@hoilab/ada-cli 0.84.2

package/dist/core/bash-executor.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"bash-executor.d.ts","sourceRoot":"","sources":["../../src/core/bash-executor.ts"],"names":[],"mappings":"AAAA;;;;;;GAMG;AAQH,OAAO,KAAK,EAAE,cAAc,EAAE,MAAM,iBAAiB,CAAC;AAOtD,MAAM,WAAW,mBAAmB;IACnC,+DAA+D;IAC/D,OAAO,CAAC,EAAE,CAAC,KAAK,EAAE,MAAM,KAAK,IAAI,CAAC;IAClC,mCAAmC;IACnC,MAAM,CAAC,EAAE,WAAW,CAAC;CACrB;AAED,MAAM,WAAW,UAAU;IAC1B,sEAAsE;IACtE,MAAM,EAAE,MAAM,CAAC;IACf,wDAAwD;IACxD,QAAQ,EAAE,MAAM,GAAG,SAAS,CAAC;IAC7B,mDAAmD;IACnD,SAAS,EAAE,OAAO,CAAC;IACnB,uCAAuC;IACvC,SAAS,EAAE,OAAO,CAAC;IACnB,yFAAyF;IACzF,cAAc,CAAC,EAAE,MAAM,CAAC;CACxB;AAMD;;;GAGG;AACH,wBAAsB,yBAAyB,CAC9C,OAAO,EAAE,MAAM,EACf,GAAG,EAAE,MAAM,EACX,UAAU,EAAE,cAAc,EAC1B,OAAO,CAAC,EAAE,mBAAmB,GAC3B,OAAO,CAAC,UAAU,CAAC,CAqGrB","sourcesContent":["/**\n * Bash command execution with streaming support and cancellation.\n *\n * This module provides a unified bash execution implementation used by:\n * - AgentSession.executeBash() for interactive and RPC modes\n * - Direct calls from modes that need bash execution\n */\n\nimport { randomBytes } from \"node:crypto\";\nimport { createWriteStream, type WriteStream } from \"node:fs\";\nimport { tmpdir } from \"node:os\";\nimport { join } from \"node:path\";\nimport { stripAnsi } from \"../utils/ansi.ts\";\nimport { sanitizeBinaryOutput } from \"../utils/shell.ts\";\nimport type { BashOperations } from \"./tools/bash.ts\";\nimport { DEFAULT_MAX_BYTES, truncateTail } from \"./tools/truncate.ts\";\n\n// ============================================================================\n// Types\n// ============================================================================\n\nexport interface BashExecutorOptions {\n\t/** Callback for streaming output chunks (already sanitized) */\n\tonChunk?: (chunk: string) => void;\n\t/** AbortSignal for cancellation */\n\tsignal?: AbortSignal;\n}\n\nexport interface BashResult {\n\t/** Combined stdout + stderr output (sanitized, possibly truncated) */\n\toutput: string;\n\t/** Process exit code (undefined if killed/cancelled) */\n\texitCode: number | undefined;\n\t/** Whether the command was cancelled via signal */\n\tcancelled: boolean;\n\t/** Whether the output was truncated */\n\ttruncated: boolean;\n\t/** Path to temp file containing full output (if output exceeded truncation threshold) */\n\tfullOutputPath?: string;\n}\n\n// ============================================================================\n// Implementation\n// ============================================================================\n\n/**\n * Execute a bash command using custom BashOperations.\n * Used for remote execution (SSH, containers, etc.).\n */\nexport async function executeBashWithOperations(\n\tcommand: string,\n\tcwd: string,\n\toperations: BashOperations,\n\toptions?: BashExecutorOptions,\n): Promise<BashResult> {\n\tconst outputChunks: string[] = [];\n\tlet outputBytes = 0;\n\tconst maxOutputBytes = DEFAULT_MAX_BYTES * 2;\n\n\tlet tempFilePath: string | undefined;\n\tlet tempFileStream: WriteStream | undefined;\n\tlet totalBytes = 0;\n\n\tconst ensureTempFile = () => {\n\t\tif (tempFilePath) {\n\t\t\treturn;\n\t\t}\n\t\tconst id = randomBytes(8).toString(\"hex\");\n\t\ttempFilePath = join(tmpdir(), `pi-bash-${id}.log`);\n\t\ttempFileStream = createWriteStream(tempFilePath);\n\t\tfor (const chunk of outputChunks) {\n\t\t\ttempFileStream.write(chunk);\n\t\t}\n\t};\n\n\tconst decoder = new TextDecoder();\n\n\tconst onData = (data: Buffer) => {\n\t\ttotalBytes += data.length;\n\n\t\t// Sanitize: strip ANSI, replace binary garbage, normalize newlines\n\t\tconst text = sanitizeBinaryOutput(stripAnsi(decoder.decode(data, { stream: true }))).replace(/\\r/g, \"\");\n\n\t\t// Start writing to temp file if exceeds threshold\n\t\tif (totalBytes > DEFAULT_MAX_BYTES) {\n\t\t\tensureTempFile();\n\t\t}\n\n\t\tif (tempFileStream) {\n\t\t\ttempFileStream.write(text);\n\t\t}\n\n\t\t// Keep rolling buffer\n\t\toutputChunks.push(text);\n\t\toutputBytes += text.length;\n\t\twhile (outputBytes > maxOutputBytes && outputChunks.length > 1) {\n\t\t\tconst removed = outputChunks.shift()!;\n\t\t\toutputBytes -= removed.length;\n\t\t}\n\n\t\t// Stream to callback\n\t\tif (options?.onChunk) {\n\t\t\toptions.onChunk(text);\n\t\t}\n\t};\n\n\ttry {\n\t\tconst result = await operations.exec(command, cwd, {\n\t\t\tonData,\n\t\t\tsignal: options?.signal,\n\t\t});\n\n\t\tconst fullOutput = outputChunks.join(\"\");\n\t\tconst truncationResult = truncateTail(fullOutput);\n\t\tif (truncationResult.truncated) {\n\t\t\tensureTempFile();\n\t\t}\n\t\tif (tempFileStream) {\n\t\t\ttempFileStream.end();\n\t\t}\n\t\tconst cancelled = options?.signal?.aborted ?? false;\n\n\t\treturn {\n\t\t\toutput: truncationResult.truncated ? truncationResult.content : fullOutput,\n\t\t\texitCode: cancelled ? undefined : (result.exitCode ?? undefined),\n\t\t\tcancelled,\n\t\t\ttruncated: truncationResult.truncated,\n\t\t\tfullOutputPath: tempFilePath,\n\t\t};\n\t} catch (err) {\n\t\t// Check if it was an abort\n\t\tif (options?.signal?.aborted) {\n\t\t\tconst fullOutput = outputChunks.join(\"\");\n\t\t\tconst truncationResult = truncateTail(fullOutput);\n\t\t\tif (truncationResult.truncated) {\n\t\t\t\tensureTempFile();\n\t\t\t}\n\t\t\tif (tempFileStream) {\n\t\t\t\ttempFileStream.end();\n\t\t\t}\n\t\t\treturn {\n\t\t\t\toutput: truncationResult.truncated ? truncationResult.content : fullOutput,\n\t\t\t\texitCode: undefined,\n\t\t\t\tcancelled: true,\n\t\t\t\ttruncated: truncationResult.truncated,\n\t\t\t\tfullOutputPath: tempFilePath,\n\t\t\t};\n\t\t}\n\n\t\tif (tempFileStream) {\n\t\t\ttempFileStream.end();\n\t\t}\n\n\t\tthrow err;\n\t}\n}\n"]}

package/dist/core/bash-executor.js

@@ -0,0 +1,111 @@
+/**
+ * Bash command execution with streaming support and cancellation.
+ *
+ * This module provides a unified bash execution implementation used by:
+ * - AgentSession.executeBash() for interactive and RPC modes
+ * - Direct calls from modes that need bash execution
+ */
+import { randomBytes } from "node:crypto";
+import { createWriteStream } from "node:fs";
+import { tmpdir } from "node:os";
+import { join } from "node:path";
+import { stripAnsi } from "../utils/ansi.js";
+import { sanitizeBinaryOutput } from "../utils/shell.js";
+import { DEFAULT_MAX_BYTES, truncateTail } from "./tools/truncate.js";
+// ============================================================================
+// Implementation
+// ============================================================================
+/**
+ * Execute a bash command using custom BashOperations.
+ * Used for remote execution (SSH, containers, etc.).
