ceo

claude-ceo

Autonomous CEO agent for Claude Code. Reads your Obsidian vault, dispatches specialized subagents, runs scheduled playbooks, and gates writes by authority tier.

Why

A solo operator has more domains than hours: code repos, research, ops, career, personal. Day-to-day, most of the work isn't doing — it's deciding what to do next, where you left off, and what's safe to ship without you. claude-ceo reads your Obsidian vault (the source of truth), runs prioritized scans on a cron schedule, and proposes work in tiers: read-only digests run automatically, low-stakes writes execute with a report, high-stakes actions queue for approval.

How it works

Cron / interactive trigger
  │
  ▼
ceo-cron.sh ──► ceo-config.sh        (resolve $CEO_VAULT)
            ──► ceo-gather.sh         (pre-gather context → env)
            ──► ceo-scan.sh           (vault diff)
            ──► ~/.ceo/registry.json  (host-local dispatch lookup)
            ──► preflight_<name>()
            ──► runner:
                 • claude (default)   — read tier = single call,
                                        write tiers = PLAN/FILTER/EXECUTE
                 • script              — deterministic shell, no LLM
            ──► ceo-report.sh         (flock-guarded append)

Output: CEO/reports/YYYY-MM-DD.md
State:  CEO/log/, CEO/approvals/, CEO/cache/

Playbooks self-register: ceo playbook scan walks CEO/playbooks/*.md, extracts each frontmatter block via yq, and rewrites the host-local ~/.ceo/registry.json. Scheduling is owned by the ceo-schedulerd daemon (which reads that registry) — scan does not touch the crontab.

Subagents

Skills under skills/ceo/agents/ define six specialized subagents the CEO can dispatch for domain-specific work. Each declares its authority and domain scope; high-stakes actions (push, publish, send) are always returned as drafts.

Subagent	Authority	Domains	Use for
analyst	read + reports	Career, Academics, all	Evaluating options, thesis planning, cost-benefit, tool comparison
code-reviewer	read + draft (post is high-stakes)	Awesome Motive, any code repo	PR review — diff analysis, CI check, draft comments
implementer	read + low-stakes write	Awesome Motive, any code repo	Bug fix — branch, commit, test; push/PR returned as recommendation
ops-manager	read + low-stakes write	NRX Research	Inventory, hiring pipeline, SOP execution, runbook compliance
researcher	read only	Academics, Career, NRX Research, any	Vault / web / claude-mem / academic source investigation
writer	read + draft	Career, Academics, NRX Research, Personal	LinkedIn posts, cover letters, emails, academic writing, docs

Playbooks

Active playbooks shipped with the plugin (live in docs/playbooks/; copy into $CEO_VAULT/CEO/playbooks/ to enable, then ceo playbook scan):

Playbook	Trigger	Tier	Runner	Purpose
morning-brief	50 8 * * 1-5	read	claude	Prioritized day overview — PR queue, top-3 tasks, blessings
morning-scan	57 8 * * 1-5	read	claude	Vault-diff digest of overnight changes
inbox	every 15 min	read	claude	Process unchecked items in CEO/inbox.md and CEO/inbox/<host>.md (preflight-gated)
pr-triage	03 10 * * 1-5	read	claude	Surface PRs needing review
pending-drip	daily	read	claude	Drip reminders from Pending.md into CEO/inbox/<host>.md
eod-summary	47 17 * * 1-5	read	claude	Recap if there are log entries after 4pm
cleanup	weekly	low-stakes-write	claude	Branch / worktree hygiene
token-intake	45 8 * * 1-5	read	script	Run ceo-token-intake.sh — token-scope snapshot to vault

tier: read runs a single claude --print --max-turns 5 --disallowedTools Bash,Write,Edit call with pre-gathered context injected as <external-data> blocks. tier: low-stakes-write and above use the three-phase PLAN → FILTER → EXECUTE pipeline; high-stakes actions are written to CEO/approvals/pending.md instead of executed.

A playbook can declare runner: script to dispatch a shell script directly, skipping the LLM call entirely (token-intake is the canonical example). The script receives CEO_VAULT, CEO_DIR, LOG_DIR, TODAY, NOW, TRIGGER as env vars; the dispatcher does not parse stdout. Exit code 0 = success.

A playbook can also declare bin: <script>.sh in frontmatter to expose itself as a shell command: ceo playbook scan creates a ~/.local/bin/<name> symlink to that script and removes stale ones when the playbook is deleted or set inactive. ceo doctor warns when a declared bin symlink is missing.

Examples

Interactive triage from a Claude Code session:

/ceo

Reads today's report, walks the vault, and proposes prioritized actions in tier order.