[BUG] Bug Report: Opencode Go Provider Incorrectly Overrides Hardcoded contextWindow with API-Returned Value 200K

[SpiritLiberation]

Bug Report: Opencode Go Provider Incorrectly Overrides Hardcoded contextWindow with API-Returned Value

Environment

• Extension: Zoo Code (community edition of Roo Code)
• Version: 3.63.100191
• File: dist/extension.js
• Provider: opencode-go
• Model: deepseek-v4-flash (natively supports 1M context window)

Summary

When using the opencode-go provider, the $Zo() function (model info merger) incorrectly allows the API response's context_window field to override the hardcoded default value from the f3i table. Since the Opencode Go API (https://opencode.ai/zen/go/v1/models) returns context_window: 200000 as a default for all models, models that are correctly defined in f3i with contextWindow: 1e6 (1M) end up being constrained to 200K.

Root Cause

Code Path

wye (opencode-go provider) → bS.fetchModel() → Rv() → R7n() → cHn() → $Zo()

1. wye class extends bS base class — handles opencode-go provider
2. bS.fetchModel() calls Rv({provider:"opencode-go", ...}) which fetches models via cHn()
3. cHn() sends GET to https://opencode.ai/zen/go/v1/models and parses each model entry through $Zo()
4. $Zo() merges hardcoded defaults (from fur() → f3i map) with API response data

The Bug (in $Zo)

// BEFORE FIX (buggy):
$Zo = t => {
  let e = fur(t.id),                         // hardcoded defaults from f3i
      r = t.context_window ?? t.context_length, // API response
      a = t.max_output_tokens ?? t.max_tokens,
      o = t.supports_images;
  return e ? {
    ...e,                                          // spread hardcoded (correct contextWindow: 1e6)
    ...r !== void 0 && {contextWindow: r},          // ← BUG: API's 200K OVERRIDES hardcoded 1M
    ...a !== void 0 && {maxTokens: a},
    ...o !== void 0 && {supportsImages: o},
    description: t.description ?? t.name ?? e.description
  } : {
    maxTokens: a ?? Lie.maxTokens,
    contextWindow: r ?? Lie.contextWindow,
    // ...
  };
};

When a model IS in the hardcoded f3i table (like deepseek-v4-flash with contextWindow: 1e6), the line ...r !== void 0 && {contextWindow: r} unconditionally overwrites the context window with whatever the API returns. The Opencode Go API returns context_window: 200000 for all models, destroying the correct 1M value.

The Fix

// AFTER FIX:
$Zo = t => {
  let e = fur(t.id),
      r = t.context_window ?? t.context_length,
      a = t.max_output_tokens ?? t.max_tokens,
      o = t.supports_images;
  return e ? {
    ...e,                                          // hardcoded contextWindow: 1e6 preserved
    ...a !== void 0 && {maxTokens: a},
    ...o !== void 0 && {supportsImages: o},
    description: t.description ?? t.name ?? e.description
  } : {
    maxTokens: a ?? Lie.maxTokens,
    contextWindow: r ?? Lie.contextWindow,          // API value still used for UNKNOWN models
    // ...
  };
};

Change: Removed ...r !== void 0 && {contextWindow: r} from the e (hardcoded model) branch. When a model has hardcoded defaults in f3i, the hardcoded contextWindow is now authoritative. The API-provided contextWindow is still used for unknown models (the fallback branch after :).

Additional Data: Cached File

The cached model file at:

~/.config/Code/User/globalStorage/zoocodeorganization.zoo-code/cache/opencode-go_models.json

contained ALL models with contextWindow: 200000, confirming the API returns a blanket 200K for every model:

{"deepseek-v4-flash": {"maxTokens":32768,"contextWindow":200000,...},...}

Impact

All models served via the opencode-go provider that are defined in the f3i hardcoded table with larger context windows than 200K are affected. This includes:

• deepseek-v4-flash (should be 1M → was 200K)
• deepseek-v4-pro (should be 1M → was 200K)
• minimax-m3 (should be 1M → was 200K)
• glm-5.2 (should be 1M → was 200K)
• mimo-v2.5 (should be 1M → was 200K)
• qwen3.6-plus / qwen3.7-plus / qwen3.7-max (should be 1M → was 200K)

Proposed Fix Upstream

In dist/extension.js, the $Zo function should be modified to not override the hardcoded contextWindow with the API-provided value when the model exists in the f3i table. The fix is a one-line removal.

Suggested patch:

- return e ? {...e, ...r !== void 0 && {contextWindow: r}, ...a !== void 0 && {maxTokens: a}, ...o !== void 0 && {supportsImages: o}, description: t.description ?? t.name ?? e.description}
+ return e ? {...e, ...a !== void 0 && {maxTokens: a}, ...o !== void 0 && {supportsImages: o}, description: t.description ?? t.name ?? e.description}

---

Report generated on 2026-06-25

[navedmerchant]

I dont get this issue? I have been using 1M context regularly on the latest version. Can you check if v 3.64 fixes your issue?

[navedmerchant]

btw https://opencode.ai/zen/go/v1/models never returns a context window. I think they want the harness to configure it