doctor: surface per-model reasoning-effort support and context-window limits

[jrob5756]

Summary

Extend conductor doctor --models (and the underlying ProviderDiagnostic/--json output) to surface, per model:

1. Reasoning effort support — which reasoning.effort levels (low/medium/high/xhigh) a model accepts, plus its default.
2. Context window — the model's prompt/output/total token limits.

Motivation

Today conductor doctor --models only lists model ids (_models_cell in src/conductor/cli/doctor.py). Users configuring reasoning.effort or worried about prompt size have no way to discover per-model limits without hitting a ValidationError at runtime or reading provider source. This data is already fetched internally for validation — it just isn't surfaced to the user.

Current state (investigated)

• Copilot provider (src/conductor/providers/copilot.py):

  • list_models() already calls client.list_models(), which returns Model objects with:
    • capabilities.supports.reasoning_effort (bool)
    • supported_reasoning_efforts: list[str] | None
    • default_reasoning_effort: str | None
    • capabilities.limits.max_prompt_tokens, max_output_tokens, max_context_window_tokens
  • _validate_reasoning_effort_for_model already reads supported_reasoning_efforts for validation — same data, not exposed to diagnostics.
  • get_max_prompt_tokens() already surfaces max_prompt_tokens for the dashboard context-window bar, but doctor doesn't call it.

• Claude provider (src/conductor/providers/claude.py):

  • list_models() / get_max_prompt_tokens() use Anthropic's models.list(), which only exposes max_input_tokens — no output/total split, no per-model reasoning-effort list.
  • Reasoning support is a static heuristic: is_claude_thinking_model() (prefix match on claude-3-7-, claude-opus-4, claude-sonnet-4, claude-haiku-4) — a boolean "supports extended thinking", not a level list. All 4 Conductor effort levels map to fixed token budgets (EFFORT_TO_BUDGET_TOKENS in src/conductor/providers/reasoning.py) when supported.

• Claude Agent SDK provider: declares reasoning_effort: None in CAPABILITIES (no concept at all) — should report "n/a" for both new columns.

Proposed approach

• Add optional provider hooks alongside the existing get_max_prompt_tokens / get_model_pricing pattern in AgentProvider (src/conductor/providers/base.py), e.g. get_model_capabilities(model) -> ModelCapabilityInfo | None, returning:
  • supported_reasoning_efforts: list[str] | None
  • default_reasoning_effort: str | None
  • max_prompt_tokens, max_output_tokens, max_context_window_tokens: int | None
  • Default implementation returns None (safe degradation to "n/a"), matching the existing optional-hook contract (never raise).
• Copilot: implement via the existing client.list_models() cache.
• Claude: implement using is_claude_thinking_model() + the 4 Conductor levels for reasoning, and max_input_tokens for prompt tokens only (no output/total available from the SDK).
• Claude Agent SDK (experimental, reasoning_effort capability = None): leave unimplemented (base default → "n/a").
• Extend ProviderDiagnostic (src/conductor/providers/diagnostics.py) and gather_provider()/gather() to populate this per-model when --models is passed.
• Extend _models_cell / add new columns in src/conductor/cli/doctor.py (_render_providers) for Rich table output, and mirror new fields in DoctorReport.to_dict() for --json.
• Degrade gracefully to n/a/— wherever a provider or model doesn't report the data — must not fail doctor or affect its exit code.

Out of scope

• Bumping the github-copilot-sdk pin (currently >=1.0.0, installed 1.0.1, latest stable 1.0.6) — not required since the needed fields already exist in 1.0.1. Tracked separately if desired.

[jrob5756]

Starting to take a look at this.