lanes

lanes coordinates independent software tasks in Git worktrees from Claude Code or Codex. The orchestrator plans, dispatches and integrates; each worker gets a short, complete task in its own worktree, and an independent reviewer checks every completed lane.

Current release: 0.2.0 · GitHub Pages · Changelog

Plain-language walkthrough and case study: lanes on bhaskarpandey.com. Built by Bhaskar Pandey, AI & Automation Solutions Engineer.

Infographics

Workflow: Plan → Brief → Lanes → Review → Integrate → Verify. Each lane gets its own independent reviewer; FIX goes back to that lane, trivial prose edits bypass review, and integrate.sh gates every merge.

Task handoff: each brief defines inputs, owned.txt, checks.sh and a stop condition; an independent reviewer returns a revision-bound verdict in review.md.

How work moves

1. The orchestrator checks the project's instructions, current tracker and worktrees, then makes a short dependency-aware plan.
2. Each brief names the task inputs and completed boundary, exact file ownership, acceptance checks, token limits and stop condition.
3. Workers implement only in assigned worktrees. DeepSeek and GLM receive simple, bounded tasks only; neither receives secrets or private data.
4. Completion events replace repeated status scans. Each completed lane goes to an independent reviewer (Opus 5.5 at medium effort in Claude Code, gpt-5.6-terra at medium in Codex), never to the orchestrator. The reviewer runs lane/scope-check.sh (changed-file hashes, out-of-scope paths, uncommitted work) and returns a revision-bound CLEAN or FIX verdict.
5. The orchestrator acts on the verdict: CLEAN for the lane's current revision is integrated; FIX findings go back to the same lane, and the new revision is reviewed again. lane/integrate.sh is the only merge path and refuses unreviewed work. Trivial prose edits (≤ 40 lines of .md/.txt-style files, no deletes, no agent-instruction, control or legal files) skip review; the script decides, not the orchestrator. The main worktree is for orchestration, planning, integration and session-detail archives before context compaction.

Model and effort routing

The per-task usage ledger records input tokens, cached input, output tokens, failures or retries, and acceptance. Default caps are 2,000/500 tokens for structural work or short drafts, 12,000/4,000 for routine work, and 24,000/8,000 for moderate work (input/output). Stop and rebrief when a cap is reached. Keep worker summaries to 10–12 lines and routine tool results to 1,000–2,000 tokens.

Requirements

• Claude Code, git, python3, curl, and bash on Linux or macOS for the Claude provider relays. The watchdog timer is Linux/systemd only.
• DEEPSEEK_API_KEY and OPENROUTER_API_KEY for external worker providers. Export them in your shell startup file; this skill never stores their values.
• Optional: Codex CLI for Codex review and Luna structural checks; Docker for disposable test databases; Beads (bd) for issue tracking.

Install

git clone https://github.com/thebpandey/lanes ~/.claude/skills/lanes
bash ~/.claude/skills/lanes/install.sh

The installer checks prerequisites, keeps backups of replaced files and tells you how to configure missing keys. Restart Claude Code after installation so its lane agents load.

It installs the DeepSeek and GLM provider relays, Codex review launcher, worktree helpers, and Claude agent definitions in ~/.claude/agents/.

Use

Start Claude Code in a Git project and use /lanes start, lane update, lane update all, pause lanes, or resume the lanes. The orchestrator checks the active tracker and resumes from existing worktrees without resetting local changes.

Direct commands:

cd <worktree> && RELAY_WAIT_S=6000 model-relay deepseek < brief.md  # or glm for a simple bounded task
cd <worktree> && RELAY_WAIT_S=6000 codex-review main owned.txt < checklist.md   # Codex lane review
~/.claude/skills/lanes/lane/scope-check.sh <base> owned.txt   # the reviewer's scope gate; checks.sh beside owned.txt holds the acceptance checks
~/.claude/skills/lanes/lane/integrate.sh [--classify] <lane-dir> <branch>   # from the main worktree; the only merge path
~/.claude/skills/lanes/lane/new-worktree.sh <branch>
~/.claude/skills/lanes/lane/lane-status.sh
~/.claude/skills/lanes/lane/secret-scan.sh <base> [head]  # required before any push; exit 0 = clean

CODEX_MODEL selects the review model (default gpt-5.6-terra); CODEX_EFFORT selects medium (default) or high. DEEPSEEK_MODEL and GLM_MODEL override provider models.

Safety

• Work in assigned worktrees. Never give external providers credentials, .env contents, customer data or private records.
• Never use shared or live databases for tests; use a task-owned disposable environment.
• Preserve user changes, stage explicit paths, and follow the project's authorization for commits, pushes, deployments and external writes.
• Before context compaction, create a full session-detail archive once per session, then incrementals from verified checkpoints before later compactions.

MIT licensed. Not affiliated with Anthropic, OpenAI, DeepSeek, Z.ai or OpenRouter.