+ */
+export async function executeBashWithOperations(command, cwd, operations, options) {
+    const outputChunks = [];
+    let outputBytes = 0;
+    const maxOutputBytes = DEFAULT_MAX_BYTES * 2;
+    let tempFilePath;
+    let tempFileStream;
+    let totalBytes = 0;
+    const ensureTempFile = () => {
+        if (tempFilePath) {
+            return;
+        }
+        const id = randomBytes(8).toString("hex");
+        tempFilePath = join(tmpdir(), `pi-bash-${id}.log`);
+        tempFileStream = createWriteStream(tempFilePath);
+        for (const chunk of outputChunks) {
+            tempFileStream.write(chunk);
+        }
+    };
+    const decoder = new TextDecoder();
+    const onData = (data) => {
+        totalBytes += data.length;
+        // Sanitize: strip ANSI, replace binary garbage, normalize newlines
+        const text = sanitizeBinaryOutput(stripAnsi(decoder.decode(data, { stream: true }))).replace(/\r/g, "");
+        // Start writing to temp file if exceeds threshold
+        if (totalBytes > DEFAULT_MAX_BYTES) {
+            ensureTempFile();
+        }
+        if (tempFileStream) {
+            tempFileStream.write(text);
+        }
+        // Keep rolling buffer
+        outputChunks.push(text);
+        outputBytes += text.length;
+        while (outputBytes > maxOutputBytes && outputChunks.length > 1) {
+            const removed = outputChunks.shift();
+            outputBytes -= removed.length;
+        }
+        // Stream to callback
+        if (options?.onChunk) {
+            options.onChunk(text);
+        }
+    };
+    try {
+        const result = await operations.exec(command, cwd, {
+            onData,
+            signal: options?.signal,
+        });
+        const fullOutput = outputChunks.join("");
+        const truncationResult = truncateTail(fullOutput);
+        if (truncationResult.truncated) {
+            ensureTempFile();
+        }
+        if (tempFileStream) {
+            tempFileStream.end();
+        }
+        const cancelled = options?.signal?.aborted ?? false;
+        return {
+            output: truncationResult.truncated ? truncationResult.content : fullOutput,
+            exitCode: cancelled ? undefined : (result.exitCode ?? undefined),
+            cancelled,
+            truncated: truncationResult.truncated,
+            fullOutputPath: tempFilePath,
+        };
+    }
+    catch (err) {
+        // Check if it was an abort
+        if (options?.signal?.aborted) {
+            const fullOutput = outputChunks.join("");
+            const truncationResult = truncateTail(fullOutput);
+            if (truncationResult.truncated) {
+                ensureTempFile();
+            }
+            if (tempFileStream) {
+                tempFileStream.end();
+            }
+            return {
+                output: truncationResult.truncated ? truncationResult.content : fullOutput,
+                exitCode: undefined,
+                cancelled: true,
+                truncated: truncationResult.truncated,
+                fullOutputPath: tempFilePath,
+            };
+        }
+        if (tempFileStream) {
+            tempFileStream.end();
+        }
+        throw err;
+    }
+}
+//# sourceMappingURL=bash-executor.js.map

package/dist/core/bash-executor.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"bash-executor.js","sourceRoot":"","sources":["../../src/core/bash-executor.ts"],"names":[],"mappings":"AAAA;;;;;;GAMG;AAEH,OAAO,EAAE,WAAW,EAAE,MAAM,aAAa,CAAC;AAC1C,OAAO,EAAE,iBAAiB,EAAoB,MAAM,SAAS,CAAC;AAC9D,OAAO,EAAE,MAAM,EAAE,MAAM,SAAS,CAAC;AACjC,OAAO,EAAE,IAAI,EAAE,MAAM,WAAW,CAAC;AACjC,OAAO,EAAE,SAAS,EAAE,MAAM,kBAAkB,CAAC;AAC7C,OAAO,EAAE,oBAAoB,EAAE,MAAM,mBAAmB,CAAC;AAEzD,OAAO,EAAE,iBAAiB,EAAE,YAAY,EAAE,MAAM,qBAAqB,CAAC;AA0BtE,+EAA+E;AAC/E,iBAAiB;AACjB,+EAA+E;AAE/E;;;GAGG;AACH,MAAM,CAAC,KAAK,UAAU,yBAAyB,CAC9C,OAAe,EACf,GAAW,EACX,UAA0B,EAC1B,OAA6B,EACP;IACtB,MAAM,YAAY,GAAa,EAAE,CAAC;IAClC,IAAI,WAAW,GAAG,CAAC,CAAC;IACpB,MAAM,cAAc,GAAG,iBAAiB,GAAG,CAAC,CAAC;IAE7C,IAAI,YAAgC,CAAC;IACrC,IAAI,cAAuC,CAAC;IAC5C,IAAI,UAAU,GAAG,CAAC,CAAC;IAEnB,MAAM,cAAc,GAAG,GAAG,EAAE,CAAC;QAC5B,IAAI,YAAY,EAAE,CAAC;YAClB,OAAO;QACR,CAAC;QACD,MAAM,EAAE,GAAG,WAAW,CAAC,CAAC,CAAC,CAAC,QAAQ,CAAC,KAAK,CAAC,CAAC;QAC1C,YAAY,GAAG,IAAI,CAAC,MAAM,EAAE,EAAE,WAAW,EAAE,MAAM,CAAC,CAAC;QACnD,cAAc,GAAG,iBAAiB,CAAC,YAAY,CAAC,CAAC;QACjD,KAAK,MAAM,KAAK,IAAI,YAAY,EAAE,CAAC;YAClC,cAAc,CAAC,KAAK,CAAC,KAAK,CAAC,CAAC;QAC7B,CAAC;IAAA,CACD,CAAC;IAEF,MAAM,OAAO,GAAG,IAAI,WAAW,EAAE,CAAC;IAElC,MAAM,MAAM,GAAG,CAAC,IAAY,EAAE,EAAE,CAAC;QAChC,UAAU,IAAI,IAAI,CAAC,MAAM,CAAC;QAE1B,mEAAmE;QACnE,MAAM,IAAI,GAAG,oBAAoB,CAAC,SAAS,CAAC,OAAO,CAAC,MAAM,CAAC,IAAI,EAAE,EAAE,MAAM,EAAE,IAAI,EAAE,CAAC,CAAC,CAAC,CAAC,OAAO,CAAC,KAAK,EAAE,EAAE,CAAC,CAAC;QAExG,kDAAkD;QAClD,IAAI,UAAU,GAAG,iBAAiB,EAAE,CAAC;YACpC,cAAc,EAAE,CAAC;QAClB,CAAC;QAED,IAAI,cAAc,EAAE,CAAC;YACpB,cAAc,CAAC,KAAK,CAAC,IAAI,CAAC,CAAC;QAC5B,CAAC;QAED,sBAAsB;QACtB,YAAY,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC;QACxB,WAAW,IAAI,IAAI,CAAC,MAAM,CAAC;QAC3B,OAAO,WAAW,GAAG,cAAc,IAAI,YAAY,CAAC,MAAM,GAAG,CAAC,EAAE,CAAC;YAChE,MAAM,OAAO,GAAG,YAAY,CAAC,KAAK,EAAG,CAAC;YACtC,WAAW,IAAI,OAAO,CAAC,MAAM,CAAC;QAC/B,CAAC;QAED,qBAAqB;QACrB,IAAI,OAAO,EAAE,OAAO,EAAE,CAAC;YACtB,OAAO,CAAC,OAAO,CAAC,IAAI,CAAC,CAAC;QACvB,CAAC;IAAA,CACD,CAAC;IAEF,IAAI,CAAC;QACJ,MAAM,MAAM,GAAG,MAAM,UAAU,CAAC,IAAI,CAAC,OAAO,EAAE,GAAG,EAAE;YAClD,MAAM;YACN,MAAM,EAAE,OAAO,EAAE,MAAM;SACvB,CAAC,CAAC;QAEH,MAAM,UAAU,GAAG,YAAY,CAAC,IAAI,CAAC,EAAE,CAAC,CAAC;QACzC,MAAM,gBAAgB,GAAG,YAAY,CAAC,UAAU,CAAC,CAAC;QAClD,IAAI,gBAAgB,CAAC,SAAS,EAAE,CAAC;YAChC,cAAc,EAAE,CAAC;QAClB,CAAC;QACD,IAAI,cAAc,EAAE,CAAC;YACpB,cAAc,CAAC,GAAG,EAAE,CAAC;QACtB,CAAC;QACD,MAAM,SAAS,GAAG,OAAO,EAAE,MAAM,EAAE,OAAO,IAAI,KAAK,CAAC;QAEpD,OAAO;YACN,MAAM,EAAE,gBAAgB,CAAC,SAAS,CAAC,CAAC,CAAC,gBAAgB,CAAC,OAAO,CAAC,CAAC,CAAC,UAAU;YAC1E,QAAQ,EAAE,SAAS,CAAC,CAAC,CAAC,SAAS,CAAC,CAAC,CAAC,CAAC,MAAM,CAAC,QAAQ,IAAI,SAAS,CAAC;YAChE,SAAS;YACT,SAAS,EAAE,gBAAgB,CAAC,SAAS;YACrC,cAAc,EAAE,YAAY;SAC5B,CAAC;IACH,CAAC;IAAC,OAAO,GAAG,EAAE,CAAC;QACd,2BAA2B;QAC3B,IAAI,OAAO,EAAE,MAAM,EAAE,OAAO,EAAE,CAAC;YAC9B,MAAM,UAAU,GAAG,YAAY,CAAC,IAAI,CAAC,EAAE,CAAC,CAAC;YACzC,MAAM,gBAAgB,GAAG,YAAY,CAAC,UAAU,CAAC,CAAC;YAClD,IAAI,gBAAgB,CAAC,SAAS,EAAE,CAAC;gBAChC,cAAc,EAAE,CAAC;YAClB,CAAC;YACD,IAAI,cAAc,EAAE,CAAC;gBACpB,cAAc,CAAC,GAAG,EAAE,CAAC;YACtB,CAAC;YACD,OAAO;gBACN,MAAM,EAAE,gBAAgB,CAAC,SAAS,CAAC,CAAC,CAAC,gBAAgB,CAAC,OAAO,CAAC,CAAC,CAAC,UAAU;gBAC1E,QAAQ,EAAE,SAAS;gBACnB,SAAS,EAAE,IAAI;gBACf,SAAS,EAAE,gBAAgB,CAAC,SAAS;gBACrC,cAAc,EAAE,YAAY;aAC5B,CAAC;QACH,CAAC;QAED,IAAI,cAAc,EAAE,CAAC;YACpB,cAAc,CAAC,GAAG,EAAE,CAAC;QACtB,CAAC;QAED,MAAM,GAAG,CAAC;IACX,CAAC;AAAA,CACD","sourcesContent":["/**\n * Bash command execution with streaming support and cancellation.\n *\n * This module provides a unified bash execution implementation used by:\n * - AgentSession.executeBash() for interactive and RPC modes\n * - Direct calls from modes that need bash execution\n */\n\nimport { randomBytes } from \"node:crypto\";\nimport { createWriteStream, type WriteStream } from \"node:fs\";\nimport { tmpdir } from \"node:os\";\nimport { join } from \"node:path\";\nimport { stripAnsi } from \"../utils/ansi.ts\";\nimport { sanitizeBinaryOutput } from \"../utils/shell.ts\";\nimport type { BashOperations } from \"./tools/bash.ts\";\nimport { DEFAULT_MAX_BYTES, truncateTail } from \"./tools/truncate.ts\";\n\n// ============================================================================\n// Types\n// ============================================================================\n\nexport interface BashExecutorOptions {\n\t/** Callback for streaming output chunks (already sanitized) */\n\tonChunk?: (chunk: string) => void;\n\t/** AbortSignal for cancellation */\n\tsignal?: AbortSignal;\n}\n\nexport interface BashResult {\n\t/** Combined stdout + stderr output (sanitized, possibly truncated) */\n\toutput: string;\n\t/** Process exit code (undefined if killed/cancelled) */\n\texitCode: number | undefined;\n\t/** Whether the command was cancelled via signal */\n\tcancelled: boolean;\n\t/** Whether the output was truncated */\n\ttruncated: boolean;\n\t/** Path to temp file containing full output (if output exceeded truncation threshold) */\n\tfullOutputPath?: string;\n}\n\n// ============================================================================\n// Implementation\n// ============================================================================\n\n/**\n * Execute a bash command using custom BashOperations.