feat(providers): add Z.AI GLM Coding Plan provider and glm-5.2 model

[yuanmingyi]

What does this PR do?

Adds a dedicated zai-coding (Z.AI / GLM Coding Plan) provider as a sibling to the existing direct-API zai provider. The two share the same API-key env vars (GLM_API_KEY/ZAI_API_KEY/Z_AI_API_KEY) and may reuse each other's stored credential, but are kept on separate billing paths — direct /api/paas/v4 vs coding-plan /api/coding/paas/v4 — so a cached or legacy GLM_BASE_URL cannot bleed across endpoints (a direct key hitting the coding path returns "Insufficient balance", and vice versa). Also adds the glm-5.2 model to both providers' catalogs and to the Z.AI endpoint probe candidates.

The Z.AI endpoint-detection logic is generalized into reusable provider-family helpers (endpoint_family_providers, base_url_matches_endpoint_family, resolve_provider_endpoint_family_base_url) backed by an official-URL canonical key that normalizes default ports (:443), trailing dots/slashes, and casing, so cross-family official URLs are reliably rejected.

Related Issue

Supersedes #1 (closed prior attempt). No tracking issue — Fixes #.

Type of Change

• ✨ New feature (non-breaking change that adds functionality)

Changes Made

• New provider: zai-coding ProviderConfig (hermes_cli/auth.py), ProviderProfile (plugins/model-providers/zai/__init__.py), HermesOverlay (hermes_cli/providers.py), with aliases (zai-coding-plan, z-ai-coding, glm-coding, glm-coding-plan, zhipu-coding, zhipu-coding-plan) and coding-plan base URL.
• Model: add glm-5.2 to the zai and zai-coding catalogs (hermes_cli/models.py, hermes_cli/setup.py) and to the Z.AI endpoint probe candidates.
• Endpoint-family isolation: split ZAI_ENDPOINTS into ZAI_DIRECT_ENDPOINTS / ZAI_CODING_ENDPOINTS; new _official_base_url_key + base_url_matches_endpoint_family guard; guards applied at all base-URL resolution sites (hermes_cli/runtime_provider.py, agent/credential_pool.py, hermes_cli/model_setup_flows.py).
• Shared credentials: shared_credential_provider_ids() lets zai/zai-coding fall back to each other's stored pool key while the endpoint still follows the selected provider's family.
• Setup flow: Z.AI endpoint picker (Global / China / custom proxy) in _select_endpoint_family_base_url, persisted to GLM_BASE_URL.
• Wiring: aux model glm-5-turbo and vision model glm-5v-turbo for zai-coding (agent/auxiliary_client.py); /coding/ URL inference → zai-coding (agent/model_metadata.py); models.dev mapping (agent/models_dev.py); /model picker + PROVIDER_GROUPS, canonical providers, labels, doctor/status labels, model_normalize prefix-strip.
• Tests: endpoint-family isolation (probe, env-override, cache), shared-credential reuse, family-guard canonicalization (:443/trailing-dot), picker fallback, and zai-coding/... normalization.

How to Test

1. scripts/run_tests.sh tests/hermes_cli/test_api_key_providers.py tests/hermes_cli/test_runtime_provider_resolution.py tests/hermes_cli/test_model_provider_persistence.py tests/hermes_cli/test_model_validation.py tests/hermes_cli/test_model_normalize.py tests/agent/test_auxiliary_named_custom_providers.py tests/providers/test_provider_profiles.py — affected zai/coding tests pass (189 passed, 0 failed).
2. hermes /model → Z.AI / GLM group → Coding Plan; confirm the endpoint picker offers Global / China / custom proxy and persists the choice to GLM_BASE_URL.
3. Configure a Coding Plan key, start a session on zai-coding, and confirm requests hit https://api.z.ai/api/coding/paas/v4 (not the direct path).
4. Switch between zai and zai-coding and confirm no cross-family base-URL bleed (stale cached/GLM_BASE_URL values are rejected with a warning and re-probed).

Checklist

Code

• My commit messages follow Conventional Commits (feat(providers): …)
• I searched for existing PRs to make sure this isn't a duplicate (supersedes closed Terminal tool #1)
• My PR contains only changes related to this fix/feature (no unrelated commits; package-lock.json excluded)
• I've run pytest tests/ -q and all tests pass — ran the affected test files (189 passed); full suite not yet run on this branch
• I've added tests for my changes
• I've tested on my platform: macOS 15 (darwin)

Documentation & Housekeeping

• I've updated relevant documentation — or N/A (provider is auto-discovered via plugins/model-providers/zai/; no new config keys)
• I've updated cli-config.yaml.example — N/A (reuses existing GLM_API_KEY / GLM_BASE_URL)
• I've updated CONTRIBUTING.md or AGENTS.md — N/A
• I've considered cross-platform impact — N/A (pure-Python provider logic; no platform-specific code)

Screenshots / Logs

N/A

[alt-glitch]

Related: #13500 and #13911 (open competing approaches that split zai into 4 explicit Global/China x direct/coding-plan providers). This PR takes a different mechanism — a single dedicated zai-coding sibling provider sharing keys but on a separate billing path, plus endpoint-family helpers and the new glm-5.2 model. Same goal, different approach; maintainer to pick the mechanism.

[yuanmingyi]

Related: #13500 and #13911 (open competing approaches that split zai into 4 explicit Global/China x direct/coding-plan providers). This PR takes a different mechanism — a single dedicated zai-coding sibling provider sharing keys but on a separate billing path, plus endpoint-family helpers and the new glm-5.2 model. Same goal, different approach; maintainer to pick the mechanism.

these two are old & duplicate and can be closed. Thanks

[kshitijk4poor]

Closing as redundant — current main already supports Z.AI Coding Plan endpoints and glm-5.2.

Verified on origin/main (E2E):

1. Coding-plan endpoints already in the probe list. ZAI_ENDPOINTS in hermes_cli/auth.py already includes coding-global (https://api.z.ai/api/coding/paas/v4) and coding-cn (https://open.bigmodel.cn/api/coding/paas/v4). A coding-plan key automatically resolves there: the probe hits direct endpoints first (401/403), then succeeds on coding-global.

2. glm-5.2 already on main. It's in ZAI_ENDPOINTS probe candidates (["glm-5.2", "glm-5.1", "glm-5v-turbo", "glm-4.7"]), in provider_model_ids("zai"), and in _SUPPORTED_MODELS["zai"] in setup.py.

3. GLM_BASE_URL env override already works for coding paths. Setting GLM_BASE_URL=https://api.z.ai/api/coding/paas/v4 resolves to that URL — env wins, probe skipped. No new provider needed to pin the coding endpoint.

4. Cache-bleed already prevented by per-key-hash cache. The detected endpoint is cached in auth.json under sha256(api_key)[:16]. Switch keys → key_hash mismatches → re-probe. The PR's own description states a direct key hitting the coding path returns "Insufficient balance" (fails) — so the same key can't silently succeed on the wrong path.

The one genuine UX gap is that the setup flow uses a plain text input for base URL rather than a picker offering Global/China/Coding Plan options — that's a small setup-flow tweak, not a 1171-line provider split with endpoint-family isolation machinery.

Thanks for the thorough work and tests — the endpoint-family guard design is thoughtful, but main's probe + env override + per-key-hash cache already covers the use case.

[farangkao]

I can confirm, it works now with Coding Plan url entered during model setup