open-multi-agent-kit 0.90.7 → 0.90.9

package/dist/core/agent-tool-settings.d.ts

@@ -0,0 +1,35 @@
+import type { AgentOptions } from "omk-agent-core";
+import type { SettingsManager } from "./settings-manager.ts";
+/** §6.3 timeout categories. `undefined` category falls through to the global setting. */
+export type ToolTimeoutCategory = "read" | "write" | "mcp" | "process" | "browser";
+/** §6.3 category default timeouts (ms): read/list/search 30s, write/edit 60s, MCP 120s, bash/process 300s, browser/computer-use 180s. */
+export declare const TOOL_CATEGORY_TIMEOUTS: Readonly<Record<ToolTimeoutCategory, number>>;
+/**
+ * Classify a tool name into its §6.3 timeout category. Conservative on
+ * purpose: unknown names return `undefined` and keep the global timeout.
+ * A browser/computer-use match wins over the generic MCP category because
+ * browser automation is the slower bound.
+ */
+export declare function resolveToolTimeoutCategory(toolName: string): ToolTimeoutCategory | undefined;
+/**
+ * Fill §6.3 category default timeouts for every active tool name that has no
+ * explicit per-name entry. Explicit entries (user settings and built-in
+ * defaults) always win and are all preserved; names without a category are
+ * omitted so they fall through to the global `toolTimeoutMs` at runtime.
+ */
+export declare function applyCategoryTimeoutDefaults(toolNames: Iterable<string>, explicit: Readonly<Record<string, number>>): Record<string, number>;
+type ToolScheduler = NonNullable<AgentOptions["toolScheduler"]>;
+export interface ResolvedAgentToolSettings {
+    readonly toolScheduler: ToolScheduler;
+    readonly maxToolConcurrency: number | undefined;
+    readonly strictExtensionClaims: boolean;
+    readonly toolTimeoutMs: number;
+    readonly toolTimeouts: Record<string, number>;
+}
+export declare class AgentToolSettingsError extends Error {
+    readonly settingName: string;
+    constructor(settingName: string);
+}
+export declare function resolveAgentToolSettings(settingsManager: SettingsManager, env?: NodeJS.ProcessEnv | Readonly<Record<string, string | undefined>>): ResolvedAgentToolSettings;
+export {};
+//# sourceMappingURL=agent-tool-settings.d.ts.map

package/dist/core/agent-tool-settings.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"agent-tool-settings.d.ts","sourceRoot":"","sources":["../../src/core/agent-tool-settings.ts"],"names":[],"mappings":"AAAA,OAAO,KAAK,EAAE,YAAY,EAAE,MAAM,gBAAgB,CAAC;AACnD,OAAO,KAAK,EAAwB,eAAe,EAAE,MAAM,uBAAuB,CAAC;AAiBnF,0FAAyF;AACzF,MAAM,MAAM,mBAAmB,GAAG,MAAM,GAAG,OAAO,GAAG,KAAK,GAAG,SAAS,GAAG,SAAS,CAAC;AAEnF,0IAAyI;AACzI,eAAO,MAAM,sBAAsB,EAAE,QAAQ,CAAC,MAAM,CAAC,mBAAmB,EAAE,MAAM,CAAC,CAM/E,CAAC;AASH;;;;;GAKG;AACH,wBAAgB,0BAA0B,CAAC,QAAQ,EAAE,MAAM,GAAG,mBAAmB,GAAG,SAAS,CAO5F;AAED;;;;;GAKG;AACH,wBAAgB,4BAA4B,CAC3C,SAAS,EAAE,QAAQ,CAAC,MAAM,CAAC,EAC3B,QAAQ,EAAE,QAAQ,CAAC,MAAM,CAAC,MAAM,EAAE,MAAM,CAAC,CAAC,GACxC,MAAM,CAAC,MAAM,EAAE,MAAM,CAAC,CAaxB;AAED,KAAK,aAAa,GAAG,WAAW,CAAC,YAAY,CAAC,eAAe,CAAC,CAAC,CAAC;AAEhE,MAAM,WAAW,yBAAyB;IACzC,QAAQ,CAAC,aAAa,EAAE,aAAa,CAAC;IACtC,QAAQ,CAAC,kBAAkB,EAAE,MAAM,GAAG,SAAS,CAAC;IAChD,QAAQ,CAAC,qBAAqB,EAAE,OAAO,CAAC;IACxC,QAAQ,CAAC,aAAa,EAAE,MAAM,CAAC;IAC/B,QAAQ,CAAC,YAAY,EAAE,MAAM,CAAC,MAAM,EAAE,MAAM,CAAC,CAAC;CAC9C;AAED,qBAAa,sBAAuB,SAAQ,KAAK;IAChD,QAAQ,CAAC,WAAW,EAAE,MAAM,CAAC;IAE7B,YAAY,WAAW,EAAE,MAAM,EAI9B;CACD;AAiFD,wBAAgB,wBAAwB,CACvC,eAAe,EAAE,eAAe,EAChC,GAAG,GAAE,MAAM,CAAC,UAAU,GAAG,QAAQ,CAAC,MAAM,CAAC,MAAM,EAAE,MAAM,GAAG,SAAS,CAAC,CAAe,GACjF,yBAAyB,CAc3B","sourcesContent":["import