TaskSquad: 3 skills for AI agents

TaskSquad - Agents

Field-tested against tasksquad.ai on 2026-05-03 using a logged-in Chrome session.

URL

https://tasksquad.ai/dashboard/agents       # Agents list for current project

Agent list

Agents are the daemon workers that execute tasks. Each agent card shows:

• Name
• Status badge (active / inactive / running / waiting_input / paused)
• Role (optional label set by maintainer)
• Last-seen timestamp

Navigate to agents via the sidebar link "Agents".

Creating an agent

Maintainer/owner only. Click "New agent" (or similar CTA in the agents view).

Provide:

• Name - display name for the agent
• Role (optional) - a freeform string describing the agent's specialty

After creation, a token is generated. Copy it immediately - it is shown only once. The daemon uses this token to authenticate.

tsq install --token <TOKEN>

Tokens are created via POST /teams/:teamId/tokens with { label, agent_id }.

Agent statuses

Agent actions (maintainer / owner)

Token management

Tokens are per-agent credentials. Each token has a label. A maintainer can:

• Generate a new token - use when setting up a new machine or rotating credentials
• Tokens cannot be viewed after creation; only revoked implicitly by deleting the agent

Gotchas

• Status is driven by daemon pings, not task state. An agent can be active while no task is running.
• inactive ≠ paused. Inactive means the daemon process is offline. Paused means a maintainer explicitly suspended it via the UI.
• Only owners and maintainers can create agents or generate tokens. Members see the list but cannot modify it.
• Token is shown once. If you miss copying it, you must generate a new one - there is no "reveal token" button.
• Reset is for recovery only. Use it when an agent is stuck in running after a daemon crash. Do not reset agents mid-task.

agents.md · updated May 4, 2026 · 436 words

TaskSquad - Authentication

Field-tested against tasksquad.ai on 2026-05-03 using a Chrome session.

URLs

https://tasksquad.ai/auth          # login page (redirects to /dashboard if already signed in)
https://tasksquad.ai/auth/cli      # CLI token auth page

Login page

The login page is a centered card with two OAuth buttons. No email/password form exists - SSO only.

goto_url("https://tasksquad.ai/auth")
wait_for_load()

Sign-in buttons

The card contains exactly two buttons:

Both are <button type="button"> inside a <div class="px-6 pb-6 flex flex-col gap-3">. Locate by label text and click coordinates returned from the rendered element:

import json
 
def click_button_by_text(label):
    """Find a <button> by exact innerText and click its center."""
    sel = json.dumps(label)
    result = js(f"""
      var b = Array.from(document.querySelectorAll("button"))
                .find(e => (e.innerText || "").trim() === {sel});
      if (!b) return null;
      var r = b.getBoundingClientRect();
      return JSON.stringify({{x: Math.round(r.x + r.width/2), y: Math.round(r.y + r.height/2)}});
    """)
    if result is None:
        raise RuntimeError(f"button not found: {label!r} — has the React app finished hydrating?")
    pos = json.loads(result)
    click_at_xy(pos["x"], pos["y"])
 
# Google
click_button_by_text("Continue with Google")
 
# GitHub
click_button_by_text("Continue with GitHub")

After clicking, a browser popup opens for the OAuth flow. Stop and ask the user to complete sign-in - do not attempt to interact with the OAuth popup window.

Auth-wall pattern

Any route under /dashboard/* redirects to /auth when the user is not signed in. If you navigate to a dashboard URL and land on /auth, the session is unauthenticated - ask the user to sign in before proceeding.

Redirect after sign-in

Firebase Auth resolves the session client-side. After the OAuth popup closes, React's onAuthStateChanged fires and automatically navigates the user to /dashboard. Do not click anything - just wait_for_load() and verify with page_info().

Sign-out

There is no dedicated sign-out page. The sign-out action is triggered from the sidebar in /dashboard. Look for a LogOut icon button in the bottom-left sidebar area.

Gotchas

• The page is a React SPA. Navigating to /auth while the app is still hydrating returns the bare <div id="root"> with no buttons yet. Always call wait_for_load() and verify the "Continue with Google" text is present before clicking.
• No error routes exist for failed OAuth. On failure, the card renders an inline <p class="text-sm text-red-500 text-center"> error message below the buttons.
• authed === null loading state. While Firebase resolves the session on /auth, the <Login> component is not rendered (the route renders null). Wait for buttons to appear before clicking.

auth.md · updated May 4, 2026 · 425 words

TaskSquad - Task Inbox

Field-tested against tasksquad.ai on 2026-05-03 using a logged-in Chrome session.

URL

https://tasksquad.ai/dashboard              # Inbox (default view)
https://tasksquad.ai/dashboard/<taskId>     # Task thread (individual task)

Inbox view

The inbox is the default view after sign-in. It shows a list of tasks assigned to agents in the currently selected project.

