Skip to content

Releases: Overbrilliant/ob-1

OB-1 v0.3.9

Choose a tag to compare

@github-actions github-actions released this 25 Aug 08:54
21f2838

Release v0.3.9

OB-1 v0.3.8

Choose a tag to compare

@github-actions github-actions released this 19 Aug 21:31
e724426

Release v0.3.8

OB-1 v0.3.7

Choose a tag to compare

@github-actions github-actions released this 07 Jul 23:38

Release v0.3.7

OB-1 v0.3.6

Choose a tag to compare

@github-actions github-actions released this 07 Jul 23:28

Release v0.3.6

OB-1 v0.3.5

Choose a tag to compare

@github-actions github-actions released this 07 Jul 23:20

Release v0.3.5

OB-1 v0.3.4

Choose a tag to compare

@github-actions github-actions released this 07 Jul 22:52

Release v0.3.4

OB-1 v0.3.3

Choose a tag to compare

@github-actions github-actions released this 07 Jul 21:58

Release v0.3.3

OB-1 v0.3.2

Choose a tag to compare

@github-actions github-actions released this 07 Jul 21:23

Release v0.3.2

OB-1 v0.3.1

Choose a tag to compare

@github-actions github-actions released this 07 Jul 20:57

Release v0.3.1

OB-1 v0.3.0

Choose a tag to compare

@github-actions github-actions released this 07 Jul 11:32

OB-1 v0.3.0 — verified multi-agent rework

OB-1 now spends extra compute only when it can prove the extra compute was needed, and it grounds every multi-agent decision in your project's own checks instead of a model's self-opinion.

  • Fusion v2 is selection-first best-of-N: it generates N candidate attempts and picks the winner by an objective verifier signal auto-detected from your project (build, typecheck, tests, linters — whatever the repo actually has), with zero configuration. When no signal exists it falls back to synthesis rather than guessing.
  • Verified escalation is ON by default: a turn runs a single agent first, and only escalates to best-of-N after automated checks prove that single-agent attempt failed. Easy work stays 1× — the extra agents are spent when, and only when, a check says the first attempt did not pass.
  • /review: an independent reviewer that reviews the working diff and tries to REFUTE each of its own findings before reporting them, so you get the surviving issues instead of a wall of speculation. It also runs automatically after an escalated apply.
  • /deep: adaptive generate-vs-refine search for hard problems, using AB-MCTS-style Thompson sampling to decide at each step whether to widen (try a new approach) or deepen (refine an existing one), with verified early-stop.
  • Removed the measured-weak modes: personas, council, fanout, ledger, and the adaptive router — our own compute-matched evaluations showed they did not beat a single agent given the same budget. /review and /deep remain because they earned their place.
  • 42-task evaluation suite (every check proven to actually discriminate), plus the policy that a mode which cannot beat compute-matched Solo gets deleted.
  • First unit-test suite: 83 tests across the multi-agent core (fusion, reviewer, deep, evaluate) and the agent loop.
  • Fusion now handles honest prose answers (a candidate that correctly says "nothing to change" is no longer penalized against candidates that edited files).
  • Free-models router now fails a fusion candidate over to the next provider on a 429 instead of aborting the candidate.
  • Fixed keyless custom endpoints: an env/custom OpenAI-compatible endpoint with no API key is now treated as reachable instead of being skipped.