Test debt: no direct/e2e coverage of forked skills under the openai-compatible provider

[griffinwork40]

Summary

Provider-parity for skills/subagents is currently proven compositionally — the child-provider factory that both the skill and subagent executors inject is tested at src/agent/providers/routing.test.ts:386-514, and the credential anti-leak matrix at src/agent/subagent.test.ts:276-416. But direct/e2e coverage of a skill running under the openai-compatible provider is thin:

1. Credential resolver — resolveCredentialForModel (the PR docs(sdk): add "Build with the SDK" guide for the programmatic library surface #640 anti-leak gate, src/agent/auth/credential-resolver.ts:74) had no dedicated unit test; it is mocked provider-blind in every skill/subagent executor test, so a regression in the OpenAI-vs-Anthropic branch would only be caught transitively. A direct unit test is being added in the working branch (src/agent/auth/credential-resolver.test.ts).
2. No integration test forks a skill subagent under a live openai-compatible session and drives a multi-round tool loop. The Responses wire in particular has zero function_call events in query-responses.test.ts — its loop seams are proven only by composition.

Proposed

• Direct unit test for resolveCredentialForModel (anti-leak matrix) — pending PR.
• Integration test: fork a skill subagent under an openai-compatible session (Chat Completions and Responses) asserting correct provider + credentials + a multi-round tool loop.

Notes

Found via a skills × provider-compatibility audit.