Agent Task Mode

kind=agent executes task body through the agent. It is single-round by default; only review: true enables the executor plus user-simulator multi-round review flow.

What It Fits

• tasks that require reasoning and iteration
• research / analysis / summarization
• human-readable reusable deliverables

What It Is Not

• not a normal chat turn
• not a database-backed task queue
• not a mode that requires external channel delivery by default

Use cases

• tasks requiring reasoning and iteration
• research / analysis / summarization
• human-readable deliverables

Execution characteristics

• single-round execution by default
• automatic multi-round review only when review: true (default max: 3)
• live progress events visible in Console UI
• final output written to output.md, with summary in result.md
• executor query / reply also written into run-directory messages.jsonl for debugging

Current Execution Logic

1. Load task.md and parse frontmatter plus body.
2. Create a new run directory: ./.downcity/task/<taskId>/<timestamp>/.
3. Write input.md and run-progress.json.
4. Create a task-specific runtime instead of reusing normal chat history.
5. Let the executor agent produce a result.
6. If review: true, let the user-simulator agent evaluate whether the result satisfies the task goal.
7. If review: true and the result is not good enough, feed the previous output plus simulator feedback back into the executor, up to 3 rounds.
8. If review is disabled, finish after the first valid output.
9. Write output.md, result.md, dialogue.*, run.json, and error.md when needed.

Success Rules

• an agent task now only requires valid single-round output text by default
• successful chat_send delivery is no longer required by default
• if review: true, the output also goes through simulator review and up to 3 rounds of revision
• if the task body explicitly asks for external delivery, that requirement belongs to the task itself

Context Model

• sessionId is stored in task definition and decides which chat receives the task result
• execution uses a separate task run context
• task execution therefore does not directly reuse normal chat message history

Diagram

Best practices

• write explicit goals, boundaries, and acceptance criteria
• keep review off by default; only set review: true when you really need review-and-revise behavior
• require auditable outputs when possible
• do not put raw shell scripts in agent tasks

Example

---
title: daily-market-research
description: daily market research summary
sessionId: research
when: 0 8 * * *
status: enabled
kind: agent
review: true
---

Generate today's market research summary including:
1. major events
2. key risks
3. next-step recommendations