v1.5.0 — Durable daytime controller

Highlights

• Durable run-level controller with bounded 1–10 Grok provider and verification pools.
• One read-only visible run supervisor; no daytime LLM review.
• Exact Board lifecycle evidence: attempt, PID, liveness, exit, heartbeat, report, reason, and next action.
• Fail-closed provider completion, report, ownership, verification, acceptance, crash, and malformed-receipt handling.
• Independent nightly Codex Sol xhigh review → Grok repair → Codex re-review remains intact.
• Claude requests use subscription authentication without per-request dollar budget caps.

Verification

• 163 Python tests
• 33 Board tests
• Shell lifecycle suites
• Real two-task detached E2E through nohup and user-systemd
• Independent Claude review: 0 Critical / 0 High after triage

Claude Lane Stack v1.4.0 — Typed autonomous night shift

Claude Lane Stack v1.4.0 turns the night workflow into a typed, receipt-driven Claude → Grok → Codex control plane.

Highlights

• Immutable schema-v2 run and task contracts with attempt-local state, exact review receipts, and deterministic acceptance.
• Bounded Grok repair lanes (1–10 slots), warm run-scoped sessions, streaming runtime receipts, and subagents disabled for write work.
• Dedicated nightly Codex reviewer profile using gpt-5.6-sol with xhigh reasoning and read-only execution.
• Standardized findings under .agents/findings/, resumable repair plans, fresh Codex re-review, and crash-safe commit/merge recovery.
• Fail-closed verification: non-empty smoke/test evidence, executable allowlists, worktree-bound arguments, package/network restrictions, and direct argv execution without a shell for schema v2.
• Updated Claude agents, Grok/Codex role contracts, shared skills, Lane Board receipt projections, installer receipts, and project-memory views.

Validation

• 132 Python tests passed.
• 23 Lane Board Node tests passed.
• Live Grok writer and Codex reviewer probes passed.
• Python compile, Bash syntax, JSON schema checks, staged secret scan, and source/install content comparison passed.

Upgrade

git pull
./install.sh

Existing running orchestrator sessions should be restarted to load the updated agents and skills.

🛡️ Claude Lane Stack v1.3.1 — Anti-join guard + detached heartbeats

Highlights

Hardening for v1.3.0 progressive accept.

lane-mode-check (anti join-wait)

Multi-task runs refuse MODE=full (exit 2 → STATUS: refused_full_on_multi_task). Implementers call it in preflight so parallel Agents cannot poll-until-done and reintroduce join-wait on the slowest task.

• OK: MODE=start / MODE=finish, or single-task MODE=full
• Emergency override (tests only): LANE_ALLOW_FULL=1

Detached heartbeats

lane-exec --heartbeat path now writes heartbeat.json on real activity (stdout/CPU, throttled). After MODE=start the Claude supervisor is gone — stall-check still sees liveness. Idle kill is unchanged (no timer-only heartbeat).

Tests

tests/test_progressive_accept.sh — mode-check refuse, accept task A while B runs, heartbeat write.

Upgrade

cd claude-lane-stack && git pull && git checkout v1.3.1
./install.sh
# fresh PM session
claude --agent dev-orchestrator

Full notes: CHANGELOG.md

⚡ Claude Lane Stack v1.3.0 — Progressive accept (no join-wait)

Highlights

⚡ Progressive accept. Multi-task runs no longer wait for the slowest concurrent lane before accepting finished ones. When task A finishes while B is still running, A is accepted immediately, its slot is freed, and the next ready task can start (still ≤3 concurrent writers).

How it works

1. Implementer MODE=start → lane-bg only, returns STATUS: started (no poll loop)
2. PM lane-poll --run-dir .agents/runs/<slug> → sees which tasks are finish_ready
3. Implementer MODE=finish → writes report.md → accept now → free slot
4. Pipeline the next ready task

Single-task / micro still uses MODE=full (start + poll + report).

Why this matters. Claude Code joins all Agent tool calls in one turn. Spawning N implementers that each poll-until-done forced a join-wait on max(t1…tn). Progressive accept breaks that without raising the concurrent slot cap.

