feat(provider): add gateway provider for OpenAI-compatible endpoints

[mabry1985]

Summary

All existing providers (codex, claude, acpx, cursor, opencode, grok, pi) shell out to a local CLI binary that has to be installed + OAuth'd on the host. That's a poor fit for containerized agent deployments where the only available LLM channel is an HTTP endpoint.

This adds a gateway provider that POSTs the already-assembled prompt to an OpenAI-compatible /chat/completions endpoint with structured outputs (response_format: json_schema). No subprocess; no extra CLI to install; same JSON contract as every other provider.

Designed for the protoLabs LiteLLM gateway:

• Inside docker network: http://gateway:4000/v1
• Externally: https://api.proto-labs.ai/v1

…but works against any OpenAI-compatible endpoint (vanilla OpenAI, vLLM, LM Studio, Ollama with the OpenAI shim).

Config

Env	Default	Notes
GATEWAY_API_KEY (preferred) or OPENAI_API_KEY	required	Bearer token
OPENAI_BASE_URL	https://api.proto-labs.ai/v1	Trailing slashes stripped
CLAWPATCH_GATEWAY_MODEL	protolabs/smart	--model on the CLI wins
CLAWPATCH_GATEWAY_TIMEOUT_MS	300000 (5 min)	Reasoning models on large features can be slow
--reasoning-effort	n/a	Forwarded as reasoning_effort body field

Why this can be the minimal provider

buildReviewPrompt / buildMapPrompt / buildFixPrompt already inline the relevant file contents as Files blocks, so the gateway provider needs zero file IO — it's effectively the smallest possible provider: prompt in, JSON out.

Smoke-test results

Verified end-to-end against api.proto-labs.ai/v1:

• 401 path: missing/wrong key returns the LiteLLM auth error verbatim ✅
• Successful request reaches the gateway and back through the wire ✅
• (Cloudflare 524 hit on full reviews because reasoning takes >100s — irrelevant for the in-cluster http://gateway:4000 path which is the actual deploy target.)

Test plan

• 11 new unit tests on the config helper covering env precedence, URL normalization, model resolution, timeout parsing, and the no-API-key auth-error path
• pnpm test --run — 724 pass / 1 pre-existing skipped
• pnpm typecheck clean

Follow-ups

• Wire clawpatch ci --since <ref> --provider gateway into Quinn's pr_review (separate PR in protoWorkstacean)
• Bake clawpatch into the workstacean Docker image