Provider: OpenAI-compatible provider seam (+ Grok/xAI reference)

[csmatar]

Summary

Generalize the OpenAI adapter into a reusable OpenAI-compatible provider seam so that providers speaking the OpenAI Chat Completions wire format (Grok/xAI, Kimi, GLM, DeepSeek, and open-weight endpoints via aggregators) can be added as a data row + pricing rows, not a new adapter. Ship Grok (xAI) as the reference provider that proves the seam.

This is the framework ticket. The per-provider issues (Kimi, GLM, DeepSeek, generic compat) depend on it.

Why this is cheap

adapters/driven/providers/openai.py already translates the OpenAI Chat Completions shape in both directions (to_wire / from_wire), and these providers emit the same messages / tools payloads and the same finish_reason values (stop, tool_calls, length). The only thing hard-coded today is the client construction: AsyncOpenAI() with no base_url. Point that at a per-provider base URL + key and the existing translation works unchanged. application/loop.py does not move (load-bearing rule).

Scope

• Extend OpenAIGateway.__init__ to accept base_url and an api_key sourced from a configurable env var; keep the default (no base_url) byte-for-byte identical to today.
• Decouple the pricing/name identity from the stop-reason mapping key: a compat gateway keeps its own name (e.g. xai, for provider:model pricing keys) but maps finish reasons through the existing "openai" table in domain/model/stop_reason.py (make map_stop_reason's provider key a parameter of the gateway, not self.name).
• Introduce a small provider registry — a data table {name: (base_url, env_var)} — consumed by composition.py::make_gateway, replacing per-provider if branches for the compat family with one lookup.
• Wire Grok (xAI) end-to-end through the registry: base_url=https://api.x.ai/v1, key from XAI_API_KEY. Add xai:<model> pricing rows to dryfire/data/pricing.yaml.
• pyproject.toml: no new SDK — the openai extra already covers compat providers. Document that in the extra's comment.
• Offline contract tests: recorded xAI payloads → from_wire / to_wire round-trips, including a tool-call turn and a malformed-arguments turn (defensive parse must hold).
• One @pytest.mark.live smoke test behind XAI_API_KEY.

Seam checklist (the 5 touchpoints, for reference)

#	File	Change
1	adapters/driven/providers/*	Reuse openai.py translation via the registry; no new adapter
2	domain/model/stop_reason.py	Map compat providers through the openai table
3	composition.py::make_gateway	Registry lookup + env-var check
4	pyproject.toml	Comment: openai extra covers compat providers
5	dryfire/data/pricing.yaml	xai:<model> rows

Acceptance

• make check passes (lint + typecheck + arch + test), all offline.
• A suite with provider: xai runs against a recorded cassette with zero new adapter code beyond the seam.
• The Anthropic and OpenAI paths are unchanged (existing v0.1/v0.2 backward-compat tests green).

Out of scope

Native non-compat wire shapes (Gemini) — tracked separately.

[csmatar]

Implemented in #79. Two decisions worth recording against the original scope:

• Registered openrouter alongside xai. OpenRouter is the single-key path that reaches every frontier + open-weight model in this milestone, so it doubles as the live/recording path (backs the live smoke test) — while xai (api.x.ai/v1) remains the named reference the ticket called for. Both are one registry row each in composition.OPENAI_COMPATIBLE.
• Pricing rows deliberately omitted. The bundled pricing.yaml carries only Anthropic rows; the OpenAI adapter itself shipped unpriced in v0.2 (advisory cost → None). Fabricating xAI/OpenRouter prices offline would break the "never a guess" rule and make _meta.source wrong. Left as a data-only follow-up / user pricing_file.