safe-bash Playwright: implement Cloudflare-supported standard CLI capabilities and output parity

[kamilio]

The current public Cloudflare consumer uses safe-bash 0.1.669. A fresh differential check against the official @playwright/cli@0.1.20 confirms that only 14 built-in commands are implemented by default. Most official commands exist only as catalog metadata or host hooks. The user now explicitly requires all Cloudflare-supported capabilities and standard help/behavior, superseding the earlier subset-only direction in #735.

Concrete mismatches:

• --help removes most original commands and appends internal Client capabilities, lifecycle and host-limit explanations. Several implemented forms are marked [limited].
• Global --json, --raw, and --version are not implemented.
• State, cookies, local/session storage, mouse actions, most element actions/navigation, eval/code, PDF, request inspection/routing and tracing are absent from the built-in implementation despite relevant Cloudflare browser APIs.
• Output/artifact behavior also diverges: the default screenshot emits raw PNG on stdout, while the standard CLI reports a saved artifact. Cosmetic help edits alone do not establish compatibility.

Required implementation and verification:

1. Inventory every command and option from the pinned official CLI; implement every capability the current Cloudflare provider can execute. Keep one standard parser/command interface, with differential help, result, error, artifact and session tests.
2. Implement standard state-save/state-load, cookies, localstorage and sessionstorage commands with exact replacement semantics. Provide host-side hooks needed for owner-scoped Workers KV persistence; provider IDs, credentials and user/agent ownership remain controlled by the consumer. IndexedDB behavior must match the selected standard options rather than being silently assumed.
3. Preserve bounded input/output, cancellation, owned-browser retirement and authenticated user-and-agent isolation. Updated user direction: remove custom URL/domain/scheme blocking, browser HTTP proxies, deny-all guardrails and WebSocket/service-worker bans. Use normal Cloudflare browser networking and standard Playwright request APIs; do not impose lifetime network or invocation quotas. Scripts must not access another owner's sessions/storage.
4. Expose required native browser capabilities through a coherent adapter. Do not require every consumer to reimplement the standard CLI as custom abilities.
5. Test actual Chromium and the actual Cloudflare adapter and published standalone package. Distinguish provider-unsupported desktop/test/video functionality from missing library implementation; do not advertise fabricated success for unsupported effects.

Root task ownership: this Codex task is handling CLI parity/capabilities and the Poe consumer's KV integration in isolated worktrees. Existing #746/#763 work remains with its current owners. Browser snapshot fix #762 is being qualified from its public release separately.

The next comment contains the exact official global help output captured from @playwright/cli@0.1.20; the current source diff and command-by-command audit are being retained with consumer evidence.

[kamilio]

Official @playwright/cli@0.1.20 stdout from playwright-cli --help (exit 0):

playwright-cli - run playwright mcp commands from terminal

Usage: playwright-cli <command> [args] [options]
Usage: playwright-cli -s=<session> <command> [args] [options]

Core:
  open [url]                  open the browser
  attach [name]               attach to a running playwright browser
  close                       close the browser
  detach                      detach from an attached browser
  goto <url>                  navigate to a url
  type <text>                 type text into editable element
  click <target> [button]     perform click on a web page
  dblclick <target> [button]  perform double click on a web page
  fill <target> <text>        fill text into editable element
  drag <startTarget> <endTarget> perform drag and drop between two elements
  drop <target>               drop files or data onto an element
  hover <target>              hover over element on page
  select <target> <val>       select an option in a dropdown
  upload <files...>           upload one or multiple files
  check <target>              check a checkbox or radio button
  uncheck <target>            uncheck a checkbox or radio button
  snapshot [target]           capture page snapshot to obtain element ref
  find [text]                 search the page snapshot for text or a regexp, returning matching nodes with surrounding context (like search snippets)
  eval <func> [target]        evaluate javascript expression on page or element
  dialog-accept [prompt]      accept a dialog
  dialog-dismiss              dismiss a dialog
  resize <w> <h>              resize the browser window
  delete-data                 delete session data

