56 lines
2.8 KiB
Markdown
56 lines
2.8 KiB
Markdown
|
---
|
||
|
|
kind: task
|
||
|
|
slug: ai-assisted-agent-builder
|
||
|
|
title: Generate structured agent drafts with AI
|
||
|
|
plan_slug: workspace-communications
|
||
|
|
epic_slug: agent-participation
|
||
|
|
status: ready
|
||
|
|
priority: P1
|
||
|
|
tenant_id: global
|
||
|
|
owner: unassigned
|
||
|
|
cursor_todo_id: null
|
||
|
|
updated_at: "2026-09-25"
|
||
|
|
---
|
||
|
|
|
||
|
|
# Task summary
|
||
|
|
|
||
|
|
Use the Agents-tab composer to turn a natural-language request into a validated structured draft that can be refined conversationally or through direct field edits.
|
||
|
|
|
||
|
|
## Description
|
||
|
|
|
||
|
|
In Agents mode, prompts such as “Create an agent that triages failed convoy runs and asks before retrying” start a builder conversation rather than a normal workspace chat. The server returns a typed agent-draft proposal: identity, purpose, instructions, capabilities, requested tools, invocation/context policy, model role, harness route, budget, and unresolved questions. Users can refine the draft conversationally, edit fields directly, and save a private draft. Permission approval, sandbox testing, installation, and activation are handled by a dependent review task. Generated configuration never silently grants tools, selects plaintext credentials, publishes, or activates an agent.
|
||
|
|
|
||
|
|
## Subtasks
|
||
|
|
|
||
|
|
- [ ] Define a versioned zod schema for agent-builder requests, clarifying questions, structured draft patches, warnings, and save readiness.
|
||
|
|
- [ ] Add a workspace-authorized streaming builder endpoint/service that uses model routing rather than a hard-coded provider.
|
||
|
|
- [ ] Reuse the shared composer shell with Agents-specific placeholder text, examples, attachments policy, and submit handling.
|
||
|
|
- [ ] Add prompt-injection boundaries, requested-tool validation, budget caps, audit logging, and malformed-model-output tests.
|
||
|
|
- [ ] Define the follow-on UI contract for transcript, structured editing, revision history, and private-draft save/resume without implementing persistence in this task.
|
||
|
|
|
||
|
|
## Owner or assignee
|
||
|
|
|
||
|
|
`unassigned`
|
||
|
|
|
||
|
|
## Status
|
||
|
|
|
||
|
|
`ready`
|
||
|
|
|
||
|
|
## Estimation
|
||
|
|
|
||
|
|
`L`
|
||
|
|
|
||
|
|
## Acceptance criteria
|
||
|
|
|
||
|
|
- [ ] Natural-language requests produce zod-validated structured drafts or clarifying questions, never unvalidated persisted configuration.
|
||
|
|
- [ ] The builder requests a model role/capability and works through the existing provider/harness abstraction.
|
||
|
|
- [ ] Initial requests and conversational refinements produce ordered clarification/draft-patch events with deterministic base revisions.
|
||
|
|
- [ ] Generated requested tools, activation, channel access, and workspace publication remain unapproved fields and cannot take effect in this task.
|
||
|
|
- [ ] This service does not persist, publish, activate, or execute an agent; UI editing and private-draft persistence are a follow-on task.
|
||
|
|
- [ ] `pnpm lint && pnpm type-check && pnpm test` pass.
|
||
|
|
|
||
|
|
## Links to related Epic / Plan
|
||
|
|
|
||
|
|
- Epic: `./Epic-agent-participation.md`
|
||
|
|
- Plan: `../Plan-workspace-communications.md`
|