@hiai-gg/hiai-opencode 0.1.1 → 0.1.3

package/src/internals/plugins/subtask2/hooks/message-hooks.ts

@@ -1,603 +1,603 @@
-import {
-  getClient,
-  getReturnState,
-  deleteReturnState,
-  setReturnState,
-  getPluginConfig,
-  getAllPendingReturns,
-  deletePendingReturn,
-  hasReturnStack,
-  peekReturnStack,
-  shiftReturnStack,
-  getCurrentReturnChain,
-  popCurrentReturnChain,
-  setHasActiveSubtask,
-  getHasActiveSubtask,
-  OPENCODE_GENERIC,
-  getSubtaskParentSession,
-  consumePendingPromptReturn,
-  setPendingStackedPromptResponse,
-  hasProcessedS2Message,
-  addProcessedS2Message,
-  getConfigs,
-} from "../core/state";
-import { log } from "../utils/logger";
-import { DEFAULT_PROMPT } from "../utils/config";
-import { S2_INLINE_INSTRUCTION } from "../utils/prompts";
-import { parseInlineSubtask } from "../parsing";
-import { executeInlineSubtask } from "../features/inline-subtasks";
-import { executeReturn } from "../features/returns";
-import {
-  getAllPendingEvaluations,
-  createEvaluationPrompt,
-  createYieldPrompt,
-} from "../loop";
-
-/**
- * Update the synthetic part in the database to make it visible in TUI.
- *
- * The v1 SDK from OpenCode exposes an internal hey-api HTTP client at client.client
- * which we use to PATCH the part. This is the only reliable way to update parts
- * from a plugin since client.part.update() doesn't exist in v1.
- *
- * Note: OpenCode's plugin system passes a v1 client. When/if they upgrade to v2,
- * we can use client.part.update() directly.
- */
-async function makePartVisible(
-  part: any,
-  msg: any,
-  newText: string
-): Promise<void> {
-  const client = getClient();
-  if (!client) {
-    log(`makePartVisible: no client available`);
-    return;
-  }
-
-  // Extract IDs from the part and message
-  const partID = part.id;
-  const messageID = part.messageID || msg?.info?.id;
-  const sessionID = part.sessionID || msg?.info?.sessionID;
-
-  if (!partID || !messageID || !sessionID) {
-    log(
-      `makePartVisible: missing IDs - partID=${partID}, messageID=${messageID}, sessionID=${sessionID}`
-    );
-    return;
-  }
-
-  log(
-    `makePartVisible: updating part ${partID} in DB to be visible with text: "${newText.substring(0, 50)}..."`
-  );
-
-  // Use the internal HTTP client from v1 SDK
-  try {
-    const httpClient = (client as any).client || (client as any)._client;
-    if (httpClient?.patch) {
-      await httpClient.patch({
-        url: `/session/${sessionID}/message/${messageID}/part/${partID}`,
-        body: {
-          id: partID,
-          messageID,
-          sessionID,
-          type: "text",
-          text: newText,
-          synthetic: false, // Make visible in TUI
-        },
-      });
-      log(`makePartVisible: success`);
-    } else {
-      log(
-        `makePartVisible: no internal HTTP client found - keys: ${Object.keys(client).join(", ")}`
-      );
-    }
-  } catch (e) {
-    log(`makePartVisible: failed - ${e}`);
-  }
-}
-
-/**
- * Helper to replace a message part with a prompt and clear any stale assistant response
- */
-function replaceAndClear(
-  part: any,
-  msg: any,
-  msgIndex: number,
-  newText: string,
-  outputMessages: any[]
-) {
-  log(
-    `Replacing generic message at index ${msgIndex} with prompt: "${newText.substring(0, 50)}..."`
-  );
-
-  // Modify in place - don't create new objects (references matter)
-  part.text = newText;
-  delete part.synthetic;
-
-  // Update the DB to make the message visible in TUI
-  makePartVisible(part, msg, newText).catch(e =>
-    log(`makePartVisible failed: ${e}`)
-  );
-
-  // CRITICAL: Remove any assistant response that came after the replaced message
-  // The LLM may have already responded to "Summarize..." before this hook fired
-  const nextIndex = msgIndex + 1;
-  if (nextIndex < outputMessages.length) {
-    const nextMsg = outputMessages[nextIndex];
-    const role = nextMsg.info?.role ?? nextMsg.role;
-    if (role === "assistant") {
-      outputMessages.splice(nextIndex, 1);
-      log(`Removed stale LLM response at index ${nextIndex}`);
-    }
-  }
-}
-
-/**
- * Hook: experimental.chat.messages.transform
- * Handles /subtask {...} inline subtasks and return prompt injection
- */
-export async function chatMessagesTransform(input: any, output: any) {
-  try {
-    // Log every invocation with message count and session info
-    const sessionIDs = [
-      ...new Set(
-        output.messages.map((m: any) => m.info?.sessionID).filter(Boolean)
-      ),
-    ];
-
-    // Log message summary for debugging
-    const msgSummary = output.messages
-      .map((m: any, i: number) => {
-        const role = m.info?.role || m.role || "?";
-        const parts = (m.parts || [])
-          .map((p: any) => {