\n * Used for remote execution (SSH, containers, etc.).\n */\nexport async function executeBashWithOperations(\n\tcommand: string,\n\tcwd: string,\n\toperations: BashOperations,\n\toptions?: BashExecutorOptions,\n): Promise<BashResult> {\n\tconst outputChunks: string[] = [];\n\tlet outputBytes = 0;\n\tconst maxOutputBytes = DEFAULT_MAX_BYTES * 2;\n\n\tlet tempFilePath: string | undefined;\n\tlet tempFileStream: WriteStream | undefined;\n\tlet totalBytes = 0;\n\n\tconst ensureTempFile = () => {\n\t\tif (tempFilePath) {\n\t\t\treturn;\n\t\t}\n\t\tconst id = randomBytes(8).toString(\"hex\");\n\t\ttempFilePath = join(tmpdir(), `pi-bash-${id}.log`);\n\t\ttempFileStream = createWriteStream(tempFilePath);\n\t\tfor (const chunk of outputChunks) {\n\t\t\ttempFileStream.write(chunk);\n\t\t}\n\t};\n\n\tconst decoder = new TextDecoder();\n\n\tconst onData = (data: Buffer) => {\n\t\ttotalBytes += data.length;\n\n\t\t// Sanitize: strip ANSI, replace binary garbage, normalize newlines\n\t\tconst text = sanitizeBinaryOutput(stripAnsi(decoder.decode(data, { stream: true }))).replace(/\\r/g, \"\");\n\n\t\t// Start writing to temp file if exceeds threshold\n\t\tif (totalBytes > DEFAULT_MAX_BYTES) {\n\t\t\tensureTempFile();\n\t\t}\n\n\t\tif (tempFileStream) {\n\t\t\ttempFileStream.write(text);\n\t\t}\n\n\t\t// Keep rolling buffer\n\t\toutputChunks.push(text);\n\t\toutputBytes += text.length;\n\t\twhile (outputBytes > maxOutputBytes && outputChunks.length > 1) {\n\t\t\tconst removed = outputChunks.shift()!;\n\t\t\toutputBytes -= removed.length;\n\t\t}\n\n\t\t// Stream to callback\n\t\tif (options?.onChunk) {\n\t\t\toptions.onChunk(text);\n\t\t}\n\t};\n\n\ttry {\n\t\tconst result = await operations.exec(command, cwd, {\n\t\t\tonData,\n\t\t\tsignal: options?.signal,\n\t\t});\n\n\t\tconst fullOutput = outputChunks.join(\"\");\n\t\tconst truncationResult = truncateTail(fullOutput);\n\t\tif (truncationResult.truncated) {\n\t\t\tensureTempFile();\n\t\t}\n\t\tif (tempFileStream) {\n\t\t\ttempFileStream.end();\n\t\t}\n\t\tconst cancelled = options?.signal?.aborted ?? false;\n\n\t\treturn {\n\t\t\toutput: truncationResult.truncated ? truncationResult.content : fullOutput,\n\t\t\texitCode: cancelled ? undefined : (result.exitCode ?? undefined),\n\t\t\tcancelled,\n\t\t\ttruncated: truncationResult.truncated,\n\t\t\tfullOutputPath: tempFilePath,\n\t\t};\n\t} catch (err) {\n\t\t// Check if it was an abort\n\t\tif (options?.signal?.aborted) {\n\t\t\tconst fullOutput = outputChunks.join(\"\");\n\t\t\tconst truncationResult = truncateTail(fullOutput);\n\t\t\tif (truncationResult.truncated) {\n\t\t\t\tensureTempFile();\n\t\t\t}\n\t\t\tif (tempFileStream) {\n\t\t\t\ttempFileStream.end();\n\t\t\t}\n\t\t\treturn {\n\t\t\t\toutput: truncationResult.truncated ? truncationResult.content : fullOutput,\n\t\t\t\texitCode: undefined,\n\t\t\t\tcancelled: true,\n\t\t\t\ttruncated: truncationResult.truncated,\n\t\t\t\tfullOutputPath: tempFilePath,\n\t\t\t};\n\t\t}\n\n\t\tif (tempFileStream) {\n\t\t\ttempFileStream.end();\n\t\t}\n\n\t\tthrow err;\n\t}\n}\n"]}

package/dist/core/cache-stats.d.ts

@@ -0,0 +1,49 @@
+import type { AssistantMessage } from "@earendil-works/pi-ai";
+import type { SessionEntry } from "./session-manager.ts";
+/**
+ * Prompt-cache TTL: idle gaps longer than this are worth mentioning as the
+ * likely cause of a miss. Anthropic's default cache TTL is 5 minutes.
+ */
+export declare const CACHE_TTL_MS: number;
+/** A counted cache miss on a single assistant message. */
+export interface CacheMiss {
+    /** Prompt tokens that were in the previous turn's prompt but not read from cache. */
+    missedTokens: number;
+    /** Extra dollars paid vs. a full cache hit; 0 when pricing is unknown. */
+    missedCost: number;
+    /** Milliseconds since the previous request (which last refreshed the cache). */
+    idleMs: number;
+    /** True when the model changed relative to the previous request. */
+    modelChanged: boolean;
+}
+export interface CacheWasteTotals {
+    missedTokens: number;
+    missedCost: number;
+    /** Number of counted misses (turns above the noise floor). */
+    missCount: number;
+}
+/** Minimal pricing lookup, satisfied by ModelRuntime. Cost is $/million tokens. */
+export interface ModelPriceSource {
+    getModel(provider: string, modelId: string): {
+        cost: {
+            cacheRead: number;
+        };
+    } | undefined;
+}
+/**
+ * Cumulative cache waste across a session: prompt tokens that should have been
+ * cache reads (they were in the previous turn's prompt) but were re-billed.
+ */
+export declare function computeCacheWaste(entries: SessionEntry[], models: ModelPriceSource): CacheWasteTotals;
+/**
+ * All counted cache misses across a session, keyed by the assistant message
+ * (by reference) that paid for them. Used to re-derive transcript notices when
+ * rebuilding the chat from entries (resume, post-compaction rebuild).
+ */
+export declare function collectCacheMisses(entries: SessionEntry[], models: ModelPriceSource): Map<AssistantMessage, CacheMiss>;
+/**
+ * Detect a cache miss on a just-completed assistant message.
+ * `entries` must not yet contain `message` (message_end fires before persistence).
+ */
+export declare function detectCacheMiss(entries: SessionEntry[], message: AssistantMessage, models: ModelPriceSource): CacheMiss | undefined;
+//# sourceMappingURL=cache-stats.d.ts.map

package/dist/core/cache-stats.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"cache-stats.d.ts","sourceRoot":"","sources":["../../src/core/cache-stats.ts"],"names":[],"mappings":"AAAA,OAAO,KAAK,EAAE,gBAAgB,EAAE,MAAM,uBAAuB,CAAC;AAC9D,OAAO,KAAK,EAAE,YAAY,EAAE,MAAM,sBAAsB,CAAC;AAEzD;;;GAGG;AACH,eAAO,MAAM,YAAY,QAAgB,CAAC;AAK1C,0DAA0D;AAC1D,MAAM,WAAW,SAAS;IACzB,qFAAqF;IACrF,YAAY,EAAE,MAAM,CAAC;IACrB,0EAA0E;IAC1E,UAAU,EAAE,MAAM,CAAC;IACnB,gFAAgF;IAChF,MAAM,EAAE,MAAM,CAAC;IACf,oEAAoE;IACpE,YAAY,EAAE,OAAO,CAAC;CACtB;AAED,MAAM,WAAW,gBAAgB;IAChC,YAAY,EAAE,MAAM,CAAC;IACrB,UAAU,EAAE,MAAM,CAAC;IACnB,8DAA8D;IAC9D,SAAS,EAAE,MAAM,CAAC;CAClB;AAED,mFAAmF;AACnF,MAAM,WAAW,gBAAgB;IAChC,QAAQ,CAAC,QAAQ,EAAE,MAAM,EAAE,OAAO,EAAE,MAAM,GAAG;QAAE,IAAI,EAAE;YAAE,SAAS,EAAE,MAAM,CAAA;SAAE,CAAA;KAAE,GAAG,SAAS,CAAC;CACzF;AAmGD;;;GAGG;AACH,wBAAgB,iBAAiB,CAAC,OAAO,EAAE,YAAY,EAAE,EAAE,MAAM,EAAE,gBAAgB,GAAG,gBAAgB,CAErG;AAED;;;;GAIG;AACH,wBAAgB,kBAAkB,CACjC,OAAO,EAAE,YAAY,EAAE,EACvB,MAAM,EAAE,gBAAgB,GACtB,GAAG,CAAC,gBAAgB,EAAE,SAAS,CAAC,CAElC;AAED;;;GAGG;AACH,wBAAgB,eAAe,CAC9B,OAAO,EAAE,YAAY,EAAE,EACvB,OAAO,EAAE,gBAAgB,EACzB,MAAM,EAAE,gBAAgB,GACtB,SAAS,GAAG,SAAS,CAEvB","sourcesContent":["import type { AssistantMessage } from \"@earendil-works/pi-ai\";\nimport type { SessionEntry } from \"./session-manager.ts\";\n\n/**\n * Prompt-cache TTL: idle gaps longer than this are worth mentioning as the\n * likely cause of a miss. Anthropic's default cache TTL is 5 minutes.\n */\nexport const CACHE_TTL_MS = 5 * 60 * 1000;\n\n/** Per-turn misses at or below this are cache breakpoint granularity noise. */\nconst NOISE_FLOOR_TOKENS = 1024;\n\n/** A counted cache miss on a single assistant message. */\nexport interface CacheMiss {\n\t/** Prompt tokens that were in the previous turn's prompt but not read from cache. */\n\tmissedTokens: number;\n\t/** Extra dollars paid vs. a full cache hit; 0 when pricing is unknown. */\n\tmissedCost: number;\n\t/** Milliseconds since the previous request (which last refreshed the cache). */\n\tidleMs: number;\n\t/** True when the model changed relative to the previous request. */\n\tmodelChanged: boolean;\n}\n\nexport interface CacheWasteTotals {\n\tmissedTokens: number;\n\tmissedCost: number;\n\t/** Number of counted misses (turns above the noise floor). */\n\tmissCount: number;\n}\n\n/** Minimal pricing lookup, satisfied by ModelRuntime. Cost is $/million tokens. */\nexport interface ModelPriceSource {\n\tgetModel(provider: string, modelId: string): { cost: { cacheRead: number } } | undefined;\n}\n\n/** The last request seen by the scan; everything in its prompt should be cached. */\ninterface PreviousRequest {\n\tpromptTokens: number;\n\tmodelKey: string;\n\ttimestamp: number;\n\t/**\n\t * Sticky: some earlier request in this scan segment reported cache activity.