| Field | Value |
|---|---|
| Decision ID | ADR-025 |
| Initiative | unsorry swarm provider portability |
| Proposed By | unsorry maintainers |
| Date | 2026-06-13 |
| Status | Accepted |
In the context of an OpenAI provider hardwired to https://api.openai.com/v1 with a fixed model allow-list, now that ADR-022 / #292 enabled coordinated --prove for openai, and a desire to run local or self-hosted OpenAI-compatible models (Ollama / vLLM / LM Studio / a proxy) and to reuse an operator’s existing pi-coder endpoint catalogue,
facing no way to point the provider at a custom endpoint or to use arbitrary local model ids, and the friction of re-declaring an endpoint, key, and model that already live in ~/.pi/agent/models.json,
we decided for (1) an OPENAI_BASE_URL override (precedence: explicit arg → env → OpenAI’s default) plus a model passthrough that bypasses the allow-list and attaches tools whenever a custom endpoint is in effect, reusing the OpenAI Chat-Completions client unchanged; and (2) a -pi toggle on agent.sh that resolves the endpoint, key, and model from ~/.pi/agent/models.json by the existing UNSORRY_MODEL name (DRY — no new model argument), via a pure-stdlib resolver tools/llm_providers/pi_config.py, then drives the same OpenAI-compatible path in both --prove-local and coordinated --prove,
and neglected parsing models.json in bash (untestable, needs jq), a --pi flag inside openai_cli.py (couples the HTTP client to pi’s schema and must be wired into both CLI codepaths), and a new --pi-model argument (violates DRY against the existing UNSORRY_MODEL),
to achieve local-model proving and translation against the same Gate A kernel-verification bar, with one-toggle reuse of an operator’s pi endpoint catalogue,
accepting that only pi api: "openai-completions" entries are usable (others error clearly); local models that lack OpenAI function-calling cannot drive the --prove tool loop (translation works on any model); apiKey $ENV references are expanded but command-style (!cmd) keys are rejected rather than executed; the pi schema is treated as observed-and-tolerated; and trust still rests solely on Gate A re-checking every proof, independent of which endpoint or model produced it.
| Reference ID | Title | Type | Location |
|————–|——-|——|———-|
| REF-1 | OpenAI-compatible endpoints and pi-config | Specification | specs/SPEC-025-A-OpenAI-Compatible-Endpoints-and-Pi-Config.md |
| REF-2 | Local provider smoke mode (the path -pi rides) | ADR | ADR-022-Local-Provider-Smoke-Mode.md |
| REF-3 | Agent loop script | Specification | specs/SPEC-007-A-Agent-Loop-Script.md |
| Status | Approver | Date | |——–|———-|——| | Proposed | unsorry maintainers | 2026-06-13 | | Accepted | unsorry maintainers | 2026-06-13 |