chief-of-staff-v2
CLAUDE.md1 starsChanged 29 days agoApache-2.0View on GitHub
A personal chief of staff: a Rust binary, cos, that drives headless Claude Code (claude -p --output-format stream-json) for one daily brief, and serves it over MCP to Claude Desktop, Codex and claude.ai. User guide: README.md. The brief works down from the goals: week (each goal's steer, the deliverables the week should hold, decisions, stop doing), then today's actions and changes. Design and measurements: docs/design.md. Slack copy for sharing the idea: docs/slack-chief-of-staff.html. START-HERE.md, plugins/chief-of-staff-starter/ (skill, references, the state.py helper and…
In this repository6 other files it gives its agents.
- AGENTS.mdAGENTS.md
- prove-itSkill
- ritual-fullSkill
- ritualSkill
- semantic-isaSkill
- prove-it-reviewSkill
The file
Files
chief-of-staff
A personal chief of staff: a Rust binary, cos, that drives headless Claude
Code (claude -p --output-format stream-json) for one daily brief, and
serves it over MCP to Claude Desktop, Codex and claude.ai. User guide:
README.md.
| Agent | Horizon | Invoked by |
tactician | today, set against this week and the goals | cos brief, cos ask "…" |
The brief works down from the goals: week (each goal's steer, the
deliverables the week should hold, decisions, stop doing), then today's
actions and changes.
Design and measurements: docs/design.md. Slack copy for sharing the idea:
docs/slack-chief-of-staff.html.
The starter kit is the authority
START-HERE.md, plugins/chief-of-staff-starter/ (skill, references, the
state.py helper and its tests), guides/ and AGENTS.md are the Chief of
Staff starter kit. Its rules win over anything cos did before: goals → KPIs
→ weekly deliverables → priorities → delegation, read-only sources, drafts
only, honest coverage, handoffs that grant nothing. When changing the kit,
keep fixtures fictional, update release-files.json, and keep both plugin
manifests at its version (tests/kit.rs).
What the brains read
<workspace>/.chief-of-staff/(the active business's workspace;--business/COS_BUSINESSpick another,--workspace/COS_WORKSPACEoverride it, default the home) —profile.json(approved sources, timezone, tracking approval) andstate.json(goals, KPIs, deliverables, commitments, sources, runs, incidents, journal). Written only throughstate::commitor the kit'sstate.py. Gitignored.context/, gitignored and seeded fromtemplates/context/:principles.md,decisions/, andprofile.md/goals.md, which matter only when there is no workspace.
The brains are read-only. They recommend; the user commits. Never commit
anything from .chief-of-staff/, context/ or out/.
Layout
| Path | What |
src/claude.rs | stream-json driver, audit log |
src/context.rs | the workspace and context/ inlined into prompts (brains have no file tools); connectors narrowed to approved sources; deny-mcp.txt |
src/state.rs | the kit's state.py in Rust: validate, transition rules, lock, atomic journaled commit |
src/track.rs | what cos commits: run receipts and source coverage (tracking approved only); the brief's proposed deliverables as a candidate |
src/handoff.rs | the kit's nine-part handoff, from a tracked deliverable |
src/proc.rs | bounded child processes: one deadline, kill the group, reap |
src/agents/ | cos agents: spawn, supervise, monitor and stop Claude Code / Codex agents in worktrees; slots.rs run slots and the build lock, seed.rs worktree seeding, judge.rs the verdict from diff, checks and scope (never the agent's summary), policy.rs hiring defaults from context/agents.json, hire.rs the plan before any spawn (repo resolution, refusal), fit.rs the pre-hire fit check over the policy's fit vocabulary, watch.rs stalled/looping flags from the stream, ledger.rs the verdict ledger and scorecard |
apps/desktop/ | Tauri 2 app over the cos lib: the report on cosreport:// in a sandboxed iframe, the Agents tab (staff roster by role, run verdicts, hire dialog); tests/ipc.rs drives its commands on the mock runtime; examples/preview.rs serves the window's UI to a browser over the same commands; task desktop-app installs ~/Applications/Chief of Staff.app |
apps/desktop/ui/tokens.css + src/design.rs | the one design token file (colour light/dark, Geist type scale, radii) over vendored Basecoat and Geist (ui/vendor/README.md); design.rs inlines it with the fonts into every page cos serves; tests/design.rs fails any page styling outside it |
src/gather.rs | cn / gh facts collected without the model |
src/pipeline.rs | brief and ask, shared by CLI and MCP |
src/reports.rs | the brief's type + its JSON Schema |
src/store.rs | out/<kind>/<day>.{md,json} |
src/mcp.rs + src/dashboard.html | MCP server (stdio = all tools, HTTP = read-only) and the MCP App view |
src/report.rs + src/report.html | the HTML report; vendor/markdy/ is inlined into it |
src/today.rs | the report's hero and Brief tab, rendered server-side (tests/today.rs: actions first, no raw URL text, dates in words) |
src/staff.rs | cos staff-sync: refresh a configured staff role from an enrolled AllStack coordinator (bounded reads, revision checks, owner parking kept); the Plan tab's Staff card reads it (docs/managed-staff.md) |
src/history.rs | the archive under out/history/, backfilled two months per run |
src/dash.rs | cos dash, the live localhost page |
src/docs.rs + src/docs.html | the documentation room: repo docs from monitoredRepos, Notion pages from cos docs notion, the user's topics, links and marks in context/docs.json; served at /docs by cos dash and the desktop app |
src/flows.rs + src/flows.html | cos flows and /flows: every ritual flow (a Chronis epic) in the main checkouts under --projects, its step inferred from its beads and ritual-* markers, its bead graph; read-only, gated on an approved tasks source, one deadline for the scan (tests/flows.rs) |
src/ritual.rs | cos ritual approve / check: the design's sha256 on the epic plus cn approve; approve is CLI-only and prompts on a terminal (--yes for an agent the user told to approve), the one place cos writes a task store (docs/ritual.md) |
.claude/skills/ritual/ | the ritual skill: Think, Spec, Plan, Build, Review, Ship, Reflect, and the ritual-* markers cos flows reads |
.claude/skills/ritual-full/, .claude/skills/semantic-isa/ | ritual with every tool on (EGM, a Semantic ISA section per design, two reviewers, cos hires); planned work in this repo runs as a ritual-full flow. All three skills are vendored byte-for-byte from decebal-claude-skills |
examples/demo.rs | the synthetic home behind every screenshot (task demo), with fictional Chronis stores and a stand-in cn for /flows |
Screenshots in docs/assets/ come only from task screenshots on the demo
home. Never capture a real COS_HOME: briefs quote mail, calendar and tickets.
Commands
task test # fmt, clippy -D warnings, tests (the kit's Python tests run inside)
task desktop # build cos, run the desktop app; `task desktop-check` lints it
task kit OUT=/abs.zip # package the starter kit from release-files.json
task setup # first run: seed context/, build, install
task install-mcp # register with Claude Desktop + Codex
task schedule BUSINESS=acme # launchd: weekday brief, the dashboard, for one business (default: the active one)
cos business list # registry: each business, the active one, its directories
cos --dry-run brief # print the prompt, call nothingAgent files carry @COS_HOME@; task install substitutes the clone path.
Review
/prove-it (.claude/skills/prove-it/) asks whether a change does what it
claims in this repo: the brains stay read-only, every child is owned and
reaped, every number comes from facts. Copilot code review reads the same
invariants from .github/skills/prove-it-review/ and
.github/instructions/review.instructions.md.
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

