VKirill/claude-lane-stack
VKirill/claude-lane-stack has 125 GitHub stars and 3 forks. GitHot measured 0 stars in 24 hours. Last checked 2026-11-15 15:40 UTC. Last push 2026-11-13.
Measured star growth
- in 24h
- 0
- in 7 days
- Update delayed
- in 15 days
- Update delayed
- in 1 month
- Update delayed
Net star changes over overlapping periods, not daily totals. Missing measurements are left blank. Stars measure attention, not software quality.
About
Multi-agent AI coding factory for one person — Claude Code PM + Codex/Qwen/Grok/Kimi/AGY writers, durable conveyor, auto-merge to main
AIREADME.en.md
Original
Claude Lane Stack
A small AI coding factory for one person
One human. One AI project manager. Real CLI coding agents on a durable conveyor.
Talk to Claude Code — it runs Codex / Qwen / Grok / Kimi / AGY, checks work, merges to main, reviews at night.
| 🏭 Factory, not five chats | 🛡️ Owned paths | 🌙 Night review | 📦 Auto-merge to main |
| One PM holds context | Writers stay in their lane | Independent Codex Sol rail | You never merge by hand |
| 🧠 Memory | 📚 Living docs | 🎨 Design | ℹ️ First hour |
| Opt-in fact corpus | docs/ from git, not essays |
docs/DESIGN.md |
/info · /app-architect |
✨ Why this exists
Working with AI coding tools usually means: five windows, copy-paste, midnight merges, and zero review discipline.
Claude Lane Stack turns that into a conveyor — plain files + git, no mandatory cloud DB.
| 😩 Typical AI coding | ✨ Lane Stack |
|---|---|
| Re-explain context every chat | One PM holds the plan |
| Models overwrite each other | owns_paths on every task |
| Nobody reviews the AI | Typed night shift (review → fix → re-review) |
| You merge branches | PM merges main after checks |
| “What were we doing?” | resume-project + memory (facts, not chat) |
| Docs rot after the sprint | Night Luna refreshes every stale docs/ page |
| Jobs die at ~2 min Bash | Detached processes (survive chat close) |
[!TIP] New here? Start with the Beginner guide — zero jargon, factory metaphors.
🔌 CLI agents we plug into
This is not “one model does everything.”
The stack is a control plane that runs real CLI coding agents as durable processes.
| CLI | Required? | Role |
|---|---|---|
| Claude Code | Yes | PM (dev-orchestrator), watchers, chat with you |
| OpenAI Codex CLI | Optional | Day writer · night Sol review · onboard · docs (Luna) · emergency |
| Qwen Code | Optional | Daytime durable writer |
| Grok (xAI CLI) | Optional | Daytime durable writer |
| Kimi CLI | Optional | Daytime durable writer (often full-profile default) |
| AGY (Gemini-oriented) | Optional | Daytime durable writer |
| Cursor / OpenCode | Optional | Extra daytime writer rails (adoc Coder) |
You ──chat──► Claude Code · dev-orchestrator ← always
│
run-supervisor + run-controller
┌─────┬─────┬─────┬─────┐
▼ ▼ ▼ ▼ ▼
Codex Qwen Grok Kimi AGY
│ │ │ │ │
└─────┴─────┴─────┴─────┘
│
owns → L1 verify → acceptance → main
- Mix and match — install what you pay for;
agents-doctor/adocbuilds the profile - One protocol — owns, verify,
acceptance.json, progressive accept - Switch writer — change
main_writein adoc (e.g.codex→qwen) - Codex for night quality when available
Profiles (examples)
| Profile | Writers | Review |
|---|---|---|
claude-only |
Claude only | light |
claude-codex |
Codex process | Codex Sol night |
claude-qwen / grok / kimi / agy |
that CLI | as configured |
full |
multi-writer ready | Codex night |
🧠 How it works (60 seconds)
You (chat) → PM (dev-orchestrator)
│ plan-critique · task YAML in .agents/runs/
▼
run-supervisor ← Claude watches one run
│
▼
run-controller ← durable OS process
│
▼
writer CLI process ← codex / qwen / grok / …
│
▼
owns → L1 tests → acceptance.json
│
▼
PM merges → main → (night) Codex review
You don’t hand-launch “Qwen the coder” for normal work.
You talk to the PM. The PM starts the conveyor. The writer is a background process, not a random brand-named subagent.
