@link-assistant/agent 0.18.0 → 0.18.2

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@link-assistant/agent",
-  "version": "0.18.0",
+  "version": "0.18.2",
   "description": "A minimal, public domain AI CLI agent compatible with OpenCode's JSON interface. Bun-only runtime.",
   "main": "src/index.js",
   "type": "module",

package/src/index.js

@@ -802,8 +802,14 @@ async function main() {
           compactJson: isCompact,
         });
 
-        // Monkey-patch globalThis.fetch for raw HTTP logging in --verbose mode.
-        // Catches ALL HTTP calls regardless of AI SDK fetch passthrough. (#217)
+        // Global fetch monkey-patch for verbose HTTP logging (#221).
+        // This catches any HTTP calls that go through globalThis.fetch directly,
+        // including non-provider calls (auth, config, tools) that may not have
+        // their own createVerboseFetch wrapper. The provider-level wrapper in
+        // provider.ts getSDK() also logs independently — both mechanisms are
+        // kept active to maximize HTTP observability in --verbose mode.
+        // See: https://github.com/link-assistant/agent/issues/221
+        // See: https://github.com/link-assistant/agent/issues/217
         if (!globalThis.__agentVerboseFetchInstalled) {
           globalThis.fetch = createVerboseFetch(globalThis.fetch, {
             caller: 'global',

package/src/provider/provider.ts

@@ -1201,11 +1201,13 @@ export namespace Provider {
         sessionID: provider.id,
       });
 
-      // Verbose HTTP logging is handled by the global fetch monkey-patch
-      // (installed in CLI middleware in index.js). The global patch catches ALL
-      // HTTP calls reliably, regardless of how the AI SDK passes fetch internally.
-      // This provider-level wrapper is kept as a fallback for environments where
-      // the global patch may not be installed (e.g., programmatic use).
+      // Verbose HTTP logging for provider API calls.
+      // This provider-level wrapper logs HTTP requests/responses independently
+      // of the global fetch monkey-patch. Both mechanisms are kept active to
+      // maximize HTTP observability — the global patch may miss calls if the
+      // AI SDK captures/resolves fetch references before it is installed,
+      // while this wrapper is injected directly into the SDK's fetch option.
+      // See: https://github.com/link-assistant/agent/issues/221
       // See: https://github.com/link-assistant/agent/issues/217
       // See: https://github.com/link-assistant/agent/issues/215
       {
@@ -1226,14 +1228,9 @@ export namespace Provider {
           init?: RequestInit
         ): Promise<Response> => {
           // Check verbose flag at call time — not at SDK creation time.
-          // When the global fetch monkey-patch is installed, it handles verbose
-          // logging for all calls. The provider wrapper is a fallback for
-          // environments without the global patch.
-          // See: https://github.com/link-assistant/agent/issues/217
-          if (
-            !Flag.OPENCODE_VERBOSE ||
-            globalThis.__agentVerboseFetchInstalled
-          ) {
+          // This ensures --verbose works even when the flag is set after SDK creation.
+          // See: https://github.com/link-assistant/agent/issues/206
+          if (!Flag.OPENCODE_VERBOSE) {
             return innerFetch(input, init);
           }
 

package/src/session/summary.ts

@@ -94,22 +94,47 @@ export namespace SessionSummary {
     const assistantMsg = messages.find((m) => m.info.role === 'assistant')!
       .info as MessageV2.Assistant;
 
-    // Use the same model as the main session (--model) instead of a small model
-    // This ensures consistent behavior and uses the model the user explicitly requested
-    // See: https://github.com/link-assistant/agent/issues/217
-    log.info(() => ({
-      message: 'loading model for summarization',
-      providerID: assistantMsg.providerID,
-      modelID: assistantMsg.modelID,
-      hint: 'Using same model as --model (not a small model)',
-    }));
-    const model = await Provider.getModel(
-      assistantMsg.providerID,
-      assistantMsg.modelID
-    ).catch(() => null);
+    // Use the compaction model (--compaction-model, e.g. gpt-5-nano) for summarization
+    // to avoid doubling rate-limit pressure on the main model.
+    // If the compaction model is unavailable, fall back to the main model.
+    // See: https://github.com/link-assistant/agent/issues/223
+    const compactionModel = userMsg.compactionModel;
+    let model: Awaited<ReturnType<typeof Provider.getModel>> | null = null;
+
+    if (compactionModel && !compactionModel.useSameModel) {
+      model = await Provider.getModel(
+        compactionModel.providerID,
+        compactionModel.modelID
+      ).catch(() => null);
+      if (model) {
+        log.info(() => ({
+          message: 'loading model for summarization',
+          providerID: model!.providerID,
+          modelID: model!.modelID,
+          hint: 'Using compaction model to reduce rate-limit pressure on main model',
+          mainModelID: assistantMsg.modelID,
+        }));
+      }
+    }
+
+    if (!model) {
+      // Fall back to the main model if compaction model is not configured or unavailable
+      log.info(() => ({
+        message: 'loading model for summarization',
+        providerID: assistantMsg.providerID,
+        modelID: assistantMsg.modelID,
+        hint: compactionModel
+          ? 'Compaction model unavailable, falling back to main model'
+          : 'Using same model as --model (no compaction model configured)',
+      }));
+      model = await Provider.getModel(
+        assistantMsg.providerID,
+        assistantMsg.modelID
+      ).catch(() => null);
+    }
     if (!model) {
       log.info(() => ({
-        message: 'could not load session model for summarization, skipping',
+        message: 'could not load model for summarization, skipping',
         providerID: assistantMsg.providerID,
         modelID: assistantMsg.modelID,
       }));