docs(providers): document Z.ai/GLM dual-endpoint model, no glm.rs needed

[Destynova2]

Verdict on roadmap B-01: keep openai_compat, do not add a dedicated GLM provider impl.

Summary

Z.ai exposes GLM models on two parallel endpoints, both already covered by grob without any new code:

Path	provider_type	base_url	Purpose
Anthropic Messages API	z.ai (existing)	https://api.z.ai/api/anthropic	Drop-in for Claude Code (Coding Plan)
OpenAI Chat Completions API	openai	https://api.z.ai/api/paas/v4	Free tier / PAYG (ultra-cheap preset)

The first path was wired in src/providers/registry.rs:233-247 (routes to AnthropicCompatibleProvider). The second path is what presets/ultra-cheap.toml uses today and works as-is through the standard OpenAIProvider translation layer.

Why no glm.rs

Investigated the OpenAI-compat endpoint quirks (docs.z.ai/api-reference/llm/chat-completion):

Quirk	Impact on grob
thinking: { type, clear_thinking }	Request-side, optional. Use provider_type = "z.ai" + Anthropic thinking blocks if needed end-to-end.
do_sample: bool	Request-side, optional. Not emitted by Claude Code / Codex CLI.
request_id: string	Request-side, optional trace ID. Not emitted by inbound clients.
tool_stream: bool (GLM-4.6+)	Request-side, optional. Not emitted by inbound clients.
Error shape { code, message }	Cosmetic — normalized at dispatch layer.
Finish reasons sensitive / model_context_window_exceeded / network_error	Cosmetic in passthrough.

None of these are emitted by Claude Code or Codex CLI (the only inbound surfaces grob proxies). A dedicated glm.rs would duplicate OpenAIProvider verbatim and ship as dead code.

Changes

• docs/how-to/providers.md: new "Z.ai / GLM" section documenting both paths, when to pick each, and the OpenAI-compat quirks grob deliberately ignores. Overview table now lists both rows.
• docs/reference/providers.md: provider creation dispatch table now documents provider_type = "openai" + base_url = api.z.ai/api/paas/v4 next to the existing z.ai row.
• tests/unit/provider_test.rs: regression test ultra_cheap_preset_zai_uses_openai_compat that parses the shipped presets/ultra-cheap.toml and asserts the zai provider entry uses provider_type = "openai" + canonical paas/v4 URL. Guards against silent breakage of the free-tier path.

Build dependency

Cherry-picks 5a34bac from PR #304 (fix/preset-mod-include-str) because main is currently broken — src/preset/mod.rs has stale include_str! references to deleted preset files (medium/local/cheap/fast.toml). Once #304 lands, that commit becomes a no-op merge.

Test plan

• cargo fmt --all -- --check passes
• cargo clippy --tests --all-targets passes (no warnings)
• cargo nextest run --test lib passes (217/217 tests)
• ultra_cheap_preset_zai_uses_openai_compat passes
• enterprise::preset_snapshot_test::snapshot_preset_ultra_cheap passes

Rollback

If a future user reports they need a dedicated GLM provider for one of the quirks above, revisit by adding the field to RequestExtensions rather than forking a new provider type.

🤖 Generated with Claude Code