feat(providers): split zai into 4 plans (Global/China × direct API/Coding Plan)

[yuanmingyi]

Summary

Z.AI ships four distinct billing/endpoint combinations (Global direct, China direct, Coding Plan Global, Coding Plan China) behind one user-facing name. The single pre-existing zai provider tried to paper over this at runtime by probing endpoints with the user's key to find the working one — slow, fragile, and gave users no way to pin a plan. This PR splits zai into four explicit Hermes providers so users pick their plan once and credentials resolve deterministically.

Hermes id	models.dev id	Endpoint	Env var	Plan
zai	zai	api.z.ai/api/paas/v4	ZAI_API_KEY	Z.AI (Global direct)
zai-cn	zhipuai	open.bigmodel.cn/api/paas/v4	GLM_API_KEY	Zhipu AI (China direct)
zai-coding-global	zai-coding-plan	api.z.ai/api/coding/paas/v4	ZAI_CODING_API_KEY	Z.AI Coding Plan (Global)
zai-coding-cn	zhipuai-coding-plan	open.bigmodel.cn/api/coding/paas/v4	GLM_CODING_API_KEY	Zhipu AI Coding Plan (China)

Display names, model catalogs, and capability metadata are sourced from models.dev's per-plan entries. Legacy provider: zai configs with only GLM_API_KEY set continue to work — see Compatibility below.

Changes

• hermes_cli/auth.py — four distinct PROVIDER_REGISTRY entries with per-plan env vars and static base URLs. Dropped detect_zai_endpoint / _resolve_zai_base_url — endpoints no longer probed. zai-cn declared before zai so auto-detect with only GLM_API_KEY still prefers the China plan.
• hermes_cli/providers.py — HERMES_OVERLAYS gains three new entries (zai-cn, zai-coding-cn, zai-coding-global). ALIASES now routes glm/zhipu → zai-cn; z-ai/z.ai → zai unchanged.
• agent/models_dev.py — PROVIDER_TO_MODELS_DEV maps each Hermes id to its distinct models.dev catalog so /model and /provider pickers show the right per-plan display name and model list.
• hermes_cli/models.py — curated _PROVIDER_MODELS:
  • direct plans (zai, zai-cn): glm-5.1, glm-5, glm-5v-turbo, glm-5-turbo, glm-4.7, glm-4.7-flashx, glm-4.6, glm-4.5, glm-4.5-air
  • coding plans (zai-coding-*): glm-5.1, glm-5-turbo, glm-4.7, glm-4.5-air
• agent/auxiliary_client.py — aux-model defaults: zai/zai-cn → glm-4.7-flashx; zai-coding-* → glm-4.7. Vision override: zai/zai-cn → glm-5v-turbo.
• run_agent.py — _anthropic_preserve_dots() extended to all four plans so glm-4.7/glm-5.1 aren't mangled at API-call time.
• agent/model_metadata.py — _PROVIDER_PREFIXES gains the two coding-plan slugs for zai-coding-*/glm-5.1 prefix stripping.
• hermes_cli/model_normalize.py — same four slugs added to _MATCHING_PREFIX_STRIP_PROVIDERS.
• hermes_cli/doctor.py — two new _apikey_providers probes (api.z.ai/api/coding/paas/v4/models, open.bigmodel.cn/api/coding/paas/v4/models); _PROVIDER_ENV_HINTS gains the six new env vars so .env with only coding-plan keys is recognised as provider-configured.
• hermes_cli/status.py — keys + apikey_providers dicts list all four plans; column widened from 12 → 20 to fit Zhipu AI Coding Plan.
• hermes_cli/dump.py / hermes_cli/setup.py / hermes_cli/config.py / hermes_cli/main.py / cli.py — per-plan env-var mappings, vision-setup labels, fallback config comments, --provider choices, docstrings updated.
• gateway/platforms/qqbot/adapter.py — STT _PROVIDER_BASE_URLS recognises all four zai slugs (plus legacy glm alias) and picks glm-asr as default model; all China/Global variants map to open.bigmodel.cn/api/paas/v4 since GLM-ASR is hosted only there.
• trajectory_compressor.py — _detect_provider now inspects URL path to distinguish Coding Plan from direct API (previous base_url_host_matches call with a path string was dead code because the helper only compares hostnames).

Compatibility

A user upgrading with model.provider: zai and only GLM_API_KEY set would otherwise break because zai's env tuple is now (ZAI_API_KEY, Z_AI_API_KEY). Rather than add a migration shim, GLM_API_KEY is kept as a last-resort fallback in zai's env tuple. On upgrade:

• provider: zai + only GLM_API_KEY → key resolves via fallback; request hits api.z.ai (may 401 if the key was issued for Zhipu/bigmodel). Real API error beats the pre-fix empty-key bailout; running hermes model re-picks the correct plan.
• provider: zai-cn + GLM_API_KEY → clean.
• Auto-detect with only GLM_API_KEY → zai-cn (registry order reversal ensures this).
• ZAI_API_KEY always wins over GLM_API_KEY when both present.

Validation

.venv/bin/python -m pytest \
  tests/hermes_cli/test_api_key_providers.py \
  tests/hermes_cli/test_model_validation.py \
  tests/hermes_cli/test_model_provider_persistence.py \
  tests/hermes_cli/test_doctor.py \
  tests/hermes_cli/test_status.py \
  tests/hermes_cli/test_model_normalize.py \
  tests/agent/test_auxiliary_client.py \
  tests/agent/test_model_metadata.py \
  tests/gateway/test_qqbot.py \
  -q

All 457 relevant tests pass. Added coverage for per-plan registry entries, legacy GLM_API_KEY fallback, auto-detect priority, and static-endpoint resolution.

🤖 Generated with Claude Code