[Bugs]: Context window is hardcoded to 200k for many models instead of resolving dynamically from actual provider/model metadata #35863

Description

OpenCode's context window handling relies on static/hardcoded values (models.generated.js / models.dev
catalog snapshots) instead of resolving each model's actual context window dynamically. This causes
auto-compaction, context tracking, and overflow checks to trigger far too early — even when the
selected model supports a much larger context (e.g. 1M).

This is not isolated to one provider — the same root cause shows up across multiple models:

• Claude Opus/Sonnet with context1m enabled still compact at ~200k instead of ~1M (Auto-compaction triggers at ~200k tokens instead of model's actual context window (1M) #15871, closed
  as not planned, but the underlying issue is unresolved)
• DeepSeek-V4-Pro shows 262K instead of the correct 1M (System prompt reports wrong context window for DeepSeek-V4-Pro (262K instead of 1M) #24119, Incorrect token count and pricing calculation for DeepSeek models #27458)
• GPT-5.5 Codex uses stale 1M metadata even after the provider dropped the effective limit to 400k (GPT-5.5 Codex context limit dropped to 400k but uses API context limit (1M) causing "exceeds the context window of this model" errors #24171)
• GitHub Copilot models use a static catalog lower than the live API-reported limit (Copilot CLI has higher context window than GitHub Copilot extension #16861)

Expected Behavior:
Context window per model should be resolved dynamically — either from the provider's live /models
endpoint when available, or from an up-to-date, easily-updatable model registry — instead of a
hardcoded fallback (200k) baked into the binary. This applies to ALL models/providers, not just
Anthropic's context1m case.

Actual Behavior:
Context window limit is compiled/cached at build time per model. When a provider updates a model's
actual context window (larger or smaller), OpenCode keeps using the stale hardcoded value until a
new release ships.

Proposed Solution:

1. Attempt to fetch context window from the provider's live model-list API at session start (with caching)
2. Fall back to a versioned, community-updatable model registry (not compiled into the binary) when
   live discovery isn't available
3. Only fall back to a generic default (e.g. 200k) as a last resort, with a visible warning in TUI/GUI
   so users know the value is an estimate, not the real limit
4. Respect user-defined limit.context overrides in opencode.json above all of the above (already
   partially supported)

Related: #15871, #24119, #27458, #24171, #16861

Plugins

N/A (core behavior)

OpenCode version

No response

Steps to reproduce

1. Select a model whose real provider-side context window differs from OpenCode's compiled/cached value
   (e.g. any model with context1m enabled, or DeepSeek-V4-Pro, or GPT-5.5 Codex)
2. Run a long session and monitor the context/compaction tracker in TUI
3. Observe compaction/overflow triggers at ~200k (or another stale hardcoded value) instead of the
   model's actual advertised context window

Screenshot and/or share link

No response

Operating System

Win11

Terminal

PowerShell 7