@cjhyy/code-shell-core 0.7.0-beta.1 → 0.7.1

package/dist/engine/turn-loop.js

@@ -6,6 +6,7 @@
  */
 import { initialTurnState, newTurnId } from "./turn-state.js";
 import { formatFriendlyError } from "./friendly-error.js";
+import { projectGoalJudgeToolResult, } from "../hooks/goal-stop-hook.js";
 import { wrapHookMessages } from "../hooks/inject.js";
 import { ContextLimitError } from "../exceptions.js";
 import { logger } from "../logging/logger.js";
@@ -19,7 +20,7 @@ import { COMPLETE_GOAL_TOOL_NAME } from "../tool-system/builtin/complete-goal.js
 import { CANCEL_GOAL_TOOL_NAME } from "../tool-system/builtin/cancel-goal.js";
 import { redactSensitiveToolResultsInMessages, toolResultForDisplay, toolResultTranscriptText, toolResultsForDisplay, } from "../tool-system/tool-result-redaction.js";
 import { addTokenUsage, cacheHitRateFromUsage, cumulativeCacheHitRate, } from "./session-usage.js";
-import { createGoalBudgetTracker, recordGoalUsage, goalBudgetExceeded, applyGoalExtension, limitProximity, GOAL_DEFAULT_MAX_STOP_BLOCKS, } from "./goal.js";
+import { createGoalBudgetTracker, recordGoalUsage, goalBudgetTerminationReason, applyGoalExtension, limitProximity, GOAL_DEFAULT_MAX_STOP_BLOCKS, } from "./goal.js";
 /**
  * 把一个 ToolResult 映射成发给 LLM 的 tool_result ContentBlock。
  * 有 contentBlocks(view_image 的图片块)就原样用作 content;否则
@@ -62,6 +63,11 @@ export class TurnLoop {
     currentCumulativeUsage;
     sensitiveToolResultRedactions = new Map();
     pendingImageMessages = new Set();
+    volatileContextMessages = new Set();
+    /** Last outer-loop turn for which the model returned a complete response. */
+    lastCompletedModelTurn = 0;
+    /** Last model turn whose transcript/persistence boundary was finalized. */
+    lastFinalizedTurn = 0;
     /**
      * Consecutive on_stop blocks (Goal mode kept the agent going). Reset to 0
      * on any unblocked completion. When it reaches config.maxStopBlocks the
@@ -126,6 +132,20 @@ export class TurnLoop {
         }
         return { ...next, maxStopBlocks: nextCap };
     }
+    /**
+     * Private Goal-judge accounting seam. Judge calls bypass ModelFacade, so the
+     * Engine forwards their provider usage here while the run-scoped tracker is
+     * live. The returned reason lets the hook suppress verdict side effects when
+     * this very request crossed a hard token/time budget.
+     */
+    recordGoalJudgeUsage(usage) {
+        if (!this.goalTracker)
+            return undefined;
+        if (usage) {
+            recordGoalUsage(this.goalTracker, (usage.promptTokens ?? 0) + (usage.completionTokens ?? 0));
+        }
+        return goalBudgetTerminationReason(this.goalTracker, Date.now());
+    }
     /**
      * Goal mode only: if the run is nearing EITHER stop ceiling (maxTurns or
      * maxStopBlocks) and we haven't announced it yet, emit one approaching_limit
@@ -160,6 +180,9 @@ export class TurnLoop {
         for (const msg of this.config.freshImageMessages ?? []) {
             this.pendingImageMessages.add(msg);
         }
+        for (const msg of this.config.volatileContextMessages ?? []) {
+            this.volatileContextMessages.add(msg);
+        }
         // Wrap onStream so a single throwing handler can't silently break
         // the channel for the rest of the run. A 2026-05-25 incident saw a
         // sub-agent's events stop reaching the renderer ~23s into its run —
@@ -212,6 +235,21 @@ export class TurnLoop {
             return;
         this.deps.transcript.appendTurnStopped();
     }
+    /**
+     * Close one completed model turn exactly once. The transcript boundary must
+     * be durable before the Engine flushes usage, heartbeat, and turnCount via
+     * onTurnBoundary, so crash recovery never folds two assistant responses into
+     * the same transcript turn or observes state ahead of the transcript.
