codex-lane-worker
Bounded Codex Lane writer protocol for one assigned task. Use only when the lane runtime assigns this worker.
dev-orchestrator only. Turns any request into a route of specialist and writer steps: which roles work, in what order, what each reads and writes. Use at the start of every new goal. SKIP: writer CLIs; specialists themselves.
$ npx -y skills add VKirill/claude-lane-stack --skill team-routes --agent claude-codeHow it fires
How this skill gets triggered: by you, by Claude, or both.
/team-routesContext preview
The summary Claude sees to decide when to auto-load this skill.
dev-orchestrator only. Turns any request into a route of specialist and writer steps: which roles work, in what order, what each reads and writes. Use at the start of every new goal. SKIP: writer CLIs; specialists themselves.
name: team-routes description: "dev-orchestrator only. Turns any request into a route of specialist and writer steps: which roles work, in what order, what each reads and writes. Use at the start of every new goal. SKIP: writer CLIs; specialists themselves."
The operator states a goal, not a staffing plan. You decide who works on it. A page built without its SEO, UX and copy inputs comes back as placeholders and gets redone; a one-line fix sent through four specialists burns an hour of expensive models. Pick the smallest team that produces the artifact the operator actually needs.
Before Score/Decompose, write one line for yourself: `deliverable: <artifact> · surface: <page/app/feature> · slug: <kebab-slug>`.
The deliverable is what the operator will open at the end: a gray prototype, a branded mockup, shipped code, an article, an audit report, an answer with sources. Ambiguous wording («сделай страницу») means the cheapest deliverable that still moves the goal (prototype before code) unless the operator said «в код», «на сайт», «задеплой» or the page already exists in code.
Start from the closest recipe in [references/recipes.md](references/recipes.md). Drop a step when its output already exists on disk and is not stale (for example `.agents/copy/pages/<slug>.md` is there, so copy only fills gaps). Add a step when the goal needs a role the recipe lacks. No recipe fits → compose a route from the role cards in [references/roles.md](references/roles.md): each step's inputs must be outputs of earlier steps or existing files.
Run steps in parallel when neither reads the other's output. Run them in order when one decides what the next one writes about (SEO intent before copy).
Not every request needs a route. Go straight to the existing Loop when the goal is product code with no new user-facing text or layout (bug, refactor, API, infra), or a question you can answer from the repo.
Post one Russian line, then start the first step in the same turn. Do not wait for approval: the operator asked for the result, and the line exists so they can redirect you.
Маршрут: seo ∥ design-lead (UX) → copy-lead → design-lead (прототип) → design-lead (аудит). Начинаю.
For routes with 2+ specialist steps, also write the route file `.agents/plans/routes/<slug>.md` (English):
goal: <operator goal, one line> deliverable: <artifact> steps: - id: seo | role: seo-specialist | reads: [...] | writes: <path> | status: todo - ... decisions: [] # conflicts you resolved, with the reason
Update `status` after each step. It is how `resume-project` and a restarted session know where the route stopped.
Every specialist prompt has the same five lines. The specialist sees only this prompt and the disk, so a missing READ path means the work is done blind.
GOAL: <operator goal + this step's part of it> SLUG: <slug> READ: <outputs of earlier steps and existing project files, as paths> WRITE: <the one path this step owns> DECIDE: <the questions this step must answer for the next step>
Close with the role's normal sentinel (`DONE <path>` / `FAILED <reason>`). Read the file from `DONE`, not the chat summary, before you start the next step.
Specialists do not talk to each other. You merge their outputs. When two outputs conflict (SEO wants a long FAQ block, design wants one short screen), decide by the deliverable's primary user action, record the call in `decisions:` with one reason, and pass it to the next step through READ/DECIDE.
Ask the operator only for what no role can produce from the disk or the web: prices, guarantees, the offer, legal claims, brand approval, money or irreversible actions. Specialists must mark unknown facts `[unknown]`, not invent them. Collect those markers into one question for the operator at the end of the route, not after each step.
End the route with the deliverable path, what each role contributed (one line each), open `[unknown]` facts, and the next sensible step (for example «в код — скажи "делай"»). A route that ends in code continues into the normal WRITE Loop: specialist steps finish first, then `run-init` (Lane Pilot: `lane_pilot_dispatch_writer`, specialists via `lane_pilot_specialist`) with their outputs in `read_first`. This keeps the Mode XOR rule: no research teammates on a goal once its write run has started.
Claude Lane Stack is a self-hosted toolkit for coordinating coding agents through CLI processes, task contracts, Git workspaces, and verification receipts (bin/run-controller:117-159, schemas/task-v2.schema.json:1-60).
Bounded Codex Lane writer protocol for one assigned task. Use only when the lane runtime assigns this worker.
Operate Claude Lane Stack schema-v2 runs from Codex. Use for lane status, start, resume, review, verification, evidence, or…
Детекция и коррекция AI-сгенерированного текста по методике DrMax LinguaForensic v3.9.4 (284 лингвистических признака + 16…
Owner-facing architect for a new app or service. Plain-language chat; living plan artifacts on disk. Use when user says…
Live browser QA + Playwright-compatible replay. Disk .agents/qa/. Use when (terminal Lane Stack): проверь в браузере, кликни…
Delegate fat file reads to the project's pm_read worker. Use when a Read/cat/head/tail/sed of a file over pm_read.min_lines was…