type { AgentOptions } from \"omk-agent-core\";\nimport type { AgentRuntimeSettings, SettingsManager } from \"./settings-manager.ts\";\n\nconst DEFAULT_TOOL_SCHEDULER = \"dag-v2\";\nconst DEFAULT_MAX_TOOL_CONCURRENCY = 4;\nconst DEFAULT_TOOL_TIMEOUT_MS = 0;\nconst MAX_TOOL_TIMEOUT_MS = 2_147_483_647;\n\nconst DEFAULT_BUILTIN_TOOL_TIMEOUTS: Readonly<Record<string, number>> = {\n\tbash: 300_000,\n\tedit: 60_000,\n\tfind: 30_000,\n\tgrep: 30_000,\n\tls: 30_000,\n\tread: 30_000,\n\twrite: 60_000,\n};\n\n/** §6.3 timeout categories. `undefined` category falls through to the global setting. */\nexport type ToolTimeoutCategory = \"read\" | \"write\" | \"mcp\" | \"process\" | \"browser\";\n\n/** §6.3 category default timeouts (ms): read/list/search 30s, write/edit 60s, MCP 120s, bash/process 300s, browser/computer-use 180s. */\nexport const TOOL_CATEGORY_TIMEOUTS: Readonly<Record<ToolTimeoutCategory, number>> = Object.freeze({\n\tread: 30_000,\n\twrite: 60_000,\n\tmcp: 120_000,\n\tprocess: 300_000,\n\tbrowser: 180_000,\n});\n\nconst READ_CATEGORY_TOOLS = new Set([\"read\", \"ls\", \"find\", \"grep\", \"glob\", \"list\", \"search\"]);\nconst WRITE_CATEGORY_TOOLS = new Set([\"write\", \"edit\"]);\nconst PROCESS_CATEGORY_TOOLS = new Set([\"bash\"]);\n// Matches direct browser/computer-use tools and browser tools bridged through\n// MCP servers (e.g. `mcp__playwright__browser_click`).\nconst BROWSER_TOOL_PATTERN = /(^|__)(browser_|computer[-_]?use$|computer_)/;\n\n/**\n * Classify a tool name into its §6.3 timeout category. Conservative on\n * purpose: unknown names return `undefined` and keep the global timeout.\n * A browser/computer-use match wins over the generic MCP category because\n * browser automation is the slower bound.\n */\nexport function resolveToolTimeoutCategory(toolName: string): ToolTimeoutCategory | undefined {\n\tif (READ_CATEGORY_TOOLS.has(toolName)) return \"read\";\n\tif (WRITE_CATEGORY_TOOLS.has(toolName)) return \"write\";\n\tif (PROCESS_CATEGORY_TOOLS.has(toolName)) return \"process\";\n\tif (BROWSER_TOOL_PATTERN.test(toolName)) return \"browser\";\n\tif (toolName.startsWith(\"mcp__\")) return \"mcp\";\n\treturn undefined;\n}\n\n/**\n * Fill §6.3 category default timeouts for every active tool name that has no\n * explicit per-name entry. Explicit entries (user settings and built-in\n * defaults) always win and are all preserved; names without a category are\n * omitted so they fall through to the global `toolTimeoutMs` at runtime.\n */\nexport function applyCategoryTimeoutDefaults(\n\ttoolNames: Iterable<string>,\n\texplicit: Readonly<Record<string, number>>,\n): Record<string, number> {\n\tconst merged: Record<string, number> = Object.create(null);\n\tfor (const toolName of toolNames) {\n\t\tif (Object.hasOwn(explicit, toolName)) continue;\n\t\tconst category = resolveToolTimeoutCategory(toolName);\n\t\tif (category !