byterover-cli 3.10.1 → 3.10.2

package/dist/agent/infra/agent/service-initializer.js

@@ -121,12 +121,18 @@ export async function createCipherAgentServices(config, agentEventBus) {
         basePath: promptsBasePath,
         validateConfig: true,
     });
-    // Register default contributors
+    // Register default contributors.
+    //
+    // Note: dateTime is intentionally NOT in the system prompt. Anthropic
+    // prompt caching does token-level prefix matching, so a per-iteration
+    // refreshed timestamp here would invalidate the cache for everything
+    // past it. dateTime is instead injected into the first user message
+    // by AgentLLMService, where it lives after the cache breakpoints and
+    // does not poison the cached prefix.
     systemPromptManager.registerContributors([
         { enabled: true, filepath: 'system-prompt.yml', id: 'base', priority: 0, type: 'file' },
         { enabled: true, id: 'env', priority: 10, type: 'environment' },
         { enabled: true, id: 'memories', priority: 20, type: 'memory' },
-        { enabled: true, id: 'datetime', priority: 30, type: 'dateTime' },
     ]);
     // Register context tree structure contributor for query/curate commands
     // This injects the .brv/context-tree structure into the system prompt,

package/dist/agent/infra/llm/agent-llm-service.d.ts

@@ -14,6 +14,15 @@ import { SessionEventBus } from '../events/event-emitter.js';
 import { ContextManager, type FileData, type ImageData } from './context/context-manager.js';
 import { type ThinkingConfig } from './thought-parser.js';
 import { type TruncationConfig } from './tool-output-processor.js';
+/**
+ * Build a `<dateTime>...</dateTime>\n\n` prefix for a user-message body.
+ *
+ * Per-call timestamps must NOT enter the system prompt (they would poison
+ * the prefix cache). They are injected into the user message instead, at
+ * the boundaries where the model legitimately needs fresh time context:
+ * the iter-0 input, and after a rolling-checkpoint history clear.
+ */
+export declare function buildDateTimePrefix(now?: Date): string;
 /**
  * Configuration for ByteRover LLM service
  */
@@ -390,15 +399,6 @@ export declare class AgentLLMService implements ILLMService {
      * @param textInput - Original user input text (for continuation prompt)
      */
     private performRollingCheckpoint;
-    /**
-     * Replace the DateTime section in a cached system prompt with a fresh timestamp.
-     * DateTimeContributor wraps its output in <dateTime>...</dateTime> XML tags,
-     * enabling reliable regex replacement without rebuilding the entire prompt.
-     *
-     * @param cachedPrompt - Previously cached system prompt
-     * @returns Updated prompt with fresh DateTime
-     */
-    private refreshDateTime;
     /**
      * Check if a rolling checkpoint should trigger.
      * Triggers every N iterations for curate/query commands, or when token utilization is high.

