Changed
deep-high runs GPT-6 Astra at xhigh, and deep-low leads with GPT-6 Sol Fast. (#8714)
The escalation lane deep-high still has one rung, GPT-6 Astra, and now runs it at xhigh instead of high. The default deep lane deep-low starts on gpt-6-sol-fast, GPT-6 Sol's Fast (priority) tier, at medium, on the OpenAI and ChatGPT subscription providers that serve that tier. It then falls back to plain gpt-6-sol at medium, so GitHub Copilot and OpenCode Zen, which do not offer the Fast tier, keep the lane. GPT-5.6 Sol is no longer a deep-low model: a setup that serves only GPT-5.6 Sol now reports deep-low as unavailable instead of running it there (#8718). ultrabrain stays on GPT-6 Astra at max, and unspecified-high stays on Claude Opus 5.5 at max first. The configuration reference's provider-chain table, which still listed GPT-6 Astra as the first unspecified-high rung, now matches the shipped chain.
explore and librarian run a six-rung chain, and every DeepSeek Flash rung is DeepSeek V4.1 Flash. (#8115)
Both agents now fall back in this order: kimi-for-coding-highspeed (off), gpt-6-luna-fast (low), deepseek-flash (max), qwen3.7-plus, minimax-m2.7, claude-haiku-4-5. The two M3 rungs (OpenCode Go and the Coding Plan) and gpt-5.4-nano are gone, so an install whose only provider is that Coding Plan no longer gets a generated explore/librarian model from it. DeepSeek renamed its API model to deepseek-flash when V4.1 Flash shipped on 2026-09-10 and only routes the retired deepseek-v4-flash name to it temporarily; the pinned engine's deepseek catalog already lists deepseek-flash and no longer lists deepseek-v4-flash, so the old rung could not resolve on OmO Native. explore, librarian and the quick category now name deepseek/deepseek-flash, and the telemetry vocabulary reads the new id while still recognizing the old one.
Fixed
Claude subscription sessions run and advertise Claude Code 2.1.280 on every install. (#8713)
Beta.85 still shipped Claude Agent SDK 0.3.278, whose bundled Claude Code 2.1.278 is older than the 2.1.280 that Claude Opus 5.5 requires. Installs made with ignore-scripts=true in ~/.npmrc, or through a Bun install that blocked the package's postinstall, also skipped the step that raises the advertised version, so they kept sending claude-cli/2.1.251. The engine now pins the SDK at 0.3.280 and declares 2.1.280 itself. The first omo launch prepares an engine that install scripts never touched and records it in the engine directory, so later launches skip the work. If that preparation fails, omo prints the reinstall command and starts anyway. A newer Claude Code on your PATH (after claude update, for example) now runs instead of the bundled copy, and CLAUDE_CODE_EXECUTABLE still overrides both.
omob can build an engine version before its workspace packages reach npm. The development build now uses the dependencies already packed into its local engine tarball instead of asking Bun to resolve their unpublished versions from the registry. Platform-specific optional packages still install normally, and an incomplete bundle fails explicitly rather than fetching a replacement.
CI update-checker tests no longer depend on sibling test order. An unnecessary module mock leaked a fixed version into the registry-channel tests, failing all five assertions when the hook tests ran first. The hook now uses only its existing injected stub; runtime update behavior is unchanged. (#8678)
- 64ab487 Merge pull request #8726 from code-yeongyu/release/v5.0.0-beta.86-source-state
- dcf0dbc Merge pull request #8715 from code-yeongyu/fix/engine-prepare-on-launch
- 013f742 test(omo-native): gate the ignore-scripts launch case on the stamp, install with copyfile
- 14dac27 fix(omo-native): prepare the installed engine on launch when postinstall never ran
- 0b7764e Merge pull request #8720 from code-yeongyu/fix/deep-low-drop-gpt-5.6-sol
- 2403251 docs: stop describing GPT-5.6 Sol as a deep-low rung
- 72c02f0 fix(routing): deep-low drops its GPT-5.6 Sol rung and gate
- 17fdabb Merge pull request #8717 from code-yeongyu/fix/8115-deepseek-v41-agent-chains
- 2ca3be2 docs: follow the six-rung explore/librarian chain and deepseek-flash
- 22968ff feat(model-core): run explore and librarian on six rungs through DeepSeek V4.1 Flash
- 378adaa Merge pull request #8716 from code-yeongyu/feat/category-defaults-refresh
- 1216f71 docs: document the deep-high xhigh and deep-low Sol Fast defaults
- f5aaa60 feat(routing): deep-high on GPT-6 Astra xhigh, deep-low led by GPT-6 Sol Fast medium
- 7124994 Merge pull request #8708 from code-yeongyu/chore/untrack-qa-evidence-8703
- d1c09e6 docs(qa): name what replaces the tracked evidence for a reviewer
- 2f99bd2 docs(qa): evidence stays local, the PR body carries the summary
- ae3c325 chore(repo): stop tracking QA evidence and guard against it coming back
- 61b805d Merge pull request #8709 from code-yeongyu/fix/omob-unpublished-bundles-20260923
- a3dc65a docs(changelog): keep pending fixes out of beta.85 notes
- fbc65aa test(auto-update-checker): stop leaking the injected version stub
- eb9f0f8 fix(omob): install unpublished workspaces from the packed engine
npm i -g omo-ai@beta