@cjhyy/code-shell-core 0.6.0-rc.9 → 0.7.0-beta.1

package/dist/engine/turn-loop.js

@@ -17,6 +17,7 @@ import { isAbortError } from "../llm/client-base.js";
 import { crossedReactiveThreshold } from "./reactive-threshold.js";
 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";
 /**
@@ -59,6 +60,7 @@ export class TurnLoop {
         cacheCreationTokens: 0,
     };
     currentCumulativeUsage;
+    sensitiveToolResultRedactions = new Map();
     pendingImageMessages = new Set();
     /**
      * Consecutive on_stop blocks (Goal mode kept the agent going). Reset to 0
@@ -237,6 +239,20 @@ export class TurnLoop {
         }
         return result.messages;
     }
+    redactConsumedSensitiveToolResults(messages) {
+        if (this.sensitiveToolResultRedactions.size === 0)
+            return messages;
+        const redacted = redactSensitiveToolResultsInMessages(messages, this.sensitiveToolResultRedactions);
+        this.sensitiveToolResultRedactions.clear();
+        return redacted;
+    }
+    modelCallRecordingOptions() {
+        if (this.sensitiveToolResultRedactions.size === 0)
+            return undefined;
+        return {
+            sensitiveToolResultRedactions: new Map(this.sensitiveToolResultRedactions),
+        };
+    }
     trackFreshImageMessage(message) {
         if (messageHasBase64ImagePayload(message)) {
             this.pendingImageMessages.add(message);
@@ -281,7 +297,12 @@ export class TurnLoop {
         if (ctx === this.lastCtxEmit)
             return;
         this.lastCtxEmit = ctx;
-        this.config.onStream({ type: "usage_update", promptTokens: ctx });
+        this.config.onStream({
+            type: "usage_update",
+            promptTokens: ctx,
+            promptTokensSource: overhead > 0 ? "calibrated_estimate" : "heuristic_estimate",
+            promptTokensConfidence: overhead > 0 ? "medium" : "low",
+        });
     }
     recordResponseUsage(usage) {
         this.currentTurnUsage = addTokenUsage(this.currentTurnUsage, usage);
@@ -320,6 +341,8 @@ export class TurnLoop {
         this.config.onStream({
             type: "usage_update",
             promptTokens,
+            promptTokensSource: "provider_usage",
+            promptTokensConfidence: "high",
             ...(usage.cacheReadTokens !== undefined ? { cacheReadTokens: usage.cacheReadTokens } : {}),
             ...(usage.cacheCreationTokens !== undefined
                 ? { cacheCreationTokens: usage.cacheCreationTokens }
@@ -349,6 +372,10 @@ export class TurnLoop {
      * Run the multi-turn agent loop until completion.
      */
     async run(initialMessages) {
+        const result = await this.runUnredacted(initialMessages);
+        return { ...result, messages: this.redactConsumedSensitiveToolResults(result.messages) };
+    }
+    async runUnredacted(initialMessages) {
         let messages = [...initialMessages];
         let finalText = "";
         const budgetTracker = createBudgetTracker();
@@ -407,13 +434,18 @@ export class TurnLoop {
                 // returns; instead, each return-causing branch logs its own terminal
                 // event (model_error, completed, etc.).
                 const turnId = newTurnId();
+                const assistantMessageId = `assistant_${turnId}`;
                 const tlog = logger.child({ turn: this.turnCount, turnId });
                 this.currentTurnLog = tlog;
                 const turnStartedAt = Date.now();
                 tlog.info("turn.start", { cat: "turn", messageCount: messages.length });
                 // Tag downstream tool-exec / permission lines with this turn's IDs.
                 this.deps.toolExecutor.setLogger(tlog);
-                this.config.onStream?.({ type: "stream_request_start", turnNumber: this.turnCount });
+                this.config.onStream?.({
+                    type: "stream_request_start",
+                    turnNumber: this.turnCount,
+                    messageId: assistantMessageId,
+                });
                 const turnStartHook = await this.emitHook("on_turn_start", {
                     turnNumber: this.turnCount,
                 });
@@ -453,16 +485,25 @@ export class TurnLoop {
                 // Pre-check: downgrade image payloads that have already had their one
                 // model-consumption turn, then run context management. Fresh images in
                 // pendingImageMessages are preserved through this next model request.
