fix(compaction): preserve agent identity across compaction boundaries

packages/opencode/src/session/compaction.ts

@@ -183,7 +183,7 @@ export namespace SessionCompaction {
         // Allow plugins to inject context or replace compaction prompt.
         const compacting = yield* plugin.trigger(
           "experimental.session.compacting",
-          { sessionID: input.sessionID },
+          { sessionID: input.sessionID, agent: userMessage.agent },
           { context: [], prompt: undefined },
         )
         const defaultPrompt = `Provide a detailed prompt for continuing our conversation above.
@@ -194,6 +194,10 @@ Respond in the same language as the user's messages in the conversation.
 
 When constructing the summary, try to stick to this template:
 ---
+## Agent Role & Constraints
+
+[If the system prompt indicates a specialized agent role (e.g. evaluator, reviewer, judge, explorer, planner), state the agent name, its role, and any behavioral constraints (read-only, no implementation, output format requirements). If the agent is a general-purpose implementor, write "Default agent — no special constraints." Frame next steps in terms appropriate to the agent's role: evaluators should evaluate, reviewers should review, planners should plan — do NOT frame all agents as implementors.]
+
 ## Goal
 
 [What goal(s) is the user trying to accomplish?]
@@ -217,9 +221,39 @@ When constructing the summary, try to stick to this template:
 ---`
 
         const prompt = compacting.prompt ?? [defaultPrompt, ...compacting.context].join("\n\n")
+        // Resolve the source agent to preserve its identity during compaction
+        const source = yield* agents.get(userMessage.agent)
+        const system: string[] = []
+        if (source?.prompt) {
+          const max = 4000
+          const truncated =
+            source.prompt.length > max ? source.prompt.slice(0, max) + "\n[...truncated]" : source.prompt
+          system.push(truncated)
+        }
         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(),
@@ -259,9 +293,10 @@ When constructing the summary, try to stick to this template:
             agent,
             sessionID: input.sessionID,
             tools: {},
-            system: [],
+            toolChoice: "none",
+            system,
             messages: [
-              ...modelMessages,
+              ...compactionMessages,
               {
                 role: "user",
                 content: [{ type: "text", text: prompt }],
@@ -283,6 +318,11 @@ When constructing the summary, try to stick to this template:
         }
 
         if (result === "continue" && input.auto) {
+          // Inject post-compaction agent identity reminder for specialized agents
+          const reminder =
+            source?.prompt && !source?.native
+              ? `<system-reminder>You are the "${userMessage.agent}" agent. Your role and constraints from your system prompt still apply after this compaction. Do not deviate from your assigned role.</system-reminder>`
+              : undefined
           if (replay) {
             const original = replay.info
             const replayMsg = yield* session.updateMessage({
@@ -310,6 +350,17 @@ When constructing the summary, try to stick to this template:
                 sessionID: input.sessionID,
               })
             }
+            if (reminder) {
+              yield* session.updatePart({
+                id: PartID.ascending(),
+                messageID: replayMsg.id,
+                sessionID: input.sessionID,
+                type: "text",
+                synthetic: true,
+                text: reminder,
+                time: { start: Date.now(), end: Date.now() },
+              })
+            }
           }
 
           if (!replay) {
@@ -332,7 +383,7 @@ When constructing the summary, try to stick to this template:
               sessionID: input.sessionID,
               type: "text",
               synthetic: true,
-              text,
+              text: reminder ? text + "\n\n" + reminder : text,
               time: {
                 start: Date.now(),
                 end: Date.now(),

packages/opencode/src/session/processor.ts

@@ -152,7 +152,8 @@ export namespace SessionProcessor {
 
             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
               }
               ctx.toolcalls[value.id] = yield* session.updatePart({
                 id: ctx.toolcalls[value.id]?.id ?? PartID.ascending(),
@@ -173,7 +174,8 @@ export namespace SessionProcessor {
 
             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
               }
               const match = ctx.toolcalls[value.toolCallId]
               if (!match) return