+     */
+    finalizeModelTurn() {
+        if (this.lastCompletedModelTurn !== this.turnCount ||
+            this.lastFinalizedTurn === this.turnCount) {
+            return;
+        }
+        this.lastFinalizedTurn = this.turnCount;
+        this.deps.transcript.appendTurnBoundary();
+        this.config.onTurnBoundary?.(this.turnCount);
+    }
     prepareMessagesForModel(messages) {
         const preserveMessages = this.pendingImageMessages.size > 0 ? this.pendingImageMessages : undefined;
         const result = downgradeImagePayloadsInHistory(messages, { preserveMessages });
@@ -224,6 +262,23 @@ export class TurnLoop {
         }
         return result.messages;
     }
+    stripVolatileContextMessages(messages) {
+        if (this.volatileContextMessages.size === 0)
+            return messages;
+        let changed = false;
+        const stripped = messages.filter((message) => {
+            const keep = !this.volatileContextMessages.has(message);
+            if (!keep)
+                changed = true;
+            return keep;
+        });
+        return changed ? stripped : messages;
+    }
+    appendVolatileContextMessages(messages) {
+        if (this.volatileContextMessages.size === 0)
+            return messages;
+        return [...this.stripVolatileContextMessages(messages), ...this.volatileContextMessages];
+    }
     markPendingImagesConsumed(messages) {
         if (this.pendingImageMessages.size === 0)
             return messages;
@@ -307,6 +362,7 @@ export class TurnLoop {
     recordResponseUsage(usage) {
         this.currentTurnUsage = addTokenUsage(this.currentTurnUsage, usage);
         this.currentCumulativeUsage = this.deps.recordCumulativeUsage?.(usage);
+        this.deps.recordCacheReadDiagnostics?.(usage);
     }
     emitCtxFromUsage(usage, messages) {
         if (!this.config.onStream)
@@ -379,13 +435,20 @@ export class TurnLoop {
         let messages = [...initialMessages];
         let finalText = "";
         const budgetTracker = createBudgetTracker();
+        // Recent execution evidence for the built-in Goal judge. Values enter this
+        // window only after irreversible sensitive-data removal, non-text omission,
+        // and per-item truncation at the current-round ToolResult boundary.
+        const goalToolResults = [];
+        const sensitiveGoalResultTools = this.config.goal
+            ? new Set(this.deps.tools.filter((tool) => tool.sensitiveResult).map((tool) => tool.name))
+            : undefined;
         // Goal-mode run-scoped budget tracker (P0). Null when no goal. Stamps a
         // wall-clock start now and accumulates prompt+completion tokens across
         // every turn; the guardrail below force-stops the run once any configured
         // budget is blown — the unattended-safety backstop.
-        this.goalTracker = this.config.goal
-            ? createGoalBudgetTracker(this.config.goal, Date.now())
-            : null;
+        if (!this.goalTracker && this.config.goal) {
+            this.goalTracker = createGoalBudgetTracker(this.config.goal, Date.now());
+        }
         const goalTracker = this.goalTracker;
         // Fresh run: re-arm the approaching-limit announcement.
         this.approachAnnounced = false;
@@ -398,6 +461,25 @@ export class TurnLoop {
         // hook emits, guards) and surfaces them as a model_error result.