+                const hasPendingSensitiveToolResults = this.sensitiveToolResultRedactions.size > 0;
                 messages = this.prepareMessagesForModel(messages);
-                // Context management (async — may trigger LLM summarization)
-                messages = await this.deps.contextManager.manageAsync(messages);
-                // 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
-                // brace: this catches an abort that landed *inside* context management.
-                if (this.config.signal?.aborted) {
-                    this.markStopped();
-                    return { text: finalText, reason: "aborted_streaming", messages };
+                if (hasPendingSensitiveToolResults) {
+                    tlog.info("turn.sensitive_tool_result_context_management_skipped", {
+                        cat: "turn",
+                        count: this.sensitiveToolResultRedactions.size,
+                    });
+                }
+                else {
+                    // Context management (async — may trigger LLM summarization)
+                    messages = await this.deps.contextManager.manageAsync(messages);
+                    // 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
+                    // brace: this catches an abort that landed *inside* context management.
+                    if (this.config.signal?.aborted) {
+                        this.markStopped();
+                        return { text: finalText, reason: "aborted_streaming", messages };
+                    }
                 }
                 // No pre-llm ctx emit here: the messages-only estimate would be ~16k
                 // smaller than the real prompt (system + tools not included), making
@@ -476,7 +517,9 @@ export class TurnLoop {
                 // call. Microcompact is lossless (just clearing redundant
                 // tool_results) so we suppress hook emits for it to keep token
                 // overhead down.
-                const pending = this.deps.consumePendingCompactInfo?.();
+                const pending = hasPendingSensitiveToolResults
+                    ? null
+                    : this.deps.consumePendingCompactInfo?.();
                 if (pending && pending.strategy !== "micro") {
                     const compactHook = await this.emitHook("post_compact", {
                         strategy: pending.strategy,
@@ -491,11 +534,12 @@ export class TurnLoop {
                 // 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();
-                // Streaming tool queue: start concurrency-safe tools during streaming
+                // Tool queue is created before the call, but enqueue happens only after
+                // the complete LLMResponse is available below.
                 const streamingQueue = new StreamingToolQueue(this.deps.toolExecutor);
                 let response;
                 try {
-                    response = await this.callModelWithFallback(messages);
+                    response = await this.callModelWithFallback(messages, assistantMessageId);
                 }
                 catch (err) {
                     if (err instanceof ContextLimitError) {
@@ -506,13 +550,14 @@ export class TurnLoop {
                             tlog.warn("turn.ptl_recovery", { cat: "turn", retry, roundsToDrop: retry });
                             messages = dropOldestRounds(messages, retry);
                             try {
-                                response = await this.callModelWithFallback(messages);
+                                response = await this.callModelWithFallback(messages, assistantMessageId);
                                 recovered = true;
                                 break;
                             }
                             catch (retryErr) {
                                 if (!(retryErr instanceof ContextLimitError)) {
                                     this.config.onStream?.({ type: "error", error: formatFriendlyError(retryErr) });
+                                    messages = this.redactConsumedSensitiveToolResults(messages);
                                     return { text: finalText, reason: "model_error", messages };
                                 }
                             }
@@ -523,6 +568,7 @@ export class TurnLoop {
                                 type: "error",
                                 error: "Context limit exceeded after 3 recovery attempts",
                             });
+                            messages = this.redactConsumedSensitiveToolResults(messages);
                             return { text: finalText, reason: "prompt_too_long", messages };
                         }
                     }
@@ -536,14 +582,17 @@ export class TurnLoop {
                         // interrupted turn folds behind the process-card header on reload).
                         this.patchOrphanedToolUses(messages);
                         this.markStopped();
+                        messages = this.redactConsumedSensitiveToolResults(messages);
                         return { text: finalText, reason: "aborted_streaming", messages };
                     }
                     else {
                         this.patchOrphanedToolUses(messages);
                         this.config.onStream?.({ type: "error", error: formatFriendlyError(err) });
+                        messages = this.redactConsumedSensitiveToolResults(messages);
                         return { text: finalText, reason: "model_error", messages };
                     }
                 }
+                messages = this.redactConsumedSensitiveToolResults(messages);
                 // Record the response once into current-turn and whole-session counters.