Added

• bin/lane-poll — multi-artifact poll (finish_ready = CLI done, no report yet)
• Implementer MODE: start | finish | full on agy / grok / codex implementers
• tests/test_lane_poll.sh

Changed

• skills/orchestrator-lanes — progressive protocol mandatory for ≥2 write tasks
• agents/claude/dev-orchestrator — PM loop: start → poll → finish/accept → refill
• Docs: LANE-EXEC, ROUTING, SOLO-ORCHESTRATION, FILE-CONTRACT, BEGINNER EN/RU
• READMEs + llms.txt → v1.3.0

Upgrade

cd claude-lane-stack && git pull && git checkout v1.3.0
./install.sh
export PATH="$HOME/.agents/bin:$PATH"
# fresh PM session required:
claude --agent dev-orchestrator

Full details: CHANGELOG.md · docs/LANE-EXEC.md

🌙 Claude Lane Stack v1.2.0 — Nightly-only review, Micro path, Lane Board

Highlights

🌙 Nightly-only review. The pre-merge review gate is gone by default (built for solo devs shipping products without live users). Everything merges after mechanical checks (verify + check-owns-paths); a cron-driven batch (night-review-all) reviews the whole day's diff across all active projects at 03:00 and writes REVIEW-<date>.md with per-run verdicts and a Morning fix plan. resume-project surfaces it at session start. Projects with real users can opt back in via gate: pre-merge.

⚡ Micro path. Score 0–2 tasks (add a button, tweak a text) skip PLAN/worktree/board/heartbeat/review entirely: minimal task YAML → one lane → owns check → commit to main with a [micro:<slug>] trail. Word-to-commit in under 3 minutes. New verify field (none | smoke | tests) lets the PM decide per task whether tests are required at all.

📋 Lane Board. From-scratch read-only dashboard (board/, bin/lane-board): zero npm dependencies, no build step — Node stdlib server + vanilla JS dark UI. Portfolio overview with a cross-project needs-attention strip, kanban with status-based scope=recent (history stays out of the way), a todos view with full idea bodies, runs timeline, night-review history, Cmd+K search, SSE live refresh.

🔥 Warm lanes. lane-session resumes run-scoped AGY/Grok conversations across related tasks (3 slots, rotation after 7 tasks), and AGY preflight is cached — no cold starts inside a run.

🎯 Scoped review. Reviewers now receive a precomputed git diff BASE_REF -- owns_paths in their SPEC and are forbidden from repo-wide exploration — small tasks get small (fast) reviews.

Also: Grok write lane defaults to --reasoning-effort low with an escalation rule; push-on-merge is now the rule (merge without push = unfinished ship); meaningful commit message policy.

Full details in CHANGELOG.md.

🤖 Generated with Claude Code

🚀 Claude Lane Stack v1.1.0 — Deep onboard + lanes that don’t die at 2 min

🚀 Claude Lane Stack v1.1.0

Your personal multi-agent coding factory — smarter onboard, lanes that don’t die at 2 minutes

---

🇬🇧 English

✨ Why this release

v1.0.0 shipped the factory. v1.1.0 makes it survive real projects and long agent runs.

Pain	Fix in 1.1.0
Onboard felt “surface only”	Deep forensic onboard (default on mature repos)
Same docs for toy & prod	Minimal vs Full auto-detect
AGY/Grok killed after ~2 min	lane-bg + lane-wait (Claude Bash host limit, not your code)
Hard timeout 570 mid-thought	lane-exec activity-aware idle/max
AGY crash on load	Ban call_mcp_tool / inheritMcp
Model soup	GPT-5.6 Sol / Terra / Luna only (no 5.5)

---

🧭 Onboard 2.0 — two axes

1) Scenario — what to seed

• 🟢 minimal — greenfield / small → CLAUDE · AGENTS · ARCHITECTURE · memory
• 🟣 full — mature (score ≥ 5 or monorepo) → + GOTCHAS · GLOSSARY · TESTING · deployment · nested package CLAUDE · SECURITY when needed

2) Depth — how hard Codex digs

• ⚡ fast — passport (default on minimal)
• 🔬 deep — forensic (default on full): entrypoints, 3–7 flows, wiki↔code audit, real cargo/npm verify, ship path, secrets surface (names only)

/project-onboard          # auto scenario + depth
/project-onboard deep     # force forensic
project-onboard . --full --deep

Auto files:

• .agents/onboard.scenario.yaml — scenario + depth + score
• .agents/runs/_onboard/artifacts/001/deep-scan.md — evidence pack for Codex

📖 Docs: docs/ONBOARD-SCENARIOS.md

---

🏃 Lanes that finish — background survival

Claude foreground Bash dies ~2 minutes. That was killing healthy AGY/Grok runs mid-edit.