Navigation:
  go-back                     go back to the previous page
  go-forward                  go forward to the next page
  reload                      reload the current page

Keyboard:
  press <key>                 press a key on the keyboard, `a`, `arrowleft`
  keydown <key>               press a key down on the keyboard
  keyup <key>                 press a key up on the keyboard

Mouse:
  mousemove <x> <y>           move mouse to a given position
  mousedown [button]          press mouse down
  mouseup [button]            press mouse up
  mousewheel <dx> <dy>        scroll mouse wheel

Save as:
  screenshot [target]         screenshot of the current page or element
  pdf                         save page as pdf

Tabs:
  tab-list                    list all tabs
  tab-new [url]               create a new tab
  tab-close [index]           close a browser tab
  tab-select <index>          select a browser tab

Storage:
  state-load <filename>       loads browser storage (authentication) state from a file
  state-save [filename]       saves the current storage (authentication) state to a file
  cookie-list                 list all cookies (optionally filtered by domain/path)
  cookie-get <name>           get a specific cookie by name
  cookie-set <name> <value>   set a cookie with optional flags
  cookie-delete <name>        delete a specific cookie
  cookie-clear                clear all cookies
  localstorage-list           list all localstorage key-value pairs
  localstorage-get <key>      get a localstorage item by key
  localstorage-set <key> <value> set a localstorage item
  localstorage-delete <key>   delete a localstorage item
  localstorage-clear          clear all localstorage
  sessionstorage-list         list all sessionstorage key-value pairs
  sessionstorage-get <key>    get a sessionstorage item by key
  sessionstorage-set <key> <value> set a sessionstorage item
  sessionstorage-delete <key> delete a sessionstorage item
  sessionstorage-clear        clear all sessionstorage

Network:
  requests                    list all network requests since loading the page. each request is numbered for use with the `request` command.
  request <index>             show full details (headers, body, response) of a single network request by its number from the `requests` command.
  request-headers <index>     print only the request headers for a single network request by its number from the `requests` command.
  request-body <index>        print only the request body for a single network request by its number from the `requests` command.
  response-headers <index>    print only the response headers for a single network request by its number from the `requests` command.
  response-body <index>       print the response body for a single network request by its number from the `requests` command. textual bodies are inlined; binary bodies are saved to a file and the path is printed.
  route <pattern>             mock network requests matching a url pattern
  route-list                  list all active network routes
  unroute [pattern]           remove routes matching a pattern (or all routes)
  network-state-set <state>   set the browser network state to online or offline

DevTools:
  console [min-level]         list console messages
  run-code [code]             run playwright code snippet
  recording-start             start recording user actions
  recording-stop              stop recording user actions and print them as playwright code
  tracing-start               start trace recording
  tracing-stop                stop trace recording
  video-start [filename]      start video recording
  video-stop                  stop video recording
  video-chapter <title>       add a chapter marker to the video recording
  video-show-actions          annotate subsequent cli/mcp actions on the page with a callout that names the action and highlights the target element
  video-hide-actions          stop annotating actions performed on the page
  show                        show playwright dashboard
  pause-at <location>         run the test up to a specific location and pause there
  resume                      resume the test execution
  step-over                   step over the next call in the test
  generate-locator <target>   generate a playwright locator for the given element
  highlight [target]          show (or with --hide, remove) a highlight overlay for an element; `--hide` without a target hides all page highlights.

WebMCP:
  webmcp-list                 list the webmcp tools registered by the page
  webmcp-call <name>          call a webmcp tool registered by the page

Install:
  install                     initialize workspace
  install-browser [browser]   install browser

Browser sessions:
  list                        list browser sessions
  close-all                   close all browser sessions
  kill-all                    forcefully kill all browser sessions (for stale/zombie processes)

Global options:
  --help [command]            print help
  --json                      output response as JSON
  --raw                       output only the result value, without status and code
  --version                   print version

[kamilio]

Verification coordination: current source and published @poe-platform/safe-bash@0.1.676 audit against @playwright/cli@0.1.20 confirms 92 catalog commands, 14 default built-ins, and 76 command-option bindings. No duplicate browser implementation edits are being made.

