Project Tasks

Manage persistent, user-visible project tasks that can be handed to the main agent or to a task agent. Only task agents run in isolated environments. Use these to track high-level deliverables and milestones that the user cares about.

Project Tasks vs Internal Task List

When to Use

• Tracking user-requested features or deliverables
• Breaking a project into visible milestones
• Communicating progress to the user
• Managing tasks that persist across sessions

When NOT to Use

• Internal agent work breakdown (use the internal task list)
• Temporary implementation steps
• Sub-tasks that only matter to the agent

Task Identifiers

Tasks are identified by

taskRef

"#1"

"#2"

"#3"

Available Functions

getProjectTask(taskRef)

Get a project task by task ref.

Parameters:

taskRef

Returns: Dict with

taskRef

title

description

state

createdAt

updatedAt

Example:

const task = await getProjectTask({ taskRef: "#1" });
// "Task #1 (Add authentication)"

updateProjectTask(taskRef, title=None, description=None, dependsOn=None)

Update an existing project task's content. All fields are optional - only provided fields are updated.

Parameters:

taskRef

title

description

dependsOn

Returns: Dict with

taskRef

title

description

state

createdAt

updatedAt

Example:

await updateProjectTask({ taskRef: "#1", title: "Updated title" });

markTaskInProgress(taskRef)

Resume work on an IMPLEMENTED task. Call this before making further changes to a task that was already marked complete.

Parameters:

taskRef

Returns: Dict with

taskRef

title

description

state

createdAt

updatedAt

Example:

await markTaskInProgress({ taskRef: "#1" });

searchProjectTasks(query, locale=None, limit=None)

Search project tasks by text query, ordered by relevance.

Parameters:

query

"exact phrase"

foo bar

foo OR bar

-foo

locale

"en"

"es"

"fr"

limit

Returns: List of dicts, each with

taskRef

title

description

state

score

matchType

createdAt

updatedAt

Example:

// Simple keyword search
const results = await searchProjectTasks({ query: "authentication" });

// Boolean syntax: find auth tasks that aren't about login
const results = await searchProjectTasks({ query: "authentication -login", limit: 5 });

// Exact phrase
const results = await searchProjectTasks({ query: '"payment integration"' });

// Non-English query — pass locale for better stemming
const results = await searchProjectTasks({ query: "autenticación", locale: "es" });

for (const r of results) {
    console.log(`${r.taskRef} (${r.title}) — score: ${r.score}`);
}

listProjectTasks(state=None, taskRefs=None, includeDescription=False)

List project tasks, optionally filtered by state or specific task refs.

Parameters:

state

taskRefs

includeDescription

Returns: List of dicts, each with

taskRef

title

state

dependsOn

createdAt

updatedAt

description

includeDescription

dependsOn

taskRefs

Example:

const allTasks = await listProjectTasks();
for (const task of allTasks) {
    console.log(`  - Task ${task.taskRef} (${task.title})`);
}

bulkCreateProjectTasks(tasks)

Create multiple tasks at once with dependency relationships. Each task is created in PROPOSED state. The plan file content becomes the task description.

By default, create one task per user request. Combine related work into a single plan rather than splitting into many tasks. Only create multiple tasks if the user explicitly asks for them or the request contains clearly independent, unrelated goals.

Parameters:

tasks

Each task object:

id

dependsOn

title

filePath

.local/tasks/payment-integration.md

dependsOn

id

"#1"

"#2"

Returns: List of created task dicts with

taskRef

title

description

state

dependsOn

createdAt

updatedAt

Examples:

// Single task (no dependencies)
const created = await bulkCreateProjectTasks({
    tasks: [
        {
            title: "Payment integration",
            filePath: ".local/tasks/payment-integration.md",
        },
    ]
});

// Multiple tasks with dependencies using id aliases
const created = await bulkCreateProjectTasks({
    tasks: [
        {
            id: "auth",
            title: "Add authentication",
            filePath: ".local/tasks/auth.md",
        },
        {
            id: "payments",
            title: "Payment integration",
            filePath: ".local/tasks/payments.md",
            dependsOn: ["auth"],
        },
    ]
});

Plan File Format

Write each project-task plan file as a plain markdown document in

.local/tasks/

By default, create one project task per user request. Combine related work into a single plan rather than splitting into many tasks. Only create multiple tasks if the user explicitly asks for them or the request contains clearly independent, unrelated goals.

