feat: add zai-coding provider for Z.AI GLM Coding Plan

[RedClaus]

Summary

Adds a new zai-coding provider that routes through the Anthropic-compatible endpoint (https://api.z.ai/api/anthropic) instead of the OpenAI-compatible /api/paas/v4 endpoint.

The Z.AI Coding Plan quota is only accessible via the Anthropic endpoint — the standard /api/paas/v4 endpoint returns error 1113 (insufficient balance) even with a valid coding plan subscription.

Changes

• New plugin: plugins/model-providers/zai-coding/ with api_mode="anthropic_messages" and fallback models (glm-5.1, glm-5-turbo, glm-4.7, glm-4.5-air)
• Registry: Added zai-coding to PROVIDER_REGISTRY in hermes_cli/auth.py with inference_base_url pointing to the Anthropic endpoint
• Aliases: zai-coding, zai-coding-plan, glm-coding, z-ai-coding

Usage

hermes config set model.provider zai-coding
hermes config set model.base_url https://api.z.ai/api/anthropic
hermes config set model.default glm-5.1

How it works

The existing Anthropic transport (anthropic_messages api_mode) already handles third-party providers — the code in agent_init.py explicitly supports non-native Anthropic providers like GLM, sending the API key via x-api-key header and skipping OAuth/Claude identity injection.

The _detect_api_mode_for_url() in runtime_provider.py auto-detects anthropic_messages mode for URLs ending in /anthropic, so the transport routes correctly.

Testing

• Verified the Anthropic endpoint accepts the coding plan key (HTTP 200)
• Confirmed hermes -z works with --provider zai-coding -m glm-5.1
• Confirmed runtime config persists via hermes config set

Co-Authored-By: Oz oz-agent@warp.dev

[alt-glitch]

Competing with open PRs #13500, #13911, and #24915 which all address Z.AI provider splitting (Global/China x direct/Coding Plan). This PR takes a narrower approach (Anthropic-endpoint-only Coding Plan provider) vs the broader 4-way split in the competing PRs.

[DeamonDev888]

👋 Superseded — please see the new PR

@RedClaus thanks for the proposal. #62080 is the new rebased PR that closes PR #61575.

Your approach adds a separate zai-coding provider. #62080 takes a different approach: instead of adding provider entries to the registry, it extends ZAI_ENDPOINTS with 6 candidates (added anthropic-global and anthropic-cn to the existing 4) and lets detect_zai_endpoint() pick the right wire format automatically.

Both approaches are valid. If maintainers prefer the 4-variant split approach from this PR (#32174) over the unified probe approach in #62080, that's a maintainer design decision — both achieve the same routing correctness. Live audit (9 production keys): both approaches would route all 9 correctly.

[teknium1]

Thanks for isolating the Coding Plan route as a small provider change. The premise remains current: hermes_cli/auth.py:623-667 only probes OpenAI-compatible Z.AI endpoints, while hermes_cli/runtime_provider.py:101-139 already supports /anthropic URLs as anthropic_messages.

Problems

• The PR diff contains no regression tests. Please cover resolve_runtime_provider("zai-coding") through the API-key path and assert the Anthropic URL and anthropic_messages mode (hermes_cli/runtime_provider.py:2033-2062).
• The provider guide still documents automatic probing only (website/docs/integrations/providers.md:279-281); it needs an entry for this explicitly selected Coding Plan route and aliases.

Suggested changes

• Add runtime/provider-alias tests and update the provider guide before salvage.

This is an automated hermes-sweeper review.

[teknium1]

Please add regression coverage for the end-to-end runtime resolution of this profile: with zai-coding selected, assert the resolved base URL remains https://api.z.ai/api/anthropic and the API mode is anthropic_messages. The runtime mode is derived from the URL at hermes_cli/runtime_provider.py:2057-2062, so profile registration alone does not exercise the routing contract.

[GottZ]

This was generated by AI during triage.

Summary

Three PRs address Z.AI Coding Plan routing, but their diffs split into two mechanisms: #32174 adds the reported Anthropic-compatible /api/anthropic route, while #34323 and #47140 add a broader dedicated provider around the OpenAI-compatible /api/coding/paas/v4 family.

Related pull requests

• #32174 related — (+47/-0) — keep open for salvage: the narrow diff directly targets the reported cause by registering zai-coding with the Anthropic endpoint and anthropic_messages mode. Consistent with the keep_open review on #32174, it still needs runtime/alias regression tests and provider documentation before merge.
• #34323 [closed] related — (+740/-82) — closed alternative/reference implementation: it adds extensive provider, setup, runtime, model, credential-pool, and test integration, but routes Coding Plan traffic through OpenAI-compatible /api/coding/paas/v4, not the Anthropic-only route asserted by #32174. It remains relevant because it demonstrates the broader dedicated-provider approach later repeated and refined by #47140.
• #47140 [closed] related — (+1171/-160) — closed as redundant with current main: it expands the same OpenAI-compatible dedicated-provider approach with endpoint-family guards, shared credentials, setup integration, tests, and glm-5.2. The contributor closure verified that main already probes Coding Plan endpoints and supports glm-5.2, but that finding does not resolve #32174's distinct /api/anthropic premise because #47140 neither implements nor tests that transport.

Duplicates

#34323 and #47140 are substantially duplicate implementations of a dedicated zai-coding provider using the OpenAI-compatible /api/coding/paas/v4 endpoint family; #47140 is the more comprehensive iteration. #32174 overlaps in user-facing goal but is not a code duplicate because it uses the Anthropic-compatible endpoint and transport.

Suggested consolidation

Consolidate on #32174, but do not merge it yet: preserve the keep_open review's required runtime/alias tests and provider-guide update, and verify the claimed Anthropic-only quota behavior against the current main implementation. Keep #34323 and #47140 closed as redundant OpenAI-compatible alternatives; if the live verification confirms that current main serves all relevant Coding Plan keys, close #32174 too, otherwise merge the tested and documented narrow #32174 route.

Cross-PR triage: Reviewed 3 pull requests and 0 issues in this complex. Each diff was read against this issue; Assessment working set: 155 kB of PR diffs, 10 kB of issue/PR text, 5 kB of discussion (8 comments), 1 verify verdict. verdicts reflect diff content, not PR titles. Part of an automated triage batch.