feat: align delegation prompts with target schema — 6 fixes

- Bob: DELEGATE→MUST DELEGATE, added MANDATORY COMPLETION GATE (Critic + Vision-for-UI)
- Bob: "delegate to Manager" → "MUST delegate to Manager" (hard rule, no self-orchestration)
- Plan config: block write/edit/apply_patch (was only bash/grep/glob/webfetch)
- Plan prompt: write scope = .bob/plans/*.md + .bob/drafts/*.md (was .bob/*.md)
- Manager: added Wave Concurrency Limits (max 7, drop to 4) + background_output pattern
- Critic: added Mandatory Escalation Gate section

Co-authored-by: CommandCodeBot <noreply@commandcode.ai>

Author: vlgalib

package.json

@@ -1,6 +1,6 @@
 {
   "name": "@hiai-gg/hiai-opencode",
-  "version": "0.4.2",
+  "version": "0.4.3",
   "description": "HiAi OpenCode Plugin — multi-agent orchestration for OpenCode with 10 agents, LSP tools, browser automation, Memory, Session management, and Caveman protocol",
   "type": "module",
   "main": "dist/index.js",

src/agents/bob.ts

@@ -65,7 +65,7 @@ Classify EVERY message before acting:
 - 2x+ effort difference or missing critical info → MUST ask
 
 ### Step 3: Delegation
-**Default: DELEGATE.**
+**Default: MUST DELEGATE.**
 > **Multi-point task (≳3 points / multi-file / open-ended)? → plan FIRST** (Key Rule 3):
 > Use \`task({subagent_type: "plan", description: "...", prompt: "..."})\` for a phased parallel plan, THEN dispatch its waves
 > (via Manager when 5+ steps or 3+ parallel). Only trivial 1-2 point work goes straight to a build/general.
@@ -121,6 +121,7 @@ When delegating via task(), use the appropriate category:
 - You NEVER execute write, edit, bash, or any mutation tool yourself.
 - Always delegate implementation to build/general.
 - Always verify with Critic before reporting completion.
+- **MANDATORY COMPLETION GATE**: After EVERY specialist completes, you MUST call task({subagent_type: "critic", ...}). If ANY UI/UX work was done, you MUST also call task({subagent_type: "vision", ...}) via agent-browser. NEVER escalate to user until Critic + Vision-for-UI have passed.
 - Fix only your own issues. Do NOT fix pre-existing.
 
 ## Subagent Handoff Protocol (CRITICAL — replaces Receiving Results)
@@ -177,7 +178,7 @@ Phase 2 (<name>): serial steps: [2.1/build then 2.2/build] — deps: Phase 1 —
 \`\`\`
 
 ### Step 2 — Decide dispatch mode
-- **5+ steps OR 3+ parallel steps** → delegate to Manager:
+- **5+ steps OR 3+ parallel steps** → MUST delegate to Manager:
   \`task({subagent_type: "manager", description: "Execute plan: <title>", prompt: "<THE FULL PLAN TEXT + Execution Graph Extract>"})\`
   Include BOTH the raw plan text (from deliverable body) AND the Execution Graph Extract in the Manager prompt.
   Manager inherits the plan text so it can dispatch each phase independently.

src/agents/critic.ts

@@ -6,6 +6,9 @@ export const CRITIC_PROMPT = `You are Critic, a review gate agent.
 ## Identity
 Quality Guardian. You verify code quality, correctness, and adherence to standards. Your verdict is final.
 
+## Mandatory Escalation Gate
+Critic review is MANDATORY before any results are escalated to the user. No agent may report completion without passing this gate. Your APPROVED verdict is the single unlock for user-facing delivery.
+
 ## Role
 - Review code changes for correctness, security, and quality
 - Verify that implementations match requirements

src/agents/manager.ts

@@ -25,6 +25,9 @@ Manager receives TWO things from Bob in the prompt:
 
 Manager's job is to dispatch phases in order, respecting the annotations — NOT to re-derive parallelism.
 
+## Wave Concurrency Limits (HARD CAP)
+Max 7 tasks per wave. When overloaded (>7 independent tasks), drop to 4 per wave and serialize overflow into a follow-on sub-wave. Never exceed 7 concurrent task() calls in a single dispatch cycle.
+
 ## Available MCP Tools
 
 **Library/API docs:** use the \`context7\` skill (CLI/HTTP) on demand — not an MCP tool.
@@ -68,7 +71,8 @@ NEVER ask 'should I continue' between steps. Just delegate next task.
 2. **Parse Phases** — Extract ordered phases from the plan text or extract. Identify parallel vs serial steps per phase.
 3. **Dispatch Phase** — For the current phase:
    - Fire ALL \`parallel: yes\` steps concurrently via \`task()\` to their \`owner\` subagent type.
-   - Collect ALL results before proceeding. If any step fails, decide: retry, escalate, or mark partial.
+   - Use \`background_output(task_id, block=true)\` to collect results from dispatched agents.
+   - If any step fails, decide: retry, escalate, or mark partial.
    - For \`parallel: no\` steps: dispatch in dependency order (step N must finish before step N+1 starts).
 4. **Advance** — Once all steps in phase N complete, move to phase N+1.
 5. **Verify** — After all phases, collect evidence and report complete/partial summary.

src/agents/plan.ts

@@ -11,7 +11,7 @@ When user says 'do X' → ALWAYS interpret as 'create a plan for X'.
 NEVER write code. NEVER edit files (except .md plans).
 
 ## Identity
-Principal Architect. You plan, you do not implement. You write ONLY .bob/*.md files and plan documents.
+Principal Architect. You plan, you do not implement. You write ONLY .bob/plans/*.md and .bob/drafts/*.md files and plan documents.
 
 ## Role
 - Analyze requirements and codebase architecture
@@ -73,7 +73,8 @@ task → ALWAYS fan out first.
 
 ## Constraints
 - You are READ-ONLY for code files. No write, edit, bash.
-- You may write plan documents to .bob/plans/*.md
+- You may write plan documents to .bob/plans/*.md and .bob/drafts/*.md
+- .bob/drafts/ is for work-in-progress plans; move to .bob/plans/ when ready for dispatch
 - You delegate research to explore (grep/glob blocked for you); do not self-explore the codebase
 - You never implement — only plan
 

src/config.ts

@@ -139,7 +139,7 @@ export const DEFAULT_CONFIG: BobConfig = {
       lsp_prepare_rename: false,
       lsp_rename: false,
     },
-    plan: { bash: false, grep: false, glob: false, webfetch: false },
+    plan: { write: false, edit: false, apply_patch: false, bash: false, grep: false, glob: false, webfetch: false },
     critic: {
       write: false,
       edit: false,