For the active implementation described below, please hand off the owning task, concrete commit SHA (or explicitly uncommitted status), reserved file/write scope, remaining work and commit ETA.

Important real-Chromium oracle result for that owner: default state-save excludes IndexedDB and has no --indexed-db flag (that flag exits 1). state-load removes stale cookies/localStorage and clears IndexedDB omitted from the saved state, but preserves existing sessionStorage entries. Replacing the whole context must not silently lose that behavior. Standard traces produce .trace/.network/resources artifacts, not a default ZIP. Source and published CLI native checks, plus the actual Cloudflare adapter cancellation fixture in local workerd/Chromium, have run; this is not deployed Cloudflare/Poe consumer qualification.

[kamilio]

Implementation is active in the dedicated poe-code worktree. The first milestone adds shared standard CLI output/help/global flags, real browser actions and storage commands, context replacement for state-load, and host persistence callbacks. Native Chromium has passed the new state/action/eval/PDF checks. Additional event/network/tracing/runtime capabilities remain in progress; this is not a full-parity claim.

Release 0.1.676 is publicly available and its publication verification succeeded (run 35339413950). A direct Cloudflare before/after fixture confirms #762: JavaScript runs in both versions, and 0.1.676 excludes hidden noscript/script/style text from snapshots.

The separate Poe page rendering failure was traced to the poe2 consumer, not to JavaScript being disabled or a defect in the library interceptor: the shared URL guard rejected Poe's public static CDN, and the 16-request pending queue dropped additional assets. Narrow browser-only asset admission and a bounded 64-request queue now load 27 static CDN assets with HTTP 200. On Cloudflare the page hydrates, the logo loads, and entering an email enables Go (no submission). Generic web_fetch and internal host restrictions remain unchanged. Those consumer fixes will be included in the consumer PR.

The owner-scoped KV profile store has passed Workers runtime checks for cold recovery, KV propagation misses, owner isolation, expiry, deletion, concurrent saves, and cancellation. Controller integration and complete cloud qualification remain outstanding. Existing #746/#763 stay with their current resolver.

[kamilio]

Cloud qualification update: an owned authenticated test Worker on the Poe Dev account now passes native JavaScript, hierarchical snapshots, cookies/localStorage, state save/load, console capture, actual trace ZIP creation, and native action recording. Recorder output contains the actual button click as Playwright code.

Owner-scoped KV profiles passed graceful controller replacement and a forced Durable Object abort: the replacement object had a different instance ID and restored the saved storage. Separate owner DOs did not see each other's cookies. Explicit close prevented implicit reopening; explicit open retained saved cookies; delete-data cleared the profile. The probe cleaned both owners afterward.

The cloud test exposed and reproduced a real disposal race: final storage checkpoint and browser release ran concurrently. The new regression holds the checkpoint pending and verifies that browser release waits for it.

Network mocks now run within the mandatory host URL policy, with context identity taken from CDP. Matching a route does not admit an internal/protected URL, and mock responses use the normal transfer/byte accounting path.

This is evidence against the local unreleased implementation, not a publication claim. Remaining work includes the complete supported options/capability matrix, isolated run-code transport, final integrated edge cases, review, and release. Production-account qualification is blocked before runtime by Cloudflare auth error 10000. The urgent Poe asset/queue repair is independently ready in poe-internal/poe2#14903 and CI is being addressed.

[kamilio]

Further real Cloudflare qualification passed: IndexedDB survives controller restart; native prompt acceptance and binary file upload work; PDF and PNG commands produce valid files. Viewport, locale, color scheme, and cookies now survive restoration together. Cold session listing exposed a missing persisted-session index; the fix has reproduce-first Workers coverage and does not allocate browsers to list saved sessions.

