fix: get_info() reports runtime-effective model, not packaged default

[Salil Das (sadlilas)]

Summary

get_info() reports the packaged static model name from config/_models.py instead of the runtime-configured effective model, violating the module's own documented model selection priority and creating an internal inconsistency within the function itself.

The bug

contracts/behaviors.md (Model Selection Policy, MUST:1) states:

MUST respect model selection priority: request.model > config["default_model"] > YAML

get_info() already applies this priority correctly for defaults["context_window"] and defaults["max_output_tokens"] — both are resolved against self._effective_default_model (which itself correctly implements the priority: self.config.get("default_model") or self._provider_config.defaults["model"]). But defaults["model"] was left untouched, so it always reported the packaged YAML default (claude-opus-4.5) regardless of what model was actually configured at runtime — even while the window fields, computed a few lines later in the same function, were correctly reporting values for the real effective model.

Reproduction (with config={"default_model": "mai-code-1.1-flash"}):

provider._effective_default_model  == "mai-code-1.1-flash"   # correct
provider.get_info().defaults["model"] == "claude-opus-4.5"   # WRONG

This is unconditional — it reproduces identically whether the model-info cache is warm or cold, since the model field was never wired to the effective-model resolution at all.

User-visible impact: Consumers that read get_info().defaults["model"] to determine the active model (e.g. to emit a provider:resolve-style event driving a live model display) show the wrong model whenever a caller overrides default_model at runtime (for example via a routing/config layer that assigns different models per sub-agent).

The fix

One line, plus an explanatory comment, in get_info():

defaults["model"] = self._effective_default_model

Placed unconditionally — before the model-info cache lookup, not inside the if info is not None: branch that corrects the window fields — so the reported model is correct even when the model cache is cold. The cold-cache fallback behavior for context_window / max_output_tokens is intentionally unchanged.

This brings the provider in line with sibling Amplifier providers (Anthropic, Chat Completions), which already report the runtime-resolved model from get_info() rather than a static default.

Tests

Added to the existing TestRuntimeConfigOverride class in tests/test_behaviors.py (the established home for Model Selection Policy tests):

1. test_get_info_reports_runtime_default_model — config={"default_model": ...} is reflected in get_info().defaults["model"].
2. test_get_info_reports_yaml_default_without_runtime_config — no default_model configured still reports the YAML default (no regression).
3. test_get_info_reports_runtime_model_on_cold_cache — with a cold model-info cache (lookup returns None), the model field is still correct, proving the fix applies outside the if info is not None: branch, while window-field cold-cache fallback stays unchanged.
4. test_get_info_does_not_mutate_shared_cached_defaults — a second provider instance with a different config is unaffected by a prior instance's get_info() call, since cfg.defaults is a process-wide @lru_cached singleton.

No existing test asserted the previous (incorrect) behavior.

Quality gates

• pytest tests/ -m "not live" — 1558 passed, 0 failed
• ruff check amplifier_module_provider_github_copilot/ tests/ — clean
• pyright . — 0 errors, 0 warnings

Scope

Out of scope (unchanged, deliberately): the cold-cache fallback behavior for context_window / max_output_tokens, get_model_info(), config/_models.py, FALLBACKS, resolve_effective_window, and the model cache implementation.

Generated with Amplifier

Co-Authored-By: Amplifier 240397093+microsoft-amplifier@users.noreply.github.com