-            if (p.type === "text") {
-              const text = p.text?.substring(0, 40) || "";
-              return `text:"${text}${p.text?.length > 40 ? "..." : ""}"${p.synthetic ? "(syn)" : ""}`;
-            }
-            if (p.type === "tool") return `tool:${p.state?.status || "?"}`;
-            if (p.type === "subtask")
-              return `subtask:"${p.prompt?.substring(0, 30) || ""}..."`;
-            return p.type;
-          })
-          .join(", ");
-        return `[${i}]${role}: ${parts}`;
-      })
-      .join(" | ");
-
-    log(
-      `message-hooks: ENTRY msgCount=${output.messages.length}, sessions=${sessionIDs.join(",") || "none"}`
-    );
-    log(`message-hooks: MESSAGES: ${msgSummary}`);
-
-    // Check for /subtask in user messages
-    // With the placeholder command file, OpenCode routes /subtask to command hook
-    // This is a fallback for when no placeholder exists
-    for (const msg of output.messages) {
-      if (msg.info?.role !== "user") continue;
-
-      // Track processed messages by ID to avoid infinite loop
-      const msgId = (msg.info as any)?.id;
-      if (msgId && hasProcessedS2Message(msgId)) continue;
-
-      for (const part of msg.parts) {
-        if (part.type !== "text") continue;
-        const text = part.text.trim();
-        const textLower = text.toLowerCase();
-
-        // Match /subtask (with space) - the recommended syntax
-        if (textLower.startsWith("/subtask ")) {
-          // If /subtask command exists, defer to command hook for instant execution
-          const configs = getConfigs();
-          if (configs["subtask"]) {
-            log(
-              `/subtask detected but deferring to command hook (subtask command exists)`
-            );
-            continue;
-          }
-
-          // Fallback: handle via message transform when no placeholder command exists
-          log(
-            `/subtask detected in message (no placeholder): "${text.substring(0, 60)}..."`
-          );
-
-          // Mark as processed BEFORE spawning
-          if (msgId) addProcessedS2Message(msgId);
-
-          // Parse: remove "/subtask " prefix
-          const toParse = text.substring("/subtask ".length);
-          // Wrap in {} if it doesn't start with { (plain prompt case)
-          const parseInput = toParse.startsWith("{")
-            ? toParse
-            : `{} ${toParse}`;
-          const parsed = parseInlineSubtask(parseInput);
-
-          if (parsed) {
-            log(
-              `/subtask inline subtask: prompt="${parsed.prompt.substring(
-                0,
-                50
-              )}...", overrides=${JSON.stringify(parsed.overrides)}`
-            );
-            // Replace with instruction to say minimal response
-            part.text = S2_INLINE_INSTRUCTION;
-            // Spawn the subtask - get sessionID from message info
-            const sessionID =
-              (msg.info as any)?.sessionID || (input as any).sessionID;
-            log(`/subtask sessionID: ${sessionID}`);
-            if (sessionID) {
-              executeInlineSubtask(parsed, sessionID).catch(console.error);
-            } else {
-              log(`/subtask ERROR: no sessionID found`);
-            }
-            return;
-          } else {
-            log(`/subtask parse failed for: "${parseInput}"`);
-          }
-        }
-      }
-    }
-
-    // Filter out summary leak messages from the assistant
-    // These are responses that summarize task tool output, which we don't want
-    const filteredMessages: any[] = [];
-
-    const isSummaryLeak = (msg: any): boolean => {
-      const parts = msg.parts || [];
-      return parts.some((p: any) => {
-        if (p.type !== "step-start") return false;
-        if (typeof p.text !== "string") return false;
-        return (
-          p.text.includes("Summary of Task Tool Output") ||
-          p.text.includes("My task was to summarize") ||
-          p.text.includes("The tool executed your command") ||
-          p.text.includes("The command was executed successfully") ||
-          p.text.includes("To complete your request")
-        );
-      });
-    };
-
-    for (let i = 0; i < output.messages.length; i++) {
-      const msg = output.messages[i];
-      const role = msg.info?.role ?? msg.role;
-
-      if (role === "assistant" && isSummaryLeak(msg)) {
-        log(`message-hooks: removing summary leak at index ${i}`);
-        continue;
-      }
-
-      filteredMessages.push(msg);
-    }
-
-    if (filteredMessages.length !== output.messages.length) {
-      log(
-        `message-hooks: filtered ${output.messages.length - filteredMessages.length} summary leak messages`
-      );
-      // CRITICAL: Must mutate the array IN PLACE, not reassign!
