feat: add context_budget_cap and context window to provider defaults

[Salil Das (sadlilas)]

Summary

Adds context_budget_cap: 200_000, context_window: 1_048_576, and max_output_tokens: 65_536 to Gemini's ProviderInfo.defaults.

Why

Budget cap: Gemini charges 2x input / 1.5x output above 200k tokens, matching Anthropic's pricing cliff. The context module (context-simple) needs a signal from the provider to cap its working budget at the cost-optimal threshold.

Context window / max output: These were missing from get_info().defaults, causing context-simple to fall through to its 200k fallback instead of calculating a dynamic budget from Gemini's actual 1M context window. While the fallback happens to land near the right number, the budget should be calculated correctly and then capped -- not accidentally correct via a fallback path.

What changes

defaults={
    "model": "gemini-2.5-flash",
    "max_tokens": 8192,
    "temperature": 0.7,
    "timeout": 600.0,
    "context_window": 1_048_576,       # NEW
    "max_output_tokens": 65_536,       # NEW
    "context_budget_cap": 200_000,     # NEW
},

Design

Part of the provider-driven budget cap approach. See amplifier-support#57 analysis for the full rationale.

Unlike Anthropic (which conditionally sets the cap only when 1M is enabled), Gemini always sets it because all Gemini models report 1M+ context windows and all have the pricing cliff.

Related PRs

• provider-anthropic#27: Same context_budget_cap: 200_000 (conditional on 1M enabled)
• context-simple: Reads context_budget_cap from provider defaults and applies budget = min(raw_budget, cap - safety_margin)

Related issues

• microsoft-amplifier/amplifier-support#57

---

🤖 Generated with Amplifier