WarmStart

The problem: slow handoffs. A fresh agent or subagent often starts cold. It rereads the repository, guesses the project's habits, and sometimes learns the one fact that mattered only after making the mistake. That adds up when you spawn a lot of them.

What WarmStart does about it: it turns repository evidence into short project notes. Its standalone review step records which notes are approved for publication.

• Short, not a dump. A handful of approved facts, not a vault of notes the agent has to wade through.
• Checked. Each fact points to evidence in the code and needs a recorded approval before publication.
• Stays current. It flags a fact when the evidence behind it changes.

Faster handoffs are the goal, not a measured result yet.

In the saved demos, the approvals were test approvals made by an agent. An approval record shows that a note was approved; on its own, it doesn’t show that a person reviewed it.

How it works

• It reads the repository without running its code.
• It finds a small number of evidence-backed facts that could change an agent's next move.
• It asks focused questions about things the code can't reveal, like what you actually intend.
• Publication requires an approval record. Approved facts are written into a marked section of the project’s AGENTS.md file. Later fault tests found that a concurrent publish can overwrite edits outside that section; this needs repair.
• It checks later whether a fact has gone stale, because the evidence behind it changed.

It runs on your machine and calls no AI model or network. Inspection and compilation can keep working state outside the repository; publishing deliberately updates its guidance file. The full evidence and the trial plan are on the WarmStart Lab page.

Read how it was built ↓

What we wanted

I’d heard of people passing work between agents through a big vault of notes. So I asked a simple question: does a newly spawned agent actually inherit any of that? In my experience, it didn’t: fresh agents kept starting cold. (This is my memory of how it began, not a timestamped record.)

The first version, and what it got wrong

The tool started life as Coldstart in early August 2026. The crew then ran it across ten real repositories, which exposed four defects in its own detection and filtering. It also showed something humbling: hand-written project guidance often beat what the tool generated. The crew kept that result and changed the design.

On one measured slice of three repositories, the filter went from letting through 231 items and holding back none to letting through 96 and holding back 70. That shows the filter changed. It doesn't show that agents got better.

The trial that couldn't count

A real-use trial ran for about 3 minutes 42 seconds. The agent in it received an empty WarmStart section, there was no comparison group, and a problem with the trial's validity was recorded. That trial did not establish whether WarmStart helped.

Getting it ready for the public

The tool was renamed WarmStart 0.2, and old commands and saved state kept working. The crew then:

• built the public version from an explicit list of files, rather than uploading the whole working folder
• wrote a clearer explanation and safety rules
• set up release checks anyone can repeat

The snag, and the repair

An independent audit said not ready. It found:

• credits that needed correcting
• the empty trial described too strongly
• private files too close to the public ones
• a sample output and licensing details that needed fixing

One round of repairs fixed all of it, and a second audit said it was ready for a license decision. Steve chose Apache-2.0. The 50-file repository and release v0.2.0 were then published and checked again after going public.

There was one small scare: the first clean-install test failed. The test itself had typed the install command wrong. With the command fixed, everything passed. It wasn't a product bug.

What you can do with it now

You can try the alpha in a disposable repository. Before trusting its publisher with shared working files, we need to fix the recovery issues found on October 9. Then comes the fair test: the same agent task, with and without WarmStart.

• Who did what:
  • Steve: product direction, and the question that exposed the problem.
  • Faye (Claude Opus): design exploration and a large part of the implementation.
  • Chad (OpenAI Codex): integration, verification and the public release; also design review.
  • Gemini 3.1 Pro: independent product and security review.
  • GPT-5.6 Sol: the one real-use trial (empty WarmStart section, no comparison group, validity problem recorded). It is not evidence that WarmStart helps.
• Timeline:
  • Coldstart development history: August 2–4, 2026.
  • Public-release cleanup and audits dated September 28, 2026.
  • Exact start time, total build time, active model time and cost: not recorded.
• Ten-repository sweep: 4 defects found in the tool; hand-written guidance often beat generated candidates. Both results are published in the repository.
• Three-repository filter change: 231 let through / 0 held back → 96 / 70. This measures the filter, not agent performance.
• Real-use trial: 221.841 seconds; empty context, no control, validity incident.
• Release checks: 234 tests from source and from the installed wheel with sockets blocked; Ruff lint and formatting; strict mypy; build and contents; alias and import checks; an empty-repository workflow run with state kept outside the target.
• Release details: 50-file public manifest; release commit a36ce46; license Apache-2.0 with DCO 1.1 sign-off.
• October 9 fault checks: six selected publication schedules in disposable repositories; three still failed after ordinary recovery. These are reproduced cases, not a failure-rate estimate. No learner was launched, and no automatic arrival connector was verified.
• Lane counts: the release records credit roles, not a full count of every agent and wave over the tool's history. We haven't guessed one.