-      // Reassigning output.messages doesn't affect the original sessionMessages in OpenCode
-      output.messages.length = 0;
-      output.messages.push(...filteredMessages);
-    }
-
-    log(`message-hooks: post-filter, msgCount=${output.messages.length}`);
-
-    // NOTE: We don't try to delete messages from DB - client.session.removeMessage doesn't exist
-    // The filtering above is sufficient - it affects the current LLM call
-
-    log(`message-hooks: about to search for OPENCODE_GENERIC`);
-
-    // Find the LAST message with OPENCODE_GENERIC
-    let lastGenericPart: any = null;
-    let lastGenericMsg: any = null;
-    let lastGenericMsgIndex: number = -1;
-
-    log(
-      `message-hooks: searching ${output.messages.length} messages for OPENCODE_GENERIC`
-    );
-
-    for (let i = 0; i < output.messages.length; i++) {
-      const msg = output.messages[i];
-      for (const part of msg.parts || []) {
-        if (part.type === "text") {
-          const isGeneric = part.text === OPENCODE_GENERIC;
-          if (part.synthetic) {
-            log(
-              `message-hooks: found synthetic text at [${i}]: "${part.text?.substring(0, 50)}..." matches=${isGeneric}`
-            );
-          }
-          if (isGeneric) {
-            lastGenericPart = part;
-            lastGenericMsg = msg;
-            lastGenericMsgIndex = i;
-          }
-        }
-      }
-    }
-
-    log(
-      `message-hooks: generic search complete, found=${!!lastGenericPart}, index=${lastGenericMsgIndex}`
-    );
-
-    // If no generic part found, check for pending prompt returns from inline subtasks
-    // Inline subtasks don't get OPENCODE_GENERIC, so we inject the pending prompt directly
-    if (!lastGenericPart) {
-      log(
-        `message-hooks: no generic part found, checking for pending prompt return`
-      );
-
-      for (const sid of sessionIDs) {
-        const pendingPrompt = consumePendingPromptReturn(sid as string);
-        if (pendingPrompt) {
-          log(`message-hooks: found pendingPromptReturn for ${sid}, injecting`);
-
-          // Clone structure from an existing user message to ensure correct format
-          const existingUserMsg = output.messages.find(
-            (m: any) => (m.info?.role ?? m.role) === "user"
-          );
-
-          if (existingUserMsg) {
-            const newMsg = JSON.parse(JSON.stringify(existingUserMsg));
-            newMsg.parts = [{ type: "text", text: pendingPrompt }];
-            // Remove IDs so OpenCode doesn't get confused
-            if (newMsg.id) delete newMsg.id;
-            if (newMsg.info?.id) delete newMsg.info.id;
-            if (newMsg.info?.createdAt) delete newMsg.info.createdAt;
-            output.messages.push(newMsg);
-            log(
-              `message-hooks: injected pending prompt: "${pendingPrompt.substring(0, 40)}..."`
-            );
-          } else {
-            // Fallback: create minimal message
-            output.messages.push({
-              info: { role: "user" },
-              parts: [{ type: "text", text: pendingPrompt }],
-            });
-            log(
-              `message-hooks: injected pending prompt (fallback): "${pendingPrompt.substring(0, 40)}..."`
-            );
-          }
-          return; // Let the LLM respond to our injected message
-        }
-
-        if (hasReturnStack(sid as string)) {
-          log(
-            `message-hooks: session ${sid} HAS returnStack but no generic part!`
-          );
-        }
-      }
-      return; // Nothing to do without generic part
-    }
-
-    // The generic message is in a subtask session, but returns are stored under the PARENT session
-    // We need to resolve the parent session to look up returns
-    const subtaskSessionID = lastGenericMsg?.info?.sessionID;
-    const resolvedParent = subtaskSessionID
-      ? getSubtaskParentSession(subtaskSessionID)
-      : undefined;
-    const parentSessionID = resolvedParent || subtaskSessionID;
-
-    log(
-      `message-hooks: subtaskSession=${subtaskSessionID}, resolvedParent=${resolvedParent || "NOT FOUND"}, using=${parentSessionID}`
-    );
-
-    // Check for deferred prompt return from session.idle
-    // This handles prompt returns that were deferred to be substituted here
-    log(
-      `message-hooks: checking pendingPromptReturn for session ${parentSessionID}`
-    );
-    if (parentSessionID) {
-      const pendingPrompt = consumePendingPromptReturn(parentSessionID);
-      if (pendingPrompt) {
-        replaceAndClear(
-          lastGenericPart,
-          lastGenericMsg,
-          lastGenericMsgIndex,
-          pendingPrompt,
-          output.messages
-        );
-
-        // Mark that we're waiting for LLM response to this prompt
-        setPendingStackedPromptResponse(parentSessionID);
-
-        return;
-      }
-    }
-
-    // Check for pending loop evaluation first (orchestrator-decides pattern)
-    log(`message-hooks: checking loop evaluations`);
-    for (const [sessionID, retryState] of getAllPendingEvaluations()) {
-      // Check if this is an unconditional loop (no until condition)
-      if (!retryState.config.until) {
-        // Unconditional loop: inject yield prompt, no evaluation needed
-        const yieldPrompt = createYieldPrompt(
-          retryState.iteration,
-          retryState.config.max
-        );
-        // Unconditional loop: modify in place
-        lastGenericPart.text = yieldPrompt;
-        log(
-          `loop: injected yield prompt (unconditional loop ${retryState.iteration}/${retryState.config.max})`
-        );
-      } else {
-        // Conditional loop: inject evaluation prompt
-        const evalPrompt = createEvaluationPrompt(
-          retryState.config.until,
-          retryState.iteration,
-          retryState.config.max
-        );
-        // Modify in place
-        lastGenericPart.text = evalPrompt;
-        log(
-          `loop: injected evaluation prompt for "${retryState.config.until}"`
-        );
-      }
-      // Don't delete yet - we need it when parsing the response
-      return;
-    }
-
-    // Check for pending return
-    const pendingReturnsEntries = [...getAllPendingReturns()];
-    log(
-      `message-hooks: checking pendingReturns, count=${pendingReturnsEntries.length}`
-    );
-    for (const [sessionID, returnPrompt] of pendingReturnsEntries) {
-      deletePendingReturn(sessionID);
-      setHasActiveSubtask(false);
-
-      // Also shift from returnState to keep them in sync
-      // This prevents session.idle from trying to re-execute the same return
-      const returnStateData = getReturnState(sessionID);
-      if (returnStateData?.length && returnStateData[0] === returnPrompt) {
-        returnStateData.shift();
-      }
