[Bug] thinkingLevelMap inconsistently populated across providers for the same model - /effort diverges by provider

[aniruddhaadak80]

Description

thinkingLevelMap - the per-model map that packages/ai/src/providers/* use to translate /effort thinking levels into API parameters (see merged PR #1258, "derive reasoning levels from provider metadata") - is present for some providers' entries of a given model and absent for other providers' entries of the same model, even when both entries use the same api. The result is provider-dependent /effort behavior for identical models: on some providers the selector shows only the levels the model actually distinguishes; on others it offers all seven levels even though most are no-ops or rejected.

Concrete case, all entries are "api": "openai-completions" in packages/ai/src/models.generated.ts (current main) for Kimi K2.5:

provider	entry line	thinkingLevelMap
openrouter (moonshotai/kimi-k2.5)	~12843	{"minimal":null,"low":null,"medium":null,"high":"high","xhigh":null,"max":null}
prime-inference (moonshotai/kimi-k2.5)	~16840	same, only high maps
moonshotai (kimi-k2.5)	~8046	absent
moonshotai-cn	~8229 block	absent
huggingface (moonshotai/Kimi-K2.5)	~6898 block	absent
opencode (kimi-k2.5)	~10097 block	absent

Moonshot's own API documents a single supported effort value for K2.5 (high), which the openrouter and prime-inference rows encode correctly. But a user running kimi-k2.5 via Moonshot direct, OpenCode Zen, or Hugging Face gets an unfiltered /effort selector where six of seven choices do nothing (the engine falls back per #1456-class behavior), while the same model through OpenRouter shows the curated set.

The same divergence pattern appears for other models (e.g., deepseek-v3.2 and gpt-oss-20b have maps on openrouter but not on several direct/gateway listings).

Steps to reproduce

1. Configure Moonshot direct (MOONSHOT_API_KEY) and select kimi-k2.5.
2. Open /effort: all levels are offered; switching between low/medium/max produces identical requests.
3. Repeat with the same model via OpenRouter: /effort shows the restricted, meaningful set.

Or statically: compare the cited entries in models.generated.ts.

Expected behavior

Model-derived metadata like thinkingLevelMap should be normalized across providers that share the same underlying model + API type, so /effort behaves identically regardless of which gateway serves the request.

Actual behavior

The map travels with whichever upstream row happened to carry it; providers whose source listing lacked it get no level filtering.

Prime Agent version

main @ e319a66

Operating system

Any

Additional context

Suggested fix: during generation, merge thinkingLevelMap across entries that normalize to the same model identity + API type (preferring the vendor-direct row), or derive it once per model rather than per provider row. This complements the per-engine z.ai fixes (#1456 / closed PRs #1075, #1565) - those fixed how maps are applied; this report is about the maps being inconsistently populated in the first place.

---

First filed as issue #1719; it was auto-closed by the contribution vouch gate within seconds, so I am re-filing it here per CONTRIBUTING.md's discussion-first process.

[awhite0030]

Investigated this report and produced a candidate fix, validated on a fork of this repository.

Root cause, fix and validation from the working notes:

Root cause: The model generation script did not merge metadata like thinkingLevelMap across identical models served by different providers on the same API, leading to inconsistent capability behavior depending on the gateway.
Fix: Updated normalizeOutliers in packages/ai/scripts/generate-models.ts to explicitly merge thinkingLevelMap instances among grouped models that share both canonical ID and API identity. Added a regression test asserting this behavior.
Validation: Ran npm run check and executed the regression test via npx tsx ../../node_modules/vitest/dist/cli.js run test/suite/regressions/kimi-k25-thinking-level-map.test.ts, both passing cleanly.
Upstream discussion: #1725

Diff: awhite0030/prime-agent@main...fix-thinking-level-map-12730448972716653604
Full pull request: awhite0030#8

The change passes npm run check and the focused regression tests against current main. If a fix for this is still wanted, I can open a pull request.