WarmStart 0.2.0 (alpha)

WarmStart 0.2.0 (alpha)

WarmStart is a local, network-free repository knowledge compiler for coding
agents. It inspects a repository without running its code, proposes a small
set of evidence-backed facts that could change what an agent does, and writes
only human-approved facts into a managed section of AGENTS.md. It can later
report facts whose evidence has gone stale.

This release renames Coldstart to WarmStart while preserving the coldstart
command and imports, .coldstart/ state directory, and legacy managed markers.

Verification

• 234 tests passed from source and from the installed wheel with network
  connections blocked.
• Ruff lint and formatting checks passed.
• Strict mypy checks passed across 18 source files.
• Wheel, source archive, CLI aliases, import aliases, and an empty-repository
  workflow were verified in fresh Python 3.12.7 environments.
• The public repository was built from an independently reviewed 50-file
  manifest; private review material and real session data are not included.

Evidence boundary

WarmStart has not yet been shown to improve agent performance. One
GPT-5.6 Sol consumer run had an empty managed WarmStart section, no matched
control, and a recorded validity incident; it is not evidence of a WarmStart
effect. The README proposes a matched trial and welcomes positive, null, and
negative results across model families, including Droid.

Licensed under Apache-2.0. Contributions use DCO 1.1 sign-off.