The isolated run-code prototype executes the actual pinned Playwright library in a Worker Loader guest with outbound network disabled. Native locators, page evaluation, cookies, new pages, context options, and context closure were verified against the parent browser. Revocation blocks delayed guest commands; implementation is now being integrated with owner limits and retirement on cancellation.

Provider limitation found by the expanded artifact probe: Cloudflare's pinned Playwright download saveAs, saveAsStream, and stream read a Worker-local filesystem path, while Chromium created the download in its remote filesystem. Standard CDP supplies a path but no generic file-reading stream handle, and the documented Browser Run API exposes no download retrieval operation. No replay or URL refetch workaround will be substituted for the original download. This needs provider transport support; the consumer will report the limitation explicitly instead of leaking an ENOENT.

Public safe-bash/safe-fs 0.1.679 is available and includes the separately owned staging fixes. #746 and #763 remain open for their remaining actual consumer/cloud acceptance. The expanded Playwright implementation in this issue is still uncommitted and unpublished; current tests do not establish a release or production deployment. Urgent Poe hydration PR: poe-internal/poe2#14903.

[kamilio]

Implementation coordination: no concrete candidate commit or write-scope handoff has arrived yet. The resolver is now implementing a bounded native milestone in its isolated issue checkout: standard global output/flags, navigation, keyboard/mouse, and supported target actions. Existing independent native acceptance/oracle fixtures are retained. This work does not rewrite network policy or Cloudflare transport. Please hand off any existing candidate SHA and reserved files so overlapping work can be integrated rather than duplicated. Full storage/network/runtime coverage and deployed qualification remain required; this is not a closure or parity claim.

[kamilio]

Coordination: the complete implementation is actively owned in /Users/kjopek/Workspace/poe-code-playwright-capabilities (base c98f143), covering packages/safe-bash/src/playwright/**, the Playwright CLI entrypoint, associated plugin/native tests, session contract, and integration-input registration. Please do not start another implementation of those files; the current patch is uncommitted while final native/config regressions are being closed. I will post the delivery SHA and release proof here.

The current focused core suite passes 310 tests; native transport checks cover cancellation, deadlines, output/page quotas, same-owner reuse, ordinary user exceptions, and emulation changes. Direct Cloudflare live tracing now passes: real .trace and .network files are published during recording, the trace grows after navigation, and stopping retains the final files/resources. The native screenshot resource basename includes @, which the old artifact validator rejected; this is fixed in producer and consumer.

Remaining active work is final standard configuration/default parity and a Cloudflare-only screen-emulation reset after a run-code interaction. A separate read-only consumer security/lifecycle review found no actionable high/medium findings. Expanded work is not yet published; public 0.1.680 is the separate mainline release. Urgent Poe asset-loading PR poe-internal/poe2#14903 has an updated review-fix commit; its preceding CI run passed.

[kamilio]

Cloudflare run-code qualification now passes the complete CLI flow: native locators and page mutation, repeated runs, ordinary script exceptions, new tabs, cookies, and mobile emulation. The repeated-connection regression was a Chromium inspector cache issue: guest disconnect clears actual device metrics but the retained inspector session still caches its earlier override. Repeating the same set operation is ineffective; clearing device metrics on that retained session immediately before restoring the native settings fixes it. Both viewport-only and exact Pixel-device cases passed through subsequent storage checkpoints/snapshots and a later check.

The root cause and passing evidence are preserved locally, and a repeated-run native regression is added. The guest's returned emulation state also has a separate 64 KiB strict validation boundary. Native code generation now uses the pinned provider's real language generators (JavaScript/TypeScript, Python, Java, C#), with exact action output tested against the official CLI.

Final checks are closing the remaining known advertised snapshot JSON and INI configuration differences before the source commit/release. No publication or production-deployment claim yet.

