test-ocas-openclaw

name: test-ocas-openclaw description: "Regression test the OpenClaw Codex App Server plugin against a live local OpenClaw instance in Telegram or Discord. Use when the user wants an end-to-end OCAS manual test pass, wants to verify binding, /cas_resume or /cas_status flags for model/fast/yolo, approvals, plan mode, review, compact, status-card controls, split default-vs-full-access permission behavior, stop-button interrupt handling, rename, or other slash-command behavior, or wants reproducible bug notes from a live chat integration run."

Test OCAS OpenClaw

Use this skill for manual regression passes of this plugin against a real local OpenClaw instance.

Preconditions

• Confirm the plugin branch under test is linked into the local OpenClaw checkout and OpenClaw is running.
• Prefer Playwright MCP or Chrome MCP for Telegram Web. It was more reliable than agent-browser in recent passes.
• Use $agent-browser only as fallback. Read references/agent-browser.md only if you intentionally choose that path.
• If Telegram Web or Discord is not logged in, stop and ask the user to complete login.
• Prefer a low-risk Codex thread unless the user asks otherwise. In this repo, discrawl and dupcanon are safe defaults for resume tests.
• Treat Permissions: toggle as switching between Default and Full Access. If Full Access is unavailable in the current Codex Desktop session, expect the status card to stay in Default mode with an explanatory note.

Canonical State Files

Use these when the UI and the plugin disagree:

• Core chat binding store: ~/.openclaw/telegram/thread-bindings-default.json
• Plugin state: ~/.openclaw/openclaw-codex-app-server/state.json

Read them when:

• a topic looks bound but /cas_status disagrees
• plan mode appears stuck
• a questionnaire should exist but Telegram Web does not render it
• slash replies appear in the wrong chat or topic
• answer buttons appear but the active pending request does not advance
• the status card claims a pending permissions migration that never applies after the run finishes

Treat the state files as diagnostic evidence, not as a thing to hand-edit.

Preferred Browser Path

• Keep the browser anchored to the requested topic or channel for the full pass.
• For Telegram topics, keep the topic fragment in the URL when possible.
• If the UI looks stale but local state shows the expected pending request or callbacks, refresh Telegram Web with Cmd+R before calling it a backend bug.
• Separate browser-navigation failures from OCAS failures. A browser can drift to General or the chat list without proving an OCAS routing bug.

Regression Flow

1. Establish the target conversation.

• Stay in the requested group, DM, channel, or topic for the entire pass.
• If the expected Telegram topic is missing, check whether /cas_resume --sync renamed it. Rename it back or create a fresh test topic if needed.
• If slash-command routing leaks into General, log that separately and move the browser back to the intended topic before continuing.

2. Baseline the conversation.

• Run /cas_status.
• If the topic is already bound, run /cas_detach first, then confirm the unbound state with /cas_status.
• For Telegram topics, the reply must appear in the same topic. If it lands in General, record a routing bug.

3. Resume a thread.

• Prefer /cas_resume or /cas_resume --all with the normal picker flow.
• Avoid /cas_resume --all <thread-id> for now. It is a known regression path and may not surface the expected approval UX.
• Cover at least one flag-driven bind path with /cas_resume --model <name>, /cas_resume --fast or /cas_resume --no-fast, and /cas_resume --yolo or /cas_resume --no-yolo.
• If the bind path falls back into project or thread pickers, verify the requested model, fast, and yolo settings still apply after the picker choice and final bind approval.
• When approval appears, use Allow Once unless the user asks otherwise.
• After binding, send who are you.
• Expect a Codex response. If the reply sounds like OpenClaw persona text instead of Codex, the bind failed.
• Run /cas_status again and verify:
  • Binding: active
  • the expected thread id or thread title
  • the expected model
  • Plugin version: 0.0.0 for local dev branch testing when that is the expected version

