c0de-agent 0.0.0 → 1.1.0

package/dist/session/shake.js

@@ -0,0 +1,304 @@
+import { estimateTokens } from './token.js';
+/** Rough token cost of a placeholder line; used only for the savings gate. */
+const PLACEHOLDER_TOKEN_ESTIMATE = 16;
+/** 识别顶层 XML 元素（小写 tag，保守策略）。 */
+const OPENING_XML = /^<([a-z_-]+)(?:\s+[^>]*)?>$/;
+const CLOSING_XML = /^<\/([a-z_-]+)>$/;
+/** Auto-shake 默认配置：保护活跃尾部，保守阈值。 */
+export const DEFAULT_SHAKE_CONFIG = {
+    protectTokens: 16_000,
+    minSavings: 4_000,
+    fenceMinTokens: 400,
+    protectedTools: [],
+};
+/**
+ * 定位 fenced 代码块和顶层 XML 元素 span。返回字符偏移 [start, end) 数组，
+ * 覆盖完整块（含围栏/标签行，不含尾换行）。围栏内抑制 XML 检测。
+ * 未闭合围栏/标签不产生 range（保守策略）。
+ */
+export function scanTextForBlockRanges(text) {
+    const ranges = [];
+    let inFence = false;
+    let fenceStart = -1;
+    const tagStack = [];
+    let xmlStart = -1;
+    let lineStart = 0;
+    for (let i = 0; i <= text.length; i++) {
+        if (i !== text.length && text[i] !== '\n')
+            continue;
+        const line = text.slice(lineStart, i);
+        const lineEnd = i;
+        const trimmedStart = line.trimStart();
+        const isFenceLine = trimmedStart.startsWith('```') || trimmedStart.startsWith('~~~');
+        if (isFenceLine) {
+            if (!inFence) {
+                inFence = true;
+                fenceStart = lineStart;
+            }
+            else {
+                inFence = false;
+                ranges.push({ start: fenceStart, end: lineEnd });
+                fenceStart = -1;
+            }
+            lineStart = i + 1;
+            continue;
+        }
+        if (!inFence) {
+            // Only recognize top-level XML (no leading whitespace)
+            if (line.length === trimmedStart.length) {
+                const openingMatch = OPENING_XML.exec(trimmedStart);
+                if (openingMatch) {
+                    const tagName = openingMatch[1];
+                    if (tagName) {
+                        if (tagStack.length === 0)
+                            xmlStart = lineStart;
+                        tagStack.push(tagName);
+                    }
+                }
+                else {
+                    const closingMatch = CLOSING_XML.exec(trimmedStart);
+                    if (closingMatch &&
+                        tagStack.length > 0 &&
+                        tagStack[tagStack.length - 1] === closingMatch[1]) {
+                        tagStack.pop();
+                        if (tagStack.length === 0 && xmlStart >= 0) {
+                            ranges.push({ start: xmlStart, end: lineEnd });
+                            xmlStart = -1;
+                        }
+                    }
+                }
+            }
+        }
+        lineStart = i + 1;
+    }
+    return mergeRanges(ranges);
+}
+/** 按 start 升序，丢弃与已保留范围重叠的（嵌套取最外层）。 */
+function mergeRanges(ranges) {
+    if (ranges.length <= 1)
+        return ranges;
+    const sorted = [...ranges].sort((a, b) => a.start - b.start);
+    const kept = [];
+    let lastEnd = -1;
+    for (const range of sorted) {
+        if (range.start < lastEnd)
+            continue;
+        kept.push(range);
+        lastEnd = range.end;
+    }
+    return kept;
+}
+/** 单条消息的 token 估算（优先用缓存的 tokenCount）。 */
+function messageTokens(m) {
+    if (m.tokenCount > 0)
+        return m.tokenCount;
+    let total = 0;
+    for (const part of m.content) {
+        switch (part._tag) {
+            case 'text':
+            case 'thinking':
+            case 'steering':
+                total += estimateTokens(part.text);
+                break;
+            case 'tool_call':
+                total += estimateTokens(JSON.stringify(part.input));
+                break;