package/dist/agent/infra/llm/agent-llm-service.js

@@ -21,6 +21,17 @@ import { OpenRouterTokenizer } from './tokenizers/openrouter-tokenizer.js';
 import { ToolOutputProcessor } from './tool-output-processor.js';
 /** Target utilization ratio for message tokens (leaves headroom for response) */
 const TARGET_MESSAGE_TOKEN_UTILIZATION = 0.7;
+/**
+ * Build a `<dateTime>...</dateTime>\n\n` prefix for a user-message body.
+ *
+ * Per-call timestamps must NOT enter the system prompt (they would poison
+ * the prefix cache). They are injected into the user message instead, at
+ * the boundaries where the model legitimately needs fresh time context:
+ * the iter-0 input, and after a rolling-checkpoint history clear.
+ */
+export function buildDateTimePrefix(now = new Date()) {
+    return `<dateTime>Current date and time: ${now.toISOString()}</dateTime>\n\n`;
+}
 /**
  * ByteRover LLM Service.
  *
@@ -652,8 +663,11 @@ export class AgentLLMService {
             this.memoryDirtyFlag = false;
         }
         else {
-            // Cache hit: reuse base prompt, only refresh the DateTime section
-            basePrompt = this.refreshDateTime(this.cachedBasePrompt);
+            // Cache hit: reuse base prompt verbatim. The cached prompt has no
+            // dateTime section to refresh — dateTime is injected into the
+            // first user message instead so the system prefix stays byte-stable
+            // across iterations and prompt caching can engage cleanly.
+            basePrompt = this.cachedBasePrompt;
         }
         let systemPrompt = basePrompt;
         // Determine which reflection prompt to add (only highest priority is chosen)
@@ -687,9 +701,13 @@ export class AgentLLMService {
         const systemPromptTokens = this.generator.estimateTokensSync(systemPrompt);
         // Add user message and compress context within mutex lock
         return this.mutex.withLock(async () => {
-            // Add user message to context only on the first iteration
+            // Add user message to context only on the first iteration. The
+            // dateTime block is prefixed here (not in the system prompt) so
+            // the cached system prefix stays byte-stable across iterations
+            // and Anthropic/OpenAI/Google prefix caches can engage cleanly.
             if (iterationCount === 0) {
-                await this.contextManager.addUserMessage(textInput, imageData, fileData);
+                const inputWithDateTime = `${buildDateTimePrefix()}${textInput}`;
+                await this.contextManager.addUserMessage(inputWithDateTime, imageData, fileData);
             }
             // Rolling checkpoint: periodically save progress and clear history for RLM commands.
             // This prevents unbounded token accumulation during long curation/query tasks.
@@ -1179,8 +1197,12 @@ export class AgentLLMService {
         this.sandboxService.setSandboxVariable(sessionId, checkpointVar, progressSummary);
         // Clear conversation history
         await this.contextManager.clearHistory();
-        // Re-inject continuation prompt with variable reference
-        const continuationPrompt = [
+        // Re-inject continuation prompt with variable reference.
+        // Prepend the dateTime block: clearHistory wiped the iter-0 user
+        // message that originally carried it, and the iter-0 guard upstream
+        // prevents re-injection. Without this, every iteration after the
+        // first checkpoint loses time context for the rest of the run.
+        const continuationPrompt = buildDateTimePrefix() + [
             `Continue task. Iteration checkpoint at turn ${iterationCount}.`,
             `Previous progress stored in variable: ${checkpointVar}`,
             `Original task: ${textInput.slice(0, 200)}${textInput.length > 200 ? '...' : ''}`,
@@ -1191,18 +1213,6 @@ export class AgentLLMService {
             message: `Rolling checkpoint at iteration ${iterationCount}: history cleared, progress saved to ${checkpointVar}`,
         });
     }
-    /**
-     * Replace the DateTime section in a cached system prompt with a fresh timestamp.
-     * DateTimeContributor wraps its output in <dateTime>...</dateTime> XML tags,
-     * enabling reliable regex replacement without rebuilding the entire prompt.
-     *
-     * @param cachedPrompt - Previously cached system prompt
-     * @returns Updated prompt with fresh DateTime
-     */
-    refreshDateTime(cachedPrompt) {
-        const freshDateTime = `<dateTime>Current date and time: ${new Date().toISOString()}</dateTime>`;
-        return cachedPrompt.replace(/<dateTime>[\S\s]*?<\/dateTime>/, freshDateTime);
-    }
     /**
      * Check if a rolling checkpoint should trigger.
      * Triggers every N iterations for curate/query commands, or when token utilization is high.

package/dist/agent/infra/llm/generators/ai-sdk-content-generator.d.ts

@@ -4,8 +4,17 @@
  * Universal IContentGenerator adapter wrapping any AI SDK LanguageModel.
  * Replaces per-provider content generators with one unified implementation.
  */
-import type { LanguageModel } from 'ai';
+import type { LanguageModel, ModelMessage } from 'ai';
 import type { GenerateContentChunk, GenerateContentRequest, GenerateContentResponse, IContentGenerator } from '../../../core/interfaces/i-content-generator.js';
+/**
+ * Prepend the system prompt as a system-role message carrying
+ * `providerOptions.anthropic.cacheControl: ephemeral`. AI SDK's top-level
+ * `system: string` parameter does not propagate providerOptions, so the
+ * only way to attach Anthropic cache_control to the system block is to
+ * pass it through the messages array. Non-Anthropic providers ignore the
+ * `anthropic` namespace.
+ */
+export declare function prependCachedSystemMessage(systemPrompt: string | undefined, messages: ModelMessage[]): ModelMessage[];
 /**
  * Configuration for AiSdkContentGenerator.
  */