                 if (response.usage?.promptTokens !== undefined) {
                     this.recordResponseUsage(response.usage);
@@ -553,7 +602,9 @@ export class TurnLoop {
                 // 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) {
-                    this.deps.contextManager.recordActualUsage(response.usage.promptTokens, messages.length);
+                    const anchor = this.deps.contextManager.recordActualUsage(response.usage.promptTokens, messages.length, messages);
+                    if (anchor)
+                        this.deps.recordContextUsageAnchor?.(anchor);
                 }
                 messages = this.markPendingImagesConsumed(messages);
                 // Truncation that cut off a TOOL CALL: the model overflowed
@@ -650,6 +701,7 @@ export class TurnLoop {
                     });
                     this.config.onStream?.({
                         type: "assistant_message",
+                        messageId: assistantMessageId,
                         message: {
                             role: "assistant",
                             content: "（Goal 预算已耗尽，强制停止。）",
@@ -662,6 +714,7 @@ export class TurnLoop {
                     // No tool use — final answer
                     this.config.onStream?.({
                         type: "assistant_message",
+                        messageId: assistantMessageId,
                         message: { role: "assistant", content: finalText },
                     });
                     await this.emitHook("on_turn_end", {
@@ -737,6 +790,7 @@ export class TurnLoop {
                         });
                         this.config.onStream?.({
                             type: "assistant_message",
+                            messageId: assistantMessageId,
                             message: {
                                 role: "assistant",
                                 content: `（Goal 续跑已达 ${maxStopBlocks} 次上限，先停下。）`,
@@ -756,6 +810,7 @@ export class TurnLoop {
                     if (this.consumeQueuedSteer(messages, "finalize_backfill")) {
                         continue;
                     }
+                    messages = this.redactConsumedSensitiveToolResults(messages);
                     return { text: finalText, reason: "completed", messages };
                 }
                 // Tool execution phase
@@ -800,8 +855,13 @@ export class TurnLoop {
                 const resultBlocks = [];
                 for (const result of results) {
                     resultBlocks.push(toolResultToBlock(result));
-                    this.deps.transcript.appendToolResult(result.id, result.toolName, result.result, result.error);
-                    this.config.onStream?.({ type: "tool_result", result });
+                    const streamResult = toolResultForDisplay(result);
+                    const transcriptResult = toolResultTranscriptText(result);
+                    if (result.sensitive && transcriptResult !== undefined) {
+                        this.sensitiveToolResultRedactions.set(result.id, transcriptResult);
+                    }
+                    this.deps.transcript.appendToolResult(result.id, result.toolName, transcriptResult, result.error, result.sensitive ? undefined : result.contentBlocks);
+                    this.config.onStream?.({ type: "tool_result", result: streamResult });
                 }
                 // Fire-and-forget tool use summary (non-blocking). The whole chain is
                 // best-effort observability — a thrown onStream handler, a failed dynamic
@@ -812,9 +872,13 @@ export class TurnLoop {
                         .then(({ generateToolUseSummary }) => {
                         if (!this.deps.model.summarize)
                             return;
-                        return generateToolUseSummary(toolCalls, results, this.deps.model.summarize).then((summary) => {
+                        return generateToolUseSummary(toolCalls, toolResultsForDisplay(results), this.deps.model.summarize).then((summary) => {
                             if (summary) {
-                                this.config.onStream?.({ type: "tool_summary", summary });
+                                this.config.onStream?.({
+                                    type: "tool_summary",
+                                    summary,
+                                    toolCallIds: toolCalls.map((toolCall) => toolCall.id),
+                                });
                             }
                         });
                     })
@@ -869,6 +933,7 @@ export class TurnLoop {
                     if (this.consumeQueuedSteer(messages, "finalize_backfill")) {
                         continue;
                     }
+                    messages = this.redactConsumedSensitiveToolResults(messages);
                     return { text: finalText, reason: "completed", messages };
                 }
                 // Goal mode: user-initiated cancellation. cancel_goal is the "strong
@@ -897,12 +962,14 @@ export class TurnLoop {
                     });
                     this.config.onStream?.({
                         type: "assistant_message",
+                        messageId: assistantMessageId,
                         message: { role: "assistant", content: finalText },
                     });
                     messages.push({ role: "assistant", content: finalText });
                     if (this.consumeQueuedSteer(messages, "finalize_backfill")) {
                         continue;
                     }
+                    messages = this.redactConsumedSensitiveToolResults(messages);
                     return { text: finalText, reason: "completed", messages };
                 }
                 if (budgetDecision === "nudge") {
@@ -963,6 +1030,7 @@ export class TurnLoop {
             // no error event, so the UI shows only the "你停止了本轮" line.