Required pattern now:

lane-bg --dir "$ARTIFACT_DIR" --label agy-frontend -- \
  lane-exec --idle 600 --max 5400 --log "$ARTIFACT_DIR/lane-exec.log" -- \
  agy --print "..." --agent lane-frontend --print-timeout 90m ...

lane-wait --dir "$ARTIFACT_DIR" --once   # poll until done

Tool	Role
lane-bg	detach (nohup) so the host can’t kill the job
lane-wait	short status polls (safe Bash)
lane-exec	idle = stuck; max = hard ceiling on the detached process

Implementers + dev-orchestrator + docs/LANE-EXEC.md updated.

---

🧠 Routing & language

• Write: Terra (xhigh) · High-risk / deep onboard / review: Sol (xhigh)
• Luna: trivia only
• Files: English · Chat with human: Russian (docs/LANGUAGE.md)

---

📦 Install / upgrade

git clone https://github.com/VKirill/claude-lane-stack.git
cd claude-lane-stack && git checkout v1.1.0
./install.sh
export PATH="$HOME/.agents/bin:$PATH"
project-onboard . --deep   # or /project-onboard deep in Claude

Already installed? pull + reinstall bins:

cd /path/to/claude-lane-stack && git pull && ./install.sh

---

🙏 Thanks

Built for solo operators who want Claude as PM, AGY/Grok as hands, Codex as reviewer — without the merge dance.

---

🇷🇺 Русский

✨ Зачем этот релиз

v1.0.0 — завод. v1.1.0 — завод, который не умирает на длинных прогонах и не врёт на онборде.

Боль	Что сделали
Онборд «по верхам»	Deep forensic (по умолчанию на взрослых репо)
Одинаковые доки для toy и prod	Minimal vs Full по score
AGY/Grok умирали через ~2 мин	lane-bg + lane-wait (лимит Bash Claude, не lane-exec)
Жёсткий timeout 570	lane-exec idle/max по активности
Падение AGY на старте	Запрет call_mcp_tool / inheritMcp
Каша моделей	Только GPT-5.6 Sol / Terra / Luna (без 5.5)

---

🧭 Онборд 2.0 — две оси

1) Сценарий — что сидится

• 🟢 minimal — маленький/новый проект
• 🟣 full — зрелый (score ≥ 5 или monorepo) + GOTCHAS / GLOSSARY / TESTING / deployment / nested CLAUDE

2) Глубина — как глубоко копает Codex

• ⚡ fast — паспорт
• 🔬 deep — forensic: entrypoints, потоки, wiki↔code, реальные тесты, ship, secrets (только имена)

/project-onboard
/project-onboard deep
project-onboard . --full --deep

---

🏃 Лейны, которые доезжают

Foreground Bash у Claude ~2 минуты — рвал живые прогоны.

Теперь только так:

lane-bg --dir "$ARTIFACT_DIR" --label agy-frontend -- \
  lane-exec --idle 600 --max 5400 -- ...

lane-wait --dir "$ARTIFACT_DIR" --once

Документация: docs/LANE-EXEC.md · implementers · dev-orchestrator.

---

📦 Установка / обновление

git clone https://github.com/VKirill/claude-lane-stack.git
cd claude-lane-stack && git checkout v1.1.0
./install.sh
export PATH="$HOME/.agents/bin:$PATH"

Уже стоит:

git pull && ./install.sh

Новая сессия dev-orchestrator / свежий spawn implementer — чтобы подтянуть agents.

