@wairon/cli 5.1.1-dev.8 → 5.1.1-dev.9

package/dist/templates/skills/sdd-delegate.md

@@ -23,16 +23,24 @@ You are the **Delegation Orchestrator**. Your job is to hand scoped work to a fo
    - Pick the agent whose `ownedPaths`/domain matches the task. If no agent fits, stop and tell the user the topology has a gap.
 2. **Fetch the LIVE brief**:
    - Call `sdd_get_agent_brief(agentId)` (or read the `wairon-agent://<agentId>` resource).
-   - The brief carries: `agentId`, `name`, `template`, `domainRoot?`, `ownedPaths`, `readPaths?`, `instructions`, `variantGuidance?`.
+   - The brief carries: `agentId`, `name`, `template`, `domainRoot?`, `ownedPaths`, `readPaths?`, `instructions`, `variantGuidance?`, and — when the project opted into `execution.tier` — `profile` and `budget`.
    - **Never reuse a brief across delegations or after a re-lock** — fetch fresh per delegation; the call is cheap and the brief is always current.
 3. **Spawn a GENERIC subagent from the brief**:
    - Prompt: `brief.instructions`, plus the concrete task description.
    - Write fence: the subagent may only modify files matching `brief.ownedPaths` (within `brief.domainRoot` when set).
    - Required first reading: `brief.readPaths` — the subagent reads these before any edit.
    - Pass `brief.variantGuidance` along when present.
-4. **The subagent does the work**:
+4. **Apply `brief.budget` when it is present** — constituting the subagent correctly is part of spawning it, not a separate concern. When the brief carries no budget the project has not opted in; spawn as you otherwise would.
+   - `modelTier` → your host's model families. On Claude Code: `small`→haiku, `standard`→sonnet, `large`→opus, `frontier`→fable. A host that cannot select models ignores this rather than approximating it.
+   - `effort`, `maxTurns` → pass through where the host supports them. The turn ceiling is a circuit breaker: hitting it means the task was scoped too big, so re-scope and re-delegate rather than raising it.
+   - `toolClass` → `read-only` grants read/search only; `implement` adds edit/write/shell; `orchestrate` is for a router that owns nothing and must not read bulk content.
+   - `allowNestedDelegation: false` → withhold the delegation tool entirely, so the subagent does its own work instead of spawning another layer.
+   - `mcp: none` → do not load MCP servers into the subagent; its brief already quotes the contract it needs.
+   - **`frontier` is never an owner tier.** Derivation never assigns it. Treat it as a sparring partner: when a subagent is genuinely stuck on something the specs do not settle, escalate that *question* to a frontier-tier helper and bring the answer back. A component whose owner truly needs frontier capability to operate is usually a component doing too much — raise that as a spec concern rather than spending the tier.
+   - Fan delegations out in ONE message when they are independent. Siblings spawned together share a cached prompt prefix; dispatched one at a time they each pay for it.
+5. **The subagent does the work**:
    - For component implementation it follows the `sdd-implement` skill (gating checks, AI-TDD, narrative coding) and reports back: what changed, test results, anything out of scope.
-5. **Review & integrate**:
+6. **Review & integrate**:
    - Read the report, verify the write fence was respected, and continue orchestrating — or delegate the next scoped task.
 
 This flow is transport-agnostic: it works identically over local stdio and the hosted data plane — the tools and `wairon-agent://` resources are the same surface.

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@wairon/cli",
-  "version": "5.1.1-dev.8",
+  "version": "5.1.1-dev.9",
   "description": "SYW Waffle AIron — CLI for managing AI coding agent topology across projects",
   "keywords": [
     "ai",