4. Verify status-card controls.

• Use the interactive /cas_status message as the primary control surface.
• Expect one status message that edits in place rather than a stream of follow-up bot messages.
• Verify the card includes:
  • Select Model
  • Reasoning: ...
  • Permissions: toggle
  • Compact
  • Stop
• Verify Fast: toggle only appears when the current model supports fast mode. If the current model does not support fast mode, expect no fast button.
• Click Select Model.
• Expect the same message to switch into a model picker.
• Choose a different model.
• Expect the same message to return to the status card and show the new model on the Model: line.
• Run /cas_status again and confirm the new model is still shown.
• Click the reasoning button.
• Expect the same message to switch into a reasoning picker.
• Choose a different reasoning level.
• Expect the same message to return to the status card and show the new reasoning on the Model: line.
• Run /cas_status again and confirm the reasoning state persisted independently of the current model.
• If the current model supports fast mode:
  • click Fast: toggle
  • expect the same message to update Fast mode: immediately
  • run /cas_status again and confirm the fast-mode state persisted
• If the current model does not support fast mode:
  • use Select Model to switch to a fast-capable model before testing the fast toggle
• Click Permissions: toggle.
• If Full Access is available and no turn is active:
  • expect the same message to update Permissions: between Default and Full Access
  • run /cas_status again and confirm the new permissions mode persisted
• If a turn is active:
  • expect the same message to stay on the current mode and add a note that the requested permissions mode will apply after the current turn ends
  • after the run finishes, run /cas_status again and confirm the pending profile migration actually applied
• If Full Access is unavailable in the current session:
  • expect the status message to remain on Permissions: Default
  • expect an explanatory note that Full Access is unavailable for the current session

5. Verify approval rendering in Default mode.

• Put the conversation back into Permissions: Default before this step.
• Send this exact prompt as plain text in the bound conversation:

I want you to run `npm view dive` and make sure to ask to exit the sandbox as it needs network access.

• Expect a real execution approval dialog, not a plain text question about whether the user wants approval.
• Expect approval buttons such as Approve or the platform-equivalent action controls.
• Expect a code-formatted command area that shows npm view dive.
• Verify the displayed command is trimmed for presentation and does not leak a shell-launcher wrapper like /bin/zsh -lc ....
• After verifying the dialog, approve the command with Approve Once or the platform-equivalent approval button so later tests are not blocked by a stale pending approval.
• If the model only asks a conversational question like Do you want to allow... without rendering execution approval controls, treat that test as invalid and rerun with the exact prompt above.

6. Verify full-access execution path when enabled.

• Use the status card to switch into Permissions: Full Access when it is available in the current session.
• Send this exact prompt as plain text:

Run `npm view dive version` and reply with only the version string.

• Expect the command to run without an approval dialog.
• Expect a direct result after a short delay rather than an approval request.
• After the command completes, switch the status card back to Permissions: Default.
• Run /cas_status again and confirm the card shows Default mode.

7. Verify long-running review behavior before plan mode.

• Run /cas_review before /cas_plan.
• Treat /cas_review as a long-running request. It may take several minutes.
• Do not interrupt /cas_review by starting /cas_plan, /cas_compact, or other long-running flows while review is active.
• After starting review, wait 30 seconds at a time and check for output again.
• Record whether review immediately skips expected desktop-style base and branch questions. That is a behavior gap, but not the same thing as a review failure.

8. Verify Stop-button behavior on a long-running run.

• Start a long-running Codex action such as /cas_review, /cas_plan, or a prompt likely to wait on approvals or take more than a few seconds.
• While the run is active, press Stop on the status card.
• Expect the active run to interrupt.
• Expect the same status message to refresh in place after the interrupt.
• Run /cas_status again and confirm there is no active-run note left behind.
• If you also want to verify deferred permission migration, start a long-running run, press Permissions: toggle during the run, confirm the pending note, then press Stop and confirm the requested mode applies only after the run ends.

