5.3 KiB
5.3 KiB
| kind | slug | title | plan_slug | status | priority | tenant_id | cursor_epic_id | updated_at |
|---|---|---|---|---|---|---|---|---|
| epic | agent-participation | AI agent studio, marketplace, and chat participation | workspace-communications | ready | P1 | global | null | 2026-10-22 |
Epic objective
Turn the existing AI rail destination into a unified AI workspace where users switch between chatting with agents and creating/managing them through an AI-assisted studio and agent marketplace. Workspace-scoped agents can then join channels, respond to explicit invocations, use approved tools, and stream normalized run events without coupling the UI to a provider or execution runtime.
In scope / out of scope
In scope
- A single
/<workspaceSlug>/aidestination with URL-backedChatandAgentstabs. - A shared composer shell whose behavior changes by tab: conversation in Chat and natural-language agent building in Agents.
- A workspace agent registry and marketplace library for templates, installed agents, workspace-created agents, drafts, and new-agent entry points.
- AI-assisted agent drafting that produces editable structured configuration before explicit save/install/publish actions.
- Agent identities, channel membership, mention resolution, and invocation policy.
- Durable agent jobs with claim, retry, cancellation, budget, and idempotency semantics.
- A normalized harness event contract for text, tools, approvals, failures, and completion.
- Streaming agent state to chat and persisting the final response and run linkage.
Out of scope
- Autonomous reading of every channel by default.
- A chat-specific model registry.
- Installing or publishing marketplace agents without an explicit review/confirmation step.
- Public ratings, billing, or cross-workspace commercial distribution in the first release.
- Bypassing workspace model policy or tenant-scoped MCP authorization.
Related tasks
| Task | Link |
|---|---|
| Build the AI Chat/Agents workspace shell | ./Task-ai-workspace-tabs.md |
| Define the versioned agent registry schema and contracts | ./Task-agent-registry-marketplace.md |
| Expose the shared catalog and private draft API | ./Task-agent-registry-api.md |
| Enforce reviewed installation and lifecycle | ./Task-agent-install-lifecycle.md |
| Build the agent marketplace library UI | ./Task-agent-marketplace-library-ui.md |
| Generate structured agent drafts with AI | ./Task-ai-assisted-agent-builder.md |
| Edit and persist AI-generated private drafts | ./Task-agent-builder-editor-drafts.md |
| Review, test, and activate agent drafts | ./Task-agent-builder-review-sandbox.md |
| Add durable chat-agent jobs | ./Task-agent-identities-and-jobs.md |
| Chat directly with saved agents | ./Task-direct-agent-chat.md |
| Add agents to workspace channels | ./Task-agent-channel-membership.md |
| Normalize harness run events | ./Task-agent-harness-chat-streaming.md |
| Stream and finalize agent runs in chat | ./Task-agent-chat-run-delivery.md |
Dependencies
- The AI workspace shell and registry can begin from the existing AI route, model routing, harness registry, template UI patterns, and current AI SDK chat endpoint.
- Registry schema → registry API → install/lifecycle are explicit server slices. Library/editor/review UIs consume those APIs. All AI selectors, builders, mentions and jobs share this registry, never a channel-local agent identity.
- Durable jobs depend on registry/lifecycle. Direct-agent chat and reviewed channel membership depend on jobs; workspace membership integration needs chat persistence, while project/object participation also waits for authoritative contextual authorization. Unified notifications consumes outcomes later rather than blocking permission enforcement. Harness execution and shared delivery follow these contracts.
- Full product includes mention-to-add with authorized confirmation, scoped history/tools, idempotency and revocation per the parent plan. Current AI convoy delivers only shell, schema/contracts and inert builder service. On 2026-10-22 the user approved kickoff and the independent shell implementer is active; no other role completion, DB transition, merge or release is asserted.
- Blocks: richer approval cards, marketplace publishing, and autonomous channel workflows.
Acceptance criteria
- The existing AI rail destination exposes accessible, deep-linkable Chat and Agents tabs without adding another rail item.
- The Agents tab presents templates, installed/workspace agents, drafts, and creation entry points as a coherent library.
- A natural-language request creates an editable agent draft; nothing is installed, published, or granted tools until the user reviews and confirms it.
- Explicitly mentioning an authorized agent creates one idempotent durable job.
- Chat and the builder request roles/capabilities and do not select providers directly.
- Harness adapters emit one normalized event protocol regardless of runtime.
- Tool calls and generated messages preserve workspace authorization and audit linkage.
- Approval-required, failed, and completed runs notify the initiating user.
Proposed timeline
| Phase | Window | Notes |
|---|---|---|
| Identity and jobs | 1 | Persistence, mention policy, queue semantics |
| Harness integration | 2 | Normalized events, streaming, finalization |