jumplane
Tell jumplane how big and how risky a task is. It answers with the tool, model, effort and exact command to run, taken from rules you accepted. When a usage limit is close, it answers with the next open lane instead.
Why
Coding agents on a subscription stop when a usage window fills, and the host does not switch for you. Claude Code's model configuration page says rate-limit errors "never trigger a switch" to its fallback model.
Evidence: the phrase "usage limit" appears in 3,575 issues on the Claude Code repository and 1,975 on the Codex repository (GitHub issue search, 3 Nov 2026; the totals include off-topic hits). One of them, claude-code issue 16157, had 727 reactions that day.
jumplane is not a proxy and holds no keys. It reads the limit numbers your tools already show, and it returns a command for you or your agent to run.
Features
- One screen,
jumplane status: each pool, percent used, reset time, age of the reading, holds, and the lane that runs next. jumplane pickmaps size, risk and complexity to one of three classes with a published table, then returns the first open lane in your order.- A pool at the hold line (90 percent by default) on any window with a valid reading is passed over, and the pick moves to the next lane in the same class.
- Every lane passed over is listed under
skippedwith its cause and the command or wait that would open it. An unverified lane is passed over for the next lane in the same class; work never moves to another class. jumplane reportrecords the outcome, tokens and receipt path. Each decision is one line in a plain log file you can read.- No proxy, no API keys, no prompt reading, no telemetry. The pick code has no network or process imports. Only the probe used by
initandverifyimports those modules directly, which a test checks. Python standard library only. - Rules and code choose the route, and model output never does.
dependencies = [], enforced by a test)Test command, from the repository root: PYTHONPATH=src python -m unittest discover -s tests -t .
How it works
Quick start
- Install uv, then jumplane from PyPI:
uv tool install jumplane. Check:jumplane --versionprints0.1.1. - Run
jumplane init. It looks for Claude Code, Codex and a local Ollama server and prints proposed rules without writing them. Check: the answer says"status": "proposed". - Read the rules, then run
jumplane init --accept. Check: the answer says"status": "accepted". - Verify each hosted route once, for example
jumplane verify codex-standard. This sends one short prompt and uses a little of that plan. Check:"status": "verified". - Run
jumplane status. Check: each class showsnextand a lane, or afix:line to run. - Run
jumplane pick --size small --risk low --complexity low, launch the returned command, thenjumplane report <pick_id> done. Check: the answer says"status": "recorded".
Claude Code shows its limits only through its status line. jumplane init prints the one settings line to add. Without it the Claude pool has no reading and stays unavailable.
Open-weight models
jumplane routes to any model your local Ollama server lists. Dense and mixture-of-experts models are launched the same way, because jumplane starts ollama run and does not load weights itself. The table lists the models launched through jumplane on one PC (RTX 4070 Super, 12 GB) on 10 Nov 2026.
| Ollama tag | Kind, from the publisher's card | Publisher's own score | Launch check through jumplane |
|---|---|---|---|
qwen3.5:9b | 9B. The card does not say dense or mixture-of-experts for this size | LiveCodeBench v6: 65.6 (card) | Verified, 24 s |
gemma4:12b | Dense, 11.95B | LiveCodeBench v6: 72.0% (card) | Verified, 19 s |
gpt-oss:20b | Mixture-of-experts, 21B with 3.6B active | SWE-bench Verified: 53.2 at medium reasoning (card) | Verified, 22 s |
qwen3.8:27b | Dense, 27B | LiveCodeBench v6: 90.3 (card) | Verified, 53 s |
qwen3.6:35b-a3b (4-bit build) | Mixture-of-experts, 35B with 3B active | SWE-bench Verified: 73.4 (card) | Verified, 38 s |
Read the scores with care. Each one is the publisher's own figure for the full-precision model. jumplane has not reproduced any of them, and a 4-bit local build can score lower. "Verified" means the model answered one short prompt when launched by jumplane verify. It says nothing about quality. init writes one local lane; more local lanes are added by editing the rules file.
Status and limits
| Built and tested | Not built | Not measured |
|---|---|---|
init, verify, pick, report, override, hold, release, status, statusline, reading; Claude status line reader; Codex log reader; local Ollama pool |
Signed rules; MCP server; OpenCode; metered API budgets in money; a tuning command that proposes rule changes from your outcomes; an audit summary command; any chat or desktop surface for people who do not use a terminal | Whether a jumplane pick beats the host default on quality, cost or completed work. No benchmark has run. Setup time for a new user. Behaviour on macOS and Linux (tests ran on Windows only) |
- jumplane is rule based. It does not predict quality.
- The Codex reader rests on an undocumented log format. Any surprise there means "no reading", and the pool is then unavailable until you record one.
- The caller states the task shape. A caller can understate it down to the minimum class your rules set.
- jumplane does not stop two callers from running the same work, and it cannot switch a session that is already running.
- The event log is a plain file. Anyone with your user account can edit it.
Sponsor
Sponsorship is not open yet. When it opens, the first money goes to the benchmark: the same task set run with and without jumplane on paid plans, with the task list, commands and raw results published whether or not they favour jumplane.
If your company makes a local-model runtime, a coding agent or a usage tool and wants that result to exist, watch the repository once it is public. A sponsor link will appear here only after the profile behind it is approved.