Skip to content
erode
Announcements

ADR-011: Intent-based provider generation profiles

Status: Accepted
Date: 2026-05-05
Authors: Anders Hassis

Context

Section titled “Context”

Erode runs the same analysis stages across Gemini, Anthropic, and OpenAI. The shared provider flow used raw token counts when it called each model. That made stage code carry provider-specific budget details.

OpenAI’s Responses API names the limit max_output_tokens, and reasoning models can spend part of that budget before producing visible output. Anthropic and Gemini expose different request parameters and have different response behavior. The shared pipeline needs to describe what it wants, not how each provider should size the request.

Decision

Section titled “Decision”

Introduce a provider-agnostic generation profile for model calls. A profile describes the expected output size and reasoning effort. BaseProvider maps each analysis phase to a default profile, and concrete providers translate the profile into SDK-specific request parameters.

Keep model tier selection unchanged. Component resolution, dependency scanning, and model patching use the fast model. Drift analysis uses the advanced model.

Rationale

Section titled “Rationale”

The analysis stages know the shape of the work. Providers know their API parameters and model behavior. Keeping those concerns separate makes it easier to tune OpenAI, Anthropic, and Gemini independently without changing the shared pipeline.

The profile also makes cost intent visible. Simple extraction work stays on cheaper models with small outputs and low reasoning effort. Drift analysis keeps the stronger model, but still defaults to low reasoning effort to keep feedback fast. Model updates use medium reasoning because they generate a concrete patch.

Consequences

Section titled “Consequences”

Positive

Section titled “Positive”
  • Provider-specific token parameters stay inside provider implementations.
  • OpenAI can use max_output_tokens and reasoning effort without leaking those names into shared analysis code.
  • Cost-aware model tier selection remains explicit in BaseProvider.

Negative

Section titled “Negative”
  • Output profile names become a shared contract that providers must translate.
  • Dynamic output hints can still affect provider cost, latency, quota usage, or model output caps.
Section titled “Related commits”
  • 9289f90 - refactor openai provider with new api