🤖 fix: retry compaction with truncation fallback

docs/agents/index.mdx

@@ -321,8 +321,6 @@ name: Compact
 description: History compaction (internal)
 ui:
   hidden: true
-providerOptions:
-  truncationMode: auto
 subagent:
   runnable: false
 ---

src/common/orpc/schemas/agentDefinition.ts

@@ -36,11 +36,6 @@ const AgentDefinitionSubagentSchema = z
   })
   .strip();
 
-const AgentDefinitionProviderOptionsSchema = z
-  .object({
-    truncationMode: z.enum(["auto", "disabled"]).optional(),
-  })
-  .strip();
 const AgentDefinitionAiDefaultsSchema = z
   .object({
     // Model identifier: full string (e.g. "anthropic:claude-sonnet-4-5") or abbreviation (e.g. "sonnet")
@@ -83,7 +78,6 @@ export const AgentDefinitionFrontmatterSchema = z
 
     subagent: AgentDefinitionSubagentSchema.optional(),
 
-    providerOptions: AgentDefinitionProviderOptionsSchema.optional(),
     ai: AgentDefinitionAiDefaultsSchema.optional(),
 
     // Tool configuration: add/remove patterns (regex).

src/node/builtinAgents/compact.md

@@ -3,8 +3,6 @@ name: Compact
 description: History compaction (internal)
 ui:
   hidden: true
-providerOptions:
-  truncationMode: auto
 subagent:
   runnable: false
 ---

src/node/services/agentDefinitions/builtInAgentContent.generated.ts

@@ -3,7 +3,7 @@
 // Source: src/node/builtinAgents/*.md
 
 export const BUILTIN_AGENT_CONTENT = {
-  "compact": "---\nname: Compact\ndescription: History compaction (internal)\nui:\n  hidden: true\nproviderOptions:\n  truncationMode: auto\nsubagent:\n  runnable: false\n---\n\nYou are running a compaction/summarization pass. Your task is to write a concise summary of the conversation so far.\n\nIMPORTANT:\n\n- You have NO tools available. Do not attempt to call any tools or output JSON.\n- Simply write the summary as plain text prose.\n- Follow the user's instructions for what to include in the summary.\n",
+  "compact": "---\nname: Compact\ndescription: History compaction (internal)\nui:\n  hidden: true\nsubagent:\n  runnable: false\n---\n\nYou are running a compaction/summarization pass. Your task is to write a concise summary of the conversation so far.\n\nIMPORTANT:\n\n- You have NO tools available. Do not attempt to call any tools or output JSON.\n- Simply write the summary as plain text prose.\n- Follow the user's instructions for what to include in the summary.\n",
   "exec": "---\nname: Exec\ndescription: Implement changes in the repository\nui:\n  color: var(--color-exec-mode)\nsubagent:\n  runnable: true\n  append_prompt: |\n    If you are running as a sub-agent in a child workspace:\n\n    - When you have a final answer, call agent_report exactly once.\n    - Do not call task/task_await/task_list/task_terminate (subagent recursion is disabled).\n    - Do not call propose_plan.\ntools:\n  add:\n    # Allow all tools by default (includes MCP tools which have dynamic names)\n    # Use tools.remove in child agents to restrict specific tools\n    - .*\n  remove:\n    # Exec mode doesn't use planning tools\n    - propose_plan\n    - ask_user_question\n---\n\nYou are in Exec mode.\n\n- Make minimal, correct, reviewable changes that match existing codebase patterns.\n- Prefer targeted commands and checks (typecheck/tests) when feasible.\n- Treat as a standing order: keep running checks and addressing failures until they pass or a blocker outside your control arises.\n",
   "explore": "---\nname: Explore\ndescription: Read-only repository exploration\nbase: exec\nui:\n  hidden: true\nsubagent:\n  runnable: true\n  append_prompt: |\n    You are an Explore sub-agent running inside a child workspace.\n\n    - Explore the repository to answer the prompt using read-only investigation.\n    - Return concise, actionable findings (paths, symbols, callsites, and facts).\n    - When you have a final answer, call agent_report exactly once.\n    - Do not call agent_report until you have completed the assigned task.\ntools:\n  # Remove editing and task tools from exec base (read-only agent)\n  remove:\n    - file_edit_.*\n    - task\n    - task_.*\n    - agent_skill_read\n    - agent_skill_read_file\n---\n\nYou are in Explore mode (read-only).\n\n=== CRITICAL: READ-ONLY MODE - NO FILE MODIFICATIONS ===\n\n- You MUST NOT create, edit, delete, move, or copy files.\n- You MUST NOT create temporary files anywhere (including /tmp).\n- You MUST NOT use redirect operators (>, >>, |) or heredocs to write to files.\n- You MUST NOT run commands that change system state (rm, mv, cp, mkdir, touch, git add/commit, installs, etc.).\n- Use bash only for read-only operations (rg, ls, cat, git diff/show/log, etc.).\n",
   "plan": "---\nname: Plan\ndescription: Create a plan before coding\nui:\n  color: var(--color-plan-mode)\nsubagent:\n  runnable: false\ntools:\n  add:\n    # Allow all tools by default (includes MCP tools which have dynamic names)\n    # Use tools.remove in child agents to restrict specific tools\n    - .*\n  # Note: file_edit_* tools ARE available but restricted to plan file only at runtime\n  # Note: task tools ARE enabled - Plan delegates to Explore sub-agents\n---\n\nYou are in Plan Mode.\n\n- Every response MUST produce or update a plan—no exceptions.\n- Simple requests deserve simple plans; a straightforward task might only need a few bullet points. Match plan complexity to the problem.\n- Keep the plan scannable; put long rationale in `<details>/<summary>` blocks.\n- When Plan Mode is requested, assume the user wants the actual completed plan; do not merely describe how you would devise one.\n\nDetailed plan mode instructions (plan file path, sub-agent delegation, propose_plan workflow) are provided separately.\n",