-
-      if (returnPrompt.startsWith("/")) {
-        // Command return: remove the summarize message entirely from history
-        // and fire the command immediately at this moment
-        if (lastGenericMsgIndex >= 0) {
-          output.messages.splice(lastGenericMsgIndex, 1);
-
-          // Also delete from DB to prevent OpenCode from re-fetching and responding to it
-          const msgId = lastGenericMsg?.id;
-          const partId = lastGenericPart?.id;
-          if (msgId && partId) {
-            const client = getClient();
-            // Delete the part (this effectively removes the message content)
-            client.session
-              .deletePart({ path: { id: partId } })
-              .catch((e: Error) => {
-                log(`Failed to delete summarize part from DB: ${e.message}`);
-              });
-          }
-
-          log(
-            `Removed summarize message at index ${lastGenericMsgIndex} for command return`
-          );
-        }
-        executeReturn(returnPrompt, sessionID).catch(console.error);
-      } else {
-        // Plain prompt return: MODIFY the existing part directly (for LLM)
-        replaceAndClear(
-          lastGenericPart,
-          lastGenericMsg,
-          lastGenericMsgIndex,
-          returnPrompt,
-          output.messages
-        );
-      }
-      return;
-    }
-
-    // No pending return found, check returnState for remaining returns
-    // This handles returns after the first one (which used pendingReturns)
-    // Use parentSessionID (resolved earlier) since returns are stored under parent session
-    if (parentSessionID) {
-      const remaining = getReturnState(parentSessionID);
-      if (remaining?.length) {
-        const nextReturn = remaining[0];
-        setHasActiveSubtask(false);
-
-        if (nextReturn.startsWith("/")) {
-          // Command return: remove the summarize message, let session.idle execute
-          if (lastGenericMsgIndex >= 0) {
-            output.messages.splice(lastGenericMsgIndex, 1);
-            log(
-              `Removed summarize message - returnState command will be processed by session.idle`
-            );
-          }
-        } else {
-          // Prompt return: REPLACE the message so LLM responds to it
-          // NOTE: We must replace the entire message, not just modify the part text,
-          // because OpenCode may deep-copy messages before sending to LLM
-          // First, check if there are stacked returns from inline subtasks
-          // and transfer them to returnState so the chain continues
-          const stackedReturns = getCurrentReturnChain(parentSessionID);
-          if (stackedReturns?.length) {
-            remaining.push(...stackedReturns);
-            popCurrentReturnChain(parentSessionID);
-            log(
-              `Transferred ${stackedReturns.length} stacked returns to returnState for continued processing`
-            );
-          }
-
-          // Shift from returnState since we're handling it here
-          remaining.shift();
-
-          replaceAndClear(
-            lastGenericPart,
-            lastGenericMsg,
-            lastGenericMsgIndex,
-            nextReturn,
-            output.messages
-          );
-
-          // Mark that we're waiting for LLM response to this prompt
-          // This prevents session.idle from processing the next return too early
-          setPendingStackedPromptResponse(parentSessionID);
-        }
-        return;
-      }
-    }
-
-    // No returns found, use generic replacement if configured
-    // BUT skip if there are stacked returns (from inline subtasks with returns)
-    // Those will be processed by session.idle hook
-    const pluginConfig = getPluginConfig();
-
-    // Check if the parent session has stacked returns
-    // Use parentSessionID (resolved earlier) since stacked returns are stored under parent session
-    let stackedSessionID: string | null = null;
-    if (parentSessionID && hasReturnStack(parentSessionID)) {
-      stackedSessionID = parentSessionID;
-    }
-
-    if (stackedSessionID) {
-      // Peek at the first stacked return to decide how to handle
-      const nextReturnArray = peekReturnStack(stackedSessionID);
-      const nextReturn = nextReturnArray?.[0];
-
-      if (nextReturn?.startsWith("/")) {
-        // Command return: remove the summarize message, let session.idle execute
-        if (lastGenericMsgIndex >= 0) {
-          output.messages.splice(lastGenericMsgIndex, 1);
-          log(
-            `Removed summarize message - stacked command return will be processed by session.idle`
-          );
-        }
-      } else if (nextReturn) {
-        // Prompt return: REPLACE the message so LLM responds to it
-        const returnPrompt = shiftReturnStack(stackedSessionID);
-        if (returnPrompt) {
-          replaceAndClear(
-            lastGenericPart,
-            lastGenericMsg,
-            lastGenericMsgIndex,
-            returnPrompt,
-            output.messages
-          );
-
-          // Mark that we're waiting for LLM response to this prompt
-          // This prevents session.idle from processing returnState too early
-          setPendingStackedPromptResponse(stackedSessionID);
-        }
-      }
-      setHasActiveSubtask(false);
-      return;
-    }
-
-    if (getHasActiveSubtask() && pluginConfig.replace_generic) {
-      const text = pluginConfig.generic_return ?? DEFAULT_PROMPT;
-      replaceAndClear(
-        lastGenericPart,
-        lastGenericMsg,
-        lastGenericMsgIndex,
-        text,
-        output.messages
-      );
-      setHasActiveSubtask(false);
-      return;
-    }
-  } catch (err) {
-    log(`message-hooks: EXCEPTION: ${err}`);
-  }
-}
+import {
+  getClient,
+  getReturnState,
+  deleteReturnState,
+  setReturnState,
+  getPluginConfig,
+  getAllPendingReturns,
+  deletePendingReturn,
+  hasReturnStack,
+  peekReturnStack,
+  shiftReturnStack,
+  getCurrentReturnChain,
+  popCurrentReturnChain,
+  setHasActiveSubtask,
+  getHasActiveSubtask,
+  OPENCODE_GENERIC,
+  getSubtaskParentSession,
+  consumePendingPromptReturn,
+  setPendingStackedPromptResponse,
+  hasProcessedS2Message,
+  addProcessedS2Message,
+  getConfigs,
+} from "../core/state";
+import { log } from "../utils/logger";
+import { DEFAULT_PROMPT } from "../utils/config";
+import { S2_INLINE_INSTRUCTION } from "../utils/prompts";
+import { parseInlineSubtask } from "../parsing";
+import { executeInlineSubtask } from "../features/inline-subtasks";
+import { executeReturn } from "../features/returns";
+import {
+  getAllPendingEvaluations,
+  createEvaluationPrompt,
+  createYieldPrompt,
+} from "../loop";
+
+/**
+ * Update the synthetic part in the database to make it visible in TUI.