---

📎 Ссылки

• 📘 README (EN)
• 📘 README (RU)
• 🧭 Onboard scenarios
• ⏱️ Lane exec + background
• 🎯 Routing Sol/Terra/Luna

---

Tag: v1.1.0 · Commit: main @ after this tag
Made with 💜 for solo builders · Сделано с 💜 для соло-разработчиков

🏭 Claude Lane Stack v1.0.0 — Your Personal AI Coding Factory

🏭 Claude Lane Stack v1.0.0 — Your Personal AI Coding Factory

One human. One project manager. A whole assembly line of AI workers.
Talk to a single Claude Code chat — it plans, delegates, reviews, and merges finished code to main for you.

This is the first public release of Claude Lane Stack: a file-based multi-agent orchestration kit for solo developers who want a conveyor, not another chat with an LLM.

---

✨ What's inside

🎛️ One chat to rule them all

• dev-orchestrator — the PM agent you actually talk to. It writes task cards, dispatches workers, checks results, and auto-merges to main. You never juggle five terminals or resolve branches by hand.

🚚 Worker lanes (all optional, auto-detected)

• ⚡ AGY (Antigravity) — fast write lane (lane-coder / lane-frontend)
• 🤖 Grok 4.5 — second write lane
• 🔍 Codex (GPT-5.6-sol) — read-only review lane, the primary quality gate
• 🧠 Claude-only mode works out of the box — no extra CLIs required

📋 File-based task contracts

• Every job is a YAML card under .agents/runs/ with owns_paths — workers can only touch files they own. No task database, no MCP server required: plain files, plain git.

🩺 agents-doctor

• One command scans your machine for available AI CLIs and writes a routing profile for the project: agents-doctor --apply . — 5 ready profiles from claude-only to full.

🚀 Onboarding & memory

• 🛂 /project-onboard — Codex (xhigh reasoning) fills CLAUDE.md, AGENTS.md and starter docs for a fresh repo
• 🔄 /resume-project — cold-start recap: board, progress, stalled tasks
• 📓 Living project memory templates: PROGRESS.md, LESSONS.md, decisions.md, session logs

🛡️ Guard rails

• Shell & code-quality hooks (guard_shell.py, guard_code_quality.py), session ledger, path-ownership checker, lane heartbeat & stall detection, night audit script

📚 Docs for humans, not just agents

• 🌍 README in 9 languages (EN, RU, 中文, 日本語, ES, DE, FR, 한국어, PT-BR)
• 🐣 Beginner guide in plain language (EN + RU) — for people who've never heard the word "orchestration"
• Deep dives: file contracts, routing, hooks, solo orchestration, MCP setups (lean & hybrid)

---

🏁 Quick start

# 1) Install the factory kit
git clone https://github.com/VKirill/claude-lane-stack.git
cd claude-lane-stack && ./install.sh
export PATH="$HOME/.agents/bin:$PATH"

# 2) In YOUR project — detect available workers
cd /path/to/your-project
agents-doctor --apply .

# 3) Start the PM and talk normally
claude --agent dev-orchestrator

---

🇷🇺 Коротко по-русски

Claude Lane Stack — маленькая ИИ-фабрика кода для одного человека.

Вы говорите с одним проект-менеджером (Claude Code). Он сам раздаёт задачи воркерам (AGY / Grok / Codex), проверяет работу и кладёт готовый код в main. Никаких пяти чатов и ручных merge.

• 📋 Задачи — файловые контракты в .agents/runs/ (никаких БД и обязательных MCP)
• 🩺 agents-doctor --apply . — сам определит, какие ИИ-инструменты установлены
• 🛂 /project-onboard — паспорт нового репозитория за одну команду
• 🐣 Гайд новичка на русском — простым языком, без жаргона
• 📖 README на русском

Автор: Кирилл Вечкасов · Telegram: Помогающий маркетолог

---

📦 Release contents

Component	Count
🤖 Agent definitions (Claude / AGY / Grok / Codex)	4 lanes
🧩 Skills	11
🔧 CLI tools in bin/	11
🎚️ Routing profiles	5
🪝 Safety hooks	4
🌍 README translations	9 languages

Full changelog: c204958...d9f0651

---

⭐ If the conveyor idea clicks for you — star the repo and tell a friend who's still copy-pasting code between five chats.