== undefined) {\n\t\t\tmerged[toolName] = TOOL_CATEGORY_TIMEOUTS[category];\n\t\t}\n\t}\n\tfor (const [toolName, timeoutMs] of Object.entries(explicit)) {\n\t\tmerged[toolName] = timeoutMs;\n\t}\n\treturn merged;\n}\n\ntype ToolScheduler = NonNullable<AgentOptions[\"toolScheduler\"]>;\n\nexport interface ResolvedAgentToolSettings {\n\treadonly toolScheduler: ToolScheduler;\n\treadonly maxToolConcurrency: number | undefined;\n\treadonly strictExtensionClaims: boolean;\n\treadonly toolTimeoutMs: number;\n\treadonly toolTimeouts: Record<string, number>;\n}\n\nexport class AgentToolSettingsError extends Error {\n\treadonly settingName: string;\n\n\tconstructor(settingName: string) {\n\t\tsuper(`Invalid ${settingName} setting`);\n\t\tthis.name = \"AgentToolSettingsError\";\n\t\tthis.settingName = settingName;\n\t}\n}\n\nfunction parseScheduler(value: unknown, settingName: string): ToolScheduler | undefined {\n\tif (value === undefined) {\n\t\treturn undefined;\n\t}\n\tif (value === \"waves-v1\" || value === \"dag-v2\") {\n\t\treturn value;\n\t}\n\tthrow new AgentToolSettingsError(settingName);\n}\n\nfunction parseTimeout(value: unknown, settingName: string, defaultValue: number): number {\n\tif (value === undefined) {\n\t\treturn defaultValue;\n\t}\n\tif (typeof value !== \"number\" || !Number.isInteger(value) || value < 0 || value > MAX_TOOL_TIMEOUT_MS) {\n\t\tthrow new AgentToolSettingsError(settingName);\n\t}\n\treturn value;\n}\n\nfunction parseConcurrency(value: unknown): number | undefined {\n\tif (value === undefined) {\n\t\treturn DEFAULT_MAX_TOOL_CONCURRENCY;\n\t}\n\tif (typeof value !== \"number\" || !Number.isInteger(value) || value < 0) {\n\t\tthrow new AgentToolSettingsError(\"agent.maxToolConcurrency\");\n\t}\n\treturn value === 0 ? undefined : value;\n}\n\nfunction parseAgentSettings(value: AgentRuntimeSettings | undefined, settingName: string): AgentRuntimeSettings {\n\tif (value === undefined) {\n\t\treturn {};\n\t}\n\tif (typeof value !== \"object\" || value === null || Array.isArray(value)) {\n\t\tthrow new AgentToolSettingsError(settingName);\n\t}\n\treturn value;\n}\n\nfunction parseToolTimeoutEntries(\n\tvalue: AgentRuntimeSettings[\"toolTimeouts\"],\n\tsettingName: string,\n): readonly (readonly [string, number])[] {\n\tif (value === undefined) {\n\t\treturn [];\n\t}\n\tif (typeof value !== \"object\" || value === null || Array.isArray(value)) {\n\t\tthrow new AgentToolSettingsError(settingName);\n\t}\n\n\treturn Object.entries(value).map(([toolName, timeoutMs]) => {\n\t\tif (toolName.length === 0) {\n\t\t\tthrow new AgentToolSettingsError(settingName);\n\t\t}\n\t\treturn [toolName, parseTimeout(timeoutMs, `${settingName}.${toolName}`, DEFAULT_TOOL_TIMEOUT_MS)] as const;\n\t});\n}\n\nfunction resolveToolTimeouts(\n\tglobalSettings: AgentRuntimeSettings,\n\tprojectSettings: AgentRuntimeSettings,\n\teffectiveSettings: AgentRuntimeSettings,\n): Record<string, number> {\n\tconst timeouts: Record<string, number> = Object.create(null);\n\tconst entryGroups = [\n\t\tObject.entries(DEFAULT_BUILTIN_TOOL_TIMEOUTS),\n\t\tparseToolTimeoutEntries(globalSettings.toolTimeouts, \"agent.toolTimeouts\"),\n\t\tparseToolTimeoutEntries(projectSettings.toolTimeouts, \"agent.toolTimeouts\"),\n\t\tparseToolTimeoutEntries(effectiveSettings.toolTimeouts, \"agent.toolTimeouts\"),\n\t];\n\tfor (const entries of entryGroups) {\n\t\tfor (const [toolName, timeoutMs] of