+ *
+ * The v1 SDK from OpenCode exposes an internal hey-api HTTP client at client.client
+ * which we use to PATCH the part. This is the only reliable way to update parts
+ * from a plugin since client.part.update() doesn't exist in v1.
+ *
+ * Note: OpenCode's plugin system passes a v1 client. When/if they upgrade to v2,
+ * we can use client.part.update() directly.
+ */
+async function makePartVisible(
+  part: any,
+  msg: any,
+  newText: string
+): Promise<void> {
+  const client = getClient();
+  if (!client) {
+    log(`makePartVisible: no client available`);
+    return;
+  }
+
+  // Extract IDs from the part and message
+  const partID = part.id;
+  const messageID = part.messageID || msg?.info?.id;
+  const sessionID = part.sessionID || msg?.info?.sessionID;
+
+  if (!partID || !messageID || !sessionID) {
+    log(
+      `makePartVisible: missing IDs - partID=${partID}, messageID=${messageID}, sessionID=${sessionID}`
+    );
+    return;
+  }
+
+  log(
+    `makePartVisible: updating part ${partID} in DB to be visible with text: "${newText.substring(0, 50)}..."`
+  );
+
+  // Use the internal HTTP client from v1 SDK
+  try {
+    const httpClient = (client as any).client || (client as any)._client;
+    if (httpClient?.patch) {
+      await httpClient.patch({
+        url: `/session/${sessionID}/message/${messageID}/part/${partID}`,
+        body: {
+          id: partID,
+          messageID,
+          sessionID,
+          type: "text",
+          text: newText,
+          synthetic: false, // Make visible in TUI
+        },
+      });
+      log(`makePartVisible: success`);
+    } else {
+      log(
+        `makePartVisible: no internal HTTP client found - keys: ${Object.keys(client).join(", ")}`
+      );
+    }
+  } catch (e) {
+    log(`makePartVisible: failed - ${e}`);
+  }
+}
+
+/**
+ * Helper to replace a message part with a prompt and clear any stale assistant response
+ */
+function replaceAndClear(
+  part: any,
+  msg: any,
+  msgIndex: number,
+  newText: string,
+  outputMessages: any[]
+) {
+  log(
+    `Replacing generic message at index ${msgIndex} with prompt: "${newText.substring(0, 50)}..."`
+  );
+
+  // Modify in place - don't create new objects (references matter)
+  part.text = newText;
+  delete part.synthetic;
+
+  // Update the DB to make the message visible in TUI
+  makePartVisible(part, msg, newText).catch(e =>
+    log(`makePartVisible failed: ${e}`)
+  );
+
+  // CRITICAL: Remove any assistant response that came after the replaced message
+  // The LLM may have already responded to "Summarize..." before this hook fired
+  const nextIndex = msgIndex + 1;
+  if (nextIndex < outputMessages.length) {
+    const nextMsg = outputMessages[nextIndex];
+    const role = nextMsg.info?.role ?? nextMsg.role;
+    if (role === "assistant") {
+      outputMessages.splice(nextIndex, 1);
+      log(`Removed stale LLM response at index ${nextIndex}`);
+    }
+  }
+}
+
+/**
+ * Hook: experimental.chat.messages.transform
+ * Handles /subtask {...} inline subtasks and return prompt injection
+ */
+export async function chatMessagesTransform(input: any, output: any) {
+  try {
+    // Log every invocation with message count and session info
+    const sessionIDs = [
+      ...new Set(
+        output.messages.map((m: any) => m.info?.sessionID).filter(Boolean)
+      ),
+    ];
+
+    // Log message summary for debugging
+    const msgSummary = output.messages
+      .map((m: any, i: number) => {
+        const role = m.info?.role || m.role || "?";
+        const parts = (m.parts || [])
+          .map((p: any) => {
+            if (p.type === "text") {
+              const text = p.text?.substring(0, 40) || "";
+              return `text:"${text}${p.text?.length > 40 ? "..." : ""}"${p.synthetic ? "(syn)" : ""}`;
+            }
+            if (p.type === "tool") return `tool:${p.state?.status || "?"}`;
+            if (p.type === "subtask")
+              return `subtask:"${p.prompt?.substring(0, 30) || ""}..."`;
+            return p.type;
+          })
+          .join(", ");
+        return `[${i}]${role}: ${parts}`;
+      })
+      .join(" | ");
+
+    log(
