escapeboy/claude-code-kit/company

/company — an IT company for one task

/company is management, not a second orchestrator. It clarifies, researches, splits the work and staffs teams; the work itself runs through skills that already exist (sprint-orchestrate for code, code-research for audits, the ops skills for hosts). The back office is the company-hq mod: it hires project specialists as agent types, enforces the optional budget cap on every spawn, and shows the org in a pane.

When to Use (and When NOT to)

If the task turns out small after intake (one part, ≤ ~5 files, one area), say so and propose the simpler path (e.g. /sprint-orchestrate quick) with a one-paragraph plan — that proposal IS stop 2: wait for the user's yes before any file is written. Never skip both stops because the task looked easy.

The flow

Track the phase with mcp__company-hq__set_phase (load the mcp__company-hq__* tools with ToolSearch first). Write every artifact under claudedocs/company/<slug>/ in the project (or ~/claudedocs/company/<slug>/ when there is no project yet).

Between stop 2 and delivery, work without check-ins unless a gate fails twice, the cap is hit, or a destructive/outward action comes up. A question that comes up is routed, never guessed: User-class → ask_user + PushNotification, only its part waits; Mechanical/Taste → decided by your recommendation and logged with record_decision (an external scorer can be plugged in; it never decides User-class questions). Details: references/questions.md.

Dashboard (optional)

Off unless asked: /company --dashboard[=local|fleetq|local+fleetq], or dashboard in ~/.claude/company-hq/config.json, or the stop-1 question. Pass the choice to open_project; its result carries the URL — tell it to the user once.

• local (default when on): a page on 127.0.0.1 served by the company-hq mod; nothing leaves the machine; no tokens are spent on it.
• fleetq: the mod also sends the snapshot and plan documents (redacted) to FleetQ; needs fleetq.url in the config and COMPANY_HQ_FLEETQ_TOKEN.
• What it shows comes from files: the company state, docs/design-<part>.md, the ## Tasks list in docs/architecture-<part>.md, Result: in docs/test-plan-<part>.md. So label every part agent <part>: … (references/workflows.md).

Defaults (edit these to your own)

• Budget cap: none by default. If the user sets one at stop 2, pass it as capUsd; the mod then refuses spawns past it, and you stop and report.
• Size: ≤10 agents per Workflow run (the session guideline). More only with an explicit yes at stop 2.
• New project: ~/projects/<slug> in your default stack unless the task implies another; init-project, then sprint-orchestrate full, then devops for CI/CD.
• Ops hosts: the hosts you list here, read-only by default. Any write, restart or other host is a separate yes.
• Specialists: hired per project (company-hq:<name>); at the end offer to save the ones that worked as permanent agents in ~/.claude/agents/.
• Language: the user's language with the user; English in agent briefs.
• Dashboard: off.

Borrowed rules (from agent-team, code-research, sprint-orchestrate)

• Every team owns a disjoint set of files; two agents never edit one file. Parallel code teams run with isolation: 'worktree'.
• Agent Teams (teammates that talk) only where talking is the point: competing-hypotheses debugging and cross-checking reviewers. Everything else is plain subagents — cheaper.
• State the team structure and the cost multiplier before spawning (3 agents ≈ 3–5× a session, 5 on a big task ≈ 8–10×).
• While teams run, the coordinator does not do their work; it waits, then synthesizes one structured report.
• Every brief carries your tool-routing block and the "verify, don't assume" rule — subagents don’t get the main conversation. See references/briefs.md.
• Work runs unattended after stop 2, so a worker never grades itself: tests by command, adversarial-verifier, /code-review.

Boundaries

Always

• Stop twice: after clarifying, and with the plan + org + cost before any agent writes code or touches a host.
• Open the project in company-hq before spawning; close it at the end, and also when the user cancels at a stop.
• Route code delivery through sprint-orchestrate (--from-design, ship --no-merge); do not re-implement its phases.
• Read command output for every "passes/works" claim; report failures as failures.
• Label every part agent <part>: …; route every question through decision-classify and record it (ask_user / record_decision).

Ask first

• Merge, deploy, push to a shared branch, any write on a server, deleting anything.
• Raising a cap the user set, more than 10 agents, or a 3rd fix round after FAIL.
• Saving a specialist as a permanent agent.

Never

• Let two agents edit the same file.
• Let a hired specialist or a mod approve a dangerous action.
• Guess the answer to a User-class question, or let an automatic scorer decide one.
• Present an estimate or a guess as a measured result.

Related

• `sprint-orchestrate` — depends-on: runs each code part (--from-design) and the single ship (--no-merge).
• `agent-team` — used for the parts where teammates must talk (debug, cross-review); lighter alternative when the whole task is one such review.
• `code-research` — the audit/research engine for phase 2 and for audit projects.
• `decision-classify`, `confidence-check`, `continuity`, ship (if you have one) — used at the points named above.
• Mod: `company-hq` — see `17-mods/guide.md` (tools open_project, hire, set_phase, close_project, ask_user, answer_question, record_decision; /company-status; the local dashboard server in server/).

Task: $ARGUMENTS