+            case 'tool_result':
+                total += estimateTokens(JSON.stringify(part.output));
+                break;
+        }
+    }
+    return total;
+}
+/** tool_result part 的输出文本。 */
+function toolResultText(output) {
+    if (output._tag === 'success')
+        return output.output;
+    if (output._tag === 'error')
+        return output.error;
+    if (output._tag === 'truncated')
+        return output.output;
+    return '';
+}
+/** 收集可 shake 的区域。纯函数，不修改输入。 */
+export function collectShakeRegions(messages, config) {
+    const n = messages.length;
+    if (n === 0)
+        return [];
+    // accumulatedAfter[i] = i 之后所有 message 的 token 总和
+    const accumulatedAfter = new Array(n);
+    let acc = 0;
+    for (let i = n - 1; i >= 0; i--) {
+        accumulatedAfter[i] = acc;
+        const m = messages[i];
+        if (m)
+            acc += messageTokens(m);
+    }
+    const regions = [];
+    for (let i = 0; i < n; i++) {
+        const msg = messages[i];
+        if (!msg)
+            continue;
+        const afterTokens = accumulatedAfter[i] ?? 0;
+        const isAfterProtectWindow = afterTokens >= config.protectTokens;
+        if (!isAfterProtectWindow)
+            continue;
+        for (let partIndex = 0; partIndex < msg.content.length; partIndex++) {
+            const part = msg.content[partIndex];
+            if (!part)
+                continue;
+            // tool_result 区域
+            if (part._tag === 'tool_result') {
+                // 已 shaken 跳过
+                if ('shakenAt' in part.output && part.output.shakenAt)
+                    continue;
+                // protectedTools 跳过
+                if (config.protectedTools.includes(part.tool))
+                    continue;
+                const text = toolResultText(part.output);
+                if (text.length === 0)
+                    continue;
+                const tokens = estimateTokens(text);
+                if (tokens < config.fenceMinTokens)
+                    continue;
+                regions.push({
+                    kind: 'toolResult',
+                    id: `${msg.id}:toolResult:${partIndex}`,
+                    messageId: msg.id,
+                    messageIndex: i,
+                    partIndex,
+                    tokens,
+                    originalText: text,
+                    label: part.tool,
+                    toolCallId: part.id,
+                });
+                continue;
+            }
+            // text/thinking block 区域
+            if (part._tag === 'text' || part._tag === 'thinking') {
+                for (const range of scanTextForBlockRanges(part.text)) {
+                    const slice = part.text.slice(range.start, range.end);
+                    if (slice.length === 0)
+                        continue;
+                    const tokens = estimateTokens(slice);
+                    if (tokens < config.fenceMinTokens)
+                        continue;
+                    regions.push({
+                        kind: 'block',
+                        id: `${msg.id}:block:${partIndex}:${range.start}`,
+                        messageId: msg.id,
+                        messageIndex: i,
+                        partIndex,
+                        start: range.start,
+                        end: range.end,
+                        tokens,
+                        originalText: slice,
+                        label: msg.role,
+                    });
+                }
+            }
+        }
+    }
+    // minSavings 门控
+    let savings = 0;
+    for (const region of regions)
+        savings += Math.max(0, region.tokens - PLACEHOLDER_TOKEN_ESTIMATE);
+    if (savings < config.minSavings)
+        return [];
+    return regions;