package/dist/agent/infra/llm/generators/ai-sdk-content-generator.js

@@ -8,6 +8,25 @@ import { generateText, streamText } from 'ai';
 import { StreamChunkType } from '../../../core/interfaces/i-content-generator.js';
 import { toAiSdkTools, toModelMessages } from './ai-sdk-message-converter.js';
 const DEFAULT_CHARS_PER_TOKEN = 4;
+/**
+ * Prepend the system prompt as a system-role message carrying
+ * `providerOptions.anthropic.cacheControl: ephemeral`. AI SDK's top-level
+ * `system: string` parameter does not propagate providerOptions, so the
+ * only way to attach Anthropic cache_control to the system block is to
+ * pass it through the messages array. Non-Anthropic providers ignore the
+ * `anthropic` namespace.
+ */
+export function prependCachedSystemMessage(systemPrompt, messages) {
+    if (!systemPrompt) {
+        return messages;
+    }
+    const systemMessage = {
+        content: systemPrompt,
+        providerOptions: { anthropic: { cacheControl: { type: 'ephemeral' } } },
+        role: 'system',
+    };
+    return [systemMessage, ...messages];
+}
 /**
  * Universal content generator that wraps any AI SDK LanguageModel.
  *
@@ -27,7 +46,7 @@ export class AiSdkContentGenerator {
         return Math.ceil(content.length / this.charsPerToken);
     }
     async generateContent(request) {
-        const messages = toModelMessages(request.contents);
+        const messages = prependCachedSystemMessage(request.systemPrompt, toModelMessages(request.contents));
         const tools = toAiSdkTools(request.tools);
         const result = await generateText({
             maxOutputTokens: request.config.maxTokens,
@@ -35,7 +54,6 @@ export class AiSdkContentGenerator {
             messages,
             model: this.model,
             temperature: request.config.temperature,
-            ...(request.systemPrompt && { system: request.systemPrompt }),
             ...(tools && { tools }),
             ...(request.config.topK !== undefined && { topK: request.config.topK }),
             ...(request.config.topP !== undefined && { topP: request.config.topP }),
@@ -68,7 +86,7 @@ export class AiSdkContentGenerator {
         };
     }
     async *generateContentStream(request) {
-        const messages = toModelMessages(request.contents);
+        const messages = prependCachedSystemMessage(request.systemPrompt, toModelMessages(request.contents));
         const tools = toAiSdkTools(request.tools);
         const result = streamText({
             maxOutputTokens: request.config.maxTokens,
@@ -76,7 +94,6 @@ export class AiSdkContentGenerator {
             messages,
             model: this.model,
             temperature: request.config.temperature,
-            ...(request.systemPrompt && { system: request.systemPrompt }),
             ...(tools && { tools }),
             ...(request.config.topK !== undefined && { topK: request.config.topK }),
             ...(request.config.topP !== undefined && { topP: request.config.topP }),

package/dist/agent/infra/llm/generators/ai-sdk-message-converter.d.ts

@@ -16,5 +16,9 @@ export declare function toModelMessages(messages: InternalMessage[]): ModelMessa
 /**
  * Convert our ToolSet to AI SDK tool definitions.
  * Tools are declared without `execute` — our agentic loop handles execution.
+ *
+ * The last tool gets `providerOptions.anthropic.cacheControl: ephemeral`,
+ * which makes Anthropic cache the entire tool block (and the system prompt
+ * before it). Non-Anthropic providers ignore the `anthropic` namespace.
  */
 export declare function toAiSdkTools(tools?: InternalToolSet): Record<string, ReturnType<typeof aiSdkTool>> | undefined;