         try {
             while (this.turnCount < this.config.maxTurns) {
+                const preTurnGoalTermination = goalTracker
+                    ? goalBudgetTerminationReason(goalTracker, Date.now())
+                    : undefined;
+                if (preTurnGoalTermination) {
+                    this.config.onStream?.({
+                        type: "assistant_message",
+                        message: {
+                            role: "assistant",
+                            content: "（Goal 预算已耗尽，强制停止。）",
+                        },
+                    });
+                    messages = this.redactConsumedSensitiveToolResults(messages);
+                    return {
+                        text: finalText,
+                        reason: "goal_budget_exhausted",
+                        messages,
+                        goalTermination: preTurnGoalTermination,
+                    };
+                }
                 this.turnCount++;
                 this.currentTurnUsage = {
                     promptTokens: 0,
@@ -425,7 +507,7 @@ export class TurnLoop {
                 // they join THIS step's request — no abort, no lost in-flight work. Same
                 // loop-top user-push pattern as turnStartInjection / turn-limit warnings
                 // below. Push to transcript too so they persist + survive resume.
-                this.consumeQueuedSteer(messages, "normal_step");
+                await this.consumeQueuedSteer(messages, "normal_step");
                 const state = initialTurnState(this.turnCount);
                 // Per-turn correlation ID. Every log written through `tlog` (or any
                 // child derived from it) is stamped with `turn` + `turnId`, so
@@ -487,6 +569,7 @@ export class TurnLoop {
                 // pendingImageMessages are preserved through this next model request.
                 const hasPendingSensitiveToolResults = this.sensitiveToolResultRedactions.size > 0;
                 messages = this.prepareMessagesForModel(messages);
+                messages = this.stripVolatileContextMessages(messages);
                 if (hasPendingSensitiveToolResults) {
                     tlog.info("turn.sensitive_tool_result_context_management_skipped", {
                         cat: "turn",
@@ -495,7 +578,7 @@ export class TurnLoop {
                 }
                 else {
                     // Context management (async — may trigger LLM summarization)
-                    messages = await this.deps.contextManager.manageAsync(messages);
+                    messages = await this.deps.contextManager.manageAsync(messages, this.config.signal);
                     // manageAsync can itself issue an LLM summarization call lasting several
                     // seconds; if the signal aborted during it, stop here rather than
                     // proceeding into the (expensive) main model call. Belt to the loop-top
@@ -531,6 +614,7 @@ export class TurnLoop {
                         messages.push(compactInjection);
                     }
                 }
+                messages = this.appendVolatileContextMessages(messages);
                 // Model call (with streaming fallback and max_output_tokens continuation)
                 // Track tool IDs streamed during this turn to avoid duplicate UI events
                 this.streamedToolIds.clear();
@@ -592,28 +676,50 @@ export class TurnLoop {
                         return { text: finalText, reason: "model_error", messages };
                     }
                 }
+                this.lastCompletedModelTurn = this.turnCount;
                 messages = this.redactConsumedSensitiveToolResults(messages);
                 // Record the response once into current-turn and whole-session counters.
                 if (response.usage?.promptTokens !== undefined) {
                     this.recordResponseUsage(response.usage);
                     this.emitCtxFromUsage(response.usage, messages);
                 }
+                // Charge every provider response to the run-scoped Goal budget before
+                // any truncation path can retry or continue the outer turn loop.
+                if (goalTracker && response.usage) {
+                    const used = (response.usage.promptTokens ?? 0) + (response.usage.completionTokens ?? 0);
+                    recordGoalUsage(goalTracker, used);
+                }
+                let budgetTermination = goalTracker
+                    ? goalBudgetTerminationReason(goalTracker, Date.now())
+                    : undefined;
                 // Feed actual token usage back to the context manager so subsequent
                 // compaction decisions use hybrid (actual + delta) estimation rather than
                 // pure heuristics. Without this the manager falls back to char/4 estimates.
                 if (response.usage?.promptTokens !== undefined) {
-                    const anchor = this.deps.contextManager.recordActualUsage(response.usage.promptTokens, messages.length, messages);
+                    const contextAnchorMessages = this.stripVolatileContextMessages(messages);
+                    const anchor = this.deps.contextManager.recordActualUsage(response.usage.promptTokens, contextAnchorMessages.length, contextAnchorMessages);
                     if (anchor)
                         this.deps.recordContextUsageAnchor?.(anchor);
                 }
                 messages = this.markPendingImagesConsumed(messages);
+                // stopBlockCount guards only the no-tool final-answer loop evaluated by
+                // on_stop (text -> block -> text -> block). A response that structurally
+                // contains tool_use has left that streak, regardless of whether the call
+                // will execute successfully. Reset here, before truncated-tool and later
+                // queued-steer paths can continue the outer loop. Repeated ineffective
+                // tool_use turns still consume turns and are bounded by maxTurns.