☀️ Day vs 🌙 night
| Day | Night | |
|---|---|---|
| Goal | Ship fast | Independent quality |
| Product code | Writer from adoc | Repair after findings |
| Watch | run-supervisor |
night-shift |
| LLM review every commit? | No | Yes — Codex Sol |
| Docs / memory | Writers do not touch docs/ |
docs-maintain-all (Luna, all stale pages) · memory-maintain-project |
| Done | acceptance.json + main |
Finding fixed + re-review |
🎭 Who is who
| Role | Kind | Does |
|---|---|---|
| You | Human | Say what you want |
dev-orchestrator |
Claude agent (PM) | Plan · dispatch · merge · chat |
run-supervisor |
Claude (watch) | One run until accepted/blocked |
run-controller |
OS process | DAG · retry · owns/verify/accept |
| Writer process | CLI (Codex/Qwen/…) | Implements the task card |
lane-supervisor |
Claude (1 action) | Typed lane-ctl recovery |
emergency-writer |
Claude → Codex | After terminal block only |
night-reviewer / night-shift |
Codex Sol | Night review + findings |
project-onboarder |
Codex | Project passport (docs/llm/*, CLAUDE) |
docs-maintainer |
Codex Luna max fast | INIT + nightly living docs/ |
memory-maintainer |
Codex (adoc) | Opt-in fact corpus (.agents/memory/) |
design-lead |
Claude | Extract / refresh / audit docs/DESIGN.md |
browser-qa |
Jev (default) | Live URL click-through via CDP + .agents/qa |
Built-ins Claude may also use: Explore, Plan, general-purpose (research / side tasks — not the daytime product conveyor).
Roster: agents/claude/README.md
📋 Task card = contract
Every unit of work lives under .agents/runs/<slug>/tasks/*.yaml:
| Field | Meaning |
|---|---|
owns_paths |
Files this task may touch |
verification[] |
Focused L1 checks |
lane: |
Matches adoc main_write |
| Receipts | report → owns → verify → acceptance.json |
[!IMPORTANT] No
acceptance.json→ not done. Chat green ≠ shipped.
🧠 Memory and 📚 living docs
Two opt-in modules in adoc (own tabs, not buried in Stages). Off until Enabled + Apply. Writers on a feature lane do not update wiki or the fact corpus.
| Memory | Docs | |
|---|---|---|
| What | Facts in .agents/memory/ (CORE + on-demand) |
Living docs/ + web.yaml + llms.txt |
| Not | docs/, LESSONS, chat logs |
Memory, wiki/, TODO/, docs/plans/ |
| First enable | lane-memory corpus |
docs-web stubs now; Luna INIT in background |
| Night | memory-maintain-project |
cron hour → docs-maintain-all --if-hour |
| Input | session ledger + git | committed git log --since ∩ owns / stubs |
| Model | from adoc Memory tab | Codex Luna max fast |
| Commit | no | no |
regular Claude / Codex / Cursor session
↓ commit
git log --since=yesterday
↓ hour (default 05:00)
docs-web → daylog → docs-stale → Luna (every stale page)
Uncommitted work is invisible to the 05:00 job. Daylog: .agents/session-log/DOCS-DAY-YYYY-MM-DD.md.
🚀 Quick start
① Install once
This repo is a Claude Code plugin marketplace. ./install.sh installs the host runtime (~/.agents) and the Claude plugin.
git clone https://github.com/VKirill/claude-lane-stack.git
cd claude-lane-stack && git checkout v1.67.4 # or: main
./install.sh
export PATH="$HOME/.agents/bin:$PATH"
After install, Claude Code has lane-stack@claude-lane-stack and the marketplace auto-updates from GitHub. Skills are /lane-stack:<name> (example: /lane-stack:orchestrator-lanes). Host ~/.agents still needs ./install.sh. Live checkout: LANE_INSTALL_LOCAL_MARKETPLACE=1 ./install.sh.
After a stack fix in this repo: commit → ./install.sh → git push. A checkout-only patch is not shipped. Version bump → tag + gh release.
Plugin-only (runtime already on the machine):
claude plugin marketplace add VKirill/claude-lane-stack
claude plugin install lane-stack@claude-lane-stack -y
From a local clone: claude plugin marketplace add . then the same install line. Reload with /reload-plugins.
Needs: Claude Code · Git · Python 3 (+ PyYAML/jsonschema) · Node · rsync · flock
Optional: Codex · Qwen · Grok · Kimi · AGY · Linux: bubblewrap
macOS: brew install flock; PyYAML/jsonschema via pip3 install --break-system-packages pyyaml jsonschema; Cursor/AGY/Codex CLIs install separately, same as on Linux. Writer lane sandboxing uses the built-in sandbox-exec (Seatbelt) — switch backends with LANE_SANDBOX_BACKEND=auto|bubblewrap|seatbelt. install.sh also appends PATH to ~/.zshrc (the default login shell). lane-bg runs on the nohup backend — no user systemd.