\n\t * Distinguishes a total miss on a cache-read-only provider (OpenAI-style,\n\t * writes unreported) from a provider that never reports caching at all.\n\t */\n\treportedCache: boolean;\n}\n\n/**\n * Compute the cache miss for one assistant message relative to the previous\n * request. Returns undefined when nothing is counted: first turn, after a\n * reset, no cache activity ever reported (provider without cache support), or\n * miss below the noise floor.\n */\nfunction detectMiss(\n\tprev: PreviousRequest | undefined,\n\tmessage: AssistantMessage,\n\tmodels: ModelPriceSource,\n): CacheMiss | undefined {\n\tconst usage = message.usage;\n\tconst promptTokens = usage.input + usage.cacheRead + usage.cacheWrite;\n\t// A zero-cache turn only counts when cache activity was reported before:\n\t// on cache-read-only providers that is a total miss, while on providers\n\t// that never report caching it means nothing.\n\tif (!prev || promptTokens <= 0 || (usage.cacheRead + usage.cacheWrite === 0 && !prev.reportedCache)) {\n\t\treturn undefined;\n\t}\n\n\tconst missedTokens = Math.min(prev.promptTokens, promptTokens) - usage.cacheRead;\n\tif (missedTokens <= NOISE_FLOOR_TOKENS) return undefined;\n\n\t// Extra cost = missed tokens billed at the actual paid rate (input/cacheWrite,\n\t// incl. write premium) instead of the cache-read rate. Missed tokens can only\n\t// land in the input or cacheWrite buckets, so the paid rate comes straight\n\t// from this message's own cost breakdown.\n\tconst paidTokens = usage.input + usage.cacheWrite;\n\tconst paidPerToken = paidTokens > 0 ? (usage.cost.input + usage.cost.cacheWrite) / paidTokens : 0;\n\tconst readPerToken =\n\t\tusage.cacheRead > 0\n\t\t\t? usage.cost.cacheRead / usage.cacheRead\n\t\t\t: (models.getModel(message.provider, message.model)?.cost.cacheRead ?? 0) / 1_000_000;\n\n\treturn {\n\t\tmissedTokens,\n\t\tmissedCost: missedTokens * Math.max(0, paidPerToken - readPerToken),\n\t\tidleMs: Math.max(0, message.timestamp - prev.timestamp),\n\t\tmodelChanged: `${message.provider}/${message.model}` !== prev.modelKey,\n\t};\n}\n\nfunction asPreviousRequest(message: AssistantMessage, reportedCache: boolean): PreviousRequest | undefined {\n\tconst usage = message.usage;\n\tconst promptTokens = usage.input + usage.cacheRead + usage.cacheWrite;\n\tif (promptTokens <= 0) return undefined;\n\treturn {\n\t\tpromptTokens,\n\t\tmodelKey: `${message.provider}/${message.model}`,\n\t\ttimestamp: message.timestamp,\n\t\treportedCache: reportedCache || usage.cacheRead + usage.cacheWrite > 0,\n\t};\n}\n\nfunction scan(\n\tentries: SessionEntry[],\n\tmodels: ModelPriceSource,\n): { prev: PreviousRequest | undefined; totals: CacheWasteTotals; misses: Map<AssistantMessage, CacheMiss> } {\n\tlet prev: PreviousRequest | undefined;\n\tconst totals: CacheWasteTotals = { missedTokens: 0, missedCost: 0, missCount: 0 };\n\tconst misses = new Map<AssistantMessage, CacheMiss>();\n\n\tfor (const entry of entries) {\n\t\tif (entry.type === \"compaction\" || entry.type === \"branch_summary\") {\n\t\t\t// The context legitimately changed; the next turn's prompt is new content,\n\t\t\t// not re-billed content. Model switches are NOT exempt: they re-bill the\n\t\t\t// full prompt and should be counted.\n\t\t\tprev = undefined;\n\t\t\tcontinue;\n\t\t}\n\t\tif (entry.type === \"message\" && entry.message.role === \"assistant\") {\n\t\t\tconst miss = detectMiss(prev, entry.message, models);\n\t\t\tif (miss) {\n\t\t\t\ttotals.missedTokens += miss.missedTokens;\n\t\t\t\ttotals.missedCost += miss.missedCost;\n\t\t\t\ttotals.missCount += 1;\n\t\t\t\tmisses.set(entry.message, miss);\n\t\t\t}\n\t\t\tprev = asPreviousRequest(entry.message, prev?.reportedCache ?? false) ?? prev;\n\t\t}\n\t}\n\treturn { prev, totals, misses };\n}\n\n/**\n * Cumulative cache waste across a session: prompt tokens that should have been\n * cache reads (they were in the previous turn's prompt) but were re-billed.\n */\nexport function computeCacheWaste(entries: SessionEntry[], models: ModelPriceSource): CacheWasteTotals {\n\treturn scan(entries, models).totals;\n}\n\n/**\n * All counted cache misses across a session, keyed by the assistant message\n * (by reference) that paid for them. Used to re-derive transcript notices when\n * rebuilding the chat from entries (resume, post-compaction rebuild).\n */\nexport function collectCacheMisses(\n\tentries: SessionEntry[],\n\tmodels: ModelPriceSource,\n): Map<AssistantMessage, CacheMiss> {\n\treturn scan(entries, models).misses;\n}\n\n/**\n * Detect a cache miss on a just-completed assistant message.\n * `entries` must not yet contain `message` (message_end fires before persistence).\n */\nexport function detectCacheMiss(\n\tentries: SessionEntry[],\n\tmessage: AssistantMessage,\n\tmodels: ModelPriceSource,\n): CacheMiss | undefined {\n\treturn detectMiss(scan(entries, models).prev, message, models);\n}\n"]}

package/dist/core/cache-stats.js

@@ -0,0 +1,101 @@
+/**
+ * Prompt-cache TTL: idle gaps longer than this are worth mentioning as the
+ * likely cause of a miss. Anthropic's default cache TTL is 5 minutes.
+ */
+export const CACHE_TTL_MS = 5 * 60 * 1000;
+/** Per-turn misses at or below this are cache breakpoint granularity noise. */
+const NOISE_FLOOR_TOKENS = 1024;
+/**
+ * Compute the cache miss for one assistant message relative to the previous
+ * request. Returns undefined when nothing is counted: first turn, after a
+ * reset, no cache activity ever reported (provider without cache support), or
+ * miss below the noise floor.
+ */
+function detectMiss(prev, message, models) {
+    const usage = message.usage;
+    const promptTokens = usage.input + usage.cacheRead + usage.cacheWrite;
+    // A zero-cache turn only counts when cache activity was reported before:
+    // on cache-read-only providers that is a total miss, while on providers
+    // that never report caching it means nothing.
+    if (!prev || promptTokens <= 0 || (usage.cacheRead + usage.cacheWrite === 0 && !prev.reportedCache)) {
+        return undefined;
+    }
+    const missedTokens = Math.min(prev.promptTokens, promptTokens) - usage.cacheRead;
+    if (missedTokens <= NOISE_FLOOR_TOKENS)
+        return undefined;
+    // Extra cost = missed tokens billed at the actual paid rate (input/cacheWrite,
+    // incl. write premium) instead of the cache-read rate. Missed tokens can only
+    // land in the input or cacheWrite buckets, so the paid rate comes straight
+    // from this message's own cost breakdown.
+    const paidTokens = usage.input + usage.cacheWrite;
+    const paidPerToken = paidTokens > 0 ? (usage.cost.input + usage.cost.cacheWrite) / paidTokens : 0;
+    const readPerToken = usage.cacheRead > 0
+        ? usage.cost.cacheRead / usage.cacheRead
+        : (models.getModel(message.provider, message.model)?.cost.cacheRead ?? 0) / 1_000_000;
+    return {
+        missedTokens,
+        missedCost: missedTokens * Math.max(0, paidPerToken - readPerToken),
+        idleMs: Math.max(0, message.timestamp - prev.timestamp),
+        modelChanged: `${message.provider}/${message.model}` !== prev.modelKey,
+    };
+}
+function asPreviousRequest(message, reportedCache) {
+    const usage = message.usage;
+    const promptTokens = usage.input + usage.cacheRead + usage.cacheWrite;
+    if (promptTokens <= 0)
+        return undefined;
+    return {
+        promptTokens,
+        modelKey: `${message.provider}/${message.model}`,
+        timestamp: message.timestamp,
+        reportedCache: reportedCache || usage.cacheRead + usage.cacheWrite > 0,
+    };
+}
+function scan(entries, models) {
+    let prev;
+    const totals = { missedTokens: 0, missedCost: 0, missCount: 0 };
+    const misses = new Map();
+    for (const entry of entries) {
+        if (entry.type === "compaction" || entry.type === "branch_summary") {
+            // The context legitimately changed; the next turn's prompt is new content,
+            // not re-billed content. Model switches are NOT exempt: they re-bill the
+            // full prompt and should be counted.