+                if (response.toolCalls.length > 0) {
+                    this.stopBlockCount = 0;
+                }
                 // Truncation that cut off a TOOL CALL: the model overflowed
                 // max_output_tokens mid tool-call, so the arg JSON is incomplete (e.g. a
                 // Write whose `content` was clipped, leaving file_path unset). Executing
                 // it raised a misleading "Missing required parameter: file_path". Instead,
                 // tell the model its output was truncated and let the next turn retry —
                 // bounded by the outer maxTurns loop.
-                if (isTruncatedStop(response.stopReason) && response.toolCalls.length > 0) {
+                if (!budgetTermination &&
+                    isTruncatedStop(response.stopReason) &&
+                    response.toolCalls.length > 0) {
                     tlog.info("turn.truncated_tool_call", {
                         cat: "turn",
                         toolCount: response.toolCalls.length,
@@ -625,13 +731,15 @@ export class TurnLoop {
                         role: "user",
                         content: "<system-reminder>Your previous response was truncated by the max output token limit before the tool call finished, so its arguments are incomplete. Do not assume it ran. Either retry with a smaller/more focused tool call (e.g. write the file in sections via Edit), or raise this model's maxOutputTokens.</system-reminder>",
                     });
+                    this.finalizeModelTurn();
                     continue;
                 }
                 // Handle max_output_tokens: if response was truncated, do continuation
                 // (up to 3 times). Truncation is reported as finish_reason "length"
                 // (OpenAI) or stop_reason "max_tokens" (Anthropic) — isTruncatedStop
                 // accepts both, so the OpenAI streaming path triggers continuation too.
-                if (isTruncatedStop(response.stopReason) &&
+                if (!budgetTermination &&
+                    isTruncatedStop(response.stopReason) &&
                     response.toolCalls.length === 0 &&
                     response.text) {
                     let combinedText = response.text;
@@ -642,6 +750,11 @@ export class TurnLoop {
                         // more model calls, emitting text after Stop.
                         if (this.config.signal?.aborted)
                             break;
+                        budgetTermination = goalTracker
+                            ? goalBudgetTerminationReason(goalTracker, Date.now())
+                            : undefined;
+                        if (budgetTermination)
+                            break;
                         tlog.info("turn.max_tokens_continuation", { cat: "turn", retry: retry + 1 });
                         const contMessages = [
                             ...messages,
@@ -653,18 +766,40 @@ export class TurnLoop {
                         ];
                         try {
                             const contResponse = await this.deps.model.call(this.deps.systemPrompt, this.prepareMessagesForModel(contMessages), this.deps.tools, this.config.onStream, this.config.signal);
+                            // Continuations are separate provider responses, so preserve the
+                            // same structural tool_use invariant before processing this one.