entries) {\n\t\t\ttimeouts[toolName] = timeoutMs;\n\t\t}\n\t}\n\treturn timeouts;\n}\n\nexport function resolveAgentToolSettings(\n\tsettingsManager: SettingsManager,\n\tenv: NodeJS.ProcessEnv | Readonly<Record<string, string | undefined>> = process.env,\n): ResolvedAgentToolSettings {\n\tconst globalSettings = parseAgentSettings(settingsManager.getGlobalSettings().agent, \"agent\");\n\tconst projectSettings = parseAgentSettings(settingsManager.getProjectSettings().agent, \"agent\");\n\tconst settings = parseAgentSettings(settingsManager.getAgentRuntimeSettings(), \"agent\");\n\tconst envScheduler = parseScheduler(env.OMK_TOOL_SCHEDULER, \"OMK_TOOL_SCHEDULER\");\n\tconst settingsScheduler = parseScheduler(settings.toolScheduler, \"agent.toolScheduler\");\n\n\treturn {\n\t\ttoolScheduler: envScheduler ?? settingsScheduler ?? DEFAULT_TOOL_SCHEDULER,\n\t\tmaxToolConcurrency: parseConcurrency(settings.maxToolConcurrency),\n\t\tstrictExtensionClaims: true,\n\t\ttoolTimeoutMs: parseTimeout(settings.toolTimeoutMs, \"agent.toolTimeoutMs\", DEFAULT_TOOL_TIMEOUT_MS),\n\t\ttoolTimeouts: resolveToolTimeouts(globalSettings, projectSettings, settings),\n\t};\n}\n"]}

package/dist/core/agent-tool-settings.js

@@ -0,0 +1,155 @@
+const DEFAULT_TOOL_SCHEDULER = "dag-v2";
+const DEFAULT_MAX_TOOL_CONCURRENCY = 4;
+const DEFAULT_TOOL_TIMEOUT_MS = 0;
+const MAX_TOOL_TIMEOUT_MS = 2_147_483_647;
+const DEFAULT_BUILTIN_TOOL_TIMEOUTS = {
+    bash: 300_000,
+    edit: 60_000,
+    find: 30_000,
+    grep: 30_000,
+    ls: 30_000,
+    read: 30_000,
+    write: 60_000,
+};
+/** §6.3 category default timeouts (ms): read/list/search 30s, write/edit 60s, MCP 120s, bash/process 300s, browser/computer-use 180s. */
+export const TOOL_CATEGORY_TIMEOUTS = Object.freeze({
+    read: 30_000,
+    write: 60_000,
+    mcp: 120_000,
+    process: 300_000,
+    browser: 180_000,
+});
+const READ_CATEGORY_TOOLS = new Set(["read", "ls", "find", "grep", "glob", "list", "search"]);
+const WRITE_CATEGORY_TOOLS = new Set(["write", "edit"]);
+const PROCESS_CATEGORY_TOOLS = new Set(["bash"]);
+// Matches direct browser/computer-use tools and browser tools bridged through
+// MCP servers (e.g. `mcp__playwright__browser_click`).
+const BROWSER_TOOL_PATTERN = /(^|__)(browser_|computer[-_]?use$|computer_)/;
+/**
+ * Classify a tool name into its §6.3 timeout category. Conservative on
+ * purpose: unknown names return `undefined` and keep the global timeout.
+ * A browser/computer-use match wins over the generic MCP category because
+ * browser automation is the slower bound.
+ */
+export function resolveToolTimeoutCategory(toolName) {
+    if (READ_CATEGORY_TOOLS.has(toolName))
+        return "read";
+    if (WRITE_CATEGORY_TOOLS.has(toolName))
+        return "write";
+    if (PROCESS_CATEGORY_TOOLS.has(toolName))
+        return "process";
+    if (BROWSER_TOOL_PATTERN.test(toolName))
+        return "browser";
+    if (toolName.startsWith("mcp__"))
+        return "mcp";
+    return undefined;
+}
+/**
+ * Fill §6.3 category default timeouts for every active tool name that has no
+ * explicit per-name entry. Explicit entries (user settings and built-in
+ * defaults) always win and are all preserved; names without a category are
+ * omitted so they fall through to the global `toolTimeoutMs` at runtime.