src/node/services/agentDefinitions/parseAgentDefinitionMarkdown.test.ts

@@ -13,8 +13,6 @@ description: Does stuff
 base: exec
 tools:
   add: ["file_read", "bash.*"]
-providerOptions:
-  truncationMode: auto
 unknownTopLevel: 123
 ui:
   hidden: false
@@ -33,7 +31,6 @@ Do the thing.
     expect(result.frontmatter.name).toBe("My Agent");
     expect(result.frontmatter.description).toBe("Does stuff");
     expect(result.frontmatter.base).toBe("exec");
-    expect(result.frontmatter.providerOptions?.truncationMode).toBe("auto");
     expect(result.frontmatter.tools).toEqual({ add: ["file_read", "bash.*"] });
     expect(result.frontmatter.ui?.hidden).toBe(false);
     expect(result.frontmatter.ui?.color).toBe("#ff00ff");

src/node/services/agentSession.ts

@@ -44,7 +44,7 @@ import type { TodoItem } from "@/common/types/tools";
 import type { PostCompactionAttachment, PostCompactionExclusions } from "@/common/types/attachment";
 import { TURNS_BETWEEN_ATTACHMENTS } from "@/common/constants/attachments";
 import { extractEditedFileDiffs } from "@/common/utils/messages/extractEditedFiles";
-import { isValidModelFormat } from "@/common/utils/ai/models";
+import { getModelName, getModelProvider, isValidModelFormat } from "@/common/utils/ai/models";
 import { materializeFileAtMentions } from "@/node/services/fileAtMentions";
 
 /**
@@ -146,6 +146,16 @@ export class AgentSession {
    * Cache the last-known experiment state so we don't spam metadata refresh
    * when post-compaction context is disabled.
    */
+  /** Track compaction requests that already retried with truncation. */
+  private readonly compactionRetryAttempts = new Set<string>();
+  /**
+   * Active compaction request metadata for retry decisions (cleared on stream end/abort).
+   */
+  private activeCompactionRequest?: {
+    id: string;
+    modelString: string;
+    options?: SendMessageOptions;
+  };
   private postCompactionContextEnabled = false;
 
   constructor(options: AgentSessionOptions) {
@@ -684,7 +694,8 @@ export class AgentSession {
 
   private async streamWithHistory(
     modelString: string,
-    options?: SendMessageOptions
+    options?: SendMessageOptions,
+    openaiTruncationModeOverride?: "auto" | "disabled"
   ): Promise<Result<void, SendMessageError>> {
     if (this.disposed) {
       return Ok(undefined);
@@ -711,6 +722,12 @@ export class AgentSession {
       );
     }
 
