Pre-release v0.1.1 on PyPI, 11 Nov 2026. MIT licence.

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.

jumplane status and pick, recorded 10 Nov 2026 Terminal capture. jumplane status shows the Claude weekly window at 15 percent, read 2 days 11 hours earlier, and the Codex weekly window at 3 percent. With the hold line set to 10 percent, the next standard lane is codex-standard. The pick for a medium task returns codex-standard and lists claude-standard as passed over, with two reasons: no valid five-hour reading, and 15 percent used against a 10 percent hold line. $ jumplane status ollama local, 6 models at last probe claude 300m 3% reset passed read 2d11h ago (statusline) claude 10080m 15% resets 9h55m read 2d11h ago (statusline) codex 10080m 3% resets 1d2h read 11h37m ago (codex-log) light: next ollama-light (qwen3.5:9b) standard: next codex-standard (gpt-6.1-terra) strong: next codex-strong (gpt-6.1-sol) A: 15% used, over the 10% hold line B: standard work goes to Codex $ jumplane pick --size medium { (excerpt of the JSON answer) "status": "picked", "lane": "codex-standard", "reason": "medium size, medium risk, medium complexity gives class standard; passed over claude-standard", "stdin": "closed", "skipped": [ { "lane": "claude-standard", "reasons": [ "pool claude has no valid reading for its 300 minute window", "pool claude at 15% of its 10080 minute window, holds at 10%" C: the hold, as a reason The returned codex command then ran with stdin closed: exit 0 after 6 s, 5,130 tokens, reported as done. Windows 11, jumplane 0.1.0.dev0 at commit 4bb382c, 10 Nov 2026 07:04 UTC. Rules copy with hold_at_percent 10.
Figure 1. Hold and move, 10 Nov 2026; hold line 10% in a rules copy. Claude's 7-day window read 15%, 2 days 11 hours old (A), so standard work went to Codex (B). The pick gave two reasons (C): over the line, no valid 5-hour reading. Limits: one PC, build 0.1.0.dev0 at 4bb382c; JSON shortened, one line wrapped.

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

78tests pass on one Windows 11 PC, 10 Nov 2026 (command below)
0runtime dependencies (dependencies = [], enforced by a test)
3hosts with a launch command: Claude Code, Codex, Ollama

Test command, from the repository root: PYTHONPATH=src python -m unittest discover -s tests -t .

How it works

How a pick is made The caller passes a task shape. A fixed table turns it into a class. jumplane walks the lanes of that class in the owner's order and checks each one: verified, not held, reading under the hold line. The first open lane is returned with its launch command. If none is open the answer is wait or unavailable with fix commands. The caller reports the outcome and every step is appended to an event log. 1. Task shape size, risk, complexity 2. Class light, standard or strong (fixed table) 3. Lanes in your order verified? not held? reading under hold line? None open wait or fix commands Open lane found lane + launch command 4. You run it, then report the outcome 5. Event log one JSON line per step Readings (inputs to step 3) Claude status line, Codex log, Ollama probe
Figure 2. How a pick is made. Solid boxes are built and tested. The dashed box is also built; it is dashed because it is the refusal path. Left out: signing of the rules file, which is planned and not built.

Quick start

  1. Install uv, then jumplane from PyPI: uv tool install jumplane. Check: jumplane --version prints 0.1.1.
  2. 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".
  3. Read the rules, then run jumplane init --accept. Check: the answer says "status": "accepted".
  4. 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".
  5. Run jumplane status. Check: each class shows next and a lane, or a fix: line to run.
  6. Run jumplane pick --size small --risk low --complexity low, launch the returned command, then jumplane 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 tagKind, from the publisher's cardPublisher's own scoreLaunch check through jumplane
qwen3.5:9b9B. The card does not say dense or mixture-of-experts for this sizeLiveCodeBench v6: 65.6 (card)Verified, 24 s
gemma4:12bDense, 11.95BLiveCodeBench v6: 72.0% (card)Verified, 19 s
gpt-oss:20bMixture-of-experts, 21B with 3.6B activeSWE-bench Verified: 53.2 at medium reasoning (card)Verified, 22 s
qwen3.8:27bDense, 27BLiveCodeBench v6: 90.3 (card)Verified, 53 s
qwen3.6:35b-a3b (4-bit build)Mixture-of-experts, 35B with 3B activeSWE-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 testedNot builtNot 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)

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.