Repository navigation
v0.61.0: your agents, your models, no dead ends
Ouroboros now runs on whatever model your agent already uses. It no longer swaps in model ids of its own, and it never sends a request your model would reject. We also took the console from an empty folder to a finished result in a real browser, and fixed every snag we hit along the way.
🧠 Your agent's models, inherited
Earlier releases pinned exact model ids for execution, such as claude-opus-5-5 and gpt-6-astra. Those ids break on Bedrock, Vertex, gateways, proxies, and accounts without access. Now each runtime picks for itself (#3021):
| Agent | What Ouroboros sends |
|---|---|
| Claude Code | haiku / sonnet / opus. Claude Code resolves each to the newest model for your provider, whether that's Anthropic, Bedrock, Vertex or a gateway. |
| Codex | Nothing. Codex runs on the model in your Codex config, unless you set tier models yourself. |
| Gemini CLI | flash / pro, Gemini's own evergreen aliases. |
- Saved configs carry over. Untouched shipped ids in old configs follow the new defaults automatically. Models you chose yourself are kept exactly as written.
- Resumes don't change course. A run you resume keeps the models it started with.
- The console shows what "default" means. For example: Codex CLI default (gpt-6-astra).
🛡 No more rejected requests
Newer Claude models (Opus 4.7 and later, Sonnet 5, Haiku 5.5, Fable, Mythos) return HTTP 400 for temperature, top_p and assistant prefill. Ouroboros now sends those only to the older Claude families documented to accept them (#3019).
- Covered: aliases,
default, provider-prefixed ids, Bedrock and Vertex forms, and models released after this version. - Also fixed: Claude routes that LiteLLM's bundled model list doesn't know yet. They no longer get
budget_tokensor a forced tool call.
🖥 The console, from goal to result
We drove the console end to end with Claude Code and Codex: Interview, Suggest an answer, Write the spec, Hand off, Result, plus Ask and Auto. These are the fixes from that run (#3022):
- Start in an empty folder. The console tells you up front when a folder isn't a git repository. At hand-off, one click creates it with an empty first commit. It never writes your git config, and it never touches an existing repository.
- Auto shows what it did. You get the outcome, a link to the run board, and the files it changed in your folder.
- Missing agents are caught before you start. If an agent isn't installed, the console says so and shows the install command, instead of failing mid-interview.
- Your project's instructions, not the console's. Interviews and specs now read your project folder's instructions, not those of the folder you launched the console from.
- Fully translated run states, and Auto jobs now appear in Activity.
- Fewer interview questions in the console (#3009). Code facts are answered from your repository. Brownfield mode starts only when you pick a project, and the readiness meter follows the server's real gate.
🧹 Also in this release
- Telemetry is lighter (#3013). One anonymous daily usage rollup replaces per-call events. It respects every opt-out, and you can read the event set in TELEMETRY.md.
- Skills and agent prompts are re-baselined for current models (#3015, #3016, #3017). Instructions are plainer and come with their reasons. Short checks run inline rather than spawning sub-agents. Worker prompts no longer name Claude-only tools.
- Interview fixes (#3007, #3011). A refused "done" now keeps the pending question instead of losing your next answer. Interviews also record which surface asked each question and why they closed, so question counts can be compared across surfaces.
Install or upgrade in one line
The same command installs fresh or upgrades in place, and it keeps your configuration. The console opens when it finishes.
# macOS / Linux / WSL 2
curl -fsSL https://raw.githubusercontent.com/Q00/ouroboros/main/scripts/install.sh | bash# Windows (PowerShell 5.1+ or pwsh 7+)
irm https://raw.githubusercontent.com/Q00/ouroboros/main/scripts/install.ps1 | iexAlready installed? Run ouroboros update, or ooo update inside your agent.
⭐ If Ouroboros saves you time, a star helps others find it.
What's Changed
- ci(release): fix Intel TUI runner and tap trust; add tag dispatch by @Q00 in #3004
- docs(agents): re-baseline interviewer and seed-architect prompts by @Q00 in #3015
- docs(skills): re-baseline QA, evaluate, unstuck and executor prompts by @Q00 in #3016
- docs(skills): re-baseline setup and footer wording for current models by @Q00 in #3017
- fix(interview): keep the pending question after a refused done by @Q00 in #3007
- fix(gui): ask fewer interview questions in the console by @Q00 in #3009
- fix(providers): never send sampling or prefill a Claude model rejects by @Q00 in #3019
- feat(models): route tiers to each runtime's own default models by @Q00 in #3021
- perf(telemetry): send one daily usage rollup instead of per-call events by @Q00 in #3013
- fix(gui): remove dead ends from the console goal-to-result flow by @Q00 in #3022
- feat(interview): record calling surface and closure gate per turn by @Q00 in #3011
- chore(release): prepare v0.61.0 by @Q00 in #3023
Full Changelog: v0.60.0...v0.61.0