+      `message-hooks: ENTRY msgCount=${output.messages.length}, sessions=${sessionIDs.join(",") || "none"}`
+    );
+    log(`message-hooks: MESSAGES: ${msgSummary}`);
+
+    // Check for /subtask in user messages
+    // With the placeholder command file, OpenCode routes /subtask to command hook
+    // This is a fallback for when no placeholder exists
+    for (const msg of output.messages) {
+      if (msg.info?.role !== "user") continue;
+
+      // Track processed messages by ID to avoid infinite loop
+      const msgId = (msg.info as any)?.id;
+      if (msgId && hasProcessedS2Message(msgId)) continue;
+
+      for (const part of msg.parts) {
+        if (part.type !== "text") continue;
+        const text = part.text.trim();
+        const textLower = text.toLowerCase();
+
+        // Match /subtask (with space) - the recommended syntax
+        if (textLower.startsWith("/subtask ")) {
+          // If /subtask command exists, defer to command hook for instant execution
+          const configs = getConfigs();
+          if (configs["subtask"]) {
+            log(
+              `/subtask detected but deferring to command hook (subtask command exists)`
+            );
+            continue;
+          }
+
+          // Fallback: handle via message transform when no placeholder command exists
+          log(
+            `/subtask detected in message (no placeholder): "${text.substring(0, 60)}..."`
+          );
+
+          // Mark as processed BEFORE spawning
+          if (msgId) addProcessedS2Message(msgId);
+
+          // Parse: remove "/subtask " prefix
+          const toParse = text.substring("/subtask ".length);
+          // Wrap in {} if it doesn't start with { (plain prompt case)
+          const parseInput = toParse.startsWith("{")
+            ? toParse
+            : `{} ${toParse}`;
+          const parsed = parseInlineSubtask(parseInput);
+
+          if (parsed) {
+            log(
+              `/subtask inline subtask: prompt="${parsed.prompt.substring(
+                0,
+                50
+              )}...", overrides=${JSON.stringify(parsed.overrides)}`
+            );
+            // Replace with instruction to say minimal response
+            part.text = S2_INLINE_INSTRUCTION;
+            // Spawn the subtask - get sessionID from message info
+            const sessionID =
+              (msg.info as any)?.sessionID || (input as any).sessionID;
+            log(`/subtask sessionID: ${sessionID}`);
+            if (sessionID) {
+              executeInlineSubtask(parsed, sessionID).catch(console.error);
+            } else {
+              log(`/subtask ERROR: no sessionID found`);
+            }
+            return;
+          } else {
+            log(`/subtask parse failed for: "${parseInput}"`);
+          }
+        }
+      }
+    }
+
+    // Filter out summary leak messages from the assistant
+    // These are responses that summarize task tool output, which we don't want
+    const filteredMessages: any[] = [];
+
+    const isSummaryLeak = (msg: any): boolean => {
+      const parts = msg.parts || [];
+      return parts.some((p: any) => {
+        if (p.type !== "step-start") return false;
+        if (typeof p.text !== "string") return false;
+        return (
+          p.text.includes("Summary of Task Tool Output") ||
+          p.text.includes("My task was to summarize") ||
+          p.text.includes("The tool executed your command") ||
+          p.text.includes("The command was executed successfully") ||
+          p.text.includes("To complete your request")
+        );
+      });
+    };
+
+    for (let i = 0; i < output.messages.length; i++) {
+      const msg = output.messages[i];
+      const role = msg.info?.role ?? msg.role;
+
+      if (role === "assistant" && isSummaryLeak(msg)) {
+        log(`message-hooks: removing summary leak at index ${i}`);
+        continue;
+      }
+
+      filteredMessages.push(msg);
+    }
+
+    if (filteredMessages.length !== output.messages.length) {
+      log(
+        `message-hooks: filtered ${output.messages.length - filteredMessages.length} summary leak messages`
+      );
+      // CRITICAL: Must mutate the array IN PLACE, not reassign!