Dependencies are not declared in the plan file. Pass them via

dependsOn

Plan body

The first line should be a short, descriptive title (3-6 words) prefixed with

#

• What & Why — Brief description of the feature/change and its purpose.
• Done looks like — Observable outcomes when complete (what the user sees, not code-level details).
• Out of scope — What is explicitly NOT included.
• Tasks — Numbered list of implementation steps within this plan. These are internal steps for the executor agent, not separate project tasks.
• Relevant files — Existing files discovered during investigation that the executor should start from. Use backtick-wrapped paths only, with no descriptions after them. Only list files you verified exist.
  • Whole file:

    src/api/billing.ts

  • Specific lines:

    src/api/billing.ts:12-85

  • Multiple ranges:

    src/api/billing.ts:12-85,200-250

  • WRONG:

    src/api/billing.ts

    — Billing API handlers (lines 12-85)

Assume features build on each other. If a new task depends on another task, declare that dependency via

dependsOn

Rules for the

## Tasks

• Each task should be describable in 1-2 sentences.
• Focus on what to build, not how to build it.
• Do not include file paths, code snippets, CSS classes, or line-level edits in task bullets. Put file references in

  ## Relevant files

  instead.
• Draw clean boundaries so parallel executors will not create conflicting changes. If two tasks would touch the same area, combine them into one project task.
• If there is a critical architectural constraint the executor must follow, add a short note.

Example

# Payment Integration

## What & Why
Add Stripe payment processing so users can upgrade to paid plans.

## Done looks like
- Users can enter payment info and subscribe to a plan
- Successful payments activate the paid tier immediately
- Failed payments show a clear error message

## Out of scope
- Invoicing and receipts (future work)
- Multiple payment methods (Stripe only for now)

## Tasks
1. **Stripe backend integration** — Set up Stripe SDK, create endpoints for creating checkout sessions and handling webhooks.

2. **Payment UI** — Build the checkout page with plan selection and Stripe Elements for card input.

3. **Tier activation** — On successful payment, upgrade the user's account to the paid tier and reflect it in the UI.

## Relevant files
- `src/api/billing.ts:12-85`
- `src/config/stripe.ts`

proposeProjectTasks(taskRefs)

Propose existing tasks for user review and approval. This pauses the agent to wait for user approval.

Parameters:

taskRefs

Returns: Dict with

proposed

{ taskRef, title }

Example:

await proposeProjectTasks({ taskRefs: created.map(t => t.taskRef) });

Task States

PROPOSED

PENDING

IN_PROGRESS

IMPLEMENTED

MERGING

MERGED

BLOCKED_BY_DRIFT

CANCELLED

User Communication Rules

Task management is how you coordinate work with the user. Follow these rules strictly:

1. Always describe tasks by ref and title: When referring to tasks, always use their task ref and title, e.g. "Task #1 (Add authentication button)".
2. Never use internal state names: When talking to the user, use these display names instead:
   • PROPOSED → "Drafts"
   • PENDING → "Active"
   • IN_PROGRESS → "Active"
   • IMPLEMENTED → "Ready"
   • MERGING → "Merging"
   • MERGED → "Merged"
   • CANCELLED → "Archived"
   • BLOCKED_BY_DRIFT → "Affected by another task that changed"
3. Never expose implementation details: Do not reveal function names (

   bulkCreateProjectTasks

   ,

   updateProjectTask

   , etc.), API surface, or internal task system mechanics to the user.

Best Practices

1. Prefer fewer tasks: Default to one task per request unless the user explicitly asks for more
2. Create tasks early: Create the project task when you understand the user's goals
3. Keep titles short: Titles should be concise and descriptive
4. Use descriptions for detail: Put implementation details in the description field

Example Workflow

// 1. Write a plan file to .local/tasks/ (using write tool beforehand)

// 2. Create the task — file content becomes the description
const created = await bulkCreateProjectTasks({
    tasks: [
        {
            title: "Payment integration",
            filePath: ".local/tasks/payment-integration.md",
        },
    ]
});

// 3. Propose for user approval (pauses agent loop)
await proposeProjectTasks({
    taskRefs: created.map(t => t.taskRef),
});

// --- After user approves, the system handles scheduling ---

// 4. Check progress
const tasks = await listProjectTasks();
for (const task of tasks) {
    console.log(`Task ${task.taskRef} (${task.title}): ${task.state}`);
}