+ */
+export function applyCategoryTimeoutDefaults(toolNames, explicit) {
+    const merged = Object.create(null);
+    for (const toolName of toolNames) {
+        if (Object.hasOwn(explicit, toolName))
+            continue;
+        const category = resolveToolTimeoutCategory(toolName);
+        if (category !== undefined) {
+            merged[toolName] = TOOL_CATEGORY_TIMEOUTS[category];
+        }
+    }
+    for (const [toolName, timeoutMs] of Object.entries(explicit)) {
+        merged[toolName] = timeoutMs;
+    }
+    return merged;
+}
+export class AgentToolSettingsError extends Error {
+    settingName;
+    constructor(settingName) {
+        super(`Invalid ${settingName} setting`);
+        this.name = "AgentToolSettingsError";
+        this.settingName = settingName;
+    }
+}
+function parseScheduler(value, settingName) {
+    if (value === undefined) {
+        return undefined;
+    }
+    if (value === "waves-v1" || value === "dag-v2") {
+        return value;
+    }
+    throw new AgentToolSettingsError(settingName);
+}
+function parseTimeout(value, settingName, defaultValue) {
+    if (value === undefined) {
+        return defaultValue;
+    }
+    if (typeof value !== "number" || !Number.isInteger(value) || value < 0 || value > MAX_TOOL_TIMEOUT_MS) {
+        throw new AgentToolSettingsError(settingName);
+    }
+    return value;
+}
+function parseConcurrency(value) {
+    if (value === undefined) {
+        return DEFAULT_MAX_TOOL_CONCURRENCY;
+    }
+    if (typeof value !== "number" || !Number.isInteger(value) || value < 0) {
+        throw new AgentToolSettingsError("agent.maxToolConcurrency");
+    }
+    return value === 0 ? undefined : value;
+}
+function parseAgentSettings(value, settingName) {
+    if (value === undefined) {
+        return {};
+    }
+    if (typeof value !== "object" || value === null || Array.isArray(value)) {
+        throw new AgentToolSettingsError(settingName);
+    }
+    return value;
+}
+function parseToolTimeoutEntries(value, settingName) {
+    if (value === undefined) {
+        return [];
+    }
+    if (typeof value !== "object" || value === null || Array.isArray(value)) {
+        throw new AgentToolSettingsError(settingName);
+    }
+    return Object.entries(value).map(([toolName, timeoutMs]) => {
+        if (toolName.length === 0) {
+            throw new AgentToolSettingsError(settingName);
+        }
+        return [toolName, parseTimeout(timeoutMs, `${settingName}.${toolName}`, DEFAULT_TOOL_TIMEOUT_MS)];
+    });
+}
+function resolveToolTimeouts(globalSettings, projectSettings, effectiveSettings) {
+    const timeouts = Object.create(null);
+    const entryGroups = [
+        Object.entries(DEFAULT_BUILTIN_TOOL_TIMEOUTS),
+        parseToolTimeoutEntries(globalSettings.toolTimeouts, "agent.toolTimeouts"),
