compaction: eliminate tool-call hallucinations from open-weight models

esrehmki · ESRE-dev · commit 3bd0eac9df90 · 2026-04-17T17:55:52.000+02:00
* Qwen3 and similar models see structured tool-call/tool-result
      markup in the conversation history and imitate it during
      compaction, either as structured API calls or raw XML text

FIX

    * convert tool-call/tool-result parts into plain text summaries
      before sending to the compaction model, so it never sees
      tool markup to imitate
    * silently ignore any remaining hallucinated tool calls in
      the processor instead of crashing the session
diff --git a/packages/opencode/src/session/compaction.ts b/packages/opencode/src/session/compaction.ts
@@ -216,6 +216,28 @@ When constructing the summary, try to stick to this template:
       const msgs = structuredClone(messages)
       yield* plugin.trigger("experimental.chat.messages.transform", {}, { messages: msgs })
       const modelMessages = yield* MessageV2.toModelMessagesEffect(msgs, model, { stripMedia: true })
+
+      // Convert structured tool-call/tool-result parts into plain text so the
+      // compaction model never sees tool markup and can't hallucinate tool calls.
+      for (const msg of modelMessages) {
+        if (!Array.isArray(msg.content)) continue
+        msg.content = msg.content.map((part: any) => {
+          if (part.type === "tool-call") {
+            const inputStr = typeof part.input === "string" ? part.input : JSON.stringify(part.input)
+            const truncatedInput = inputStr.length > 300 ? inputStr.slice(0, 300) + "... [truncated]" : inputStr
+            return { type: "text" as const, text: `[Called tool: ${part.toolName}]\n[Input: ${truncatedInput}]` }
+          }
+          if (part.type === "tool-result") {
+            const outputStr = typeof part.output === "string" ? part.output : JSON.stringify(part.output)
+            const truncatedOutput = outputStr.length > 500 ? outputStr.slice(0, 500) + "... [truncated]" : outputStr
+            return { type: "text" as const, text: `[Tool result: ${part.toolName}]\n${truncatedOutput}` }
+          }
+          return part
+        })
+      }
+      // Remove tool-role messages that are now empty or redundant after transformation
+      const compactionMessages = modelMessages.filter((msg) => msg.role !== "tool")
+
       const ctx = yield* InstanceState.context
       const msg: MessageV2.Assistant = {
         id: MessageID.ascending(),
@@ -256,7 +278,7 @@ When constructing the summary, try to stick to this template:
         tools: {},
         system: [],
         messages: [
-          ...modelMessages,
+          ...compactionMessages,
           {
             role: "user",
             content: [{ type: "text", text: prompt }],
diff --git a/packages/opencode/src/session/processor.ts b/packages/opencode/src/session/processor.ts
@@ -258,7 +258,8 @@ export const layer: Layer.Layer<
 
           case "tool-input-start":
             if (ctx.assistantMessage.summary) {
-              throw new Error(`Tool call not allowed while generating summary: ${value.toolName}`)
+              log.warn("ignoring tool call during summary generation", { tool: value.toolName })
+              return
             }
             const part = yield* session.updatePart({
               id: ctx.toolcalls[value.id]?.partID ?? PartID.ascending(),
@@ -286,7 +287,8 @@ export const layer: Layer.Layer<
 
           case "tool-call": {
             if (ctx.assistantMessage.summary) {
-              throw new Error(`Tool call not allowed while generating summary: ${value.toolName}`)
+              log.warn("ignoring tool call during summary generation", { tool: value.toolName })
+              return
             }
             yield* updateToolCall(value.toolCallId, (match) => ({
               ...match,