Task list

Tasks are rendered as rows. Each row shows:

• Subject line
• Agent name
• Status badge
• Relative timestamp

Status values

Filters

Two dropdowns appear above the task list:

Status (left): All / Pending / Queued / Running / Waiting / Done / Failed / Scheduled

Origin (right): All / System / Mine / From Note / Critique / Scheduled

• "System" = tasks created by automated conveyors
• "Mine" = tasks you composed yourself
• "From Note" = tasks spawned from Notes
• "Critique" = tasks that are note critiques

Both dropdowns use shadcn <Select>. Click the trigger to open, then click the item by text.

Refresh

A RefreshCw icon button sits next to the "Inbox" heading. Click it to reload the task list without full navigation. Alternatively, wait_for_load() after any state change - the app auto-polls when active tasks exist (interval depends on plan; see Gotchas).

Composing a new task

Click the "New message" button (top-right of Inbox). A dialog opens.

Dialog fields

Optional advanced toggles (revealed via UI controls):

• Schedule: set a future delivery time
• Auto-close: task closes automatically when the agent finishes
• Save tokens: compress the task context (lite / full / ultra)
• Close steps: newline-separated post-completion steps

Submit with the "Send" button inside the dialog.

import json
 
def _coords_by_text(label, tag="button"):
    """Find an element by exact innerText and return click coords, or raise."""
    sel = json.dumps(tag)
    txt = json.dumps(label)
    result = js(f"""
      var el = Array.from(document.querySelectorAll({sel}))
                .find(e => (e.innerText || "").trim() === {txt});
      if (!el) return null;
      var r = el.getBoundingClientRect();
      return JSON.stringify({{x: Math.round(r.x + r.width/2), y: Math.round(r.y + r.height/2)}});
    """)
    if result is None:
        raise RuntimeError(f"{tag!r} with text {label!r} not found")
    p = json.loads(result)
    return p["x"], p["y"]
 
def _coords_by_selector(selector):
    """Find an element by CSS selector and return click coords, or raise."""
    sel = json.dumps(selector)
    result = js(f"""
      var el = document.querySelector({sel});
      if (!el) return null;
      var r = el.getBoundingClientRect();
      return JSON.stringify({{x: Math.round(r.x + r.width/2), y: Math.round(r.y + r.height/2)}});
    """)
    if result is None:
        raise RuntimeError(f"selector not found: {selector!r}")
    p = json.loads(result)
    return p["x"], p["y"]
 
# Open the compose dialog
click_at_xy(*_coords_by_text("New message"))
wait(1.0)  # let the dialog mount; agents may still be loading
 
# Open the agent <Select> trigger and pick an agent.
# The trigger renders the placeholder "Select agent…" until an agent is chosen.
click_at_xy(*_coords_by_text("Select agent…"))
wait(0.5)
# Then click the agent name from the opened dropdown:
click_at_xy(*_coords_by_text("YOUR_AGENT_NAME"))
 
# Focus the subject input, then type
click_at_xy(*_coords_by_selector('input[id="subject"]'))
type_text("Check the build status")
 
# (Optional) focus the message textarea and type the body
# click_at_xy(*_coords_by_selector('textarea'))
# type_text("Detailed instructions go here.")
 
# Submit
click_at_xy(*_coords_by_text("Send"))

Task thread

Clicking a task row opens the thread at /dashboard/<taskId>.

The thread is an email-style conversation. Messages alternate between:

• User messages (right-aligned or labeled with sender name)
• Agent messages (left-aligned)

Replying

When a task is in waiting_input status, a reply box appears at the bottom. Type your reply and press Enter or click the send button.

A scheduled reply pending delivery blocks the reply box - a cancel option appears.

Task actions

Actions available from the thread view (usually via icon buttons or a menu):

• Close - mark the task done manually
• Delete - permanently remove the task
• Forward - reassign to a different agent (appears as "Forward to agent" with agent selector and optional instructions)

Gotchas

• Inbox auto-polls while active tasks exist (pending / running / waiting_input). Free plan polls every 5 s, Pro every 2 s. For scraping, add a wait_for_load() between reads if statuses are changing.
• queued is a derived client-side status - the server returns pending, but the UI shows queued when the target agent is itself running or waiting_input. Filter accordingly.
• Task list is per-project. Switching projects (teams) reloads the list. The active project is stored in localStorage as tsq_team_id.
• Dialog does not open until agents are loaded. If "Select agent…" placeholder is missing after clicking "New message", the agents list is still fetching - retry after a short wait.
• Free plan project limit: free accounts have a project limit (currently 1).

tasks.md · updated May 4, 2026 · 873 words