fix(provider): default a stream-idle timeout for openai-compatible providers

[beatakouchnir]

Issue for this PR

Fixes #46581

Type of change

• Bug fix
• New feature
• Refactor / code improvement
• Documentation

What does this PR do?

Custom @ai-sdk/openai-compatible providers never got a chunkTimeout or headerTimeout default. wrapSSE() is a no-op when chunkTimeout is unset (if (typeof ms !== "number" || ms <= 0) return res), so these providers stream with no idle guard at all — a backend that stalls mid-stream hangs the session indefinitely. Named providers (openai, anthropic) already default a timeout; this gives openai-compatible providers the same protection via an overridable chunkTimeout default (300s, matching the header-timeout default).

I moved the existing inline includeUsage defaulting into a small exported applyOpenAICompatibleDefaults() and added the chunkTimeout default there. includeUsage behavior is unchanged.

This covers the stream-stall mode (no SSE bytes for the timeout window). The other mode in #46581 — the stream closing on [DONE] while the AI SDK's fullStream still awaits a usage chunk — is not addressed here, because a byte-idle timeout can't fire once the final byte has arrived; that one needs a change at the AI-SDK adapter and I'd rather keep it separate.

I hit this running unattended opencode run sessions against a local llama.cpp server; a couple stalled and hung until an external timeout. Forensics are in the comment on #46581.

How did you verify your code works?

• New test/provider/openai-compatible-defaults.test.ts: default applied for openai-compatible, not applied to other providers, never overrides an explicit chunkTimeout, preserves includeUsage: false.
• test/provider/header-timeout.test.ts already verifies that a set chunkTimeout turns a stalled SSE stream into a ResponseStreamError; this change makes that apply by default.
• Full test/provider/ suite passes (593 tests). typecheck and prettier clean.

Screenshots / recordings

n/a — no UI change.

Checklist

• I have tested my changes locally
• I have not included unrelated changes in this PR

[github-actions]

Thanks for your contribution!

This PR doesn't have a linked issue. All PRs must reference an existing issue.

Please:

1. Open an issue describing the bug/feature (if one doesn't exist)
2. Add Fixes #<number> or Closes #<number> to this PR description

See CONTRIBUTING.md for details.

[holny]

Nice catch — the ai-sdk layer already had the machinery, it just never got a default for custom openai-compatible providers. The no-op guard in wrapSSE makes an unset timeout a silent gap, so this is a real fix; the includeUsage extraction preserves behavior and the explicit-chunkTimeout override is covered by tests.

Context that's relevant here: the V2 native route path (packages/ai) has no chunk-idle guard at all — not just a missing default. That's tracked in #46692 (being worked separately), so no overlap with this PR; just noting the ai-sdk layer won't be the only surface needing this.

[beatakouchnir]

Thanks for the review @holny. Looks like #46890 has since landed a global chunkTimeout default (also 5 min, disable-per-provider via chunkTimeout: false), which covers the openai-compatible case here and then some — so this PR is redundant now. Closing in favor of it. Glad the stall gap is addressed; the [DONE]-without-usage variant from #46581 is a separate mode that a byte-idle timeout won't catch, so that one's still open.