package/dist/agent/infra/llm/generators/ai-sdk-message-converter.js

@@ -46,16 +46,23 @@ export function toModelMessages(messages) {
 /**
  * Convert our ToolSet to AI SDK tool definitions.
  * Tools are declared without `execute` — our agentic loop handles execution.
+ *
+ * The last tool gets `providerOptions.anthropic.cacheControl: ephemeral`,
+ * which makes Anthropic cache the entire tool block (and the system prompt
+ * before it). Non-Anthropic providers ignore the `anthropic` namespace.
  */
 export function toAiSdkTools(tools) {
     if (!tools || Object.keys(tools).length === 0) {
         return undefined;
     }
+    const entries = Object.entries(tools);
     const result = {};
-    for (const [name, def] of Object.entries(tools)) {
+    for (const [index, [name, def]] of entries.entries()) {
+        const isLast = index === entries.length - 1;
         result[name] = aiSdkTool({
             description: def.description ?? '',
             inputSchema: jsonSchema(def.parameters),
+            ...(isLast && { providerOptions: { anthropic: { cacheControl: { type: 'ephemeral' } } } }),
         });
     }
     return result;

package/dist/agent/infra/map/abstract-generator.d.ts

@@ -8,6 +8,16 @@ export interface AbstractGenerateResult {
     /** L1: key points + structure (~1500 tokens) */
     overviewContent: string;
 }
+/**
+ * Result from a batched abstract generation. One entry per input item, in
+ * input order. Empty string fields signal the model failed to produce content
+ * for that path — the caller's existing fail-open semantics still apply.
+ */
+export interface BatchedAbstractItem {
+    abstractContent: string;
+    contextPath: string;
+    overviewContent: string;
+}
 /**
  * Generate L0 abstract and L1 overview for a knowledge file.
  *
@@ -20,3 +30,22 @@ export interface AbstractGenerateResult {
  * @returns Abstract and overview content strings
  */
 export declare function generateFileAbstracts(fullContent: string, generator: IContentGenerator): Promise<AbstractGenerateResult>;
+/**
+ * Generate L0 abstracts and L1 overviews for N knowledge files in two batched
+ * LLM calls (one batch for all L0s, one for all L1s) instead of 2N per-file
+ * calls.
+ *
+ * Two parallel calls; each call carries all input files in an XML envelope
+ * and the model is instructed to return one element per file. Output is
+ * parsed by path tag and matched back to the input order. Files the model
+ * fails to produce content for receive empty strings (caller's existing
+ * fail-open semantics still apply).
+ *
+ * Caller is responsible for capping batch size; this function does not split
+ * its input. Recommended cap is 5 files per call to keep the L1 batch's
+ * output budget under ~8K tokens.
+ */
+export declare function generateFileAbstractsBatch(items: ReadonlyArray<{
+    contextPath: string;
+    fullContent: string;
+}>, generator: IContentGenerator): Promise<BatchedAbstractItem[]>;