Jev provider. Jev calls (fast-jev compaction, the router, opencode-lane, winnow, jev_decisions) go to OpenLux (jev-1.13.0:stable, +1.5 s timeout) when OPENLUX_API_KEY is set, from the environment or ~/secrets/openlux.env. Otherwise they go to TypeSafe (jev-latest) as before, via TYPESAFE_API_KEY / JEV_API_KEY. JEV_PROVIDER=openlux|typesafe forces one. Keys never go into the repo; merge_claude_settings.py copies OPENLUX_API_KEY into Claude's env without overwriting an existing value.
② Prepare your project once
cd /path/to/your-project
agents-doctor --apply . # or: adoc
③ Start the PM
lane-pm # boot + session name <agent>-<folder>-DD-MM-YYYY
# Claude Code 2.1+ auto-submits the agent initialPrompt
| In chat | When |
|---|---|
/project-onboard |
First time on a repo |
/resume-project |
After a break |
/info |
Conveyor cheat sheet |
/app-architect |
New app / service talk |
| “Add dark mode to settings” | Normal feature work |
⚙️ adoc (simple)
Tabs: Coder · Stages · Memory · Docs · Work · Night · UI · Status · Info · Apply.
| Knob | Meaning |
|---|---|
Writer main_write |
codex / qwen / grok / kimi / agy / cursor / opencode |
| Model / effort | e.g. Codex luna + max |
| Stages | plan-critique · write · night review · specialist · onboard |
| Memory | opt-in fact corpus · inject · maintain |
| Docs | opt-in living docs/ · hour · all stale pages (cap 0) |
| Info | ? — roles + first-hour commands |
| Fast mode | Codex / Cursor service_tier: fast |
| Workspace | in_place · worktree · auto |
🧰 Commands cheat sheet
| Command | Purpose |
|---|---|
agents-doctor / adoc |
Detect CLIs · write profile |
resume-project . |
Now / Blocked / Next |
run-init · run-validate · run-board |
Runs (usually via PM) |
run-controller start|watch|status |
Durable lifecycle |
lane-ctl … |
Typed control plane |
night-shift · night-shift-all |
Night review/repair |
plan-critique --run-dir … |
Plan quality gate |
docs-maintain-project · docs-maintain-all |
Living docs (INIT / night / lint) |
docs-web · docs-stale |
Stubs / INDEX / stale map (no LLM) |
lane-memory · memory-maintain-project |
Fact corpus |
❓ FAQ
Plugin or ~/.agents?
Claude-facing (PM agents, playbook skills, /project-onboard) ships as plugin lane-stack in marketplace claude-lane-stack. Host runtime (bin/, board, writer profiles) still lives in ~/.agents. ./install.sh does both. Do not copy stack agents into ~/.claude/agents — those copies override the plugin.
Do I need every CLI?
No. Claude Code alone works (claude-only). Each extra CLI is a pluggable writer (or Codex review rail).
What is general-purpose?
A built-in Claude Code subagent for multi-step side tasks (docs). PM may use it for research. Product features still go through the conveyor.
Why ~2 minute Bash death?
Claude Code host limit. Writers run detached. Don’t keep long jobs in the PM foreground.
Who merges?
The PM. You never merge. If it asks you to — the PM is wrong.
Language?
Chat: any language. Agent-written files: English (LANGUAGE.md).
Does a normal Claude/Codex chat update docs?
Commits feed the night job. Enable Docs in adoc on that repo, Apply. Uncommitted edits are invisible. Feature lanes do not write docs/.
📚 Documentation
| Doc | For |
|---|---|
| BEGINNER.md | Zero-jargon tour |
| SOLO-ORCHESTRATION.md | Day / night rules |
| ROUTING.md | Models & profiles |
| LANE-EXEC.md | Process survival |
| PLATFORM-CAPABILITIES.md | Claude Code + Codex features |
| FILE-CONTRACT.md | Layout & receipts |
| DOCS-STAGE.md | Living docs blueprint |
| PROJECT-MEMORY.md | Memory vs docs |
| agents/claude/README.md | Role agent names |
| plugins/lane-stack/README.md | Claude plugin layout |
| CHANGELOG.md | Releases |
🤝 Contributing & community
- Contributing · Code of Conduct · Security
- Issues: use the templates · PRs welcome with tests + docs
- Channel: Помогающий маркетолог
MIT © VKirill and contributors
If this factory helps you — ★ star the repo
It helps others find a calmer way to multi-agent coding.