oh-my-pi/packages/coding-agent/src/prompts/tools/task.md

Task

Launch a new agent to handle complex, multi-step tasks autonomously. Each agent type has specific capabilities and tools available to it.

This matters. Get it right.

Subagents have NO access to conversation history. They only see:

1. Their agent-specific system prompt
2. The context string you provide
3. The task string you provide

Use a single Task call with multiple tasks entries when parallelizing. Multiple concurrent Task calls bypass coordination.

For code changes, have subagents write files directly with Edit/Write. Do not ask them to return patches for you to apply.

Agents with output="structured" enforce their own schema; the output parameter is ignored for those agents.

{{#list agents join="\n"}} {{description}} {{default (join tools ", ") "All tools"}} {{/list}} This matters. Be thorough.

1. Plan before acting. Define the goal, acceptance criteria, and scope per task.
2. Put shared constraints and decisions in context; keep each task request short and unambiguous.
3. State whether each task is research-only or should modify files.
4. Provide an output schema whenever possible. Do not repeat the schema in context; the agent does not need it there.
5. Assign distinct file scopes per task to avoid conflicts.
6. Trust the returned data, then verify with tools when correctness matters.

- `agent`: Agent type to use for all tasks - `context`: Template with `\{{placeholders}}` for multi-task. Each placeholder is filled from task args. `\{{id}}` and `\{{description}}` are always available. - `isolated`: (optional) Run each task in its own git worktree and return patches; patches are applied only if all apply cleanly. - `tasks`: Array of `{id, description, args}` - tasks to run in parallel - `id`: Short CamelCase identifier (max 32 chars, e.g., "SessionStore", "LspRefactor") - `description`: Short human-readable description of what the task does - `args`: Object with keys matching `\{{placeholders}}` in context (always include this, even if empty) - `skills`: (optional) Array of skill names to preload into this task's system prompt. When set, the skills index section is omitted and the full SKILL.md contents are embedded. - `output`: (optional) JTD schema for structured subagent output (used by the submit_result tool). Do not duplicate this schema in `context`. Returns task results for each spawned agent: - Truncated preview of agent output (use `read agent://` for full content if truncated) - Summary with line/character counts - For agents with `output` schema: structured JSON accessible via `agent://?q=` or `agent:///`

Results are keyed by task id (e.g., "AuthProvider", "AuthApi").

user: "Looks good, execute the plan" assistant: I'll execute the refactoring plan. assistant: Uses the Task tool: { "agent": "task", "context": "Refactoring the auth module into separate concerns.\n\nPlan:\n1. AuthProvider - Extract React context and provider from src/auth/index.tsx\n2. AuthApi - Extract API calls to src/auth/api.ts, use existing fetchJson helper\n3. AuthTypes - Move types to types.ts, re-export from index\n\nConstraints:\n- Preserve all existing exports from src/auth/index.tsx\n- Use project's fetchJson (src/utils/http.ts), don't use raw fetch\n- No new dependencies\n\nTask: \{{step}}\n\nFiles: \{{files}}", "output": { "properties": { "summary": { "type": "string" }, "decisions": { "elements": { "type": "string" } }, "concerns": { "elements": { "type": "string" } } } }, "tasks": [ { "id": "AuthProvider", "description": "Extract React context", "args": { "step": "Execute step 1: Extract AuthProvider and AuthContext", "files": "src/auth/index.tsx" } }, { "id": "AuthApi", "description": "Extract API layer", "args": { "step": "Execute step 2: Extract API calls to api.ts", "files": "src/auth/api.ts" } }, { "id": "AuthTypes", "description": "Extract types", "args": { "step": "Execute step 3: Move types to types.ts", "files": "src/auth/types.ts" } } ] } - Confirmation bias: ask for factual discovery instead of yes/no exploration prompts - Reading a specific file path → Use Read tool instead - Finding files by pattern/name → Use Find tool instead - Searching for a specific class/function definition → Use Grep tool instead - Searching code within 2-3 specific files → Use Read tool instead - Tasks unrelated to the agent descriptions above