v0.1.0-alpha.22 — Reasoning-model architecture cleanup + adapter-google httpOptions (additive)

[baabakk]

Released 2026-06-19. Install: pnpm add @llm-ports/adapter-openai@alpha @llm-ports/adapter-google@alpha

Two themes, both empirically driven by external work this week.

1. Reasoning-model architecture cleanup (adapter-openai)

Driven by ADW's 2026-06-19 instrumentation of their multi-team agentic build loop. DeepInfra-hosted openai/gpt-oss-120b was silently failing as a reasoning model in alpha.21 — same model file, same weights, same behavior as Cerebras's gpt-oss-120b, but the namespace prefix broke the catalog regex (/^gpt-oss-/i anchored at ^). The model wasn't recognized as reasoning → budget multiplier never applied → first call burned its budget on hidden CoT → empty visible output → loop terminated as completed.

The fix isn't another regex. The fix is architectural.

What shipped

• normalizeModelId helper — strips a <owner>/ namespace prefix, returning the canonical name. Used at every entry point to the capability learner. Catalog patterns stay anchored at ^; canonical names match regardless of which provider serves the model. The catalog stops needing per-(model × provider) regex variants forever.
• Runtime detection broadening — learnFromResponse reads message.reasoning_content (DeepInfra's harmony field) in addition to the existing fields. reasoningStarvedResponse accepts finish_reason: "stop" in addition to "length" (DeepInfra returns stop for gpt-oss harmony serving). Also guards against rescuing successful tool-use.
• Xiaomi MiMo catalog entry — MiMo-V\d+ family (distinct from MiniMax). Filed after observing mimo-parasail starvation in ADW production this morning.

What's still NOT fixed

DeepInfra-served gpt-oss tool-use still doesn't execute the model's tool-call intent. With alpha.22 the budget is correct, the starvation rescue fires, but the tool-call intent often lands in reasoning_content rather than tool_calls. Until the adapter parses the harmony channel for tool calls — a separate research-first workstream — for tool-use workloads against gpt-oss, route to Cerebras.

2. adapter-google httpOptions pass-through

createGoogleAdapter now forwards opts.httpOptions to the underlying @google/genai GoogleGenAI constructor. Closes the Dramma backend-proxy plan blocker: the browser bundle can now set baseUrl to a backend proxy that holds the real GEMINI_API_KEY, instead of inlining the key into the bundle.

const adapter = createGoogleAdapter({
  apiKey: process.env.YOUR_BACKEND_BEARER!,
  httpOptions: {
    baseUrl: "https://your-app.example/api/llm/google",
    // also: apiVersion, headers, timeout, retryOptions
  },
});

HttpOptions is re-exported from @llm-ports/adapter-google so consumers can type their override without an extra peer dep.

Backwards compatibility

All changes are additive. Existing callers see no behavior change.

Tests

763 total (was 736), +29 new across the two adapters. 0 regressions in the other 6 packages.

What's NOT in this release

• onValidationRetry Registry-level emission (still type-only)
• Streamed cost surfacing
• LP-REQ-01 resilient fallback preset
• Harmony-channel tool-call parser (research-first; pending more empirical evidence — see release notes)

These slide to alpha.23 or later. Scope discipline: alpha.21 was structured-output reliability; alpha.22 is reasoning-model handling. Each release stays themed.

Coming in alpha.23 (planned)

• Comprehensive reasoning-model coverage across Parasail, DeepInfra, SambaNova
• Architectural redesign of capability detection so the static catalog becomes a perf optimization, not a load-bearing correctness mechanism (designed away from per-model regex maintenance)

Full release notes.