+            prev = undefined;
+            continue;
+        }
+        if (entry.type === "message" && entry.message.role === "assistant") {
+            const miss = detectMiss(prev, entry.message, models);
+            if (miss) {
+                totals.missedTokens += miss.missedTokens;
+                totals.missedCost += miss.missedCost;
+                totals.missCount += 1;
+                misses.set(entry.message, miss);
+            }
+            prev = asPreviousRequest(entry.message, prev?.reportedCache ?? false) ?? prev;
+        }
+    }
+    return { prev, totals, misses };
+}
+/**
+ * Cumulative cache waste across a session: prompt tokens that should have been
+ * cache reads (they were in the previous turn's prompt) but were re-billed.
+ */
+export function computeCacheWaste(entries, models) {
+    return scan(entries, models).totals;
+}
+/**
+ * All counted cache misses across a session, keyed by the assistant message
+ * (by reference) that paid for them. Used to re-derive transcript notices when
+ * rebuilding the chat from entries (resume, post-compaction rebuild).
+ */
+export function collectCacheMisses(entries, models) {
+    return scan(entries, models).misses;
+}
+/**
+ * Detect a cache miss on a just-completed assistant message.
+ * `entries` must not yet contain `message` (message_end fires before persistence).
+ */
+export function detectCacheMiss(entries, message, models) {
+    return detectMiss(scan(entries, models).prev, message, models);
+}
+//# sourceMappingURL=cache-stats.js.map

package/dist/core/cache-stats.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"cache-stats.js","sourceRoot":"","sources":["../../src/core/cache-stats.ts"],"names":[],"mappings":"AAGA;;;GAGG;AACH,MAAM,CAAC,MAAM,YAAY,GAAG,CAAC,GAAG,EAAE,GAAG,IAAI,CAAC;AAE1C,+EAA+E;AAC/E,MAAM,kBAAkB,GAAG,IAAI,CAAC;AAuChC;;;;;GAKG;AACH,SAAS,UAAU,CAClB,IAAiC,EACjC,OAAyB,EACzB,MAAwB,EACA;IACxB,MAAM,KAAK,GAAG,OAAO,CAAC,KAAK,CAAC;IAC5B,MAAM,YAAY,GAAG,KAAK,CAAC,KAAK,GAAG,KAAK,CAAC,SAAS,GAAG,KAAK,CAAC,UAAU,CAAC;IACtE,yEAAyE;IACzE,wEAAwE;IACxE,8CAA8C;IAC9C,IAAI,CAAC,IAAI,IAAI,YAAY,IAAI,CAAC,IAAI,CAAC,KAAK,CAAC,SAAS,GAAG,KAAK,CAAC,UAAU,KAAK,CAAC,IAAI,CAAC,IAAI,CAAC,aAAa,CAAC,EAAE,CAAC;QACrG,OAAO,SAAS,CAAC;IAClB,CAAC;IAED,MAAM,YAAY,GAAG,IAAI,CAAC,GAAG,CAAC,IAAI,CAAC,YAAY,EAAE,YAAY,CAAC,GAAG,KAAK,CAAC,SAAS,CAAC;IACjF,IAAI,YAAY,IAAI,kBAAkB;QAAE,OAAO,SAAS,CAAC;IAEzD,+EAA+E;IAC/E,8EAA8E;IAC9E,2EAA2E;IAC3E,0CAA0C;IAC1C,MAAM,UAAU,GAAG,KAAK,CAAC,KAAK,GAAG,KAAK,CAAC,UAAU,CAAC;IAClD,MAAM,YAAY,GAAG,UAAU,GAAG,CAAC,CAAC,CAAC,CAAC,CAAC,KAAK,CAAC,IAAI,CAAC,KAAK,GAAG,KAAK,CAAC,IAAI,CAAC,UAAU,CAAC,GAAG,UAAU,CAAC,CAAC,CAAC,CAAC,CAAC;IAClG,MAAM,YAAY,GACjB,KAAK,CAAC,SAAS,GAAG,CAAC;QAClB,CAAC,CAAC,KAAK,CAAC,IAAI,CAAC,SAAS,GAAG,KAAK,CAAC,SAAS;QACxC,CAAC,CAAC,CAAC,MAAM,CAAC,QAAQ,CAAC,OAAO,CAAC,QAAQ,EAAE,OAAO,CAAC,KAAK,CAAC,EAAE,IAAI,CAAC,SAAS,IAAI,CAAC,CAAC,GAAG,SAAS,CAAC;IAExF,OAAO;QACN,YAAY;QACZ,UAAU,EAAE,YAAY,GAAG,IAAI,CAAC,GAAG,CAAC,CAAC,EAAE,YAAY,GAAG,YAAY,CAAC;QACnE,MAAM,EAAE,IAAI,CAAC,GAAG,CAAC,CAAC,EAAE,OAAO,CAAC,SAAS,GAAG,IAAI,CAAC,SAAS,CAAC;QACvD,YAAY,EAAE,GAAG,OAAO,CAAC,QAAQ,IAAI,OAAO,CAAC,KAAK,EAAE,KAAK,IAAI,CAAC,QAAQ;KACtE,CAAC;AAAA,CACF;AAED,SAAS,iBAAiB,CAAC,OAAyB,EAAE,aAAsB,EAA+B;IAC1G,MAAM,KAAK,GAAG,OAAO,CAAC,KAAK,CAAC;IAC5B,MAAM,YAAY,GAAG,KAAK,CAAC,KAAK,GAAG,KAAK,CAAC,SAAS,GAAG,KAAK,CAAC,UAAU,CAAC;IACtE,IAAI,YAAY,IAAI,CAAC;QAAE,OAAO,SAAS,CAAC;IACxC,OAAO;QACN,YAAY;QACZ,QAAQ,EAAE,GAAG,OAAO,CAAC,QAAQ,IAAI,OAAO,CAAC,KAAK,EAAE;QAChD,SAAS,EAAE,OAAO,CAAC,SAAS;QAC5B,aAAa,EAAE,aAAa,IAAI,KAAK,CAAC,SAAS,GAAG,KAAK,CAAC,UAAU,GAAG,CAAC;KACtE,CAAC;AAAA,CACF;AAED,SAAS,IAAI,CACZ,OAAuB,EACvB,MAAwB,EACoF;IAC5G,IAAI,IAAiC,CAAC;IACtC,MAAM,MAAM,GAAqB,EAAE,YAAY,EAAE,CAAC,EAAE,UAAU,EAAE,CAAC,EAAE,SAAS,EAAE,CAAC,EAAE,CAAC;IAClF,MAAM,MAAM,GAAG,IAAI,GAAG,EAA+B,CAAC;IAEtD,KAAK,MAAM,KAAK,IAAI,OAAO,EAAE,CAAC;QAC7B,IAAI,KAAK,CAAC,IAAI,KAAK,YAAY,IAAI,KAAK,CAAC,IAAI,KAAK,gBAAgB,EAAE,CAAC;YACpE,2EAA2E;YAC3E,yEAAyE;YACzE,qCAAqC;YACrC,IAAI,GAAG,SAAS,CAAC;YACjB,SAAS;QACV,CAAC;QACD,IAAI,KAAK,CAAC,IAAI,KAAK,SAAS,IAAI,KAAK,CAAC,OAAO,CAAC,IAAI,KAAK,WAAW,EAAE,CAAC;YACpE,MAAM,IAAI,GAAG,UAAU,CAAC,IAAI,EAAE,KAAK,CAAC,OAAO,EAAE,MAAM,CAAC,CAAC;YACrD,IAAI,IAAI,EAAE,CAAC;gBACV,MAAM,CAAC,YAAY,IAAI,IAAI,CAAC,YAAY,CAAC;gBACzC,MAAM,CAAC,UAAU,IAAI,IAAI,CAAC,UAAU,CAAC;gBACrC,MAAM,CAAC,SAAS,IAAI,CAAC,CAAC;gBACtB,MAAM,CAAC,GAAG,CAAC,KAAK,CAAC,OAAO,EAAE,IAAI,CAAC,CAAC;YACjC,CAAC;YACD,IAAI,GAAG,iBAAiB,CAAC,KAAK,CAAC,OAAO,EAAE,IAAI,EAAE,aAAa,IAAI,KAAK,CAAC,IAAI,IAAI,CAAC;QAC/E,CAAC;IACF,CAAC;IACD,OAAO,EAAE,IAAI,EAAE,MAAM,EAAE,MAAM,EAAE,CAAC;AAAA,CAChC;AAED;;;GAGG;AACH,MAAM,UAAU,iBAAiB,CAAC,OAAuB,EAAE,MAAwB,EAAoB;IACtG,OAAO,IAAI,CAAC,OAAO,EAAE,MAAM,CAAC,CAAC,MAAM,CAAC;AAAA,CACpC;AAED;;;;GAIG;AACH,MAAM,UAAU,kBAAkB,CACjC,OAAuB,EACvB,MAAwB,EACW;IACnC,OAAO,IAAI,CAAC,OAAO,EAAE,MAAM,CAAC,CAAC,MAAM,CAAC;AAAA,CACpC;AAED;;;GAGG;AACH,MAAM,UAAU,eAAe,CAC9B,OAAuB,EACvB,OAAyB,EACzB,MAAwB,EACA;IACxB,OAAO,UAAU,CAAC,IAAI,CAAC,OAAO,EAAE,MAAM,CAAC,CAAC,IAAI,EAAE,OAAO,EAAE,MAAM,CAAC,CAAC;AAAA,CAC/D","sourcesContent":["import type { AssistantMessage } from \"@earendil-works/pi-ai\";\nimport type { SessionEntry } from \"./session-manager.ts\";\n\n/**\n * Prompt-cache TTL: idle gaps longer than this are worth mentioning as the\n * likely cause of a miss. Anthropic's default cache TTL is 5 minutes.\n */\nexport const CACHE_TTL_MS = 5 * 60 * 1000;\n\n/** Per-turn misses at or below this are cache breakpoint granularity noise. */\nconst NOISE_FLOOR_TOKENS = 1024;\n\n/** A counted cache miss on a single assistant message. */\nexport interface CacheMiss {\n\t/** Prompt tokens that were in the previous turn's prompt but not read from cache. */\n\tmissedTokens: number;\n\t/** Extra dollars paid vs. a full cache hit; 0 when pricing is unknown. */\n\tmissedCost: number;\n\t/** Milliseconds since the previous request (which last refreshed the cache). */\n\tidleMs: number;\n\t/** True when the model changed relative to the previous request. */\n\tmodelChanged: boolean;\n}\n\nexport interface CacheWasteTotals {\n\tmissedTokens: number;\n\tmissedCost: number;\n\t/** Number of counted misses (turns above the noise floor). */\n\tmissCount: number;\n}\n\n/** Minimal pricing lookup, satisfied by ModelRuntime. Cost is $/million tokens. */\nexport interface ModelPriceSource {\n\tgetModel(provider: string, modelId: string): { cost: { cacheRead: number } } | undefined;\n}\n\n/** The last request seen by the scan; everything in its prompt should be cached. */\ninterface PreviousRequest {\n\tpromptTokens: number;\n\tmodelKey: string;\n\ttimestamp: number;\n\t/**\n\t * Sticky: some earlier request in this scan segment reported cache activity.\n\t * Distinguishes a total miss on a cache-read-only provider (OpenAI-style,\n\t * writes unreported) from a provider that never reports caching at all.\n\t */\n\treportedCache: boolean;\n}\n\n/**\n * Compute the cache miss for one assistant message relative to the previous\n * request. Returns undefined when nothing is counted: first turn, after a\n * reset, no cache activity ever reported (provider without cache support), or\n * miss below the noise floor.