[BUG] Rig agent ignores `[llm].model` and `CODEGRAPH_MODEL`, defaulting Anthropic to Sonnet

[gudzenkov]

Summary

The Rig agent path does not use the configured [llm].model or the CODEGRAPH_MODEL environment variable. With Anthropic selected, it defaults to claude-sonnet-4-20250514 unless CODEGRAPH_LLM_MODEL or CODEGRAPH_AGENT_MODEL is set.

This makes CLI/config output misleading and can unexpectedly route requests to a more constrained or more expensive model.

Related

• No matching issue or PR found for CODEGRAPH_MODEL, CODEGRAPH_LLM_MODEL, Rig model selection, or Anthropic defaulting to Sonnet.

Clean Reproduction

Config file:

[llm]
provider = "anthropic"
model = "claude-haiku-4-5"
context_window = 200000
max_retries = 1

Environment:

CODEGRAPH_LLM_PROVIDER=anthropic
CODEGRAPH_MODEL=claude-haiku-4-5

Official MCP call with correct project id:

{
  "name": "agentic_context",
  "arguments": {
    "query": "Find where this feature is implemented. Return concrete files and symbols.",
    "focus": "search",
    "limit": 1
  }
}

Observed error shows the request went to Sonnet, not Haiku:

429 Too Many Requests ...
model: claude-sonnet-4-20250514
rate limit of 30,000 input tokens per minute

When rerun with:

CODEGRAPH_LLM_MODEL=claude-haiku-4-5
CODEGRAPH_AGENT_MODEL=claude-haiku-4-5

the same path used Haiku:

model: claude-haiku-4-5-20251001
rate limit of 50,000 input tokens per minute

Code Evidence

crates/codegraph-mcp-rig/src/adapter/llm_adapter.rs:

pub fn get_model_name() -> String {
    env::var("CODEGRAPH_LLM_MODEL")
        .or_else(|_| env::var("CODEGRAPH_AGENT_MODEL"))
        .unwrap_or_else(|_| default_model_for_provider())
}

fn default_model_for_provider() -> String {
    match RigProvider::from_env() {
        Ok(RigProvider::Anthropic) => "claude-sonnet-4-20250514".to_string(),
        // ...
    }
}

The Rig path does not read config.llm.model from ConfigManager and does not read CODEGRAPH_MODEL, even though ConfigManager supports CODEGRAPH_MODEL.

Expected

Model selection should be consistent across config and runtime paths:

• [llm].model should affect the Rig MCP path.
• CODEGRAPH_MODEL should either work consistently or be documented as not applicable.
• If Rig requires CODEGRAPH_LLM_MODEL, the CLI/config docs should make that explicit.

Actual

The official MCP Rig path silently defaults Anthropic to Sonnet despite config/env selecting Haiku.

Impact

Users can unknowingly run a different model from the one configured, changing cost, quality, context limits, and provider rate-limit behavior.