwright-cli snapshot after navigation or major DOM changes.
5. For assertions, prefer playwright-cli eval and parse the ### Result value from stdout.
6. Treat ### Error output as command failure, even when process exit code is 0.
7. For screenshots, capture PNG first, then compress to JPEG with ffmpeg before final save/return.
8. Use explicit filenames when saving artifacts (screenshots, snapshots, PDFs, videos, storage state).
9. Return executed commands and key results (URLs, counts, file paths, errors).

Command Mapping Rules

• Map intent verbs directly:
• open/launch -> open
• navigate -> goto
• click/tap -> click
• double click/double tap -> dblclick
• type/input -> type
• fill/set field -> fill
• back/forward/refresh -> go-back/go-forward/reload
• Preserve user-provided optional flags and values.
• Omit optional args when unspecified.
• Use -s=<name> when the user asks for a named/session-isolated browser.
• Prefer native commands over eval and run-code; use eval or run-code only if no first-class command exists.
• Keep session names short and filesystem-safe (lowercase letters + numbers, no punctuation, target <= 12 chars) to avoid local socket path errors on macOS.

Reliability Guardrails (Must Follow)

• Do not classify pass/fail from shell exit codes for eval or run-code.
• Always inspect command output:
• If output contains ### Error, treat as failure.
• If output contains ### Result, use that value for assertions.
• Prefer eval for boolean checks and DOM assertions.
• Use run-code for imperative Playwright flows only.
• run-code expects a valid JS expression; statements like bare return ... or throw ... are invalid.
• For JS snippets passed to eval, prefer expression-safe patterns:
• Good: Array.from(...).some(function(x){ ... })
• Risky in this CLI context: snippets that are parsed as non-expression wrappers.
• When user requests pass and fail screenshots:
• Capture one screenshot per scenario after assertion.
• Capture to temporary PNG, then convert with ffmpeg to JPEG as the final artifact.
• Name final files with scenario status prefix, e.g. pass-<name>.jpg, fail-<name>.jpg.
• Save a machine-readable index file (results.txt or JSON) mapping scenario -> status -> final JPEG artifact path.

Screenshot Compression Rule (Required)

• Always return compressed JPEG screenshots, not raw PNG screenshots.
• Required flow:

1. Capture screenshot to temporary PNG path.
2. Compress to JPEG with ffmpeg.
3. Return/report only the JPEG path.

• Default command pattern:

  playwright-cli screenshot --full-page --filename "$TMP_PNG"
  ffmpeg -y -i "$TMP_PNG" -q:v 4 "$FINAL_JPG"

• If ffmpeg is unavailable, stop and report the missing dependency instead of returning an uncompressed PNG.

Session and Artifact Conventions

• Use default session unless the user requests named sessions.
• Keep multi-step flows in one session for continuity.
• Place generated artifacts in predictable paths when the user does not specify:
• output/playwright/*.jpg
• output/playwright/*.pdf
• output/playwright/*.json
• output/playwright/*.zip

Full Command Coverage

Load references/command-map.md when mapping or validating command usage. It contains the complete command catalog for:

• Core
• Navigation
• Keyboard
• Mouse
• Save as
• Tabs
• Storage (state, cookies, localStorage, sessionStorage)
• Network
• DevTools
• Install
• Configuration
• Sessions