+        parseToolTimeoutEntries(projectSettings.toolTimeouts, "agent.toolTimeouts"),
+        parseToolTimeoutEntries(effectiveSettings.toolTimeouts, "agent.toolTimeouts"),
+    ];
+    for (const entries of entryGroups) {
+        for (const [toolName, timeoutMs] of entries) {
+            timeouts[toolName] = timeoutMs;
+        }
+    }
+    return timeouts;
+}
+export function resolveAgentToolSettings(settingsManager, env = process.env) {
+    const globalSettings = parseAgentSettings(settingsManager.getGlobalSettings().agent, "agent");
+    const projectSettings = parseAgentSettings(settingsManager.getProjectSettings().agent, "agent");
+    const settings = parseAgentSettings(settingsManager.getAgentRuntimeSettings(), "agent");
+    const envScheduler = parseScheduler(env.OMK_TOOL_SCHEDULER, "OMK_TOOL_SCHEDULER");
+    const settingsScheduler = parseScheduler(settings.toolScheduler, "agent.toolScheduler");
+    return {
+        toolScheduler: envScheduler ?? settingsScheduler ?? DEFAULT_TOOL_SCHEDULER,
+        maxToolConcurrency: parseConcurrency(settings.maxToolConcurrency),
+        strictExtensionClaims: true,
+        toolTimeoutMs: parseTimeout(settings.toolTimeoutMs, "agent.toolTimeoutMs", DEFAULT_TOOL_TIMEOUT_MS),
+        toolTimeouts: resolveToolTimeouts(globalSettings, projectSettings, settings),
+    };
+}
+//# sourceMappingURL=agent-tool-settings.js.map

package/dist/core/agent-tool-settings.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"agent-tool-settings.js","sourceRoot":"","sources":["../../src/core/agent-tool-settings.ts"],"names":[],"mappings":"AAGA,MAAM,sBAAsB,GAAG,QAAQ,CAAC;AACxC,MAAM,4BAA4B,GAAG,CAAC,CAAC;AACvC,MAAM,uBAAuB,GAAG,CAAC,CAAC;AAClC,MAAM,mBAAmB,GAAG,aAAa,CAAC;AAE1C,MAAM,6BAA6B,GAAqC;IACvE,IAAI,EAAE,OAAO;IACb,IAAI,EAAE,MAAM;IACZ,IAAI,EAAE,MAAM;IACZ,IAAI,EAAE,MAAM;IACZ,EAAE,EAAE,MAAM;IACV,IAAI,EAAE,MAAM;IACZ,KAAK,EAAE,MAAM;CACb,CAAC;AAKF,0IAAyI;AACzI,MAAM,CAAC,MAAM,sBAAsB,GAAkD,MAAM,CAAC,MAAM,CAAC;IAClG,IAAI,EAAE,MAAM;IACZ,KAAK,EAAE,MAAM;IACb,GAAG,EAAE,OAAO;IACZ,OAAO,EAAE,OAAO;IAChB,OAAO,EAAE,OAAO;CAChB,CAAC,CAAC;AAEH,MAAM,mBAAmB,GAAG,IAAI,GAAG,CAAC,CAAC,MAAM,EAAE,IAAI,EAAE,MAAM,EAAE,MAAM,EAAE,MAAM,EAAE,MAAM,EAAE,QAAQ,CAAC,CAAC,CAAC;AAC9F,MAAM,oBAAoB,GAAG,IAAI,GAAG,CAAC,CAAC,OAAO,EAAE,MAAM,CAAC,CAAC,CAAC;AACxD,MAAM,sBAAsB,GAAG,IAAI,GAAG,CAAC,CAAC,MAAM,CAAC,CAAC,CAAC;AACjD,8EAA8E;AAC9E,uDAAuD;AACvD,MAAM,oBAAoB,GAAG,8CAA8C,CAAC;AAE5E;;;;;GAKG;AACH,MAAM,UAAU,0BAA0B,CAAC,QAAgB,EAAmC;IAC7F,IAAI,mBAAmB,CAAC,GAAG,CAAC,QAAQ,CAAC;QAAE,OAAO,MAAM,CAAC;IACrD,IAAI,oBAAoB,CAAC,GAAG,CAAC,QAAQ,CAAC;QAAE,OAAO,OAAO,CAAC;IACvD,IAAI,sBAAsB,CAAC,GAAG,CAAC,QAAQ,CAAC;QAAE,OAAO,SAAS,CAAC;IAC3D,IAAI,oBAAoB,CAAC,IAAI,CAAC,QAAQ,CAAC;QAAE,OAAO,SAAS,CAAC;IAC1D,IAAI,QAAQ,CAAC,UAAU,CAAC,OAAO,CAAC;QAAE,OAAO,KAAK,CAAC;IAC/C,OAAO,SAAS,CAAC;AAAA,CACjB;AAED;;;;;GAKG;AACH,MAAM,UAAU,4BAA4B,CAC3C,SAA2B,EAC3B