domain-skills: manus + perplexity — task workflows

[muqsitnawaz]

Adds two new domain skills for long-running agent products.

manus/tasks.md

• URL patterns: /app/<taskId> for a run, /share/<taskId> for a public link (same 22-char id).
• Private API: Connect RPC at api.manus.im, pattern /<package>.v<N>.<Service>/<Method>. Lists the services + methods observed on the wire (SessionService/{ListSessions,GetSession,ShareSession,UpdateReadPosition}, etc.).
• Composer: the input is TipTap/ProseMirror (div.ProseMirror[contenteditable="true"]) — .value setting does nothing, placeholder lives on the inner <p> via data-placeholder.
• Submission: coordinate-locate + type_text() + Enter / send-button click; the submit button is disabled while the editor is empty.
• Task state: only durable DOM marker for completion is the literal text Task completed; response body lives in a Shiki-rendered markdown block.
• Share: clicking Share fires SessionService/ShareSession, popover uses Floating UI portal. The Copy-link button writes via document.execCommand('copy') rather than navigator.clipboard.writeText — the share URL is https://manus.im/share/<taskId>, derivable from the path.

perplexity/computer.md

• URL patterns: /computer/tasks dashboard, /computer/tasks/<slug>-<id> detail, same URL + ?view=thread for public share (no separate /share path).
• Wire format note: __NEXT_DATA__ is null on detail pages and there are no visible api.perplexity.ai fetches on page load — task state hydrates via streaming RSC. wss://suggest.perplexity.ai/suggest/ws is typeahead only, not the task stream.
• Composer: Lexical ([data-ask-input-container] + [data-lexical-editor="true"]) — distinct from Manus. Same contenteditable caveats.
• Dashboard: CSS-grid table with proper ARIA (div[role="row"] / div[role="cell"]), rows are clickable but not <a>.
• Detail page: tool invocations use group/tool-wrapper, the Todo panel is a floating popover with plan items, LM-rendered markdown blocks carry data-renderer="lm".
• Artifacts: persistent URLs on pplx-res.cloudinary.com, presigned (expiring) URLs on CloudFront — archive via Cloudinary when possible.
• Share: privacy levels carry stable data-testid="access-level-{private,specific-people,public}" hooks; Copy Link button uses navigator.clipboard.writeText (not execCommand).

Process notes

• Both skills were built by driving the live sites through browser-harness end-to-end (Manus task submitted and completed + shared; Perplexity Computer dashboard, an existing completed task, and the share popover were explored).
• No raw pixel coordinates in code examples (per SKILL.md) — selectors compute bounding rects at runtime.
• No user-specific ids, session cookies, or presigned tokens are included.

---

Summary by cubic

Adds two domain skills for Manus and Perplexity Computer to run long tasks end-to-end with durable, coordinate-free selectors and stable URL/ID patterns. Improves robustness for editor mounting, ID parsing (including trailing slashes), and completion detection.

• New Features

  • Manus: api.manus.im Connect RPC map; TipTap composer selectors; /app/<taskId> and /share/<taskId>; completion via “Task completed”; Share popover copies via document.execCommand('copy').
  • Perplexity Computer: /computer/tasks dashboard and /computer/tasks/<slug>-<id> detail; Lexical composer; ARIA role="row"/"cell" task list; tool steps via group/tool-wrapper; LM blocks data-renderer="lm"; artifacts on Cloudinary vs presigned CloudFront; share is ?view=thread using navigator.clipboard.writeText with data-testid privacy controls.

• Refinements

  • Connect transport: request JSON is wrapped in Uint8Array, but on-the-wire content is JSON.
  • Composer handling: no pixel clicks; compute rects, and add a bounded retry for late TipTap mount on Manus.
  • Perplexity IDs: extract the 22-char id from the URL pathname (trailing-slash safe) to avoid ?view=thread/hash corruption.
  • Todo checks: scope to the Radix popover via aria-controls; all_todo_done() returns None when closed/hydrating and True only when all icons are #pplx-icon-check.
  • Removed real task IDs/slugs and tightened SKILL.md compliance.

^{Written for commit 106075e. Summary will update on new commits.}

[browser-harness-review]

✅ Skill review passed

Reviewed 2 file(s) — no findings.

[cubic-dev-ai]

2 issues found across 2 files

Prompt for AI agents (unresolved issues)

Check if these issues are valid — if so, understand the root cause of each and fix them. If appropriate, use sub-agents to investigate and fix each issue separately.


