Repository navigation

feat(training): train a bundle's own agents with TRL's harness rollouts - #434

Merged

earakely-scale merged 10 commits into

mainfrom

Nov 2, 2026

Merged

feat(training): train a bundle's own agents with TRL's harness rollouts#434
earakely-scale merged 10 commits into
mainfrom
feat/harness

Commits on Nov 1, 2026
- feat(training): train a bundle's own agents with TRL's harness rollouts
  
  earakely-scale
  and
  claude
  committed
  73aa7f4
- fix(training): give the solver alone the capture proxy as its model endpoint
  
  earakely-scale
  and
  claude
  committed
  ec820d0
- test(training): give the unprompted agent its own endpoint
  
  earakely-scale
  and
  claude
  committed
  3bf983e
- fix(training): leave out untrainable tasks before preparing the bundle, and a solver that judges itself
  
  earakely-scale
  and
  claude
  committed
  9246926
- test(training): name the self-judged task's judge agent
  
  earakely-scale
  and
  claude
  committed
  1b6ddad
- test(training): the shipped harnesses' smoke tasks through the capture proxy, on Docker
  
  earakely-scale
  and
  claude
  committed
  98d3a3f
- fix(training): stop and tear down a rollout in flight on close, and run only on local Docker
  
  earakely-scale
  and
  claude
  committed
  46d6d9b
Commits on Nov 2, 2026