Describe the bug

When using a model configured in GOOSE_PREDEFINED_MODELS with a custom context_limit (e.g., 1M for Claude Opus 4.6 with the context-1m beta header), the context limit resolves to 200k instead of 1M. This causes auto-compaction to fire at ~160k tokens (80% of 200k), destroying conversation history mid-session.

Two issues combine:

1. GOOSE_CONTEXT_LIMIT in config.yaml is silently ignored. The code reads it via std::env::var("GOOSE_CONTEXT_LIMIT") in ModelConfig::new_base() (crates/goose/src/model.rs), which checks the process environment — not config.yaml. The desktop app (Electron) doesn't source shell profiles, so the value never reaches the code.

2. Canonical registry overrides predefined models. In ModelConfig::with_canonical_limits(), the canonical registry lookup resolves goose-claude-4-6-opus → anthropic/claude-opus-4.6 → context: 200000 and sets self.context_limit. The predefined models fallback (context_limit: 1000000) is only checked if self.context_limit.is_none() — but it's already set to 200k by the canonical registry.

Resolution chain:

To Reproduce

1. Set GOOSE_CONTEXT_LIMIT: 1000000 in ~/.config/goose/config.yaml
2. Select goose-claude-4-6-opus as the model (via UI dropdown or config)
3. Confirm GOOSE_PREDEFINED_MODELS env var includes the model with context_limit: 1000000 and request_params: { anthropic_beta: ["context-1m-2025-08-07"] }
4. Start a session and check the info-msg — context limit reports as 200k, not 1M
5. Auto-compaction fires at ~160k tokens

Expected behavior

Context limit should be 1,000,000 tokens. Either:

• GOOSE_CONTEXT_LIMIT from config.yaml should be respected (priority 1), OR
• context_limit from the matching GOOSE_PREDEFINED_MODELS entry should take priority over the canonical registry default

Screenshots

N/A — reproducible via info-msg token reporting.

Please provide the following information

• OS & Arch: macOS (Apple Silicon)
• Interface: UI (desktop app)
• Version: 1.27.2-block
• Extensions enabled: Computer Controller, Memory, Extension Manager
• Provider & Model: Databricks – goose-claude-4-6-opus

Additional context

Suggested fix (Option A — minimal): In with_canonical_limits(), check predefined models BEFORE the canonical registry. Predefined models are explicitly configured by the deployment team and should take priority over generic canonical defaults:

// Check predefined models first (explicit config > generic defaults)
if self.context_limit.is_none() {
    if let Some(pm) = find_predefined_model(&self.model_name) {
        self.context_limit = pm.context_limit;
    }
}

// Fall back to canonical registry
if self.context_limit.is_none() {
    if let Some(canonical) = ... {
        self.context_limit = Some(canonical.limit.context);
    }
}

Suggested fix (Option B — also needed): Make GOOSE_CONTEXT_LIMIT in config.yaml work by reading it via Config::global().get_param() (like GOOSE_MAX_TOKENS already does) instead of std::env::var().

Workaround: Set GOOSE_CONTEXT_LIMIT as a real environment variable visible to GUI apps via launchctl setenv GOOSE_CONTEXT_LIMIT 1000000, then restart the app.

Other reports: This was also reported in #goose-help on Slack (March 10, 2026) without resolution. Related: GitHub issue #4357 (closed October 2025) addressed a different surface of the same underlying problem.