             if (isAbortError(err) || this.config.signal?.aborted) {
                 this.markStopped();
+                messages = this.redactConsumedSensitiveToolResults(messages);
                 return { text: finalText, reason: "aborted_streaming", messages };
             }
             this.currentTurnLog.error("turn.unhandled_error", {
@@ -971,6 +1039,7 @@ export class TurnLoop {
                 stack: err.stack?.split("\n").slice(0, 4).join("\n"),
             });
             this.config.onStream?.({ type: "error", error: formatFriendlyError(err) });
+            messages = this.redactConsumedSensitiveToolResults(messages);
             return { text: finalText, reason: "model_error", messages };
         }
         // Max turns reached — do one final summarization call (no tools)
@@ -980,8 +1049,18 @@ export class TurnLoop {
             turnCount: this.turnCount,
         });
         this.consumeQueuedSteer(messages, "finalize_backfill");
+        const hasPendingSensitiveToolResults = this.sensitiveToolResultRedactions.size > 0;
         messages = this.prepareMessagesForModel(messages);
-        messages = this.deps.contextManager.manage(messages);
+        if (hasPendingSensitiveToolResults) {
+            logger.info("turn.sensitive_tool_result_context_management_skipped", {
+                cat: "turn",
+                count: this.sensitiveToolResultRedactions.size,
+                phase: "max_turns_summary",
+            });
+        }
+        else {
+            messages = this.deps.contextManager.manage(messages);
+        }
         messages.push({
             role: "user",
             content: "<system-reminder>Turn limit reached. Provide a final summary of what you accomplished and what remains to be done. Do NOT call any tools.</system-reminder>",
@@ -989,7 +1068,7 @@ export class TurnLoop {
         this.emitCtxFromMessages(messages);
         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.config.onStream, this.config.signal, this.modelCallRecordingOptions());
             messages = this.markPendingImagesConsumed(messages);
             if (summaryResponse.text) {
                 finalText = summaryResponse.text;
@@ -1006,14 +1085,14 @@ export class TurnLoop {
             });
             messages.push({ role: "assistant", content: finalText });
         }
-        this.config.onStream?.({ type: "turn_complete", reason: "max_turns" });
+        messages = this.redactConsumedSensitiveToolResults(messages);
         return { text: finalText, reason: "max_turns", messages };
     }
     /**
      * Call model with streaming fallback.
      * If streaming fails, emit tombstone and retry non-streaming.
      */
-    async callModelWithFallback(messages) {
+    async callModelWithFallback(messages, assistantMessageId) {
         // Wrap stream callback to track tool_use_start events and reactive compaction
         let streamingResponseTokens = 0;
         let reactiveBucket = -1;
@@ -1045,7 +1124,7 @@ export class TurnLoop {
             }
             : undefined;
         try {
-            return await this.deps.model.call(this.deps.systemPrompt, messages, this.deps.tools, wrappedStream, this.config.signal);
+            return await this.deps.model.call(this.deps.systemPrompt, messages, this.deps.tools, wrappedStream, this.config.signal, this.modelCallRecordingOptions());
         }
         catch (err) {
             // If it's a context or rate limit error, don't fallback — propagate
@@ -1065,14 +1144,25 @@ export class TurnLoop {
             // really just a cancel. The signal is the authoritative cancel source.
             if (isAbortError(err) || this.config.signal?.aborted)
                 throw err;
+            // Sensitive tool results are model-facing exactly once. A streaming
+            // fallback would re-send the same pending plaintext in a second request,
+            // so fail the turn and let the unified exit redact returned history.
+            if (this.sensitiveToolResultRedactions.size > 0) {
+                this.config.onStream?.({ type: "tombstone", messageId: assistantMessageId });
+                this.currentTurnLog.warn("turn.streaming_fallback_skipped_sensitive", {
+                    cat: "turn",
+                    error: err.message,
+                });
+                throw err;
+            }
             // Streaming might have partially emitted — send tombstone to revoke
-            this.config.onStream?.({ type: "tombstone", messageId: `turn_${this.turnCount}` });
+            this.config.onStream?.({ type: "tombstone", messageId: assistantMessageId });
             this.currentTurnLog.warn("turn.streaming_fallback", {
                 cat: "turn",
                 error: err.message,
             });
             // Retry without streaming
-            return await this.deps.model.callWithoutStreaming(this.deps.systemPrompt, messages, this.deps.tools, this.config.signal);
+            return await this.deps.model.callWithoutStreaming(this.deps.systemPrompt, messages, this.deps.tools, this.config.signal, this.modelCallRecordingOptions());
         }
     }
     get currentTurn() {

package/dist/engine/types.d.ts

@@ -75,6 +75,8 @@ export interface EngineConfig {
     /** Browser automation bridge (browser_* tools). Wired by the host (desktop)
      *  after construction via setBrowserBridge. Undefined → tools degrade. */
     browserBridge?: import("../tool-system/browser-bridge.js").BrowserBridge;
+    /** Host-backed workspace switch bridge (desktop only). */
+    workspaceBridge?: import("../tool-system/workspace-bridge.js").WorkspaceBridge;
     /** Inject a cookie credential into the built-in browser (InjectCredential
      *  tool). Wired by the host after construction via setInjectCredential.
