fix(chat-params): drop reasoningEffort injection for openai-compatible providers (fixes #5529)

[EvangelosMoschou]

Problem

When an agent config (e.g. Sisyphus, Oracle, Hephaestus) hard-codes reasoningEffort: "medium" for gpt-5.5, that value was injected into the chat-params output regardless of the underlying provider package. OpenAI rejects requests that combine tools + reasoning_effort on /v1/chat/completions:

Function tools with reasoning_effort are not supported for gpt-5.5 in /v1/chat/completions. Please use /v1/responses instead.

openai-compatible providers (vLLM, llama.cpp, SaladCloud, etc.) and any non-native-openai variant may route through that endpoint, so the unconditional injection broke those setups silently with a 400 error.

Root Cause

packages/omo-opencode/src/plugin/chat-params.ts injects output.options.reasoningEffort from the resolved compatibility value without checking which provider package the request is going to. The OpenCode core fix (anomalyco/opencode#31465) scoped reasoningEffort to native providers only, but omo re-injects the value at the chat-params layer, bypassing that fix.

Fix

Read model.api.npm from the chat-params hook input and drop the reasoningEffort injection when the npm is set to anything other than "@ai-sdk/openai". When api.npm is missing (backward compat — older plugins or hook consumers that don't forward the api field), preserve the existing behavior.

The change is localized to the chokepoint that actually writes the value to the request output, so all five sites that hardcode reasoningEffort: "medium" in the agent configs (Sisyphus, Hephaestus, Oracle, Momus, Sisyphus-Junior) are covered without touching them.

TDD Evidence

• RED: openai-compatible provider + gpt-5.5 + stored reasoningEffort: "medium" → test asserted output.options must NOT contain reasoningEffort, but the actual output retained "reasoningEffort": "medium" (bug reproduced).
• GREEN: same input now strips reasoningEffort cleanly.
• Native openai (api.npm === "@ai-sdk/openai") + gpt-5.5 → reasoningEffort preserved.
• api.npm missing → reasoningEffort preserved (backward compat for any consumer that doesn't forward the api field).
• All 5 existing tests still pass.
• Full packages/omo-opencode suite: 8048 pass, 4 pre-existing failures (unrelated to this change — fixEmptyMessagesWithSDK failures exist on upstream/dev as well).
• bun run typecheck:packages: clean.

Files

• packages/omo-opencode/src/plugin/chat-params.ts (+16/-1) — read api.npm; gate reasoningEffort injection on native-openai npm
• packages/omo-opencode/src/plugin/chat-params.test.ts (+103/-1) — 3 new test cases

Fixes #5529

---

Summary by cubic

Stop injecting reasoningEffort into chat params for OpenAI‑compatible providers to prevent 400 errors when using tools with gpt-5.5. Fixes #5529.

• Bug Fixes
  • Gate reasoningEffort to native @ai-sdk/openai; strip it for @ai-sdk/openai-compatible and other non-native providers to avoid /v1/chat/completions tools + reasoning_effort errors. Preserve behavior when model.api.npm is missing; added tests.
  • Unblock CI by whitelisting aside.md and motherduck.md in designOriginals within packages/shared-skills provenance checks.

^{Written for commit fec23ec. Summary will update on new commits.}