recommended-models: reorder the default ladder (Opus 5.5 -> Fable 5.1 -> Kimi K3 -> Astra -> Sol -> GLM 5.3) and rank provider lanes

[code-yeongyu]

Summary

The shipped recommended-model ladder (packages/coding-agent/src/core/extensions/builtin/recommended-models/index.ts) no longer matches the intended default order, and it picks a provider for a recommended model in registry order instead of ranking the provider lanes the way the category and profile chains do.

Current behavior

RECOMMENDED_DEFAULT_MODELS on main:

1. kimi-k3 max
2. gpt-6-astra high
3. gpt-6-sol medium
4. gpt-5.6-sol medium
5. claude-fable-5-1 high
6. claude-opus-5-5 max
7. glm-5.2 max

findAvailableRecommendation returns the first registry entry whose canonical id matches, so when a machine holds both an Anthropic API key and the Claude subscription, whichever provider the registry lists first wins (the subscription lane is registered by an extension, i.e. usually later).

Expected (ideal state)

The ladder is, in order:

1. claude-opus-5-5 medium
2. claude-fable-5-1 xhigh
3. kimi-k3 max
4. gpt-6-astra xhigh
5. gpt-6-sol medium
6. glm-5.3 max

For each rung, candidate providers are ranked like the existing Claude/Kimi/GPT/GLM chains (subscription lanes first): Claude anthropic-subscription, anthropic, anthropic-api, github-copilot, opencode; Kimi kimi-coding, kimi-for-coding, moonshotai, opencode-go; GPT chatgpt-subscription, openai, github-copilot, opencode; GLM zai-coding-plan, opencode-go. A provider outside the ranking still serves the rung after the ranked ones, in registry order.

Gateway aggregators (opengateway, openrouter, vercel-ai-gateway) are never pulled in by the ladder: they list models under vendor-prefixed ids (anthropic/claude-opus-5-5), and the exact canonical-id match must keep rejecting them.

Acceptance criteria

• RECOMMENDED_DEFAULT_MODELS equals the six rungs above with the listed thinking levels.
• With both anthropic/claude-opus-5-5 and anthropic-subscription/claude-opus-5-5 available (registry order API first), the subscription model is selected at medium.
• A registry that only has opengateway/anthropic/claude-opus-5-5 plus kimi-coding/kimi-k3 selects kimi-k3.
• settings.recommendedModels overrides keep working.
• CHANGELOG [Unreleased] and builtin changes.md record the new order.

Related

• feat: Claude Opus 5.5 support — catalog, request compat, prompt preset, recommended default at max #2022 / feat: GPT-6 Sol and GPT-6 Luna - catalog rows, family prompt preset, recommended default #2024 (last edits to the ladder), feat(coding-agent): recommend GPT-6 Astra ahead of GPT-5.6 Sol #1375, feat(coding-agent): add recommended-models builtin with default-model priority and warnings #466, fix(coding-agent): respect explicitly saved system default in recommended-models #625, fix(coding-agent): stop headless recommended-model auto-switch from clobbering user defaults #473.
• A follow-up in oh-my-openagent makes the same ladder the unset-profile default for OmO Native sessions outside the TUI.

[code-yeongyu]

Fixed by #2075, merged as 817984ddeb807660d471c2016ba5cd00e109bdaa.

Evidence:

• bun run --cwd packages/coding-agent test test/suite/recommended-models-extension.test.ts test/suite/recommended-models-session-scope.test.ts test/suite/regressions/8700-claude-code-promoted-model-support.test.ts: 3 files, 24 tests passed.
• Mutation check: replacing the ranked provider pick with candidates[0] fails 2 tests (subscription lane over an earlier-listed API lane; unranked provider skipped). Restored afterwards.
• bunx tsc --noEmit exit 0. PR CI: every check passed.
• Acceptance: the shipped ladder is opus-5-5 medium -> fable-5-1 xhigh -> kimi-k3 max -> gpt-6-astra xhigh -> gpt-6-sol medium -> glm-5.3 max. A gateway-only Claude (opengateway/anthropic/claude-opus-5-5) falls through to kimi-k3. recommendedModels overrides are unchanged. CHANGELOG [Unreleased] > Changed and builtin changes.md are updated.

The change ships in the next senpi release; oh-my-openagent adopts it through its senpi pin.

[code-yeongyu]

Released in v2026.9.24. npm latest = 2026.9.24, tarball 200, the GitHub release is non-draft, and provenance is recorded on rekor. oh-my-openagent adopted it in #8782 (merge f83066d80) and shipped v5.0.0-beta.89. The desktop runtime pins both in omo-desktop-app#1107 (merge 3bfaf0c4b).