+    this.activeCompactionRequest = this.resolveCompactionRequest(
+      historyResult.data,
+      modelString,
+      options
+    );
+
     // Check for external file edits (timestamp-based polling)
     const changedFileAttachments = await this.getChangedFileAttachments();
 
@@ -728,7 +745,7 @@ export class AgentSession {
     // Bind recordFileState to this session for the propose_plan tool
     const recordFileState = this.recordFileState.bind(this);
 
-    return this.aiService.streamMessage(
+    const streamResult = await this.aiService.streamMessage(
       historyResult.data,
       this.workspaceId,
       modelString,
@@ -745,10 +762,147 @@ export class AgentSession {
       postCompactionAttachments,
       options?.experiments,
       options?.disableWorkspaceAgents,
-      () => !this.messageQueue.isEmpty()
+      () => !this.messageQueue.isEmpty(),
+      openaiTruncationModeOverride
+    );
+
+    if (!streamResult.success) {
+      this.activeCompactionRequest = undefined;
+    }
+
+    return streamResult;
+  }
+
+  private resolveCompactionRequest(
+    history: MuxMessage[],
+    modelString: string,
+    options?: SendMessageOptions
+  ): { id: string; modelString: string; options?: SendMessageOptions } | undefined {
+    for (let index = history.length - 1; index >= 0; index -= 1) {
+      const message = history[index];
+      if (message.role !== "user") {
+        continue;
+      }
+      if (!isCompactionRequestMetadata(message.metadata?.muxMetadata)) {
+        return undefined;
+      }
+      return {
+        id: message.id,
+        modelString,
+        options,
+      };
+    }
+    return undefined;
+  }
+
+  private async finalizeCompactionRetry(messageId: string): Promise<void> {
+    this.activeCompactionRequest = undefined;
+    this.emitChatEvent({
+      type: "stream-abort",
+      workspaceId: this.workspaceId,
+      messageId,
+    });
+    await this.clearFailedCompaction(messageId);
+  }
+
+  private async clearFailedCompaction(messageId: string): Promise<void> {
+    const [partialResult, deleteMessageResult] = await Promise.all([
+      this.partialService.deletePartial(this.workspaceId),
+      this.historyService.deleteMessage(this.workspaceId, messageId),
+    ]);
+
+    if (!partialResult.success) {
+      log.warn("Failed to clear partial before compaction retry", {
+        workspaceId: this.workspaceId,
+        error: partialResult.error,
+      });
+    }
+
+    if (
+      !deleteMessageResult.success &&
+      !(
+        typeof deleteMessageResult.error === "string" &&
+        deleteMessageResult.error.includes("not found in history")
+      )
+    ) {
+      log.warn("Failed to delete failed compaction placeholder", {
+        workspaceId: this.workspaceId,
+        error: deleteMessageResult.error,
+      });
+    }
+  }
+
+  private isGptClassModel(modelString: string): boolean {
+    return (
+      getModelProvider(modelString) === "openai" &&
+      getModelName(modelString).toLowerCase().startsWith("gpt-")
     );
   }
 
