fix(auxiliary): prepend zai coding-plan vision endpoint before metered paas/v4 (fixes #55112)

[Morad37]

Closes #55112

Prepend the coding-plan quota endpoint as the first URL tried for ZAI vision requests, so subscription users hit the quota-billed endpoint before falling through to the metered pay-as-you-go endpoints.

[Caiofwp]

Thanks for the quick fix — the ordering is exactly right, prepending /api/coding/paas/v4 so subscription users hit the quota-billed endpoint first. 🙏

One small nit: the new open.bigmodel.cn line is indented with 8 spaces while the rest of the list uses 12, so the block is misaligned:

        zai_openai_urls = [
            "https://api.z.ai/api/coding/paas/v4",   # coding-plan quota
        "https://open.bigmodel.cn/api/paas/v4",  # metered fallthrough   <-- 8 spaces
            "https://api.z.ai/api/paas/v4",
        ]

It's valid Python (indentation is free inside the list literal), but black/ruff in CI will likely flag or reformat it. Re-aligning to 12 spaces should keep the lint check green:

        zai_openai_urls = [
            "https://api.z.ai/api/coding/paas/v4",   # coding-plan quota
            "https://open.bigmodel.cn/api/paas/v4",  # metered fallthrough
            "https://api.z.ai/api/paas/v4",
        ]

[DeamonDev888]

🟢 Superseded — please close in favor of the new PR

@Morad37 thanks for the original PR. The rebased and consolidated fix is now #62080. The vision helper zai_openai_urls list ordering is implemented there with the indentation bug from your original fixed.

The new PR also includes:

• Layer 1: _is_payment_error() Z.AI Coding Plan exemption
• Layer 2: vision helper (this fix)
• Layer 3: runtime pool re-resolution
• Layer 4: config.yaml precedence
• Layer 5: Anthropic-wire endpoints in ZAI_ENDPOINTS

If maintainers agree, please close this PR in favor of #62080. Credit remains in git history.

[DeamonDev888]

Superseded by #62467. Your vision coding-endpoint ordering fix is now handled by the unified resolver Step 4 (provider-specific resolver) which probes coding endpoints before metered for all Z.AI calls, including vision. Credit for identifying the ordering issue.

[teknium1]

Thanks for isolating the Z.AI auxiliary-vision routing issue. The current fallback list still begins with metered endpoints at agent/auxiliary_client.py:5434-5437, so the proposed ordering change addresses a live defect.

Problems

• The modified open.bigmodel.cn entry is misindented in the PR diff; align it with the neighboring list entries.
• Please add a regression test for the ordering. Current searches found no test for zai_openai_urls or the Z.AI vision fallback sequence; agent/auxiliary_client.py:5438-5447 is the behavior under change.

Suggested changes

• Mock _get_cached_client() and assert that the first attempted base URL is https://api.z.ai/api/coding/paas/v4.

Automated hermes-sweeper review.

[Caiofwp]

Gentle bump (2026-08-12)

Issue #55112 is still reproducible on current main: the Z.AI vision fallback list in agent/auxiliary_client.py ~L7145–7148 remains metered-only (open.bigmodel.cn/api/paas/v4 + api.z.ai/api/paas/v4); coding-plan is still missing.

This PR is still the smallest fix for that root cause. Earlier feedback that is still outstanding (from review / nit):

1. Align indentation on the open.bigmodel.cn list entry (12 spaces like the siblings).
2. Add a regression test that mocks _get_cached_client() and asserts the first attempted base URL is https://api.z.ai/api/coding/paas/v4.

@Morad37 — if you are busy, I am happy to open a tiny follow-up (or take over the branch with credit) to land (1)+(2) so this can merge cleanly. Maintainers: if you prefer a different vehicle (#62080 bundle / #74071), a quick signal would help avoid further duplicate effort.

Thanks!