[kamilio]

Candidate implementation is committed locally as 7d0897d, rebased onto current main a57fe89. It is not pushed or published yet. Reserved scope remains the Playwright source, CLI entrypoint, associated tests/contracts and third-party attribution; please do not duplicate this implementation.

Validation: 324 focused tests, strict source types, maintained workspace build, guarded ESLint (0 errors), and the rebased runner checks pass. Direct Cloudflare JSON/INI configuration, native JSON accessibility trees, Python action output, custom test IDs, configured scripts and restart/cookie retention pass. The consumer also fixes guest-created page/context init scripts disappearing on disconnect; native regression covers duplicate registrations, ordinary errors, last-page closure and bounded metadata (13 native cases pass).

Final cloud acceptance exposed one remaining timeout interaction: a 1.5-second native action timeout currently bounds the entire isolated init-page module startup. Checking this against the official behavior before publication. The separate Poe page-loading PR poe-internal/poe2#14903 is fully green and both exact-head reviews are clean.

[kamilio]

Implementation delivered to remote main in 7d0897d, followed by the independently bounded code-execution timeout fix in 75d1fa3. Remote main was verified after push. The scoped safe-package release has been dispatched; publication is not yet verified.

Final combined Cloudflare checks pass: JSON/INI configuration, native JSON snapshots, Python code generation, custom test IDs, CJS init-page scripts, cookie/config restoration, and native timeout state in both directions. A script that changes defaults and then throws preserves them: a later isolated call reports the exact native 25 ms click timeout and 40 ms navigation timeout. Guest-created init scripts survive disconnect with duplicate multiplicity and context semantics preserved; 14 native runner tests pass.

Final guarded ESLint completes with 0 errors (4 unrelated existing warnings), strict source types and focused tests pass, and all 546 maintained runner checks pass after rebase. Remaining host/provider constraints are explicitly recorded in the capability matrix; this is not a claim that local dashboard/OS facilities or the provider’s missing download/video transport exist.

[kamilio]

Native browser follow-up: removing our network proxy confirms ordinary HTTP navigation, the Poe development login page, and the Cloudflare API are reachable through Cloudflare's native browser network. Public poe.com can still return its own Cloudflare challenge/403; JavaScript executes successfully, so that response is distinct from our former URL/CDN blocking.

One compatibility bug reproduced independently in native Chromium: a same-document SPA navigation during snapshot capture makes safe-bash throw Snapshot stale during capture, even when the native page.ariaSnapshot({ mode: 'ai' }) completes with a valid current tree.

Reproduction: serve <button>Ready</button> through a local Playwright route, open that HTTPS page with the controller, and call history.pushState(null, '', '#spa-settled') while the snapshot hook is active. The controller's framenavigated listener increments its snapshot epoch; native capture succeeds, but our post-capture epoch comparison rejects it. Direct native capture returns the Ready button tree. This matches the development-site navigation failure observed in the Cloudflare probe.

The original CLI0.1.20 calls the native snapshot directly. Proposed narrow correction: preserve stale-reference rejection and retry only a generation-invalidated, read-only whole-page capture once within the original deadline. Do not replay navigation/actions, and do not reuse an explicit subtree root across a navigation. Source correction is being coordinated before the next release; this comment does not claim it is published.

[kamilio]

Tracked the separate native binary-upload hang upstream: cloudflare/playwright#229.

A standalone @cloudflare/playwright 1.3.6 + Miniflare 4.20260708.1 reproduction sends all 6 MiB of zero bytes to a local HTTP server, but page.evaluate() stays pending beyond 20 seconds. A 1 MiB binary control and 33 × 2 MiB printable uploads complete. This is established only in local Miniflare Browser Rendering; it is not a proven hosted Cloudflare failure.

The public report contains only the standalone SDK/local HTTP reproduction and version/byte-count evidence. The separate safe-bash lifetime diagnostic-retention failure on larger cumulative traffic remains tracked here and is being fixed with bounded rolling history.