+      // Reassigning output.messages doesn't affect the original sessionMessages in OpenCode
+      output.messages.length = 0;
+      output.messages.push(...filteredMessages);
+    }
+
+    log(`message-hooks: post-filter, msgCount=${output.messages.length}`);
+
+    // NOTE: We don't try to delete messages from DB - client.session.removeMessage doesn't exist
+    // The filtering above is sufficient - it affects the current LLM call
+
+    log(`message-hooks: about to search for OPENCODE_GENERIC`);
+
+    // Find the LAST message with OPENCODE_GENERIC
+    let lastGenericPart: any = null;
+    let lastGenericMsg: any = null;
+    let lastGenericMsgIndex: number = -1;
+
+    log(
+      `message-hooks: searching ${output.messages.length} messages for OPENCODE_GENERIC`
+    );
+
+    for (let i = 0; i < output.messages.length; i++) {
+      const msg = output.messages[i];
+      for (const part of msg.parts || []) {
+        if (part.type === "text") {
+          const isGeneric = part.text === OPENCODE_GENERIC;
+          if (part.synthetic) {
+            log(
+              `message-hooks: found synthetic text at [${i}]: "${part.text?.substring(0, 50)}..." matches=${isGeneric}`
+            );
+          }
+          if (isGeneric) {
+            lastGenericPart = part;
+            lastGenericMsg = msg;
+            lastGenericMsgIndex = i;
+          }
+        }
+      }
+    }
+
+    log(
+      `message-hooks: generic search complete, found=${!!lastGenericPart}, index=${lastGenericMsgIndex}`
+    );
+
+    // If no generic part found, check for pending prompt returns from inline subtasks
+    // Inline subtasks don't get OPENCODE_GENERIC, so we inject the pending prompt directly
+    if (!lastGenericPart) {
+      log(
+        `message-hooks: no generic part found, checking for pending prompt return`
+      );
+
+      for (const sid of sessionIDs) {
+        const pendingPrompt = consumePendingPromptReturn(sid as string);
+        if (pendingPrompt) {
+          log(`message-hooks: found pendingPromptReturn for ${sid}, injecting`);
+
+          // Clone structure from an existing user message to ensure correct format
+          const existingUserMsg = output.messages.find(
+            (m: any) => (m.info?.role ?? m.role) === "user"
+          );
+
+          if (existingUserMsg) {
+            const newMsg = JSON.parse(JSON.stringify(existingUserMsg));
+            newMsg.parts = [{ type: "text", text: pendingPrompt }];
+            // Remove IDs so OpenCode doesn't get confused
+            if (newMsg.id) delete newMsg.id;
+            if (newMsg.info?.id) delete newMsg.info.id;
+            if (newMsg.info?.createdAt) delete newMsg.info.createdAt;
+            output.messages.push(newMsg);
+            log(
+              `message-hooks: injected pending prompt: "${pendingPrompt.substring(0, 40)}..."`
+            );
+          } else {
+            // Fallback: create minimal message
+            output.messages.push({
+              info: { role: "user" },
+              parts: [{ type: "text", text: pendingPrompt }],
+            });
+            log(
+              `message-hooks: injected pending prompt (fallback): "${pendingPrompt.substring(0, 40)}..."`
+            );
+          }
+          return; // Let the LLM respond to our injected message
+        }
+
+        if (hasReturnStack(sid as string)) {
+          log(
+            `message-hooks: session ${sid} HAS returnStack but no generic part!`
+          );
+        }
+      }
+      return; // Nothing to do without generic part
+    }
+
+    // The generic message is in a subtask session, but returns are stored under the PARENT session
+    // We need to resolve the parent session to look up returns
+    const subtaskSessionID = lastGenericMsg?.info?.sessionID;
+    const resolvedParent = subtaskSessionID
+      ? getSubtaskParentSession(subtaskSessionID)
+      : undefined;
+    const parentSessionID = resolvedParent || subtaskSessionID;
+
+    log(
+      `message-hooks: subtaskSession=${subtaskSessionID}, resolvedParent=${resolvedParent || "NOT FOUND"}, using=${parentSessionID}`
+    );
+
+    // Check for deferred prompt return from session.idle
+    // This handles prompt returns that were deferred to be substituted here
+    log(
+      `message-hooks: checking pendingPromptReturn for session ${parentSessionID}`
+    );
+    if (parentSessionID) {
+      const pendingPrompt = consumePendingPromptReturn(parentSessionID);
+      if (pendingPrompt) {
+        replaceAndClear(
+          lastGenericPart,
+          lastGenericMsg,
+          lastGenericMsgIndex,
+          pendingPrompt,
+          output.messages
+        );
+
+        // Mark that we're waiting for LLM response to this prompt
+        setPendingStackedPromptResponse(parentSessionID);
+
+        return;
+      }
+    }
+
+    // Check for pending loop evaluation first (orchestrator-decides pattern)
+    log(`message-hooks: checking loop evaluations`);
+    for (const [sessionID, retryState] of getAllPendingEvaluations()) {
+      // Check if this is an unconditional loop (no until condition)
+      if (!retryState.config.until) {
+        // Unconditional loop: inject yield prompt, no evaluation needed
+        const yieldPrompt = createYieldPrompt(
+          retryState.iteration,
+          retryState.config.max
+        );
+        // Unconditional loop: modify in place
+        lastGenericPart.text = yieldPrompt;
+        log(
+          `loop: injected yield prompt (unconditional loop ${retryState.iteration}/${retryState.config.max})`
+        );
+      } else {
+        // Conditional loop: inject evaluation prompt
+        const evalPrompt = createEvaluationPrompt(
+          retryState.config.until,
+          retryState.iteration,
+          retryState.config.max
+        );
+        // Modify in place
+        lastGenericPart.text = evalPrompt;
+        log(
+          `loop: injected evaluation prompt for "${retryState.config.until}"`
+        );
+      }
+      // Don't delete yet - we need it when parsing the response
+      return;
+    }
+
+    // Check for pending return
+    const pendingReturnsEntries = [...getAllPendingReturns()];
+    log(