package/dist/agent/infra/map/abstract-generator.js

@@ -31,6 +31,110 @@ ${content}
 }
 /** Truncate content before embedding in LLM prompts to avoid exceeding model context windows during bulk ingest. */
 const MAX_ABSTRACT_CONTENT_CHARS = 20_000;
+/**
+ * Per-file truncation when N files share a single batched call. Matches the
+ * non-batched cap (20 KB) so each file gets the same view of its content
+ * regardless of batched vs per-file mode — total batched user content scales
+ * linearly with N. Avoids quality regression on long-file curates that batched
+ * mode would otherwise see.
+ */
+const MAX_BATCHED_CONTENT_CHARS_PER_FILE = MAX_ABSTRACT_CONTENT_CHARS;
+/** L0 batch output budget: 5 files × ~80 tokens + framing tags ≈ 600 tokens. */
+const BATCH_L0_MAX_OUTPUT_TOKENS = 800;
+/** L1 batch output budget: 5 files × ~1500 tokens + framing tags ≈ 8000 tokens. */
+const BATCH_L1_MAX_OUTPUT_TOKENS = 8500;
+const BATCHED_ABSTRACT_SYSTEM_PROMPT = `You are a technical documentation assistant.
+You produce precise one-line summaries of knowledge documents in a strict XML format.
+Output ONLY the XML — no preamble, no commentary, no markdown fences.`;
+const BATCHED_OVERVIEW_SYSTEM_PROMPT = `You are a technical documentation assistant.
+You produce structured overviews of knowledge documents in a strict XML format.
+Output ONLY the XML — no preamble, no commentary, no markdown fences.`;
+function escapeXmlAttr(value) {
+    return value.replaceAll('&', '&amp;').replaceAll('"', '&quot;').replaceAll('<', '&lt;').replaceAll('>', '&gt;');
+}
+/**
+ * Wrap raw file content in a CDATA section so XML/HTML/JSX/markdown that
+ * mentions `</document>` or `</file>` (perfectly normal for docs that describe
+ * those formats) cannot terminate the envelope and conflate files. The inner
+ * `]]>` escape is the standard CDATA-in-CDATA trick: split the sequence so it
+ * never appears verbatim inside the active section.
+ */
+function wrapCdata(content) {
+    return `<![CDATA[${content.replaceAll(']]>', ']]]]><![CDATA[>')}]]>`;
+}
+function buildBatchedAbstractPrompt(items) {
+    const filesXml = items.map((it) => `<file path="${escapeXmlAttr(it.contextPath)}">
+<document>${wrapCdata(it.content)}</document>
+</file>`).join('\n');
+    return `For each of the following knowledge documents, produce a ONE-LINE summary (max 80 tokens) that is a complete sentence capturing the core topic and key insight.
+
+Output format — emit exactly one <file> element per input file, with the same path attribute:
+<file path="<path>"><abstract>One-line summary.</abstract></file>
+
+Output only these XML elements, in any order. No preamble, no markdown fences.
+
+<files>
+${filesXml}
+</files>`;
+}
+function buildBatchedOverviewPrompt(items) {
+    const filesXml = items.map((it) => `<file path="${escapeXmlAttr(it.contextPath)}">
+<document>${wrapCdata(it.content)}</document>
+</file>`).join('\n');
+    return `For each of the following knowledge documents, produce a structured overview (markdown, under 1500 tokens) that includes:
+- Key points (3-7 bullet points)
+- Structure / sections summary
+- Any notable entities, patterns, or decisions mentioned
+
+Output format — emit exactly one <file> element per input file, with the same path attribute:
+<file path="<path>"><overview>
+- bullet 1
+- bullet 2
+...
+</overview></file>
+
+Output only these XML elements, in any order. No preamble, no markdown fences.
+
+<files>
+${filesXml}
+</files>`;
+}
+/**
+ * Extract <abstract>...</abstract> per <file path="..."> from the model output.
+ * Tolerant: ignores extra whitespace, supports nested newlines inside the inner
+ * tag. Returns a Map keyed by path. Paths that don't appear are absent.
+ *
+ * Anchored on `<file path="...">` openers (not `</file>` closers) so a model
+ * overview that mentions `</file>` literally in prose — perfectly normal for
+ * docs about XML, JSX, or build systems — cannot prematurely terminate the
+ * outer match and orphan the inner tag. Each opener owns the response slice
+ * up to the next opener (or end-of-string), and the inner regex extracts
+ * the payload from that slice.
+ */
+function parseBatchedTags(response, innerTag) {
+    const result = new Map();
+    const fileOpenerRe = /<file\s+path="([^"]*)"[^>]*>/g;
+    const innerRe = new RegExp(`<${innerTag}>([\\s\\S]*?)<\\/${innerTag}>`);
+    const openers = [];
+    let m;
+    while ((m = fileOpenerRe.exec(response)) !== null) {