+                            if (contResponse.toolCalls.length > 0) {
+                                this.stopBlockCount = 0;
+                            }
                             if (contResponse.usage?.promptTokens !== undefined) {
                                 this.recordResponseUsage(contResponse.usage);
                                 continuedResponse = true;
                             }
+                            if (goalTracker && contResponse.usage) {
+                                const used = (contResponse.usage.promptTokens ?? 0) +
+                                    (contResponse.usage.completionTokens ?? 0);
+                                recordGoalUsage(goalTracker, used);
+                            }
+                            budgetTermination = goalTracker
+                                ? goalBudgetTerminationReason(goalTracker, Date.now())
+                                : undefined;
                             combinedText += contResponse.text;
-                            if (!isTruncatedStop(contResponse.stopReason) || contResponse.toolCalls.length > 0) {
+                            if (budgetTermination ||
+                                !isTruncatedStop(contResponse.stopReason) ||
+                                contResponse.toolCalls.length > 0) {
                                 response = { ...contResponse, text: combinedText };
                                 break;
                             }
                         }
-                        catch {
-                            break;
+                        catch (err) {
+                            if (isAbortError(err) || this.config.signal?.aborted) {
+                                this.markStopped();
+                                this.finalizeModelTurn();
+                                return { text: finalText, reason: "aborted_streaming", messages };
+                            }
+                            this.config.onStream?.({ type: "error", error: formatFriendlyError(err) });
+                            this.finalizeModelTurn();
+                            return { text: finalText, reason: "model_error", messages };
                         }
                     }
                     response = { ...response, text: combinedText };
@@ -672,16 +807,10 @@ export class TurnLoop {
                         this.emitCtxFromUsage(response.usage, messages);
                     }
                 }
-                // Goal-mode run-scoped accounting: add this turn's total token usage
-                // (prompt + completion) to the running total. Done after continuation so
-                // continued output is counted against the budget.
-                if (goalTracker && response.usage) {
-                    const used = (response.usage.promptTokens ?? 0) + (response.usage.completionTokens ?? 0);
-                    recordGoalUsage(goalTracker, used);
-                }
                 // Aborted?
                 if (this.config.signal?.aborted) {
                     this.markStopped();
+                    this.finalizeModelTurn();
                     return { text: finalText, reason: "aborted_streaming", messages };
                 }
                 // Accumulate text
@@ -692,7 +821,10 @@ export class TurnLoop {
                 // is force-stopped regardless of what the model wants to do next (stop OR
                 // continue with tool calls). This is the unattended-safety backstop, so
                 // it sits BEFORE the "tool calls?" branch — both paths pass the gate.
-                if (goalTracker && goalBudgetExceeded(goalTracker, Date.now())) {
+                budgetTermination = goalTracker
+                    ? (budgetTermination ?? goalBudgetTerminationReason(goalTracker, Date.now()))
+                    : undefined;
+                if (goalTracker && budgetTermination) {
                     tlog.info("turn.goal_budget_exhausted", {
                         cat: "goal",
                         tokensUsed: goalTracker.tokensUsed,
@@ -707,7 +839,13 @@ export class TurnLoop {
                             content: "（Goal 预算已耗尽，强制停止。）",
                         },
                     });
-                    return { text: finalText, reason: "goal_budget_exhausted", messages };
+                    this.finalizeModelTurn();
+                    return {
+                        text: finalText,
+                        reason: "goal_budget_exhausted",
+                        messages,
+                        goalTermination: budgetTermination,
+                    };
                 }
                 // Post-check: tool calls?
                 if (response.toolCalls.length === 0) {
@@ -722,7 +860,8 @@ export class TurnLoop {
                         hasToolUse: false,
                     });
                     messages.push({ role: "assistant", content: finalText });
-                    if (this.consumeQueuedSteer(messages, "finalize_backfill")) {
+                    this.finalizeModelTurn();
+                    if (await this.consumeQueuedSteer(messages, "finalize_backfill")) {
                         continue;
                     }
                     // on_stop seam: the model wants to stop. Give handlers (Goal mode)
@@ -731,11 +870,79 @@ export class TurnLoop {
                     // we run another turn instead of returning. Bounded by
                     // maxStopBlocks (consecutive) and the outer maxTurns ceiling.
