Skip to content MVP Plan Summary
User Flows
- Desktop/TUI/UI spawn:
- Client calls
POST /agent-team/spawn.
- Server applies
SpawnPolicy, skill validation, budget inheritance, capability defaults.
- On allow, engine creates child session + instance and emits spawn lifecycle events.
- Tool-driven spawn:
- LLM calls
spawn_agent.
- Engine hook routes to the same runtime gate as UI spawn.
- Denials are machine-readable (
code, requiresUserApproval).
- Orchestrator runtime spawn:
POST /mission/{id}/event with mission_started triggers role-mapped spawns.
mission_canceled propagates cancellation to mission instances.
- Approval surfaces:
GET /agent-team/approvals returns pending spawn-approval records and pending tool approvals for agent-team sessions.
Definition Of Done
- Spawn gate cannot be bypassed:
spawn_agent, orchestrator bridge, and UI spawn all pass one server-side policy evaluator.
- Safe defaults enforced:
- Missing policy denies spawn.
- Network disabled unless explicitly enabled.
- Git push disabled unless enabled; optional approval requirement asks through permission flow.
- Secrets hidden unless alias is listed in
secretsScopes.
- Budgets enforced:
- Per-instance limits enforced for tokens, steps, tool calls, wall time, and optional cost.
- Child budgets capped from parent remaining limits.
- Optional mission total budget enforced and emits mission exhaustion event.
- Skills enforced:
- Required skills enforced per role.
- Skill source policy and pinned hashes enforced.
skillHash included in spawn responses, instance metadata, events, and audit rows.
- Visibility:
- SSE includes spawn, capability denials, budget usage/exhaustion, completion/failure/cancellation.
- HTTP exposes templates, instances, mission summaries, approvals, spawn, and cancellation APIs.
- Tests:
- Unit + integration coverage for gating, budgets, skill validation, and mission bridge behavior.
Phase Breakdown
P0: Safe Core Gate
- Deliver:
- Spawn policy and template registry loading.
- Unified spawn gate used by UI/tool/orchestrator.
- Spawn lifecycle SSE + audit log rows with
skillHash.
- Base endpoints (
templates, instances, spawn, cancel).
- Acceptance:
- Spawn denied when policy missing.
- Spawn denied on role-edge/required-skill violations.
- Tool and UI paths produce equivalent deny/allow decisions.
P1: Runtime Enforcement
- Deliver:
- Budget supervisor (tokens/steps/tool calls/time/cost).
- Mission total budget and parent-remaining inheritance capping.
- Capability runtime checks (tools/fs/net/secrets/git).
- Approval handoff for push when
push_requires_approval=true.
- Dedicated approvals view endpoint.
- Completion/failure lifecycle events.
- Acceptance:
- Budget exhaustion triggers cancel-first behavior and events.
- Mission budget exhaustion cancels mission instances and emits mission event.
- Disallowed fs/net/git/secrets operations are blocked with machine-readable reasons.
P2: Command Center Surfaces
- Deliver:
- Desktop/TUI surfaces for instances, mission rollups, and approvals.
- Live event-based status and budget telemetry.
- Filtering by mission, status, parent/child lineage.
- Acceptance:
- Operator can observe who spawned whom, active status, budget burn, denials, and failures.
- Operator can cancel instance/mission and resolve approvals from dedicated UI workflows.