+  private async maybeRetryCompactionOnContextExceeded(data: {
+    messageId: string;
+    errorType?: string;
+  }): Promise<boolean> {
+    if (data.errorType !== "context_exceeded") {
+      return false;
+    }
+
+    const context = this.activeCompactionRequest;
+    if (!context) {
+      return false;
+    }
+
+    if (!this.isGptClassModel(context.modelString)) {
+      return false;
+    }
+
+    if (this.compactionRetryAttempts.has(context.id)) {
+      return false;
+    }
+
+    this.compactionRetryAttempts.add(context.id);
+
+    log.info("Compaction hit context limit; retrying once with OpenAI truncation", {
+      workspaceId: this.workspaceId,
+      model: context.modelString,
+      compactionRequestId: context.id,
+    });
+
+    await this.finalizeCompactionRetry(data.messageId);
+
+    const retryResult = await this.streamWithHistory(context.modelString, context.options, "auto");
+    if (!retryResult.success) {
+      log.error("Compaction retry failed to start", {
+        workspaceId: this.workspaceId,
+        error: retryResult.error,
+      });
+      return false;
+    }
+
+    return true;
+  }
+
+  private async handleStreamError(data: {
+    workspaceId: string;
+    messageId: string;
+    error: string;
+    errorType?: string;
+  }): Promise<void> {
+    if (await this.maybeRetryCompactionOnContextExceeded(data)) {
+      return;
+    }
+
+    this.activeCompactionRequest = undefined;
+
+    const streamError: StreamErrorMessage = {
+      type: "stream-error",
+      messageId: data.messageId,
+      error: data.error,
+      errorType: (data.errorType ?? "unknown") as StreamErrorMessage["errorType"],
+    };
+    this.emitChatEvent(streamError);
+  }
+
   private attachAiListeners(): void {
     const forward = (
       event: string,
@@ -792,10 +946,14 @@ export class AgentSession {
     forward("reasoning-delta", (payload) => this.emitChatEvent(payload));
     forward("reasoning-end", (payload) => this.emitChatEvent(payload));
     forward("usage-delta", (payload) => this.emitChatEvent(payload));
-    forward("stream-abort", (payload) => this.emitChatEvent(payload));
+    forward("stream-abort", (payload) => {
+      this.activeCompactionRequest = undefined;
+      this.emitChatEvent(payload);
+    });
     forward("runtime-status", (payload) => this.emitChatEvent(payload));
 
     forward("stream-end", async (payload) => {
+      this.activeCompactionRequest = undefined;
       const handled = await this.compactionHandler.handleCompletion(payload as StreamEndEvent);
       if (!handled) {
         this.emitChatEvent(payload);
@@ -824,13 +982,7 @@ export class AgentSession {
         error: string;
         errorType?: string;
       };
-      const streamError: StreamErrorMessage = {
-        type: "stream-error",
-        messageId: data.messageId,
-        error: data.error,
-        errorType: (data.errorType ?? "unknown") as StreamErrorMessage["errorType"],
-      };
-      this.emitChatEvent(streamError);
+      void this.handleStreamError(data);
     };
 
     this.aiListeners.push({ event: "error", handler: errorHandler });

src/node/services/aiService.ts

@@ -1023,6 +1023,7 @@ export class AIService extends EventEmitter {
    * @param changedFileAttachments Optional attachments for files that were edited externally
    * @param postCompactionAttachments Optional attachments to inject after compaction
    * @param disableWorkspaceAgents When true, read agent definitions from project path instead of workspace worktree
+   * @param openaiTruncationModeOverride Optional OpenAI truncation override (e.g., compaction retry)
    * @returns Promise that resolves when streaming completes or fails
    */
   async streamMessage(
@@ -1042,7 +1043,8 @@ export class AIService extends EventEmitter {
     postCompactionAttachments?: PostCompactionAttachment[] | null,
     experiments?: { programmaticToolCalling?: boolean; programmaticToolCallingExclusive?: boolean },
     disableWorkspaceAgents?: boolean,
-    hasQueuedMessage?: () => boolean
+    hasQueuedMessage?: () => boolean,
+    openaiTruncationModeOverride?: "auto" | "disabled"
   ): Promise<Result<void, SendMessageError>> {
     // Support interrupts during startup (before StreamManager emits stream-start).
     // We register an AbortController up-front and let stopStream() abort it.
@@ -1852,7 +1854,7 @@ export class AIService extends EventEmitter {
       }
 
       // Build provider options based on thinking level and message history
-      const truncationMode = agentDefinition.frontmatter.providerOptions?.truncationMode;
+      const truncationMode = openaiTruncationModeOverride;
       // Pass filtered messages so OpenAI can extract previousResponseId for persistence
       // Also pass callback to filter out lost responseIds (OpenAI invalidated them)
       // Pass workspaceId to derive stable promptCacheKey for OpenAI caching