fix(agent-core): only send prompt_cache_key to official OpenAI endpoints

[creatiVision]

Related Issue

Tracking / backup of #2203 (original by @B143KC47).
Resolves #2166.
Related: #2611 (Azure Foundry 400 rejection), #2761.

If #2203 is merged first, this PR can be closed with no further action.

Problem

Since v0.29.0 (#1970), every OpenAI-compatible provider receives the session prompt_cache_key in the request body. Strictly-validating endpoints (e.g. NVIDIA NIM at integrate.api.nvidia.com (#2166) or Azure Foundry (#2611)) reject the unknown parameter with:

400 Validation: Unsupported parameter(s): `prompt_cache_key`

That makes custom openai / openai_responses providers unusable for agent sessions.

What changed

Same fix as #2203, rebased onto current main:

• Only send prompt_cache_key when the effective base URL targets the official OpenAI API (api.openai.com or *.api.openai.com, or unset → client default).
• v1: gate in toKosongProviderConfig for openai and openai_responses (isOfficialOpenAIBaseUrl).
• v2: same gate on the bare cacheKey → prompt_cache_key fallback in OpenAI chat-completions and Responses bases.
• Kimi protocol branch untouched (Kimi endpoints support the field).
• Tests: custom endpoint omits the field; official endpoint keeps it.

Checklist

• Based on upstream fix(agent-core): only send prompt_cache_key to official OpenAI endpoints #2203
• Rebased on latest main
• Changeset included (from fix(agent-core): only send prompt_cache_key to official OpenAI endpoints #2203)
• Tests included (from fix(agent-core): only send prompt_cache_key to official OpenAI endpoints #2203)

[changeset-bot]

🦋 Changeset detected

Latest commit: 30e00d4

The changes in this PR will be included in the next version bump.

This PR includes changesets to release 2 packages

Name	Type
@moonshot-ai/agent-core-v2	Patch
@moonshot-ai/kimi-code	Patch

Not sure what this means? Click here to learn what changesets are.

Click here if you're a maintainer who wants to add another changeset to this PR

[TaylorTWBrown]

I'd love to see this merged soon. I can't use Kimi + Kimi Code on Azure as it is.

[creatiVision]

Rebased and synced with latest main. The merge conflicts resulting from the recent kosong package extraction under packages/agent-core-v2 have been resolved by dropping the obsolete packages/agent-core-v2/src/kosong files.

The PR diff is now clean and focused strictly on gating prompt_cache_key behind isOfficialOpenAIBaseUrl(baseUrl) in packages/agent-core/src/session/provider-manager.ts for both openai and openai_responses providers, along with comprehensive unit tests in runtime-provider.test.ts. All test suites pass.

[creatiVision]

Rebased and synced with latest main. With upstream's removal of the legacy packages/agent-core engine (#3542), the gating logic has been cleanly applied to the active packages/agent-core-v2 engine:

1. packages/agent-core-v2/src/human/llm/requester/bases/openai-base-url.ts: Implements isOfficialOpenAIBaseUrl() helper targeting api.openai.com / *.api.openai.com (or default unset).
2. packages/agent-core-v2/src/human/llm/requester/bases/openai/format.ts & openai-responses/format.ts: Gates prompt_cache_key behind isOfficialOpenAIBaseUrl(ctx.model.baseUrl) on fallback when no trait hook overrides it, resolving HTTP 400 rejections on strict third-party endpoints like NVIDIA NIM (Error with models not supporting prompt cache #2166) and Azure Foundry (MS Foundry + Kimi Code 2.7 - 400 Unrecognized request argument supplied: prompt_cache_key #2611).
3. packages/agent-core-v2/src/human/test/llm/cache-key.test.ts: Verified with unit tests that prompt_cache_key is sent to official endpoints and omitted for third-party endpoints (all tests green, no comments in agent-core-v2).
4. Changeset included.

All merge conflicts are resolved and the PR is clean, mergeable, and ready for review.