Ceo Delegate

Read skills dispatch table — read $VAULT/CEO/SKILLS.md.

Match task to playbook — compare the user's task description against the Task Type Index in SKILLS.md:

• Match by keywords (e.g., "review PR" → pr-review, "triage" → pr-triage, "brief" → morning-brief)
• If ambiguous, ask the user: "This could match [type A] or [type B]. Which playbook should I follow?"
• If no match: say "No playbook matches this task. I can attempt it without a playbook, or you can create one. What would you prefer?"

Read matched playbook — read $VAULT/CEO/playbooks/<matched-type>.md.

• If the playbook file doesn't exist yet, tell the user: "Playbook <type>.md is listed in SKILLS.md but hasn't been created yet. Want me to create a starter playbook based on your task?"

Execute playbook with subagent dispatch — read the playbook steps. For each step:

a. Determine if the step needs a specialist. Match by keywords:

• Diff analysis, code review, CI check → Code Reviewer
• Write code, fix bug, run tests → Implementer
• Search, investigate, literature review → Researcher
• Draft text, write post, compose email → Writer
• Check inventory, run SOP, review hiring → Ops Manager
• Compare options, analyze data, evaluate → Analyst
• No specialist needed → CEO handles directly

b. Choose dispatch mode:

• Will Nathan want to review the subagent's output later? → vault-mediated
• Quick subtask within a larger flow? → inline

c. Dispatch subagent (if specialist needed):

• Read the role template from the plugin's skills/ceo/agents/<role>.md
• Replace TASK_DESCRIPTION with the specific step's task
• Replace SCOPED_CONTEXT with:
  • $VAULT/CEO/AGENTS.md (global rules)
  • Relevant training file for the domain
  • Task-specific context (repo path, PR number, vault files, etc.)
• For vault-mediated: create $VAULT/CEO/delegations/<timestamp>-<role>-<slug>.md with task details
• Dispatch via Agent tool
• Collect the result
• For vault-mediated: write result to the delegation record's ## Result section

d. Process subagent result:

• If result contains high-stakes recommendations → propose to user (interactive) or write to pending.md (cron)
• If result is actionable low-stakes → continue playbook with the result
• If result is a draft (Writer, Code Reviewer) → present to user for review

e. Handle directly (if no specialist needed):

• Execute the step respecting authority tiers (unchanged from current behavior)

Log results — append to $VAULT/CEO/log/YYYY-MM-DD.md:

## HH:MM — <task-type>

**Status:** completed | failed | partial
**Playbook:** playbooks/<type>.md
**Delegated task:** <user's original description>
**Actions:**
- <what was done>
**Audibles:**
- <deviations from playbook, if any>
**Errors:**
- <errors encountered, if any>
**Delegations:**
- <role>: <task summary> — <status>

Report back — summarize what was done to the user.