+}
+/** 为区域生成占位符文本。 */
+function placeholderFor(region) {
+    if (region.kind === 'toolResult') {
+        return `[shaken: ${region.label}, ${region.tokens} tokens]`;
+    }
+    return '[shaken]';
+}
+/** 原位替换选中区域。返回新数组，不修改原数组。 */
+export function applyShakeRegions(messages, regions) {
+    if (regions.length === 0)
+        return messages;
+    // 按 messageId 分组
+    const byMessage = new Map();
+    for (const region of regions) {
+        const list = byMessage.get(region.messageId) ?? [];
+        list.push(region);
+        byMessage.set(region.messageId, list);
+    }
+    return messages.map((msg) => {
+        const msgRegions = byMessage.get(msg.id);
+        if (!msgRegions)
+            return msg;
+        // 深拷贝 content（保证不修改原对象）
+        const newContent = structuredClone(msg.content);
+        const now = Date.now();
+        // toolResult 区域：替换 output 文本
+        for (const region of msgRegions) {
+            if (region.kind !== 'toolResult')
+                continue;
+            const part = newContent[region.partIndex];
+            if (!part)
+                continue;
+            if (part._tag !== 'tool_result')
+                continue;
+            const placeholder = placeholderFor(region);
+            if (part.output._tag === 'success') {
+                part.output = { ...part.output, output: placeholder, shakenAt: now };
+            }
+            else if (part.output._tag === 'error') {
+                part.output = { ...part.output, error: placeholder, shakenAt: now };
+            }
+            else if (part.output._tag === 'truncated') {
+                part.output = { ...part.output, output: placeholder, shakenAt: now };
+            }
+        }
+        // block 区域：按 partIndex 分组，同一 text 内按 start 降序 splice
+        const blockByPart = new Map();
+        for (const region of msgRegions) {
+            if (region.kind !== 'block')
+                continue;
+            const list = blockByPart.get(region.partIndex) ?? [];
+            list.push(region);
+            blockByPart.set(region.partIndex, list);
+        }
+        for (const [partIndex, blockRegions] of blockByPart) {
+            const part = newContent[partIndex];
+            if (!part || (part._tag !== 'text' && part._tag !== 'thinking'))
+                continue;
+            const sorted = [...blockRegions].sort((a, b) => b.start - a.start);
+            let text = part.text;
+            for (const br of sorted) {
+                text = text.slice(0, br.start) + placeholderFor(br) + text.slice(br.end);
+            }
+            part.text = text;
+        }
+        return { ...msg, content: newContent };
+    });
+}
+/** 区域转 API 视图。protectWindow 窗口内的标记 isAfterProtectWindow=false。 */
+export function toRegionViews(regions, config, messages) {
+    const n = messages.length;
+    const accumulatedAfter = new Array(n);
+    let acc = 0;
+    for (let i = n - 1; i >= 0; i--) {
+        accumulatedAfter[i] = acc;
+        const m = messages[i];
+        if (m)
+            acc += messageTokens(m);
+    }
+    return regions.map((region) => ({
+        id: region.id,
+        kind: region.kind,
+        messageId: region.messageId,
+        messageIndex: region.messageIndex,
+        partIndex: region.partIndex,
+        tokens: region.tokens,
+        label: region.label,
+        preview: region.originalText.slice(0, 200),
+        placeholder: region.kind === 'toolResult'
+            ? `[shaken: ${region.label}, ${region.tokens} tokens]`
+            : '[shaken]',
+        isAfterProtectWindow: (accumulatedAfter[region.messageIndex] ?? 0) >= config.protectTokens,
+        ...(region.kind === 'toolResult' ? { toolCallId: region.toolCallId } : {}),
+    }));
+}