+      `message-hooks: checking pendingReturns, count=${pendingReturnsEntries.length}`
+    );
+    for (const [sessionID, returnPrompt] of pendingReturnsEntries) {
+      deletePendingReturn(sessionID);
+      setHasActiveSubtask(false);
+
+      // Also shift from returnState to keep them in sync
+      // This prevents session.idle from trying to re-execute the same return
+      const returnStateData = getReturnState(sessionID);
+      if (returnStateData?.length && returnStateData[0] === returnPrompt) {
+        returnStateData.shift();
+      }
+
+      if (returnPrompt.startsWith("/")) {
+        // Command return: remove the summarize message entirely from history
+        // and fire the command immediately at this moment
+        if (lastGenericMsgIndex >= 0) {
+          output.messages.splice(lastGenericMsgIndex, 1);
+
+          // Also delete from DB to prevent OpenCode from re-fetching and responding to it
+          const msgId = lastGenericMsg?.id;
+          const partId = lastGenericPart?.id;
+          if (msgId && partId) {
+            const client = getClient();
+            // Delete the part (this effectively removes the message content)
+            client.session
+              .deletePart({ path: { id: partId } })
+              .catch((e: Error) => {
+                log(`Failed to delete summarize part from DB: ${e.message}`);
+              });
+          }
+
+          log(
+            `Removed summarize message at index ${lastGenericMsgIndex} for command return`
+          );
+        }
+        executeReturn(returnPrompt, sessionID).catch(console.error);
+      } else {
+        // Plain prompt return: MODIFY the existing part directly (for LLM)
+        replaceAndClear(
+          lastGenericPart,
+          lastGenericMsg,
+          lastGenericMsgIndex,
+          returnPrompt,
+          output.messages
+        );
+      }
+      return;
+    }
+
+    // No pending return found, check returnState for remaining returns
+    // This handles returns after the first one (which used pendingReturns)
+    // Use parentSessionID (resolved earlier) since returns are stored under parent session
+    if (parentSessionID) {
+      const remaining = getReturnState(parentSessionID);
+      if (remaining?.length) {
+        const nextReturn = remaining[0];
+        setHasActiveSubtask(false);
+
+        if (nextReturn.startsWith("/")) {
+          // Command return: remove the summarize message, let session.idle execute
+          if (lastGenericMsgIndex >= 0) {
+            output.messages.splice(lastGenericMsgIndex, 1);
+            log(
+              `Removed summarize message - returnState command will be processed by session.idle`
+            );
+          }
+        } else {
+          // Prompt return: REPLACE the message so LLM responds to it
+          // NOTE: We must replace the entire message, not just modify the part text,
+          // because OpenCode may deep-copy messages before sending to LLM
+          // First, check if there are stacked returns from inline subtasks
+          // and transfer them to returnState so the chain continues
+          const stackedReturns = getCurrentReturnChain(parentSessionID);
+          if (stackedReturns?.length) {
+            remaining.push(...stackedReturns);
+            popCurrentReturnChain(parentSessionID);
+            log(
+              `Transferred ${stackedReturns.length} stacked returns to returnState for continued processing`
+            );
+          }
+
+          // Shift from returnState since we're handling it here
+          remaining.shift();
+
+          replaceAndClear(
+            lastGenericPart,
+            lastGenericMsg,
+            lastGenericMsgIndex,
+            nextReturn,
+            output.messages
+          );
+
+          // Mark that we're waiting for LLM response to this prompt
+          // This prevents session.idle from processing the next return too early
+          setPendingStackedPromptResponse(parentSessionID);
+        }
+        return;
+      }
+    }
+
+    // No returns found, use generic replacement if configured
+    // BUT skip if there are stacked returns (from inline subtasks with returns)
+    // Those will be processed by session.idle hook
+    const pluginConfig = getPluginConfig();
+
+    // Check if the parent session has stacked returns
+    // Use parentSessionID (resolved earlier) since stacked returns are stored under parent session
+    let stackedSessionID: string | null = null;
+    if (parentSessionID && hasReturnStack(parentSessionID)) {
+      stackedSessionID = parentSessionID;
+    }
+
+    if (stackedSessionID) {
+      // Peek at the first stacked return to decide how to handle
+      const nextReturnArray = peekReturnStack(stackedSessionID);
+      const nextReturn = nextReturnArray?.[0];
+
+      if (nextReturn?.startsWith("/")) {
+        // Command return: remove the summarize message, let session.idle execute
+        if (lastGenericMsgIndex >= 0) {
+          output.messages.splice(lastGenericMsgIndex, 1);
+          log(
+            `Removed summarize message - stacked command return will be processed by session.idle`
+          );
+        }
+      } else if (nextReturn) {
+        // Prompt return: REPLACE the message so LLM responds to it
+        const returnPrompt = shiftReturnStack(stackedSessionID);
+        if (returnPrompt) {
+          replaceAndClear(
+            lastGenericPart,
+            lastGenericMsg,
+            lastGenericMsgIndex,
+            returnPrompt,
+            output.messages
+          );
+
+          // Mark that we're waiting for LLM response to this prompt
+          // This prevents session.idle from processing returnState too early
+          setPendingStackedPromptResponse(stackedSessionID);
+        }
+      }
+      setHasActiveSubtask(false);
+      return;
+    }
+
+    if (getHasActiveSubtask() && pluginConfig.replace_generic) {
+      const text = pluginConfig.generic_return ?? DEFAULT_PROMPT;
+      replaceAndClear(
+        lastGenericPart,
+        lastGenericMsg,
+        lastGenericMsgIndex,
+        text,
+        output.messages
+      );
+      setHasActiveSubtask(false);
+      return;
+    }
+  } catch (err) {
+    log(`message-hooks: EXCEPTION: ${err}`);
+  }
+}