,QAA0C,EACjB;IACzB,MAAM,MAAM,GAA2B,MAAM,CAAC,MAAM,CAAC,IAAI,CAAC,CAAC;IAC3D,KAAK,MAAM,QAAQ,IAAI,SAAS,EAAE,CAAC;QAClC,IAAI,MAAM,CAAC,MAAM,CAAC,QAAQ,EAAE,QAAQ,CAAC;YAAE,SAAS;QAChD,MAAM,QAAQ,GAAG,0BAA0B,CAAC,QAAQ,CAAC,CAAC;QACtD,IAAI,QAAQ,KAAK,SAAS,EAAE,CAAC;YAC5B,MAAM,CAAC,QAAQ,CAAC,GAAG,sBAAsB,CAAC,QAAQ,CAAC,CAAC;QACrD,CAAC;IACF,CAAC;IACD,KAAK,MAAM,CAAC,QAAQ,EAAE,SAAS,CAAC,IAAI,MAAM,CAAC,OAAO,CAAC,QAAQ,CAAC,EAAE,CAAC;QAC9D,MAAM,CAAC,QAAQ,CAAC,GAAG,SAAS,CAAC;IAC9B,CAAC;IACD,OAAO,MAAM,CAAC;AAAA,CACd;AAYD,MAAM,OAAO,sBAAuB,SAAQ,KAAK;IACvC,WAAW,CAAS;IAE7B,YAAY,WAAmB,EAAE;QAChC,KAAK,CAAC,WAAW,WAAW,UAAU,CAAC,CAAC;QACxC,IAAI,CAAC,IAAI,GAAG,wBAAwB,CAAC;QACrC,IAAI,CAAC,WAAW,GAAG,WAAW,CAAC;IAAA,CAC/B;CACD;AAED,SAAS,cAAc,CAAC,KAAc,EAAE,WAAmB,EAA6B;IACvF,IAAI,KAAK,KAAK,SAAS,EAAE,CAAC;QACzB,OAAO,SAAS,CAAC;IAClB,CAAC;IACD,IAAI,KAAK,KAAK,UAAU,IAAI,KAAK,KAAK,QAAQ,EAAE,CAAC;QAChD,OAAO,KAAK,CAAC;IACd,CAAC;IACD,MAAM,IAAI,sBAAsB,CAAC,WAAW,CAAC,CAAC;AAAA,CAC9C;AAED,SAAS,YAAY,CAAC,KAAc,EAAE,WAAmB,EAAE,YAAoB,EAAU;IACxF,IAAI,KAAK,KAAK,SAAS,EAAE,CAAC;QACzB,OAAO,YAAY,CAAC;IACrB,CAAC;IACD,IAAI,OAAO,KAAK,KAAK,QAAQ,IAAI,CAAC,MAAM,CAAC,SAAS,CAAC,KAAK,CAAC,IAAI,KAAK,GAAG,CAAC,IAAI,KAAK,GAAG,mBAAmB,EAAE,CAAC;QACvG,MAAM,IAAI,sBAAsB,CAAC,WAAW,CAAC,CAAC;IAC/C,CAAC;IACD,OAAO,KAAK,CAAC;AAAA,CACb;AAED,SAAS,gBAAgB,CAAC,KAAc,EAAsB;IAC7D,IAAI,KAAK,KAAK,SAAS,EAAE,CAAC;QACzB,OAAO,4BAA4B,CAAC;IACrC,CAAC;IACD,IAAI,OAAO,KAAK,KAAK,QAAQ,IAAI,CAAC,MAAM,CAAC,SAAS,CAAC,KAAK,CAAC,IAAI,KAAK,GAAG,CAAC,EAAE,CAAC;QACxE,MAAM,IAAI,sBAAsB,CAAC,0BAA0B,CAAC,CAAC;IAC9D,CAAC;IACD,OAAO,KAAK,KAAK,CAAC,CAAC,CAAC,CAAC,SAAS,CAAC,CAAC,CAAC,KAAK,CAAC;AAAA,CACvC;AAED,SAAS,kBAAkB,CAAC,KAAuC,EAAE,WAAmB,EAAwB;IAC/G,IAAI,KAAK,KAAK,SAAS,EAAE,CAAC;QACzB,OAAO,EAAE,CAAC;IACX,CAAC;IACD,IAAI,OAAO,KAAK,KAAK,QAAQ,IAAI,KAAK,KAAK,IAAI,IAAI,KAAK,CAAC,OAAO,CAAC,KAAK,CAAC,EAAE,CAAC;QACzE,MAAM,IAAI,sBAAsB,CAAC,WAAW,CAAC,CAAC;IAC/C,CAAC;IACD,OAAO,KAAK,CAAC;AAAA,CACb;AAED,SAAS,uBAAuB,CAC/B,KAA2C,EAC3C,WAAmB,EACsB;IACzC,IAAI,KAAK,KAAK,SAAS,EAAE,CAAC;QACzB,OAAO,EAAE,CAAC;IACX,CAAC;IACD,IAAI,OAAO,KAAK,KAAK,QAAQ,IAAI,KAAK,KAAK,IAAI,IAAI,KAAK,CAAC,OAAO,CAAC,KAAK,CAAC,EAAE,CAAC;QACzE,MAAM,IAAI,sBAAsB,CAAC,WAAW,CAAC,CAAC;IAC/C,CAAC;IAED,OAAO,MAAM,CAAC,OAAO,CAAC,KAAK,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC,QAAQ,EAAE,SAAS,CAAC,EAAE,EAAE,CAAC;QAC3D,IAAI,QAAQ,CAAC,MAAM,KAAK,CAAC,EAAE,CAAC;YAC3B,MAAM,IAAI,sBAAsB,CAAC,WAAW,CAAC,CAAC;QAC/C,CAAC;QACD,OAAO,CAAC,QAAQ,EAAE,YAAY,CAAC,SAAS,EAAE,GAAG,WAAW,IAAI,QAAQ,EAAE,EAAE,uBAAuB,CAAC,CAAU,CAAC;IAAA,CAC3G,CAAC,CAAC;AAAA,CACH;AAED,SAAS,mBAAmB,CAC3B,cAAoC,EACpC,eAAqC,EACrC,iBAAuC,EACd;IACzB,MAAM,QAAQ,GAA2B,MAAM,CAAC,MAAM,CAAC,IAAI,CAAC,CAAC;IAC7D,MAAM,WAAW,GAAG;QACnB,MAAM,CAAC,OAAO,CAAC,6BAA6B,CAAC;QAC7C,uBAAuB,CAAC,cAAc,CAAC,YAAY,EAAE,oBAAoB,CAAC;QAC1E,uBAAuB,CAAC,eAAe,CAAC,YAAY,EAAE,oBAAoB,CAAC;QAC3E,uBAAuB,CAAC,iBAAiB,CAAC,YAAY,EAAE,oBAAoB