feat(ai): point the platform at any model, not just Anthropic #5561

Closes the one objection that would have cost us developers at launch.

The objection was portability, not uptime

Git, CI and merges contain zero AI references, so an Anthropic outage could never take the platform down. But the question a developer evaluating Gluecron actually asks is "can I point this at my own model?" — and the answer was no, not without editing source.

That is what "solely dependent on Anthropic" would have come to mean in a launch thread. Not availability. Portability.

It was a one-function problem

Despite 53 files using AI, the entire codebase contained exactly one runtime construction of the Anthropic SDK, in ai-client.getAnthropic(). The nine import type Anthropic lines are erased at compile time, and the only other match lives inside a template string of generated user code.

So a provider only has to satisfy the slice those 53 files actually consume:

client.messages.create({ model, max_tokens, messages, system? })
  -> { content: [{type:"text", text}], usage: {input_tokens, output_tokens} }

ai-provider.ts speaks that shape over an OpenAI-compatible /chat/completions endpoint. Not a single call site changed.

AI_PROVIDER=openai            # or inferred from AI_BASE_URL alone
AI_BASE_URL=http://localhost:11434
AI_MODEL=llama3
AI_API_KEY=                   # local runtimes need none

Because the OpenAI chat shape is the de-facto standard, one adapter reaches Ollama, LM Studio, vLLM, llama.cpp, OpenRouter, Together, Groq, Fireworks, Azure OpenAI and OpenAI itself — including models running entirely on the operator's own hardware, with no key and no egress.

Two details that decide whether this is real or nominal

• isAiAvailable() does not require a key on the OpenAI path. A model served by Ollama on the same box authenticates nothing. Demanding a key there would report "AI unavailable" on a working local install — exactly the false negative that makes self-hosting feel broken.
• modelForTask() / internalModel() yield to AI_MODEL. The light/standard split is a Claude-pricing optimisation that means nothing to a runtime serving one model, and without this every request would 400 on an unknown model id.

Testing the failure that would otherwise hide

A dropped system prompt or a mis-read response returns an empty string — so every AI feature would appear to work while producing confident blanks. That is the failure mode the 12 new tests exist to catch: system-prompt placement, content flattening, legacy text fields, malformed/empty responses, absent auth headers, and error bodies surfaced rather than swallowed.

Usage counts floor to 0 rather than NaN, because a NaN propagates into the spend ledger and poisons every later total.

Verification

• 12 new tests, all passing
• bun run typecheck clean
• Existing AI consumers unaffected: ai-ci-healer, ai-quota-metering, copilot — 35 pass, 0 fail

Follow-on

The launch page should now demo this rather than claim it: AI_BASE_URL pointed at a local Ollama, running a real PR review, is a far stronger answer than any paragraph.

🤖 Generated with Claude Code