lanes

01 How it works

One orchestrator. Many lanes. The same gates.

The orchestrator plans, dispatches, acts on review verdicts and integrates; it never reviews lane work itself. Workers implement only in assigned worktrees and return compact evidence. The main worktree handles orchestration, planning and integration.

• Lanes are worktrees

  Each lane is a git worktree with one agent in it. Lanes run in parallel, up to a cap you choose (default 6 concurrent agents, reviews included).

• Route by task

  Sol at medium effort handles routine coordination. Opus 5.5 defaults to medium in Claude Code. Use high effort for legal meaning, security, difficult defects and final acceptance; use Luna or scripts for structural checks.

• Gates don't change

  Each completion event goes to an independent reviewer, never the author or the orchestrator. It runs a deterministic scope check (hashes, out-of-scope paths, uncommitted work) and returns CLEAN or FIX for the exact revision.

v0.2.0 Release notes

Independent review. Enforced merges.

The orchestrator no longer reviews its own lanes, and unreviewed work can no longer reach the main branch.

• A reviewer per lane

  Every completed lane goes to an independent reviewer: lane-reviewer (Opus 5.5, medium) in Claude Code or gpt-5.6-terra (medium) in Codex. FIX findings go back to the same lane.

• Gated in code

  lane/integrate.sh is the only merge path. It requires a CLEAN verdict for the exact lane head; small prose edits it classifies as trivial skip review.

• Hardened relays

  External models refuse the main checkout and .env files, the relay guard blocks command smuggling, relays run on macOS, and pushes require a clean secret scan.

Read the full 0.2.0 changelog

02 Routing

Triage first. The tier picks the model.

Use Sol at medium effort for routine coordination and Opus 5.5 at medium effort as the Claude Code default. Raise effort for legal meaning, security, difficult defects and final acceptance. Use Luna or scripts for structural checks, and reserve DeepSeek and GLM for simple tasks.

Recommended orchestration defaults are Codex Sol at medium effort and Claude Opus 5.5 at medium effort. Use high effort only for legal meaning, security, difficult defects and final acceptance. Each task brief states a token cap; short drafts stop at 500 output tokens. Record input, cached input, output, failures and acceptance in the lane usage ledger. Completion events trigger an independent review instead of repeated status scans. Model overrides: DEEPSEEK_MODEL, GLM_MODEL, CODEX_MODEL, CODEX_EFFORT.

• Review before mergeEvery lane gets an independent reviewer, never the orchestrator. FIX findings go back to the same lane. lane/integrate.sh is the only merge path and refuses anything without a CLEAN verdict for the current revision; only small prose edits it classifies as trivial skip review.
• Workers stay in worktreesmodel-relay refuses the main checkout and any worktree holding .env files.
• Either host reviews.Claude Code uses lane-reviewer; Codex uses codex-review.

03 Install

Clone it into your skills folder.

The installer checks prerequisites and both keys, tells you exactly what to add if something is missing, and backs up any file it replaces.

git clone https://github.com/thebpandey/lanes ~/.claude/skills/lanes
bash ~/.claude/skills/lanes/install.sh            # installs relays, configs, and the opus-lane/sonnet-lane/lane-reviewer/deepseek/glm/codex subagents
bash ~/.claude/skills/lanes/install.sh --smoke    # optional: one tiny task through each provider
bash ~/.claude/skills/lanes/install.sh --watchdog # optional (Linux): timer that cleans orphaned agent work

04 Use

Say it in any git project.

Start Claude Code in the project and talk to the orchestrator. Status updates focus on completion events and changed lane states.

• /lanes start

  Audits the project (instructions, git history, worktrees, tracker), shows what's done, in progress, ready and blocked, triages each task into a tier, asks a few setup questions one at a time, then runs the lanes.

• lane update

  Blockers, completed events and changed lane states in a compact 10–12-line summary.

• lane update all

  A concise view of every workstream task and its current state.

• pause lanes · resume the lanes

  Safe pause notes per lane, and resuming into the same worktrees.

Direct commands no Claude tokens

Run from the shell. Model overrides: DEEPSEEK_MODEL, GLM_MODEL, CODEX_MODEL; set CODEX_EFFORT=high for high-risk review. To add another Anthropic-compatible provider, drop a ~/.claude/<name>.json next to the others.

cd <worktree> && RELAY_WAIT_S=6000 model-relay deepseek < brief.md   # or: model-relay glm
cd <worktree> && RELAY_WAIT_S=6000 codex-review main owned.txt < checklist.md
~/.claude/skills/lanes/lane/new-worktree.sh <branch>                  # worktree + monorepo node_modules wiring
~/.claude/skills/lanes/lane/lane-status.sh                            # evidence table for "lane update"

05 Safety notes

Read these before the first run.

• Data leaves your machine

  DeepSeek and GLM receive whatever the task gives them, and run with full shell access inside their worktree, with .env reads denied. Run them only in worktrees, and never give them secrets, credentials or customer data.

• Costs

  total_cost_usd in a worker's output is priced as if the run were Claude, so ignore it. Check your real spend on the DeepSeek balance page and on https://openrouter.ai/api/v1/key.

• Self-identification

  Workers call themselves "Claude" because Claude Code's system prompt says so. The job record's model: line shows the real model.

• Approvals

  The skill asks before pushes, deploys, external writes, and anything hard to undo.

06 Uninstall

Three lines to remove it all.

rm -f ~/.local/bin/{claude-via,claude-deepseek,claude-glm,model-relay,relay-guard,relay-spawn,codex-review}
rm -f ~/.claude/{deepseek,glm}.json ~/.claude/agents/{deepseek,glm,codex,opus-lane,sonnet-lane,lane-reviewer}.md
systemctl --user disable --now lanes-watchdog.timer 2>/dev/null; rm -rf ~/.claude/skills/lanes