                     const maxStopBlocks = this.config.maxStopBlocks ?? GOAL_DEFAULT_MAX_STOP_BLOCKS;
+                    if (goalTracker) {
+                        this.deps.updateGoalJudgeContext?.({
+                            toolResults: goalToolResults.map((item) => ({
+                                ...item,
+                            })),
+                            progress: {
+                                turnCount: this.turnCount,
+                                stopRound: this.stopBlockCount + 1,
+                                elapsedMs: Math.max(0, Date.now() - goalTracker.startedAtMs),
+                                tokensUsed: goalTracker.tokensUsed,
+                                tokenBudget: goalTracker.goal.tokenBudget,
+                                timeBudgetMs: goalTracker.goal.timeBudgetMs,
+                                maxTurns: this.config.maxTurns,
+                                maxStopBlocks,
+                            },
+                        });
+                    }
                     const stopHook = await this.emitHook("on_stop", {
                         goal: this.config.goal,
                         finalText,
                         turnCount: this.turnCount,
                     });
+                    // The primary-model judge is a real billed request. Its hook forwards
+                    // usage through recordGoalJudgeUsage above; re-check immediately so an
+                    // over-budget verdict cannot authorize another main-model turn.
+                    const hookGoalTermination = stopHook.goalTermination;
+                    const hookBudgetTermination = hookGoalTermination === "token_budget_exhausted" ||
+                        hookGoalTermination === "time_budget_exhausted"
+                        ? hookGoalTermination
+                        : undefined;
+                    const postJudgeBudgetTermination = goalTracker
+                        ? (hookBudgetTermination ?? goalBudgetTerminationReason(goalTracker, Date.now()))
+                        : undefined;
+                    if (goalTracker && postJudgeBudgetTermination) {
+                        tlog.info("turn.goal_budget_exhausted_after_judge", {
+                            cat: "goal",
+                            reason: postJudgeBudgetTermination,
+                            tokensUsed: goalTracker.tokensUsed,
+                            tokenBudget: goalTracker.goal.tokenBudget,
+                            timeBudgetMs: goalTracker.goal.timeBudgetMs,
+                        });
+                        this.config.onStream?.({
+                            type: "assistant_message",
+                            messageId: assistantMessageId,
+                            message: {
+                                role: "assistant",
+                                content: "（Goal 预算已耗尽，强制停止。）",
+                            },
+                        });
+                        return {
+                            text: finalText,
+                            reason: "goal_budget_exhausted",
+                            messages,
+                            goalTermination: postJudgeBudgetTermination,
+                        };
+                    }
+                    if (goalTracker && hookGoalTermination === "judge_prompt_too_large") {
+                        tlog.warn("turn.goal_judge_prompt_too_large", { cat: "goal" });
+                        this.config.onStream?.({
+                            type: "assistant_message",
+                            messageId: assistantMessageId,
+                            message: {
+                                role: "assistant",
+                                content: "（Goal 裁判输入超过安全上限，无法判定；本次已停止，目标已保留。）",
+                            },
+                        });
+                        return {
+                            text: finalText,
+                            reason: "completed",
+                            messages,
+                            goalTermination: hookGoalTermination,
+                        };
+                    }
                     // The judge's structured verdict (set by GoalStopHook in result.data)
                     // rides back here so we can show goal progress WITHOUT a second LLM
                     // call — `gaps` is whatever the judge already computed.
@@ -775,6 +982,7 @@ export class TurnLoop {
                         });
                         continue;
                     }
+                    let goalTermination;
                     if (stopHook.continueSession && this.stopBlockCount >= maxStopBlocks) {
                         // Cap hit: stop anyway, but tell the user why we're not looping
                         // forever on an unsatisfiable goal.
@@ -788,6 +996,7 @@ export class TurnLoop {
                             status: "exhausted",
                             round: this.stopBlockCount,
                         });
+                        goalTermination = "stop_blocks_exhausted";
                         this.config.onStream?.({
                             type: "assistant_message",
                             messageId: assistantMessageId,
@@ -807,11 +1016,11 @@ export class TurnLoop {
                         });
                     }
                     this.stopBlockCount = 0;
-                    if (this.consumeQueuedSteer(messages, "finalize_backfill")) {
+                    if (await this.consumeQueuedSteer(messages, "finalize_backfill")) {
                         continue;
                     }
                     messages = this.redactConsumedSensitiveToolResults(messages);
-                    return { text: finalText, reason: "completed", messages };
+                    return { text: finalText, reason: "completed", messages, goalTermination };
                 }
                 // Tool execution phase
                 tlog.info("turn.tool_use", {
@@ -854,6 +1063,16 @@ export class TurnLoop {
                 // Record results in transcript and stream
                 const resultBlocks = [];
                 for (const result of results) {
+                    if (goalTracker) {
+                        goalToolResults.push(projectGoalJudgeToolResult(result, this.turnCount, sensitiveGoalResultTools?.has(result.toolName) === true));
+                        // Never evict the front of one legal batch: the default executor may
+                        // return 25 results in a turn. Keep at least maxToolCallsPerTurn so
+                        // the judge renderer can preserve metadata for the entire newest batch.