package/dist/session/snapshot.d.ts

@@ -0,0 +1,11 @@
+import type { DB } from '../db/client.js';
+import type { FileSnapshot } from './types.js';
+/** Create or update a file snapshot. Always creates a new version. Returns the snapshot id. */
+declare function upsertFileSnapshot(handle: DB, sessionId: string, filePath: string, content: string): Promise<string>;
+/** Get all snapshots for a session. */
+declare function getFileSnapshots(handle: DB, sessionId: string): Promise<FileSnapshot[]>;
+/** Get the latest snapshot for a specific file, or null. */
+declare function getLatestFileSnapshot(handle: DB, sessionId: string, filePath: string): Promise<FileSnapshot | null>;
+/** Quick check: return cached file content if a snapshot exists, else null. */
+declare function checkFileSnapshot(handle: DB, sessionId: string, filePath: string): Promise<string | null>;
+export { checkFileSnapshot, getFileSnapshots, getLatestFileSnapshot, upsertFileSnapshot };

package/dist/session/snapshot.js

@@ -0,0 +1,60 @@
+import { createHash } from 'node:crypto';
+import { and, desc, eq } from 'drizzle-orm';
+import { fileSnapshots } from '../db/schema.js';
+import { estimateTokens } from './token.js';
+function rowToSnapshot(row) {
+    const createdAt = row.createdAt instanceof Date ? row.createdAt.getTime() : new Date(row.createdAt).getTime();
+    return {
+        id: row.id,
+        sessionId: row.sessionId,
+        filePath: row.filePath,
+        content: row.content,
+        contentHash: row.contentHash,
+        tokenCount: row.tokenCount ?? 0,
+        version: row.version ?? 1,
+        createdAt,
+    };
+}
+/** Create or update a file snapshot. Always creates a new version. Returns the snapshot id. */
+async function upsertFileSnapshot(handle, sessionId, filePath, content) {
+    const contentHash = createHash('sha256').update(content).digest('hex');
+    const tokenCount = estimateTokens(content);
+    const [existing] = await handle.db
+        .select()
+        .from(fileSnapshots)
+        .where(and(eq(fileSnapshots.sessionId, sessionId), eq(fileSnapshots.filePath, filePath)))
+        .orderBy(desc(fileSnapshots.version))
+        .limit(1);
+    const version = (existing?.version ?? 0) + 1;
+    const [row] = await handle.db
+        .insert(fileSnapshots)
+        .values({ sessionId, filePath, content, contentHash, tokenCount, version })
+        .returning();
+    if (!row)
+        throw new Error('Failed to insert file snapshot');
+    return row.id;
+}
+/** Get all snapshots for a session. */
+async function getFileSnapshots(handle, sessionId) {
+    const rows = await handle.db
+        .select()
+        .from(fileSnapshots)
+        .where(eq(fileSnapshots.sessionId, sessionId));
+    return rows.map(rowToSnapshot);
+}
+/** Get the latest snapshot for a specific file, or null. */
+async function getLatestFileSnapshot(handle, sessionId, filePath) {
+    const [row] = await handle.db
+        .select()
+        .from(fileSnapshots)
+        .where(and(eq(fileSnapshots.sessionId, sessionId), eq(fileSnapshots.filePath, filePath)))
+        .orderBy(desc(fileSnapshots.version))
+        .limit(1);
+    return row ? rowToSnapshot(row) : null;
+}
+/** Quick check: return cached file content if a snapshot exists, else null. */
+async function checkFileSnapshot(handle, sessionId, filePath) {
+    const snapshot = await getLatestFileSnapshot(handle, sessionId, filePath);
+    return snapshot?.content ?? null;
+}
+export { checkFileSnapshot, getFileSnapshots, getLatestFileSnapshot, upsertFileSnapshot };

package/dist/session/squash.d.ts

@@ -0,0 +1,9 @@
+import type { DB } from '../db/client.js';
+import type { CompactionResult, SquashConfig, Summarizer } from './types.js';
+/**
+ * Squash the most recent `count` messages into a summary, keeping `keepRecent` verbatim.
+ * Similar to `git rebase -i + squash`: compresses recent interactions while preserving
+ * the conversation prefix and a small tail for cache stability.
+ */
+declare function squashRecent(handle: DB, sessionId: string, count: number, summarizer: Summarizer, config?: Partial<SquashConfig>): Promise<CompactionResult>;
+export { squashRecent };

