[heydryft]

What and why

Adds runcrate as an explicit PGBOT_AI_PROVIDER alias for explain / ask, following the same shape as the opper alias from #118. Runcrate serves open models over an OpenAI-compatible /chat/completions API, so the existing OpenAIProvider covers it and only the endpoint, key variable and default model differ.

• PGBOT_AI_PROVIDER=runcrate sends /chat/completions to https://api.runcrate.ai/v1 with RUNCRATE_API_KEY (or PGBOT_AI_API_KEY).
• It is never auto-detected and never picks up an OpenAI or OpenRouter key, the same as requesty and opper.
• Runcrate does not serve the OpenAI default gpt-5.6-terra, so the alias defaults to deepseek-ai/DeepSeek-V4.1-Flash when neither PGBOT_AI_MODEL nor PGBOT_OPENAI_MODEL is set. Any id from https://api.runcrate.ai/v1/models works through PGBOT_AI_MODEL.

Files: internal/ai/openai.go (endpoint and default model constants), internal/ai/resolve.go (alias, key variable, default model, keyVarsFor), internal/ai/resolve_test.go (TestResolve_runcrate), README.md, CHANGELOG.md (Unreleased, Added).

Keys are created at https://www.runcrate.ai/dashboard/api-keys?utm_source=pgbot&utm_medium=directory&utm_campaign=gpu_pricing .

Testing

• gofmt -l .: no output
• go vet ./...: clean
• go test ./internal/ai/ -run TestResolve -v: all 19 TestResolve_* pass, including the new TestResolve_runcrate (both key variables, endpoint, default model, PGBOT_AI_MODEL override, refuses an OpenAI key, not auto-detected)
• scripts/gate.sh: HEAD builds, vets, and tests clean (6 targets)
• Live check: Resolve() with PGBOT_AI_PROVIDER=runcrate and a real RUNCRATE_API_KEY, then Generate() with a system and user turn (throwaway test, not committed). deepseek-ai/DeepSeek-V4.1-Flash (the default), zai-org/GLM-5.3-Flash and Qwen/Qwen3.8-Flash each returned "pong" with finish_reason=stop.

Checklist

• scripts/gate.sh passes (builds HEAD, not just the working tree)
• New SQL is read-only; no EXPLAIN ANALYZE; findings stay deterministic (computed in Go) (no SQL touched)
• No PII enters a model.Context / --json / the store (unchanged)
• --json change is additive, or model.SchemaVersion bumped + schema regenerated (no --json change)
• A new finding has a docs/findings/<id>.md page + catalog entry (no new finding)

This PR was prepared with AI assistance: an AI coding agent wrote the change and ran the checks above.

Disclosure: I work on Runcrate.