+                        const residencyLimit = Math.max(20, this.config.maxToolCallsPerTurn);
+                        if (goalToolResults.length > residencyLimit) {
+                            goalToolResults.splice(0, goalToolResults.length - residencyLimit);
+                        }
+                    }
                     resultBlocks.push(toolResultToBlock(result));
                     const streamResult = toolResultForDisplay(result);
                     const transcriptResult = toolResultTranscriptText(result);
@@ -930,7 +1149,8 @@ export class TurnLoop {
                     tlog.info("turn.goal_self_reported_complete", { cat: "goal" });
                     this.stopBlockCount = 0;
                     this.deps.clearPersistedGoal?.();
-                    if (this.consumeQueuedSteer(messages, "finalize_backfill")) {
+                    this.finalizeModelTurn();
+                    if (await this.consumeQueuedSteer(messages, "finalize_backfill")) {
                         continue;
                     }
                     messages = this.redactConsumedSensitiveToolResults(messages);
@@ -946,7 +1166,8 @@ export class TurnLoop {
                     tlog.info("turn.goal_user_cancelled", { cat: "goal" });
                     this.stopBlockCount = 0;
                     this.deps.clearPersistedGoal?.();
-                    if (this.consumeQueuedSteer(messages, "finalize_backfill")) {
+                    this.finalizeModelTurn();
+                    if (await this.consumeQueuedSteer(messages, "finalize_backfill")) {
                         continue;
                     }
                     return { text: finalText, reason: "completed", messages };
@@ -966,7 +1187,8 @@ export class TurnLoop {
                         message: { role: "assistant", content: finalText },
                     });
                     messages.push({ role: "assistant", content: finalText });
-                    if (this.consumeQueuedSteer(messages, "finalize_backfill")) {
+                    this.finalizeModelTurn();
+                    if (await this.consumeQueuedSteer(messages, "finalize_backfill")) {
                         continue;
                     }
                     messages = this.redactConsumedSensitiveToolResults(messages);
@@ -1008,9 +1230,7 @@ export class TurnLoop {
                     hasToolUse: true,
                     toolCallCount: toolCalls.length,
                 });
-                // Record turn boundary
-                this.deps.transcript.appendTurnBoundary();
-                this.config.onTurnBoundary?.(this.turnCount);
+                this.finalizeModelTurn();
                 tlog.info("turn.end", {
                     cat: "turn",
                     duration_ms: Date.now() - turnStartedAt,
@@ -1025,6 +1245,9 @@ export class TurnLoop {
             // engine's post-run saveState records model_error instead of leaving
             // the session stuck at "active".
             this.patchOrphanedToolUses(messages);
+            // If the model already returned, this is still a completed model turn.
+            // The idempotent helper is a no-op when its normal branch finalized first.
+            this.finalizeModelTurn();
             // A user-initiated Stop also bubbles here (an AbortError thrown from the
             // per-turn scaffolding). Treat it as a clean abort, not a model_error —
             // no error event, so the UI shows only the "你停止了本轮" line.