\n */\nfunction detectMiss(\n\tprev: PreviousRequest | undefined,\n\tmessage: AssistantMessage,\n\tmodels: ModelPriceSource,\n): CacheMiss | undefined {\n\tconst usage = message.usage;\n\tconst promptTokens = usage.input + usage.cacheRead + usage.cacheWrite;\n\t// A zero-cache turn only counts when cache activity was reported before:\n\t// on cache-read-only providers that is a total miss, while on providers\n\t// that never report caching it means nothing.\n\tif (!prev || promptTokens <= 0 || (usage.cacheRead + usage.cacheWrite === 0 && !prev.reportedCache)) {\n\t\treturn undefined;\n\t}\n\n\tconst missedTokens = Math.min(prev.promptTokens, promptTokens) - usage.cacheRead;\n\tif (missedTokens <= NOISE_FLOOR_TOKENS) return undefined;\n\n\t// Extra cost = missed tokens billed at the actual paid rate (input/cacheWrite,\n\t// incl. write premium) instead of the cache-read rate. Missed tokens can only\n\t// land in the input or cacheWrite buckets, so the paid rate comes straight\n\t// from this message's own cost breakdown.\n\tconst paidTokens = usage.input + usage.cacheWrite;\n\tconst paidPerToken = paidTokens > 0 ? (usage.cost.input + usage.cost.cacheWrite) / paidTokens : 0;\n\tconst readPerToken =\n\t\tusage.cacheRead > 0\n\t\t\t? usage.cost.cacheRead / usage.cacheRead\n\t\t\t: (models.getModel(message.provider, message.model)?.cost.cacheRead ?? 0) / 1_000_000;\n\n\treturn {\n\t\tmissedTokens,\n\t\tmissedCost: missedTokens * Math.max(0, paidPerToken - readPerToken),\n\t\tidleMs: Math.max(0, message.timestamp - prev.timestamp),\n\t\tmodelChanged: `${message.provider}/${message.model}` !== prev.modelKey,\n\t};\n}\n\nfunction asPreviousRequest(message: AssistantMessage, reportedCache: boolean): PreviousRequest | undefined {\n\tconst usage = message.usage;\n\tconst promptTokens = usage.input + usage.cacheRead + usage.cacheWrite;\n\tif (promptTokens <= 0) return undefined;\n\treturn {\n\t\tpromptTokens,\n\t\tmodelKey: `${message.provider}/${message.model}`,\n\t\ttimestamp: message.timestamp,\n\t\treportedCache: reportedCache || usage.cacheRead + usage.cacheWrite > 0,\n\t};\n}\n\nfunction scan(\n\tentries: SessionEntry[],\n\tmodels: ModelPriceSource,\n): { prev: PreviousRequest | undefined; totals: CacheWasteTotals; misses: Map<AssistantMessage, CacheMiss> } {\n\tlet prev: PreviousRequest | undefined;\n\tconst totals: CacheWasteTotals = { missedTokens: 0, missedCost: 0, missCount: 0 };\n\tconst misses = new Map<AssistantMessage, CacheMiss>();\n\n\tfor (const entry of entries) {\n\t\tif (entry.type === \"compaction\" || entry.type === \"branch_summary\") {\n\t\t\t// The context legitimately changed; the next turn's prompt is new content,\n\t\t\t// not re-billed content. Model switches are NOT exempt: they re-bill the\n\t\t\t// full prompt and should be counted.\n\t\t\tprev = undefined;\n\t\t\tcontinue;\n\t\t}\n\t\tif (entry.type === \"message\" && entry.message.role === \"assistant\") {\n\t\t\tconst miss = detectMiss(prev, entry.message, models);\n\t\t\tif (miss) {\n\t\t\t\ttotals.missedTokens += miss.missedTokens;\n\t\t\t\ttotals.missedCost += miss.missedCost;\n\t\t\t\ttotals.missCount += 1;\n\t\t\t\tmisses.set(entry.message, miss);\n\t\t\t}\n\t\t\tprev = asPreviousRequest(entry.message, prev?.reportedCache ?? false) ?? prev;\n\t\t}\n\t}\n\treturn { prev, totals, misses };\n}\n\n/**\n * Cumulative cache waste across a session: prompt tokens that should have been\n * cache reads (they were in the previous turn's prompt) but were re-billed.\n */\nexport function computeCacheWaste(entries: SessionEntry[], models: ModelPriceSource): CacheWasteTotals {\n\treturn scan(entries, models).totals;\n}\n\n/**\n * All counted cache misses across a session, keyed by the assistant message\n * (by reference) that paid for them. Used to re-derive transcript notices when\n * rebuilding the chat from entries (resume, post-compaction rebuild).\n */\nexport function collectCacheMisses(\n\tentries: SessionEntry[],\n\tmodels: ModelPriceSource,\n): Map<AssistantMessage, CacheMiss> {\n\treturn scan(entries, models).misses;\n}\n\n/**\n * Detect a cache miss on a just-completed assistant message.\n * `entries` must not yet contain `message` (message_end fires before persistence).\n */\nexport function detectCacheMiss(\n\tentries: SessionEntry[],\n\tmessage: AssistantMessage,\n\tmodels: ModelPriceSource,\n): CacheMiss | undefined {\n\treturn detectMiss(scan(entries, models).prev, message, models);\n}\n"]}

package/dist/core/compaction/branch-summarization.d.ts

@@ -0,0 +1,98 @@
+/**
+ * Branch summarization for tree navigation.
+ *
+ * When navigating to a different point in the session tree, this generates
+ * a summary of the branch being left so context isn't lost.
+ */
+import type { AgentMessage, StreamFn } from "@earendil-works/pi-agent-core";
+import type { RetryCallbacks, RetryPolicy } from "@earendil-works/pi-ai";
+import type { Model, Usage } from "@earendil-works/pi-ai/compat";
+import type { ReadonlySessionManager, SessionEntry } from "../session-manager.ts";
+import { type FileOperations } from "./utils.ts";
+export interface BranchSummaryResult {
+    summary?: string;
+    usage?: Usage;
+    readFiles?: string[];
+    modifiedFiles?: string[];
+    aborted?: boolean;
+    error?: string;
+}
+/** Details stored in BranchSummaryEntry.details for file tracking */
+export interface BranchSummaryDetails {
+    readFiles: string[];
+    modifiedFiles: string[];
+}
+export type { FileOperations } from "./utils.ts";
+export interface BranchPreparation {
+    /** Messages extracted for summarization, in chronological order */
+    messages: AgentMessage[];
+    /** File operations extracted from tool calls */
+    fileOps: FileOperations;
+    /** Total estimated tokens in messages */
+    totalTokens: number;
+}
+export interface CollectEntriesResult {
+    /** Entries to summarize, in chronological order */
+    entries: SessionEntry[];
+    /** Common ancestor between old and new position, if any */
+    commonAncestorId: string | null;
+}
+export interface GenerateBranchSummaryOptions {
+    /** Model to use for summarization */
+    model: Model<any>;
+    /** API key for the model */
+    apiKey?: string;
+    /** Request headers for the model */
+    headers?: Record<string, string>;
+    /** Provider-scoped environment values for the model */
+    env?: Record<string, string>;
+    /** Abort signal for cancellation */
+    signal: AbortSignal;
+    /** Optional custom instructions for summarization */
+    customInstructions?: string;
+    /** If true, customInstructions replaces the default prompt instead of being appended */
+    replaceInstructions?: boolean;
+    /** Tokens reserved for prompt + LLM response (default 16384) */
+    reserveTokens?: number;
+    /** Optional session stream function. Used to preserve SDK request behavior without mutating agent state. */
+    streamFn?: StreamFn;
+    /** Retry policy for transient summarization errors. Reuses coding-agent's `settings.retry`. */
+    retry?: RetryPolicy;
+    /** Optional callbacks for retry reporting (e.g. TUI retry indicators). */
+    callbacks?: RetryCallbacks;
+}
+/**
+ * Collect entries that should be summarized when navigating from one position to another.
+ *
+ * Walks from oldLeafId back to the common ancestor with targetId, collecting entries
+ * along the way. Does NOT stop at compaction boundaries - those are included and their
+ * summaries become context.