9. Verify plan mode only after review has finished.

• Use /cas_plan <prompt>. Do not test plan mode by sending the raw prompt without the command.
• Use a short questionnaire prompt first. A breakfast-choice prompt is good because it exercises question rendering and answer submission.
• Expect:
  • questionnaire text
  • answer buttons
  • question progression when answering
  • final Implement this plan controls
• If question 1 renders but button clicks or plain-text answers do not advance the active pending request in state.json, record that as a real failure.
• If plan mode gets stuck, /cas_plan off should exit it.

10. Verify long-plan truncation.

• Ask for a final plan longer than 4000 characters.
• Expect the Telegram preview to truncate and the plugin to attach the full Markdown plan when needed.

11. Cover other control commands as needed.

• /cas_resume
  • Verify --model, --fast or --no-fast, and --yolo or --no-yolo paths on direct bind, picker bind, and --new flows.
  • Expect the chosen settings to show up immediately in /cas_status after the bind completes.
• /cas_mcp
  • Expect a list of configured MCP servers.
• /cas_skills
  • Expect installed skills.
  • If the response shows both a full text list and buttons for the same skills, record a display bug rather than a blocker.
• /cas_status
  • Verify in-place message edits for model, reasoning, fast when supported, permissions, compact, and stop controls.
  • Verify the card does not post extra bot messages for each button press unless the platform requires it.
  • Verify /cas_status --model, /cas_status --fast or /cas_status --no-fast, and /cas_status --yolo or /cas_status --no-yolo update the bound conversation settings and refresh the shown status.
• /cas_compact
  • Expect progress keepalives and a final context-usage report.
• /cas_rename
  • Verify thread rename and, if requested, --sync topic rename behavior.
• /cas_diff, /cas_init
  • Verify current forward-or-placeholder behavior instead of assuming full implementation.
  • Treat known-placeholder behavior as neutral, not as a regression, unless the user says the command should already work.
• Do not look for /cas_model, /cas_fast, or /cas_permissions.
  • Those controls were folded into /cas_resume, /cas_status, and the status card.

12. Clean up.

• End by running /cas_detach.
• Confirm the conversation is no longer bound.
• Confirm the topic has no lingering pending request in state.json if plan mode, stop, or approvals were exercised.

Result Legend

• ✅ Pass. Behavior matched what should work now.
• ❌ Fail. Behavior should work now and did not.
• ➖ Neutral. The command is known to be unimplemented, placeholder-only, intentionally incomplete, or not configured in this environment.

Use a flat results table while testing:

Gotchas

• Telegram Web can go stale. If a simple request appears hung for more than about one minute, and the local state file already shows the pending questionnaire or callbacks, refresh Telegram Web with Cmd+R before calling it a backend bug.
• Plain text can still route correctly even when slash commands from a topic leak to General. Record both facts separately.
• A core binding can exist in thread-bindings-default.json while plugin-local state is stale or missing. Compare both files before concluding which side is wrong.
• Telegram can leave stale questionnaire buttons visible in the DOM. If a click produces No active Codex run is waiting for input. but pendingRequests still shows the active questionnaire unchanged, record that the visible button press did not reach the active request.
• Starting /cas_plan while /cas_review is still running can interrupt the review. Avoid that sequence during normal regression passes.
• Permissions migration is edge-triggered on turn completion. If you toggle permissions during an active run, do not expect the profile switch to apply until the run ends or is stopped.
• A full-access test only proves the split-profile path when the command runs without approval and later Default-mode commands still trigger approval again.

Evidence To Capture

• Exact command sent
• Exact chat or topic where the reply appeared
• Whether the reply was plain text, buttons, or an attachment
• Whether the status card edited in place or created a new bot message
• Which permissions mode was selected before the command under test
• Relevant snippets from state.json and thread-bindings-default.json when UI behavior is suspicious
• Known-good versus observed behavior
• The result-table row for each command or flow tested

Formal Bug Notes

If the user wants the findings turned into issues, PR notes, or project-board items, use $project-manager.