@@ -1048,7 +1271,7 @@ export class TurnLoop {
             maxTurns: this.config.maxTurns,
             turnCount: this.turnCount,
         });
-        this.consumeQueuedSteer(messages, "finalize_backfill");
+        await this.consumeQueuedSteer(messages, "finalize_backfill");
         const hasPendingSensitiveToolResults = this.sensitiveToolResultRedactions.size > 0;
         messages = this.prepareMessagesForModel(messages);
         if (hasPendingSensitiveToolResults) {
@@ -1069,12 +1292,28 @@ export class TurnLoop {
         try {
             const summaryResponse = await this.deps.model.call(this.deps.systemPrompt, this.prepareMessagesForModel(messages), [], // No tools available for summary turn
             this.config.onStream, this.config.signal, this.modelCallRecordingOptions());
+            if (summaryResponse.usage?.promptTokens !== undefined) {
+                this.recordResponseUsage(summaryResponse.usage);
+            }
+            if (goalTracker && summaryResponse.usage) {
+                recordGoalUsage(goalTracker, (summaryResponse.usage.promptTokens ?? 0) + (summaryResponse.usage.completionTokens ?? 0));
+            }
+            if (this.config.signal?.aborted) {
+                this.markStopped();
+                messages = this.redactConsumedSensitiveToolResults(messages);
+                return { text: finalText, reason: "aborted_streaming", messages };
+            }
             messages = this.markPendingImagesConsumed(messages);
             if (summaryResponse.text) {
                 finalText = summaryResponse.text;
             }
         }
-        catch {
+        catch (err) {
+            if (isAbortError(err) || this.config.signal?.aborted) {
+                this.markStopped();
+                messages = this.redactConsumedSensitiveToolResults(messages);
+                return { text: finalText, reason: "aborted_streaming", messages };
+            }
             // If even summary fails, just return what we have
             this.currentTurnLog.warn("turn.summary_failed", { cat: "turn" });
         }
@@ -1086,7 +1325,12 @@ export class TurnLoop {
             messages.push({ role: "assistant", content: finalText });
         }
         messages = this.redactConsumedSensitiveToolResults(messages);
-        return { text: finalText, reason: "max_turns", messages };
+        return {
+            text: finalText,
+            reason: "max_turns",
+            messages,
+            ...(this.config.goal ? { goalTermination: "max_turns_exhausted" } : {}),
+        };
     }
     /**
      * Call model with streaming fallback.
@@ -1168,10 +1412,12 @@ export class TurnLoop {
     get currentTurn() {
         return this.turnCount;
     }
-    consumeQueuedSteer(messages, source) {
+    async consumeQueuedSteer(messages, source) {
         const steered = this.deps.consumeSteer?.(source) ?? [];
         let consumed = false;
-        for (const { id, text, clientMessageId } of steered) {
+        for (let index = 0; index < steered.length; index++) {
+            const item = steered[index];
+            const { id, text, clientMessageId } = item;
             if (!text)
                 continue;
             if (clientMessageId && this.deps.claimClientMessageId?.(clientMessageId, "steer") === false) {
@@ -1183,9 +1429,33 @@ export class TurnLoop {
                 });
                 continue;
             }
+            let content;
+            try {
+                content = this.deps.buildSteerUserMessageContent
+                    ? await this.deps.buildSteerUserMessageContent(item)
+                    : text;
+            }
+            catch (err) {
+                if (clientMessageId)
+                    this.deps.releaseClientMessageId?.(clientMessageId);
+                const pendingSuffix = steered.slice(index);
+                this.deps.restoreSteer?.(pendingSuffix);
+                logger.warn("steer.prepare.failed_requeued", {
+                    clientMessageId,
+                    steerId: id,
+                    sessionId: this.deps.sessionId,
+                    source,
+                    restoredCount: pendingSuffix.length,
+                    error: err.message,
+                });
+                break;
+            }
             consumed = true;
-            messages.push({ role: "user", content: text });
-            this.deps.transcript.appendMessage("user", text, { steerId: id, clientMessageId });
+            this.deps.setOriginClientMessageId?.(clientMessageId);
+            const message = { role: "user", content };
+            messages.push(message);
+            this.trackFreshImageMessage(message);
+            this.deps.transcript.appendMessage("user", content, { steerId: id, clientMessageId });
             this.config.onStream?.({ type: "steer_injected", text, id });
         }
         return consumed;