+ *
+ * @param session - Session manager (read-only access)
+ * @param oldLeafId - Current position (where we're navigating from)
+ * @param targetId - Target position (where we're navigating to)
+ * @returns Entries to summarize and the common ancestor
+ */
+export declare function collectEntriesForBranchSummary(session: ReadonlySessionManager, oldLeafId: string | null, targetId: string): CollectEntriesResult;
+/**
+ * Prepare entries for summarization with token budget.
+ *
+ * Walks entries from NEWEST to OLDEST, adding messages until we hit the token budget.
+ * This ensures we keep the most recent context when the branch is too long.
+ *
+ * Also collects file operations from:
+ * - Tool calls in assistant messages
+ * - Existing branch_summary entries' details (for cumulative tracking)
+ *
+ * @param entries - Entries in chronological order
+ * @param tokenBudget - Maximum tokens to include (0 = no limit)
+ */
+export declare function prepareBranchEntries(entries: SessionEntry[], tokenBudget?: number): BranchPreparation;
+/**
+ * Generate a summary of abandoned branch entries.
+ *
+ * @param entries - Session entries to summarize (chronological order)
+ * @param options - Generation options
+ */
+export declare function generateBranchSummary(entries: SessionEntry[], options: GenerateBranchSummaryOptions): Promise<BranchSummaryResult>;
+//# sourceMappingURL=branch-summarization.d.ts.map

package/dist/core/compaction/branch-summarization.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"branch-summarization.d.ts","sourceRoot":"","sources":["../../../src/core/compaction/branch-summarization.ts"],"names":[],"mappings":"AAAA;;;;;GAKG;AAEH,OAAO,KAAK,EAAE,YAAY,EAAE,QAAQ,EAAE,MAAM,+BAA+B,CAAC;AAC5E,OAAO,KAAK,EAAE,cAAc,EAAE,WAAW,EAAE,MAAM,uBAAuB,CAAC;AAEzE,OAAO,KAAK,EAAE,KAAK,EAAuB,KAAK,EAAE,MAAM,8BAA8B,CAAC;AAOtF,OAAO,KAAK,EAAE,sBAAsB,EAAE,YAAY,EAAE,MAAM,uBAAuB,CAAC;AAElF,OAAO,EAIN,KAAK,cAAc,EAInB,MAAM,YAAY,CAAC;AAMpB,MAAM,WAAW,mBAAmB;IACnC,OAAO,CAAC,EAAE,MAAM,CAAC;IACjB,KAAK,CAAC,EAAE,KAAK,CAAC;IACd,SAAS,CAAC,EAAE,MAAM,EAAE,CAAC;IACrB,aAAa,CAAC,EAAE,MAAM,EAAE,CAAC;IACzB,OAAO,CAAC,EAAE,OAAO,CAAC;IAClB,KAAK,CAAC,EAAE,MAAM,CAAC;CACf;AAED,qEAAqE;AACrE,MAAM,WAAW,oBAAoB;IACpC,SAAS,EAAE,MAAM,EAAE,CAAC;IACpB,aAAa,EAAE,MAAM,EAAE,CAAC;CACxB;AAED,YAAY,EAAE,cAAc,EAAE,MAAM,YAAY,CAAC;AAEjD,MAAM,WAAW,iBAAiB;IACjC,mEAAmE;IACnE,QAAQ,EAAE,YAAY,EAAE,CAAC;IACzB,gDAAgD;IAChD,OAAO,EAAE,cAAc,CAAC;IACxB,yCAAyC;IACzC,WAAW,EAAE,MAAM,CAAC;CACpB;AAED,MAAM,WAAW,oBAAoB;IACpC,mDAAmD;IACnD,OAAO,EAAE,YAAY,EAAE,CAAC;IACxB,2DAA2D;IAC3D,gBAAgB,EAAE,MAAM,GAAG,IAAI,CAAC;CAChC;AAED,MAAM,WAAW,4BAA4B;IAC5C,qCAAqC;IACrC,KAAK,EAAE,KAAK,CAAC,GAAG,CAAC,CAAC;IAClB,4BAA4B;IAC5B,MAAM,CAAC,EAAE,MAAM,CAAC;IAChB,oCAAoC;IACpC,OAAO,CAAC,EAAE,MAAM,CAAC,MAAM,EAAE,MAAM,CAAC,CAAC;IACjC,uDAAuD;IACvD,GAAG,CAAC,EAAE,MAAM,CAAC,MAAM,EAAE,MAAM,CAAC,CAAC;IAC7B,oCAAoC;IACpC,MAAM,EAAE,WAAW,CAAC;IACpB,qDAAqD;IACrD,kBAAkB,CAAC,EAAE,MAAM,CAAC;IAC5B,wFAAwF;IACxF,mBAAmB,CAAC,EAAE,OAAO,CAAC;IAC9B,gEAAgE;IAChE,aAAa,CAAC,EAAE,MAAM,CAAC;IACvB,4GAA4G;IAC5G,QAAQ,CAAC,EAAE,QAAQ,CAAC;IACpB,+FAA+F;IAC/F,KAAK,CAAC,EAAE,WAAW,CAAC;IACpB,0EAA0E;IAC1E,SAAS,CAAC,EAAE,cAAc,CAAC;CAC3B;AAMD;;;;;;;;;;;GAWG;AACH,wBAAgB,8BAA8B,CAC7C,OAAO,EAAE,sBAAsB,EAC/B,SAAS,EAAE,MAAM,GAAG,IAAI,EACxB,QAAQ,EAAE,MAAM,GACd,oBAAoB,CAkCtB;AAoCD;;;;;;;;;;;;GAYG;AACH,wBAAgB,oBAAoB,CAAC,OAAO,EAAE,YAAY,EAAE,EAAE,WAAW,GAAE,MAAU,GAAG,iBAAiB,CAoDxG;AAwCD;;;;;GAKG;AACH,wBAAsB,qBAAqB,CAC1C,OAAO,EAAE,YAAY,EAAE,EACvB,OAAO,EAAE,4BAA4B,GACnC,OAAO,CAAC,mBAAmB,CAAC,CAgF9B","sourcesContent":["/**\n * Branch summarization for tree navigation.\n *\n * When navigating to a different point in the session tree, this generates\n * a summary of the branch being left so context isn't lost.\n */\n\nimport type { AgentMessage, StreamFn } from \"@earendil-works/pi-agent-core\";\nimport type { RetryCallbacks, RetryPolicy } from \"@earendil-works/pi-ai\";\nimport { contentText } from \"@earendil-works/pi-ai\";\nimport type { Model, SimpleStreamOptions, Usage } from \"@earendil-works/pi-ai/compat\";\nimport {\n\tconvertToLlm,\n\tcreateBranchSummaryMessage,\n\tcreateCompactionSummaryMessage,\n\tcreateCustomMessage,\n} from \"../messages.ts\";\nimport type { ReadonlySessionManager, SessionEntry } from \"../session-manager.ts\";\nimport { completeSummarization, estimateTokens } from \"./compaction.ts\";\nimport {\n\tcomputeFileLists,\n\tcreateFileOps,\n\textractFileOpsFromMessage,\n\ttype FileOperations,\n\tformatFileOperations,\n\tSUMMARIZATION_SYSTEM_PROMPT,\n\tserializeConversation,\n} from \"./utils.ts\";\n\n// ============================================================================\n// Types\n// ============================================================================\n\nexport interface BranchSummaryResult {\n\tsummary?: string;\n\tusage?: Usage;\n\treadFiles?: string[];\n\tmodifiedFiles?: string[];\n\taborted?: boolean;\n\terror?: string;\n}\n\n/** Details stored in BranchSummaryEntry.details for file tracking */\nexport interface BranchSummaryDetails {\n\treadFiles: string[];\n\tmodifiedFiles: string[];\n}\n\nexport type { FileOperations } from \"./utils.ts\";\n\nexport interface BranchPreparation {\n\t/** Messages extracted for summarization, in chronological order */\n\tmessages: AgentMessage[];\n\t/** File operations extracted from tool calls */\n\tfileOps: FileOperations;\n\t/** Total estimated tokens in messages */\n\ttotalTokens: number;\n}\n\nexport interface CollectEntriesResult {\n\t/** Entries to summarize, in chronological order */\n\tentries: SessionEntry[];\n\t/** Common ancestor between old and new position, if any */\n\tcommonAncestorId: string | null;\n}\n\nexport interface GenerateBranchSummaryOptions {\n\t/** Model to use for summarization */\n\tmodel: Model<any>;\n\t/** API key for the model */\n\tapiKey?: string;\n\t/** Request headers for the model */\n\theaders?: Record<string, string>;\n\t/** Provider-scoped environment values for the model */\n\tenv?: Record<string, string>;\n\t/** Abort signal for cancellation */\n\tsignal: AbortSignal;\n\t/** Optional custom instructions for summarization */\n\tcustomInstructions?: string;\n\t/** If true, customInstructions replaces the default prompt instead of being appended */\n\treplaceInstructions?: boolean;\n\t/** Tokens reserved for prompt + LLM response (default 16384) */\n\treserveTokens?: number;\n\t/** Optional session stream function. Used to preserve SDK request behavior without mutating agent state. */\n\tstreamFn?: StreamFn;\n\t/** Retry policy for transient summarization errors. Reuses coding-agent's `settings.retry`. */\n\tretry?: RetryPolicy;\n\t/** Optional callbacks for retry reporting (e.g. TUI retry indicators). */\n\tcallbacks?: RetryCallbacks;\n}\n\n// ============================================================================\n// Entry Collection\n// ============================================================================\n\n/**\n * Collect entries that should be summarized when navigating from one position to another.\n *\n * Walks from oldLeafId back to the common ancestor with targetId, collecting entries\n * along the way. Does NOT stop at compaction boundaries - those are included and their\n * summaries become context.\n *\n * @param session - Session manager (read-only access)\n * @param oldLeafId - Current position (where we're navigating from)\n * @param targetId - Target position (where we're navigating to)\n * @returns Entries to summarize and the common ancestor\n */\nexport function collectEntriesForBranchSummary(\n\tsession: ReadonlySessionManager,\n\toldLeafId: string | null,\n\ttargetId: string,\n): CollectEntriesResult {\n\t// If no old position, nothing to summarize\n\tif (!oldLeafId) {\n\t\treturn { entries: [], commonAncestorId: null };\n\t}\n\n\t// Find common ancestor (deepest node that's on both paths)\n\tconst oldPath = new Set(session.getBranch(oldLeafId).map((e) => e.id));\n\tconst targetPath = session.getBranch(targetId);\n\n\t// targetPath is root-first, so iterate backwards to find deepest common ancestor\n\tlet commonAncestorId: string | null = null;\n\tfor (let i = targetPath.length - 1; i >= 0; i--) {\n\t\tif (oldPath.has(targetPath[i].id)) {\n\t\t\tcommonAncestorId = targetPath[i].id;\n\t\t\tbreak;\n\t\t}\n\t}\n\n\t// Collect entries from old leaf back to common ancestor\n\tconst entries: SessionEntry[] = [];\n\tlet current: string | null = oldLeafId;\n\n\twhile (current && current !