Goose Custom Provider Context Limit Bug

Date: 2026-05-13
Issue: Custom provider context_limit is ignored; goose always shows 0/128k instead of the configured value (e.g., 204800).

Symptom

When using a custom provider like custom_minimax with model MiniMax-M2.7 and context_limit: 204800, goose displays:

  ╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌ 0% 0/128k

instead of the expected 0/200k.

Root Cause

ModelConfig::with_canonical_limits at crates/goose/src/model.rs:113-136 only resolves context limits from two sources:

1. Canonical model registry — built-in providers (openai, anthropic, google, etc.)
2. Predefined models — from the GOOSE_PREDEFINED_MODELS environment variable

It never checks the models array inside declarative/custom provider JSON files (~/.config/goose/custom_providers/*.json).

When the model name (MiniMax-M2.7) is unknown to both sources, context_limit stays None, and ModelConfig::context_limit() falls back to DEFAULT_CONTEXT_LIMIT = 128_000 (crates/goose/src/model.rs:8).

Affected code paths

All of these call ModelConfig::new(...).with_canonical_limits(provider_name), which bypasses custom provider configs.

Where models IS used (and where it is not)

The models array in ~/.config/goose/custom_providers/custom_minimax.json is not completely useless, but its context_limit is ignored at runtime.

Used

NOT used

Workarounds (no code changes)

Option 1: GOOSE_CONTEXT_LIMIT environment variable

Highest priority. Overrides everything, including canonical limits.

export GOOSE_CONTEXT_LIMIT=204800
goose session

Validation: must be > 4096.

Option 2: GOOSE_PREDEFINED_MODELS environment variable

Model-scoped fallback. Checked after canonical registry, before the 128k default.

export GOOSE_PREDEFINED_MODELS='[{"name":"MiniMax-M2.7","context_limit":204800}]'
goose session

Multiple models:

export GOOSE_PREDEFINED_MODELS='[
  {"name":"MiniMax-M2.7","context_limit":204800},
  {"name":"MiniMax-Other","context_limit":128000}
]'

With provider-specific request parameters:

export GOOSE_PREDEFINED_MODELS='[{
  "name": "MiniMax-M2.7",
  "context_limit": 204800,
  "request_params": {
    "anthropic_beta": "token-efficient-tools-2025-02-19"
  }
}]'

Note: GOOSE_PREDEFINED_MODELS is read via std::env::var directly (model.rs:21). It cannot be set in ~/.config/goose/config.yaml; it must be an environment variable.

Option 3: Wrapper script

#!/bin/bash
export GOOSE_CONTEXT_LIMIT=204800
exec goose "$@"

Relevant source locations

Proper fix (requires code change)

In ModelConfig::with_canonical_limits, after the predefined-models check, add a fallback that loads the declarative provider config and searches its models array:

#![allow(unused)]
fn main() {
// Try filling remaining gaps from declarative/custom providers
if self.context_limit.is_none() {
    if let Ok(loaded_provider) =
        crate::config::declarative_providers::load_provider(provider_name)
    {
        if let Some(model_info) = loaded_provider
            .config
            .models
            .iter()
            .find(|m| m.name == self.model_name)
        {
            self.context_limit = Some(model_info.context_limit);
        }
    }
}
}

This would fix the bug for CLI, server, and default-model creation paths in one place.