feat(goose): honor GOOSE_FAST_MODEL env var in ModelConfig::with_fast

[vampik33]

Closes #9295.

Summary

Adds GOOSE_FAST_MODEL env var, honored at the single call site inside ModelConfig::with_fast in crates/goose/src/model.rs. When set and non-empty it replaces the provider-supplied fast-model default; when unset/empty (or whitespace-only) behavior is unchanged.

Mirrors the existing GOOSE_MODEL env-var pattern for the primary model.

Motivation

Each provider's from_env() calls with_fast(...) with a hardcoded constant (e.g. OPENROUTER_DEFAULT_FAST_MODEL = "google/gemini-2.5-flash", OPEN_AI_DEFAULT_FAST_MODEL = "gpt-4o-mini"). use_fast_model() is invoked by Goose internals for auxiliary calls (tool-selection, classification, session-title generation) independent of GOOSE_MODEL — so today users have no way to redirect that traffic.

Concrete impact: cost-controlled deployments that intentionally pin a single model on OpenRouter still bleed spend on Gemini 2.5 Flash because the auxiliary path is unreachable from config. Privacy/compliance setups have the same gap. Closely related to the closed #4350 / #4318, which addressed canonical lookups but left the override gap in place.

Design

Single point of override at the existing with_fast call site means all 8 providers (openrouter, openai, anthropic, google, gemini_oauth, ollama, databricks, kimicode) pick it up via existing call sites without per-provider edits.

let resolved = std::env::var("GOOSE_FAST_MODEL")
    .ok()
    .map(|v| v.trim().to_string())
    .filter(|v| !v.is_empty())
    .unwrap_or_else(|| fast_model_name.to_string());
let fast_config = ModelConfig::new(&resolved)?.with_canonical_limits(provider_name);

The trim+filter handles both unset and accidentally-empty/whitespace values so a stray GOOSE_FAST_MODEL= in .env doesn't break the auxiliary path.

Tests

Three unit tests in the existing model::tests module using the same env_lock pattern that covers GOOSE_MAX_TOKENS / GOOSE_TEMPERATURE:

• test_with_fast_uses_default_when_env_unset — preserves existing behavior
• test_with_fast_honors_env_override — env var wins over provider default
• test_with_fast_ignores_empty_env — whitespace-only value is treated as unset

$ cargo test -p goose --lib model::tests::test_with_fast
running 3 tests
test model::tests::test_with_fast_honors_env_override ... ok
test model::tests::test_with_fast_ignores_empty_env ... ok
test model::tests::test_with_fast_uses_default_when_env_unset ... ok
test result: ok. 3 passed; 0 failed

cargo fmt -p goose --check and cargo clippy -p goose --lib --no-deps -- -D warnings both clean.

Backwards compatibility

No behavior change when GOOSE_FAST_MODEL is unset or empty. No new dependencies. No public API change beyond a new env var honored by an existing method.

Out of scope (follow-ups, happy to address separately)

• Config-file field on ModelConfig for non-env-var users — chose env-var-only for this PR per CONTRIBUTING.md "start small" guidance.
• Per-provider override (OPENROUTER_FAST_MODEL, etc.) — 8x the surface area for the same outcome when users typically run one provider per session.
• A GOOSE_DISABLE_FAST_MODEL=1 flag to skip the fast-model optimization entirely.

Checklist

• Code compiles with cargo build
• Tests pass with cargo test
• Clippy clean with -D warnings
• cargo fmt --check clean
• Linked to issue (Feature: GOOSE_FAST_MODEL env var to override hardcoded per-provider fast-model defaults #9295)

[DOsinga]

Pushed a cleanup commit: dropped the verbose comment, simplified the trim/filter chain (avoids unnecessary allocation), removed the three trivial tests, and added GOOSE_FAST_MODEL to the environment variables documentation.

One thing left: your original commit is missing the DCO sign-off (Signed-off-by line). The DCO check will fail without it. You can fix this with:

git commit --amend -s
git push --force-with-lease

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: d02655204e

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

Trim GOOSE_FAST_MODEL before constructing fast model config

The new override path checks v.trim().is_empty() but then passes the untrimmed String into ModelConfig::new, so values like " gpt-4o-mini " are treated as non-empty yet produce a model name with embedded whitespace. In that case auxiliary fast-model calls can fail with an invalid model identifier even though the intended model is correct; trimming the value before ModelConfig::new would make this override behave consistently with the emptiness check.

Useful? React with 👍 / 👎.