package/dist/session/squash.js

@@ -0,0 +1,88 @@
+import { generateId } from '../shared/index.js';
+import { archiveOriginalEntries } from './archive.js';
+import { extractHotFiles } from './compaction.js';
+import { deleteEntriesByIds, getMessages, insertEntry } from './message.js';
+import { upsertFileSnapshot } from './snapshot.js';
+import { estimateTokens } from './token.js';
+const DEFAULT_SQUASH_CONFIG = {
+    keepRecent: 2,
+    preserveFileSnapshots: true,
+    archiveOriginal: true,
+};
+/**
+ * Squash the most recent `count` messages into a summary, keeping `keepRecent` verbatim.
+ * Similar to `git rebase -i + squash`: compresses recent interactions while preserving
+ * the conversation prefix and a small tail for cache stability.
+ */
+async function squashRecent(handle, sessionId, count, summarizer, config) {
+    const cfg = { ...DEFAULT_SQUASH_CONFIG, ...config };
+    const messages = await getMessages(handle, sessionId);
+    if (messages.length < count + cfg.keepRecent) {
+        return { compacted: false, reason: 'too_few_messages' };
+    }
+    const tailStart = messages.length - cfg.keepRecent;
+    const squashStart = messages.length - count;
+    const toSquash = messages.slice(squashStart, tailStart);
+    const keepTail = messages.slice(tailStart);
+    if (toSquash.length === 0) {
+        return { compacted: false, reason: 'nothing_to_compact' };
+    }
+    const history = toSquash
+        .map((m) => `[${m.role}] ${m.content.map((p) => (p._tag === 'text' ? p.text : JSON.stringify(p))).join(' ')}`)
+        .join('\n');
+    const prompt = `将以下最近的交互压缩为简洁摘要，保留关键决策和上下文，丢弃冗余细节。
+重点保留：修改了哪些文件、做了什么决策、当前进度、下一步计划。
+
+## 最近操作
+[简要描述做了什么]
+
+## 修改的文件
+- path: 变更描述
+
+## 关键决策
+[做出的重要决策]
+
+## 当前状态
+[进度和下一步]
+
+---
+交互历史：
+${history}`;
+    const summary = await summarizer(prompt);
+    const squashEntryId = generateId();
+    const archiveId = cfg.archiveOriginal
+        ? await archiveOriginalEntries(handle, sessionId, toSquash, 'squash', summary, squashEntryId)
+        : generateId();
+    const fileSnapshotIds = [];
+    if (cfg.preserveFileSnapshots) {
+        const hotFiles = extractHotFiles(toSquash);
+        for (const file of hotFiles) {
+            const id = await upsertFileSnapshot(handle, sessionId, file.path, file.content);
+            fileSnapshotIds.push(id);
+        }
+    }
+    await deleteEntriesByIds(handle, toSquash.map((m) => m.id));
+    await insertEntry(handle, {
+        id: squashEntryId,
+        sessionId,
+        tag: 'squash',
+        content: {
+            summary,
+            squashedEntryIds: toSquash.map((m) => m.id),
+            archiveId,
+        },
+        tokenCount: estimateTokens(summary),
+        // Position at the first squashed message's timestamp so it sorts between
+        // the prefix and the kept tail under createdAt-ascending order.
+        createdAt: toSquash[0] ? new Date(toSquash[0].createdAt) : new Date(),
+    });
+    return {
+        compacted: true,
+        summary,
+        archiveId,
+        fileSnapshots: fileSnapshotIds,
+        compactedCount: toSquash.length,
+        keptCount: keepTail.length,
+    };
+}
+export { squashRecent };

package/dist/session/token.d.ts

@@ -0,0 +1,10 @@
+import type { MessageContent } from '../shared/types/message.js';
+/**
+ * CJK-aware token estimate.
+ * Chinese/CJK characters ≈ 2 tokens each (denser encoding).
+ * Other characters ≈ 4 chars/token (standard heuristic).
+ */
+declare const estimateTokens: (text: string) => number;
+/** Sum token estimates across all parts of a message's content array. */
+declare const estimateMessageTokens: (content: MessageContent[]) => number;
+export { estimateMessageTokens, estimateTokens };

package/dist/session/token.js

@@ -0,0 +1,33 @@
+/**
+ * CJK-aware token estimate.
+ * Chinese/CJK characters ≈ 2 tokens each (denser encoding).
+ * Other characters ≈ 4 chars/token (standard heuristic).
+ */
+const estimateTokens = (text) => {
+    if (text.length === 0)
+        return 0;
+    const cjkCount = text.match(/[\u4e00-\u9fff\u3040-\u30ff\uac00-\ud7af]/g)?.length ?? 0;
+    const otherCount = text.length - cjkCount;
+    return Math.ceil(cjkCount * 2 + otherCount / 4);
+};
+/** Sum token estimates across all parts of a message's content array. */
+const estimateMessageTokens = (content) => {
+    let total = 0;
+    for (const part of content) {
+        switch (part._tag) {
+            case 'text':
+            case 'thinking':
+            case 'steering':
+                total += estimateTokens(part.text);
+                break;
+            case 'tool_call':
+                total += estimateTokens(JSON.stringify(part.input));
+                break;
+            case 'tool_result':
+                total += estimateTokens(JSON.stringify(part.output));
+                break;
+        }
+    }
+    return total;
+};
+export { estimateMessageTokens, estimateTokens };

