Summary

The opencode-go plugin defines a static model catalog with 19+ models including correct contextWindow values (e.g., deepseek-v4-pro: 1_000_000). However, the plugin does NOT register these into the runtime model context window cache, so resolveContextTokensForModel() falls back to 200K.

Impact

Every agent using opencode-go models has premature compaction — the threshold is calculated as contextWindow - reserveTokens - softThresholdTokens, which becomes 156K instead of 956K (6x too frequent).

$ openclaw models list | grep deepseek
opencode-go/deepseek-v4-pro   text   195k   # should be 977k (1M)
opencode-go/deepseek-v4-flash text   195k   # should be 977k (1M)

Root Cause

1. Plugin catalog (provider-catalog.js) defines correct contextWindow: 1e6
2. resolveContextTokensForModel() looks up lookupContextTokens("opencode-go/deepseek-v4-pro") → miss
3. Falls back to ?? 2e5 (200K) in resolveMemoryFlushContextWindowTokens()

The plugin openclaw.plugin.json includes modelCatalog.providers.opencode-go but the onboarding pipeline does not populate the MODEL_CONTEXT_TOKEN_CACHE from the plugin catalog.

Evidence

Source (provider-catalog-BakzcJID.js:12-35):

const OPENCODE_GO_MODELS = [
  { id: "deepseek-v4-pro",  contextWindow: 1e6,   maxTokens: 384e3 },
  { id: "deepseek-v4-flash", contextWindow: 1e6,   maxTokens: 384e3 },
  // ... 17 more models, all with correct contextWindow
];

Runtime (context-fjzFpTQH.js:215-252):

function resolveContextTokensForModel(params) {
  const bareResult = lookupContextTokens(params.model, ...);
  // "deepseek-v4-pro" -> NOT in cache -> undefined
  return params.fallbackContextTokens; // not passed
}

Caller (agent-runner.runtime-Duta-cpW.js:244-251):

function resolveMemoryFlushContextWindowTokens(params) {
  return resolveContextTokensForModel({...}) ?? 2e5; // 200K fallback
}

Real-world data from live gateway (agent:main session, 2026-06-14):

• deepseek-v4-pro shows contextTokens: 1000000 in systemPromptReport
• Compaction fired at 204K tokens instead of ~956K (confirmed from compactionCheckpoints)
• 2 compactions in 24 minutes with high-density conversation

Workaround

Manually add models.providers.opencode-go in openclaw.json with full model list including contextWindow.

Proposed Fix

The plugin onboarding pipeline should register model contextWindow values into the runtime cache so lookupContextTokens returns correct values for all plugin-provided models, without requiring manual config overrides.