      *  Undefined → the tool degrades with a clear "no browser" error. */
@@ -127,6 +129,12 @@ export interface EngineConfig {
      * automation shown; tui hidden). Sub-agents override to "subagent".
      */
     origin?: SessionOrigin;
+    /**
+     * Host-specific builtin tool surface. Desktop replaces the legacy core
+     * worktree tools with SwitchSessionWorkspace because Electron main owns the
+     * authoritative workspace switch service and UI refresh path.
+     */
+    builtinToolHost?: "desktop";
     /**
      * Optional shared EngineRuntime providing pre-constructed shared resources
      * (modelPool, toolRegistry, etc.). When provided, Engine uses these instead

package/dist/git/worktree/crud.d.ts

@@ -0,0 +1,69 @@
+import type { SandboxBackend } from "../../tool-system/sandbox/index.js";
+export interface WorktreeSession {
+    originalCwd: string;
+    worktreePath: string;
+    worktreeName: string;
+    worktreeBranch: string;
+    originalBranch?: string;
+    sessionId: string;
+    createdAt: number;
+}
+/** Per-platform setup/cleanup scripts (a project's localEnvironment). */
+export interface PlatformScripts {
+    default?: string;
+    macos?: string;
+    linux?: string;
+    windows?: string;
+}
+export interface CreateWorktreeOptions {
+    prefix?: string;
+}
+/**
+ * Pick the setup/cleanup script for the running platform, falling back to
+ * `default`. Empty/whitespace-only scripts are treated as absent so a project
+ * can leave a platform key blank without spawning an empty shell. `platform`
+ * defaults to `process.platform` so callers usually omit it; tests pass a
+ * fixed value.
+ */
+export declare function selectPlatformScript(scripts: PlatformScripts | undefined, platform?: NodeJS.Platform): string | undefined;
+/**
+ * Create an isolated git worktree for an agent session.
+ */
+export declare function createWorktree(cwd: string, slug: string, sessionId: string, opts?: CreateWorktreeOptions): Promise<WorktreeSession>;
+export interface WorktreeSetupResult {
+    /** True if no setup script was configured for this platform (nothing ran). */
+    skipped: boolean;
+    /** True if a script ran and exited 0. */
+    ok: boolean;
+    /** Combined stdout/stderr, for surfacing to the user on failure. */
+    output: string;
+    /** Exit code when a script ran; undefined when skipped. */
+    exitCode?: number | null;
+}
+/**
+ * Run a project's `localEnvironment.setupScripts` once, in the freshly-created
+ * worktree's root, right after `git worktree add`.
+ */
+export declare function runWorktreeSetup(worktreePath: string, script: string | undefined, opts?: {
+    sandbox?: SandboxBackend;
+    shellEnv?: Record<string, string>;
+    timeoutMs?: number;
+    signal?: AbortSignal;
+}): Promise<WorktreeSetupResult>;
+export interface RemoveWorktreeResult {
+    /** True once `git worktree remove` completed and the directory is gone. */
+    dirRemoved: boolean;
+    /** The branch targeted for deletion when removeBranch=true. */
+    branch?: string;
+    /** True when removeBranch=true and branch deletion completed. */
+    branchDeleted?: boolean;
+    /** Non-empty when the worktree directory is gone but branch deletion failed. */
+    branchError?: string;
+}
+export interface RemoveWorktreeOptions {
+    prefix?: string;
+}
+/**
+ * Remove a worktree and optionally its branch.
+ */
+export declare function removeWorktree(worktreePath: string, removeBranch?: boolean, opts?: RemoveWorktreeOptions): RemoveWorktreeResult;