Environment variables

Variable	Default	Purpose
AI_API_KEY	None	Provider-neutral bearer token. Takes precedence when set.
OPENCODE_API_KEY	None	OpenCode Zen token used when AI_API_KEY is unset.
AI_BASE_URL	https://opencode.ai/zen/v1	Base URL for an OpenAI-compatible API.

When neither key variable is set, the default endpoint receives public as its bearer credential. Other providers normally require AI_API_KEY. The conversation model is pinned in app/ai-agent/model.ts; Requireganizer does not silently change models because model identity affects output quality, caching, and reproducibility.

Copy .env.example to .env.local for local development. Do not commit credentials.

Compatibility requirements

The configured service and model must support OpenAI-compatible streaming chat completions with:

• request tools entries of type: "function" containing JSON Schema parameters;
• streamed deltas at choices[0].delta, including reasoning_content for thinking-capable models;
• usage reported on the final chunk without stream_options (the gateway hangs when stream_options.include_usage is sent);
• no reliance on the store request field.

The model must accept reasoning effort values of low, high, or max; these are mapped from pi thinking levels in app/ai-agent/model.ts.

Retry and timeout behavior

The proxy route performs a single streaming attempt per turn and performs no application-level retries. Transient provider failures surface as an error event in the conversation stream and then as a validation error in the UI, where the command can be reissued.

Proposal repair happens inside the conversation: validation failures of a submitted proposal are returned to the model as tool results so it can correct and resubmit within the same turn.