Changelog
Readable release notes for a project that ships in public.
OB-1 releases should be easy to inspect: what changed and why it matters.
OB-1 0.3.7
- Free-model catalog refresh now runs in the background when Free models are active, so free and paid catalogs stay current without waiting for a model call.
- Clarified the catalog timing: free users get newly released free models after 30 days; hosted plans get them immediately.
OB-1 0.3.5
- Free users now get newly released free models after the 30-day promotion window, while hosted plans get them immediately.
- The CLI verifies signed catalog updates before activating refreshed models and keeps separate free/paid caches.
- Updated pricing and plan copy around immediate access to new free models.
OB-1 0.3.4
- Polished the first-run onboarding messages for the Free, custom endpoint, and Hosted frontier paths.
- Hosted frontier setup now sends you straight to account signup from the CLI instead of asking a second in-terminal account question.
- Updated the fresh-install smoke check for the new free-models activation summary.
OB-1 0.3.3
- Changed /mode into the execution picker: Auto, Act, and Plan now live together in one place.
- Added Auto mode for no-questions-asked execution: edits and commands run without approval prompts.
- Moved deliberate best-of-N orchestration to /fusion, with /solo to return to the normal single-agent path.
OB-1 0.3.2
- Fixed installed /login being hijacked by stale local-server shell exports.
- Localhost OB1_SERVER overrides now require an explicit OB1_ALLOW_LOCAL_SERVER=1 opt-in.
OB-1 0.3.1
- Added in-session /login and /logout commands so browser authentication can happen without leaving the OB-1 TUI.
- Added /subscribe as a visible subscription-page command, alongside /upgrade.
- Changed /plan into a Plan/Act toggle: run /plan once for read-only planning, then run /plan again to return to normal Act mode.
OB-1 0.3.0
- Multi-agent, rebuilt around one rule: a mode ships only if it beats a single agent at equal compute. We measured our own heavier modes against compute-matched Solo on our own eval suite — and deleted every one that lost. The mixed-panel modes that ran several reviewer styles as a committee couldn't beat a single strong agent at the same token budget; a weak member just dragged the result down while multiplying the cost. What survived is what measured a real win.
- Modes are now just Solo and Fusion. Fusion v2 is best-of-N with a real selector: it runs several candidates and selects the one that passes your project's actual tests and checks, verbatim — candidate selection grounded in your project's real tests, not model vibes.
- Verified escalation, on by default: Solo escalates to a Fusion best-of-N pass only after your automated checks prove the single-agent attempt failed, at most once per turn. You never pay for a committee you didn't need. Toggle with /escalation.
- New /review: an independent reviewer that tries to refute its own findings first and reports only the correctness bugs that survive — each with a concrete failure scenario and a file:line citation — then offers to fix them.
- New /deep: an adaptive generate-vs-refine search (AB-MCTS-style) that widens or deepens based on the real verifier signal, then applies the best result through the normal gated apply loop.
- New /eval: run the compute-matched test yourself — does a mode beat Solo at equal tokens? Read-only subagents for parallel investigation are unchanged and still on by default.
OB-1 0.2.0
- Free models are now built into the CLI: an in-process router with no second process, local server, or dashboard to run.
- Keyless providers answer the first message with no key and no account. Keys live in one editable file at ~/.ob1/keys.env and activate or deactivate on your next message, no restart.
- Automatic failover, rate-limit cooldowns, provider health checks, and routing strategies (balanced, smartest, fastest, reliable, priority) are built in.
- New /free command manages the pool — status, keys, strategy, and health — and /models can pin a specific free model or leave it on auto.
- Existing FreeLLMAPI setups migrate automatically to the built-in router on next launch; the external FreeLLMAPI service is no longer needed or bundled.
OB-1 0.1.5
- Trust gate on by default: the first run in a folder OB-1 has not seen starts in ask mode, so risky actions wait for your approval.
- Onboarding fixes: pressing Esc at the first picker starts the free path, and the free path no longer dead-ends when no model is selected.
- Friendlier crashes: unexpected errors now show a readable message instead of a raw stack trace.
- Process cleanup: foreground shell commands are group-killed and worktree processes are reaped by working directory, so nothing lingers after you stop a run.
- npm installs now work without Bun, via a files allowlist and a Node shim binary.
- Browser-opened sign-in and checkout links are tagged with their source for launch attribution.
OB-1 0.1.4
- FreeLLMAPI became the default free path: OB-1 pooled provider free tiers and accepted your provider keys.
- Runtime BYOK routing now recognizes OpenRouter, OpenAI, Gemini, Groq, and custom OpenAI-compatible endpoints.
- Named presets cover OpenRouter, Ollama, LM Studio, llama.cpp, vLLM, Groq, and custom endpoints.
- Launch docs, demo assets, update-check plumbing, Biome linting, and contributor templates are in place.
OB-1 0.1.3
- Pointed the default managed server at the production endpoint.
- Hardened workspace boundaries, SQL checks, secret handling, and SSRF protections.
- Improved terminal behavior around viewport bounds, history recall, and stop handling.
OB-1 0.1.2
- Fixed standalone binary startup.
- Added release-install checks for packaged binaries.
- Improved installer and runtime packaging paths.
OB-1 0.1.1
- Moved release links and package metadata to the Overbrilliant organization.
- Updated install scripts and documentation for the public repo location.