+        openers.push({ bodyStart: fileOpenerRe.lastIndex, rawPath: m[1] });
+    }
+    for (const [i, opener] of openers.entries()) {
+        // Each opener's slice runs from its end to the start of the next opener
+        // (or end-of-string). Within that slice, the inner regex picks up the
+        // payload. A literal `</file>` in prose has no special meaning here.
+        const sliceEnd = i + 1 < openers.length ? openers[i + 1].bodyStart : response.length;
+        const slice = response.slice(opener.bodyStart, sliceEnd);
+        const inner = innerRe.exec(slice);
+        if (inner) {
+            const path = opener.rawPath
+                .replaceAll('&amp;', '&').replaceAll('&quot;', '"').replaceAll('&lt;', '<').replaceAll('&gt;', '>');
+            result.set(path, inner[1].trim());
+        }
+    }
+    return result;
+}
 /**
  * Generate L0 abstract and L1 overview for a knowledge file.
  *
@@ -65,3 +169,60 @@ export async function generateFileAbstracts(fullContent, generator) {
         overviewContent: overviewText.trim(),
     };
 }
+/**
+ * Generate L0 abstracts and L1 overviews for N knowledge files in two batched
+ * LLM calls (one batch for all L0s, one for all L1s) instead of 2N per-file
+ * calls.
+ *
+ * Two parallel calls; each call carries all input files in an XML envelope
+ * and the model is instructed to return one element per file. Output is
+ * parsed by path tag and matched back to the input order. Files the model
+ * fails to produce content for receive empty strings (caller's existing
+ * fail-open semantics still apply).
+ *
+ * Caller is responsible for capping batch size; this function does not split
+ * its input. Recommended cap is 5 files per call to keep the L1 batch's
+ * output budget under ~8K tokens.
+ */
+export async function generateFileAbstractsBatch(items, generator) {
+    if (items.length === 0)
+        return [];
+    // Dedup by contextPath, keeping the LAST occurrence's content. The queue is
+    // FIFO so later items carry the most recent fullContent — and the disk file
+    // already reflects that write, so the abstract must summarize the latest
+    // state rather than an intermediate one. Without this dedup, duplicate paths
+    // emit two `<file path>` blocks the model may answer in either order; the
+    // tag parser keys on path and Map-collapses, leaving non-deterministic
+    // results for the duplicates.
+    const byPath = new Map();
+    for (const it of items) {
+        byPath.set(it.contextPath, {
+            content: it.fullContent.slice(0, MAX_BATCHED_CONTENT_CHARS_PER_FILE),
+            contextPath: it.contextPath,
+        });
+    }
+    const truncated = [...byPath.values()];
+    const [abstractText, overviewText] = await Promise.all([
+        streamToText(generator, {
+            config: { maxTokens: BATCH_L0_MAX_OUTPUT_TOKENS, temperature: 0 },
+            contents: [{ content: buildBatchedAbstractPrompt(truncated), role: 'user' }],
+            model: 'default',
+            systemPrompt: BATCHED_ABSTRACT_SYSTEM_PROMPT,
+            taskId: randomUUID(),
+        }),
+        streamToText(generator, {
+            config: { maxTokens: BATCH_L1_MAX_OUTPUT_TOKENS, temperature: 0 },
+            contents: [{ content: buildBatchedOverviewPrompt(truncated), role: 'user' }],
+            model: 'default',
+            systemPrompt: BATCHED_OVERVIEW_SYSTEM_PROMPT,
+            taskId: randomUUID(),
+        }),
+    ]);
+    const abstracts = parseBatchedTags(abstractText, 'abstract');
+    const overviews = parseBatchedTags(overviewText, 'overview');
+    return items.map((it) => ({
+        abstractContent: (abstracts.get(it.contextPath) ?? '').trim(),
+        contextPath: it.contextPath,
+        overviewContent: (overviews.get(it.contextPath) ?? '').trim(),
+    }));
+}

package/dist/agent/infra/map/abstract-queue.d.ts

@@ -20,6 +20,13 @@ export interface AbstractQueueStatus {
 export declare class AbstractGenerationQueue {
     private readonly projectRoot;
     private readonly maxAttempts;
+    /**
+     * When true, scheduleNext fires the next batch even if pending is below
+     * BATCH_SIZE_CAP. Set by drain(); reset once the queue is fully idle.
+     * Without this, items below the cap would be buffered indefinitely with
+     * no flush trigger when a curate writes fewer files than the cap.
+     */
+    private drainRequested;
     private drainResolvers;
     private failed;
     private generator;