== commonAncestorId) {\n\t\tconst entry = session.getEntry(current);\n\t\tif (!entry) break;\n\t\tentries.push(entry);\n\t\tcurrent = entry.parentId;\n\t}\n\n\t// Reverse to get chronological order\n\tentries.reverse();\n\n\treturn { entries, commonAncestorId };\n}\n\n// ============================================================================\n// Entry to Message Conversion\n// ============================================================================\n\n/**\n * Extract AgentMessage from a session entry.\n * Similar to getMessageFromEntry in compaction.ts but also handles compaction entries.\n */\nfunction getMessageFromEntry(entry: SessionEntry): AgentMessage | undefined {\n\tswitch (entry.type) {\n\t\tcase \"message\":\n\t\t\t// Skip tool results - context is in assistant's tool call\n\t\t\tif (entry.message.role === \"toolResult\") return undefined;\n\t\t\treturn entry.message;\n\n\t\tcase \"custom_message\":\n\t\t\treturn createCustomMessage(entry.customType, entry.content, entry.display, entry.details, entry.timestamp);\n\n\t\tcase \"branch_summary\":\n\t\t\treturn createBranchSummaryMessage(entry.summary, entry.fromId, entry.timestamp);\n\n\t\tcase \"compaction\":\n\t\t\treturn createCompactionSummaryMessage(entry.summary, entry.tokensBefore, entry.timestamp);\n\n\t\t// These don't contribute to conversation content\n\t\tcase \"thinking_level_change\":\n\t\tcase \"model_change\":\n\t\tcase \"custom\":\n\t\tcase \"label\":\n\t\tcase \"session_info\":\n\t\t\treturn undefined;\n\t}\n}\n\n/**\n * Prepare entries for summarization with token budget.\n *\n * Walks entries from NEWEST to OLDEST, adding messages until we hit the token budget.\n * This ensures we keep the most recent context when the branch is too long.\n *\n * Also collects file operations from:\n * - Tool calls in assistant messages\n * - Existing branch_summary entries' details (for cumulative tracking)\n *\n * @param entries - Entries in chronological order\n * @param tokenBudget - Maximum tokens to include (0 = no limit)\n */\nexport function prepareBranchEntries(entries: SessionEntry[], tokenBudget: number = 0): BranchPreparation {\n\tconst messages: AgentMessage[] = [];\n\tconst fileOps = createFileOps();\n\tlet totalTokens = 0;\n\n\t// First pass: collect file ops from ALL entries (even if they don't fit in token budget)\n\t// This ensures we capture cumulative file tracking from nested branch summaries\n\t// Only extract from pi-generated summaries (fromHook !== true), not extension-generated ones\n\tfor (const entry of entries) {\n\t\tif (entry.type === \"branch_summary\" && !entry.fromHook && entry.details) {\n\t\t\tconst details = entry.details as BranchSummaryDetails;\n\t\t\tif (Array.isArray(details.readFiles)) {\n\t\t\t\tfor (const f of details.readFiles) fileOps.read.add(f);\n\t\t\t}\n\t\t\tif (Array.isArray(details.modifiedFiles)) {\n\t\t\t\t// Modified files go into both edited and written for proper deduplication\n\t\t\t\tfor (const f of details.modifiedFiles) {\n\t\t\t\t\tfileOps.edited.add(f);\n\t\t\t\t}\n\t\t\t}\n\t\t}\n\t}\n\n\t// Second pass: walk from newest to oldest, adding messages until token budget\n\tfor (let i = entries.length - 1; i >= 0; i--) {\n\t\tconst entry = entries[i];\n\t\tconst message = getMessageFromEntry(entry);\n\t\tif (!message) continue;\n\n\t\t// Extract file ops from assistant messages (tool calls)\n\t\textractFileOpsFromMessage(message, fileOps);\n\n\t\tconst tokens = estimateTokens(message);\n\n\t\t// Check budget before adding\n\t\tif (tokenBudget > 0 && totalTokens + tokens > tokenBudget) {\n\t\t\t// If this is a summary entry, try to fit it anyway as it's important context\n\t\t\tif (entry.type === \"compaction\" || entry.type === \"branch_summary\") {\n\t\t\t\tif (totalTokens < tokenBudget * 0.9) {\n\t\t\t\t\tmessages.unshift(message);\n\t\t\t\t\ttotalTokens += tokens;\n\t\t\t\t}\n\t\t\t}\n\t\t\t// Stop - we've hit the budget\n\t\t\tbreak;\n\t\t}\n\n\t\tmessages.unshift(message);\n\t\ttotalTokens += tokens;\n\t}\n\n\treturn { messages, fileOps, totalTokens };\n}\n\n// ============================================================================\n// Summary Generation\n// ============================================================================\n\nconst BRANCH_SUMMARY_PREAMBLE = `The user explored a different conversation branch before returning here.\nSummary of that exploration:\n\n`;\n\nconst BRANCH_SUMMARY_PROMPT = `Create a structured summary of this conversation branch for context when returning later.\n\nUse this EXACT format:\n\n## Goal\n[What was the user trying to accomplish in this branch?]\n\n## Constraints & Preferences\n- [Any constraints, preferences, or requirements mentioned]\n- [Or \"(none)\" if none were mentioned]\n\n## Progress\n### Done\n- [x] [Completed tasks/changes]\n\n### In Progress\n- [ ] [Work that was started but not finished]\n\n### Blocked\n- [Issues preventing progress, if any]\n\n## Key Decisions\n- **[Decision]**: [Brief rationale]\n\n## Next Steps\n1. [What should happen next to continue this work]\n\nKeep each section concise. Preserve exact file paths, function names, and error messages.`;\n\n/**\n * Generate a summary of abandoned branch entries.\n *\n * @param entries - Session entries to summarize (chronological order)\n * @param options - Generation options\n */\nexport async function generateBranchSummary(\n\tentries: SessionEntry[],\n\toptions: GenerateBranchSummaryOptions,\n): Promise<BranchSummaryResult> {\n\tconst {\n\t\tmodel,\n\t\tapiKey,\n\t\theaders,\n\t\tenv,\n\t\tsignal,\n\t\tcustomInstructions,\n\t\treplaceInstructions,\n\t\treserveTokens = 16384,\n\t\tstreamFn,\n\t\tretry,\n\t\tcallbacks,\n\t} = options;\n\n\t// Token budget = context window minus reserved space for prompt + response\n\tconst contextWindow = model.contextWindow || 128000;\n\tconst tokenBudget = contextWindow - reserveTokens;\n\n\tconst { messages, fileOps } = prepareBranchEntries(entries, tokenBudget);\n\n\tif (messages.length === 0) {\n\t\treturn { summary: \"No content to summarize\" };\n\t}\n\n\t// Transform to LLM-compatible messages, then serialize to text\n\t// Serialization prevents the model from treating it as a conversation to continue\n\tconst llmMessages = convertToLlm(messages);\n\tconst conversationText = serializeConversation(llmMessages);\n\n\t// Build prompt\n\tlet instructions: string;\n\tif (replaceInstructions && customInstructions) {\n\t\tinstructions = customInstructions;\n\t} else if (customInstructions) {\n\t\tinstructions = `${BRANCH_SUMMARY_PROMPT}\\n\\nAdditional focus: ${customInstructions}`;\n\t} else {\n\t\tinstructions = BRANCH_SUMMARY_PROMPT;\n\t}\n\tconst promptText = `<conversation>\\n${conversationText}\\n</conversation>\\n\\n${instructions}`;\n\n\tconst summarizationMessages = [\n\t\t{\n\t\t\trole: \"user\" as const,\n\t\t\tcontent: [{ type: \"text\" as const, text: promptText }],\n\t\t\ttimestamp: Date.now(),\n\t\t},\n\t];\n\n\t// Call LLM for summarization. Prefer the session stream function so SDK\n\t// request behavior (timeouts, retries, attribution headers) stays consistent\n\t// without running through agent state/events. Retried via completeSummarization\n\t// so transient stream drops reuse the configured retry policy.\n\tconst context = { systemPrompt: SUMMARIZATION_SYSTEM_PROMPT, messages: summarizationMessages };\n\tconst requestOptions: SimpleStreamOptions = { apiKey, headers, env, signal, maxTokens: 2048 };\n\tconst response = await completeSummarization(model, context, requestOptions, streamFn, retry, callbacks);\n\n\t// Check if aborted or errored\n\tif (response.stopReason === \"aborted\") {\n\t\treturn { aborted: true };\n\t}\n\tif (response.stopReason === \"error\") {\n\t\treturn { error: response.errorMessage || \"Summarization failed\" };\n\t}\n\n\tlet summary = contentText(response.content);\n\n\t// Prepend preamble to provide context about the branch summary\n\tsummary = BRANCH_SUMMARY_PREAMBLE + summary;\n\n\t// Compute file lists and append to summary\n\tconst { readFiles, modifiedFiles } = computeFileLists(fileOps);\n\tsummary += formatFileOperations(readFiles, modifiedFiles);\n\n\treturn {\n\t\tsummary: summary || \"No summary generated\",\n\t\tusage: response.usage,\n\t\treadFiles,\n\t\tmodifiedFiles,\n\t};\n}\n"]}