<file name="domain-skills/manus/tasks.md">

<violation number="1" location="domain-skills/manus/tasks.md:16">
P2: Connect protocol description is inaccurate/contradictory: it claims `application/json` requests are binary-encoded, but example and Connect docs indicate JSON payloads for unary JSON codec.</violation>
</file>

<file name="domain-skills/perplexity/computer.md">

<violation number="1" location="domain-skills/perplexity/computer.md:153">
P2: Todo completion detection uses a global icon query instead of scoping to the Todo panel, which can produce incorrect completion state.</violation>
</file>

_{Reply with feedback, questions, or to request a fix. Tag @cubic-dev-ai to re-run a review, or fix all with cubic.}

[muqsitnawaz]

Thanks — both are fair. Pushed 253887b:

• manus/tasks.md:16 — the Connect description was wrong. Connect's JSON codec sends JSON text both ways; my naive fetch-hook logged bodies as <<bytes:N>> because the JS client wraps the JSON in a Uint8Array before fetch, and I mis-read that as a separate binary encoding. Rewritten to explain the hook artifact and confirm the wire content matches content-type: application/json.
• perplexity/computer.md:153 — agreed, the global svg use query was junk. The Todo button is a Radix trigger, so the panel id = button.aria-controls once data-state === "open". all_todo_done() now resolves the panel via that, scopes the icon query to it, and returns None / True instead of a boolean that lies when the panel is closed.

The same commit also cleans up three SKILL.md violations I missed in the first push (one raw click coord, one run-narration example, two real taskIds in the URL-pattern docstrings).

[cubic-dev-ai]

3 issues found across 2 files (changes from recent commits).

Prompt for AI agents (unresolved issues)

Check if these issues are valid — if so, understand the root cause of each and fix them. If appropriate, use sub-agents to investigate and fix each issue separately.


<file name="domain-skills/manus/tasks.md">

<violation number="1" location="domain-skills/manus/tasks.md:59">
P2: DOM-submit example dereferences a possibly missing editor element without a null guard or retry, so the helper can throw and block task submission.</violation>
</file>

<file name="domain-skills/perplexity/computer.md">

<violation number="1" location="domain-skills/perplexity/computer.md:13">
P2: ID extraction guidance is incorrect/ambiguous and can produce wrong task IDs (especially when query params are present).</violation>

<violation number="2" location="domain-skills/perplexity/computer.md:175">
P3: Documentation contradicts the helper implementation: hydration returns `false`, not `None`.</violation>
</file>

_{Reply with feedback, questions, or to request a fix. Tag @cubic-dev-ai to re-run a review, or fix all with cubic.}

[muqsitnawaz]

Round 2 — pushed 01fdbd7. All three are fair:

• manus/tasks.md:59 — the ProseMirror lookup had no retry and no null guard. On a cold Manus tab the editor mounts after the initial wait(1.5) about 1 in 5 runs, so ce.getBoundingClientRect() would throw on a null. Rewrapped in a bounded retry loop (10 × 0.5s) with a final assert rect so the failure mode is a clear message instead of a TypeError mid-script.
• perplexity/computer.md:13 — "strip the trailing 22 chars" implicitly assumed the caller passed pathname, but the common case is pasting a full href with ?view=thread. Replaced with explicit URL-parse-then-slice and showed both JS (new URL(href).pathname.split('/').pop().slice(-22)) and Python (urlparse(href).path.rsplit('/', 1)[-1][-22:]) forms.
• perplexity/computer.md:175 — real contradiction. The prose said the helper returns None on hydrate, the JS returned false. Changed the helper to return null when icons.length === 0, so callers can distinguish "panel closed or still hydrating" (None) from "panel mounted, some rows not yet checked" (False).

[cubic-dev-ai]

1 issue found across 2 files (changes from recent commits).

Prompt for AI agents (unresolved issues)

Check if these issues are valid — if so, understand the root cause of each and fix them. If appropriate, use sub-agents to investigate and fix each issue separately.


<file name="domain-skills/perplexity/computer.md">

<violation number="1" location="domain-skills/perplexity/computer.md:15">
P2: ID extraction examples are brittle: they return an empty ID when the task URL has a trailing slash.</violation>
</file>

_{Reply with feedback, questions, or to request a fix. Tag @cubic-dev-ai to re-run a review, or fix all with cubic.}

[muqsitnawaz]

Pushed one more — trailing-slash handling in the id-extraction snippet (JS .filter(Boolean), Python .rstrip('/')). That's the last cubic chase from me; happy to keep iterating if a maintainer has substantive feedback. Ready for human review.