[henrikdukefoss]

Add Standard Compute as a built-in provider.
It uses the same pattern as the Cheaper Inference provider in #879.

Standard Compute is an OpenAI-compatible LLM gateway.
One API key gives access to open and frontier models.

It is wired on rig's OpenAI completions client.
The default endpoint is https://api.stdcmpt.com/v1.
The key is STANDARDCOMPUTE_API_KEY.
The default model is StandardCompute, which is routed: each request goes to the cheapest enabled model that can handle it, with automatic fallback.
Pinned model ids use the vendor/model form listed by GET https://api.stdcmpt.com/v1/models.
Reasoning effort goes out as a top-level reasoning_effort.
It is not part of key autodetection.

Usage:

export STANDARDCOMPUTE_API_KEY="..."
dirge --provider standardcompute  # defaults to StandardCompute

Details

Files mirror #879: src/provider/{resolve,client,dispatch,mod,build,adapter,summarize,stream_dispatch,mod_tests}.rs, src/compression.rs, src/config/mod.rs, src/agent/agent_loop/{h7_smoke,rig_stream_factory}.rs, plus README.md, docs/config.md, docs/features.md, CHANGELOG.md.

Tests: new unit tests for parse, default model, base URL, key lookup, built-in name protection, and the top-level reasoning_effort shape, mirroring the Cheaper Inference ones. The h7 smoke scenarios skip when STANDARDCOMPUTE_API_KEY is unset.

I did not build or run the tests locally; I'm relying on CI here.

Disclosure: I work on Standard Compute.

🤖 Generated with Claude Code