fix(copilot): capture resolved model for auto-routed runs

[ldavidgomez]

Summary

This PR captures the concrete model resolved by the Copilot SDK for auto-routed runs.

Previously, when an agent was configured with model: auto, Conductor recorded AgentOutput.model as "auto" even though the Copilot SDK had resolved the request to a concrete model. Token usage was available, but downstream accounting could not attribute that usage to the actual model.

This change captures event.data.model from the SDK assistant.usage event, stores it as SDKResponse.resolved_model, preserves it through response reconstruction paths, and uses it for AgentOutput.model when available.

What changed

• Add resolved_model to SDKResponse.
• Capture the resolved model from assistant.usage.
• Preserve resolved_model through _execute_sdk_call.
• Use the resolved model when constructing AgentOutput.
• Keep existing fallback behavior when no resolved model is available.
• Add tests for:
  • resolved model propagation;
  • fallback behavior;
  • cost calculation when the resolved model is priceable;
  • SDK assistant.usage event extraction.

Validation

Unit tests:

• uv run pytest tests/test_providers/test_copilot.py::TestCopilotProviderResolvedModel -v — passed
• uv run pytest tests/test_providers/test_copilot.py -v — passed

Local smoke tests with Copilot:

• model: auto structured-output run resolved to a concrete model instead of "auto".
• explicit resolved-model run succeeded.
• model: auto no-schema/raw-output run resolved to a concrete model instead of "auto".

Observed resolved models included gpt-5.3-codex / gpt-5.4-mini depending on the run.

Notes

This PR does not attempt to calculate GitHub Copilot credit or USD cost. Copilot pricing/accounting is plan and policy dependent and may change independently from the SDK. The change records the concrete resolved model so downstream pricing or accounting logic can make an informed decision instead of receiving "auto".

[jrob5756]

Reviewed at the PR head in an isolated checkout. Lint, format, typecheck, and the full copilot suite (88 tests) pass, and the approach matches how claude_agent_sdk.py already reads msg.model from the SDK, so nothing here blocks merge.

My main note is one behavioral point. The capture isn't limited to model: auto: assistant.usage.model is always set, so the resolved name replaces the configured model on every run. For an explicitly pinned priceable model that can swap in a name the pricing table doesn't recognize (some dotted Copilot IDs like claude-sonnet-4.5 and gemini-2.5-pro aren't in it), and then cost_usd goes None and quietly drops out of the budget total in usage.py, so spend can be undercounted with no error. Suggestion inline.

Two smaller things. claude.py still reports the configured model rather than the SDK-served one, so the three providers now diverge on this. Anthropic has no auto-routing so it's a reasonable choice, but AGENTS.md treats model as a parity-tracked field, so a line in the PR description would make the divergence intentional. And the parsed-JSON branch that agents with an output schema take isn't covered by the new tests, so its resolved_model propagation could regress unnoticed.

[ldavidgomez]

Updated.

The main execution path now only uses the SDK resolved_model when the configured agent model is None or "auto", so explicitly configured models remain authoritative for reporting and pricing.

I also threaded the configured agent model into the Copilot follow-up path and applied the same precedence there:

• auto / default-routed agents use the SDK resolved model when available;
• explicitly configured agent models are preserved;
• provider default remains the final fallback.

I also updated the resolved_model docstring to clarify that it can be None when the usage event is absent or when the event does not carry a usable model name.

Validation:

• uv run pytest tests/test_providers/test_copilot.py -k resolved_model -q
• uv run pytest tests/test_providers/test_copilot.py -q

[jrob5756]

All five review comments are addressed (auto-gating, docstring, follow-up path threads agent.model, and tests now exercise execute() including the explicit-priceable + unpriceable-resolved precedence case). Pushed db188c8 to restore the dropped cache_write_tokens on the main-path AgentOutput (fixed the F841 lint + a silent cache-write accounting regression) and a format fix. CI is green.

[codecov-commenter]

Codecov Report

❌ Patch coverage is 75.00000% with 2 lines in your changes missing coverage. Please review.
⚠️ Please upload report for BASE (main@5d769eb). Learn more about missing BASE report.

Files with missing lines	Patch %	Lines
src/conductor/providers/copilot.py	75.00%	2 Missing ⚠️

Additional details and impacted files

@@           Coverage Diff           @@
##             main     #268   +/-   ##
=======================================
  Coverage        ?   86.48%           
=======================================
  Files           ?       69           
  Lines           ?    12157           
  Branches        ?        0           
=======================================
  Hits            ?    10514           
  Misses          ?     1643           
  Partials        ?        0           

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.
• 📦 JS Bundle Analysis: Save yourself from yourself by tracking and limiting bundle sizes in JS merges.