package/dist/session/types.d.ts

@@ -0,0 +1,122 @@
+import type { MessageRole } from '../shared/types/base.js';
+import type { ChatMessage } from '../shared/types/llm.js';
+import type { Message, MessageContent, Session, SessionMetadata } from '../shared/types/message.js';
+export type { MessageRole } from '../shared/types/base.js';
+export type { ChatMessage, Message, MessageContent, Session, SessionMetadata };
+/** A compaction summary entry — replaces compacted messages with a summary. */
+type CompactionEntry = {
+    _tag: 'compaction';
+    id: string;
+    sessionId: string;
+    summary: string;
+    originalEntryIds: string[];
+    archiveId: string;
+    tokenCount: number;
+    createdAt: number;
+};
+/** A squash entry — compresses recent interactions into a summary. */
+type SquashEntry = {
+    _tag: 'squash';
+    id: string;
+    sessionId: string;
+    summary: string;
+    squashedEntryIds: string[];
+    archiveId: string;
+    tokenCount: number;
+    createdAt: number;
+};
+/** Records that a session was forked from another. */
+type BranchSummaryEntry = {
+    _tag: 'branch_summary';
+    id: string;
+    sessionId: string;
+    summary: string;
+    sourceSessionId: string;
+    createdAt: number;
+};
+/** A steering instruction injected mid-conversation. */
+type SteeringEntry = {
+    _tag: 'steering';
+    id: string;
+    sessionId: string;
+    content: string;
+    createdAt: number;
+};
+/** All session entries, ordered chronologically. Narrow via `'_tag' in entry`. */
+type SessionEntry = Message | CompactionEntry | SquashEntry | BranchSummaryEntry | SteeringEntry;
+/** Input for `appendMessage` — role + content; id/timestamps auto-generated. */
+type MessageInput = {
+    role: MessageRole;
+    content: MessageContent[];
+    tokenCount?: number;
+};
+/** A hot file detected from tool-call history. */
+type HotFile = {
+    path: string;
+    content: string;
+    tokenCount: number;
+    accessCount: number;
+};
+/** Configuration for compaction. */
+type CompactionConfig = {
+    /** Token budget for the recent-message keep window. Messages are retained
+     *  from the newest backward until their cumulative token count exceeds this
+     *  value; everything older is summarized. Drives compaction by tokens, not
+     *  by a fixed message count. */
+    keepRecentTokens: number;
+    preserveSnapshots: boolean;
+};
+/** Result of a compaction or squash operation. */
+type CompactionResult = {
+    compacted: true;
+    summary: string;
+    archiveId: string;
+    fileSnapshots: string[];
+    compactedCount: number;
+    keptCount: number;
+} | {
+    compacted: false;
+    reason: 'too_few_messages' | 'nothing_to_compact';
+};
+/** Configuration for squash. */
+type SquashConfig = {
+    keepRecent: number;
+    preserveFileSnapshots: boolean;
+    archiveOriginal: boolean;
+};
+/** Injected summarizer — the session layer does NOT import the LLM package. */
+type Summarizer = (prompt: string) => Promise<string>;
+/** Parsed `@[archive:<id>]` / `@[squash:<n>]` reference. */
+type ArchiveRef = {
+    type: 'archive' | 'squash';
+    id: string;
+};
+/** A node in the session tree. */
+type SessionTreeNode = {
+    session: Session;
+    children: SessionTreeNode[];
+};
+/** A decoded compaction archive row. */
+type CompactionArchive = {
+    id: string;
+    sessionId: string;
+    compactionId: string;
+    archiveType: 'compaction' | 'squash' | 'shake';
+    originalEntries: SessionEntry[];
+    summary: string;
+    tokenCount: number;
+    searchableText: string;
+    createdAt: number;
+};
+/** A decoded file snapshot row. */
+type FileSnapshot = {
+    id: string;
+    sessionId: string;
+    filePath: string;
+    content: string;
+    contentHash: string;
+    tokenCount: number;
+    version: number;
+    createdAt: number;
+};
+export type { ArchiveRef, BranchSummaryEntry, CompactionArchive, CompactionConfig, CompactionEntry, CompactionResult, FileSnapshot, HotFile, MessageInput, SessionEntry, SessionTreeNode, SquashConfig, SquashEntry, SteeringEntry, Summarizer, };

package/dist/session/types.js

@@ -0,0 +1 @@
+export {};

package/dist/shared/index.d.ts

@@ -0,0 +1,2 @@
+export type * from './types/index.js';
+export { generateId, now } from './utils.js';

package/dist/shared/index.js

@@ -0,0 +1 @@
+export { generateId, now } from './utils.js';