,CAAC;KAC7E,CAAC;IACF,KAAK,MAAM,OAAO,IAAI,WAAW,EAAE,CAAC;QACnC,KAAK,MAAM,CAAC,QAAQ,EAAE,SAAS,CAAC,IAAI,OAAO,EAAE,CAAC;YAC7C,QAAQ,CAAC,QAAQ,CAAC,GAAG,SAAS,CAAC;QAChC,CAAC;IACF,CAAC;IACD,OAAO,QAAQ,CAAC;AAAA,CAChB;AAED,MAAM,UAAU,wBAAwB,CACvC,eAAgC,EAChC,GAAG,GAAqE,OAAO,CAAC,GAAG,EACvD;IAC5B,MAAM,cAAc,GAAG,kBAAkB,CAAC,eAAe,CAAC,iBAAiB,EAAE,CAAC,KAAK,EAAE,OAAO,CAAC,CAAC;IAC9F,MAAM,eAAe,GAAG,kBAAkB,CAAC,eAAe,CAAC,kBAAkB,EAAE,CAAC,KAAK,EAAE,OAAO,CAAC,CAAC;IAChG,MAAM,QAAQ,GAAG,kBAAkB,CAAC,eAAe,CAAC,uBAAuB,EAAE,EAAE,OAAO,CAAC,CAAC;IACxF,MAAM,YAAY,GAAG,cAAc,CAAC,GAAG,CAAC,kBAAkB,EAAE,oBAAoB,CAAC,CAAC;IAClF,MAAM,iBAAiB,GAAG,cAAc,CAAC,QAAQ,CAAC,aAAa,EAAE,qBAAqB,CAAC,CAAC;IAExF,OAAO;QACN,aAAa,EAAE,YAAY,IAAI,iBAAiB,IAAI,sBAAsB;QAC1E,kBAAkB,EAAE,gBAAgB,CAAC,QAAQ,CAAC,kBAAkB,CAAC;QACjE,qBAAqB,EAAE,IAAI;QAC3B,aAAa,EAAE,YAAY,CAAC,QAAQ,CAAC,aAAa,EAAE,qBAAqB,EAAE,uBAAuB,CAAC;QACnG,YAAY,EAAE,mBAAmB,CAAC,cAAc,EAAE,eAAe,EAAE,QAAQ,CAAC;KAC5E,CAAC;AAAA,CACF","sourcesContent":["import type { AgentOptions } from \"omk-agent-core\";\nimport type { AgentRuntimeSettings, SettingsManager } from \"./settings-manager.ts\";\n\nconst DEFAULT_TOOL_SCHEDULER = \"dag-v2\";\nconst DEFAULT_MAX_TOOL_CONCURRENCY = 4;\nconst DEFAULT_TOOL_TIMEOUT_MS = 0;\nconst MAX_TOOL_TIMEOUT_MS = 2_147_483_647;\n\nconst DEFAULT_BUILTIN_TOOL_TIMEOUTS: Readonly<Record<string, number>> = {\n\tbash: 300_000,\n\tedit: 60_000,\n\tfind: 30_000,\n\tgrep: 30_000,\n\tls: 30_000,\n\tread: 30_000,\n\twrite: 60_000,\n};\n\n/** §6.3 timeout categories. `undefined` category falls through to the global setting. */\nexport type ToolTimeoutCategory = \"read\" | \"write\" | \"mcp\" | \"process\" | \"browser\";\n\n/** §6.3 category default timeouts (ms): read/list/search 30s, write/edit 60s, MCP 120s, bash/process 300s, browser/computer-use 180s. */\nexport const TOOL_CATEGORY_TIMEOUTS: Readonly<Record<ToolTimeoutCategory, number>> = Object.freeze({\n\tread: 30_000,\n\twrite: 60_000,\n\tmcp: 120_000,\n\tprocess: 300_000,\n\tbrowser: 180_000,\n});\n\nconst READ_CATEGORY_TOOLS = new Set([\"read\", \"ls\", \"find\", \"grep\", \"glob\", \"list\", \"search\"]);\nconst WRITE_CATEGORY_TOOLS = new Set([\"write\", \"edit\"]);\nconst PROCESS_CATEGORY_TOOLS = new Set([\"bash\"]);\n// Matches direct browser/computer-use tools and browser tools bridged through\n// MCP servers (e.g. `mcp__playwright__browser_click`).\nconst BROWSER_TOOL_PATTERN = /(^|__)(browser_|computer[-_]?use$|computer_)/;\n\n/**\n * Classify a tool name into its §6.3 timeout category. Conservative on\n * purpose: unknown names return `undefined` and keep the global timeout.\n * A browser/computer-use match wins over the generic MCP category because\n * browser automation is the slower bound.\n */\nexport function resolveToolTimeoutCategory(toolName: string): ToolTimeoutCategory | undefined {\n\tif (READ_CATEGORY_TOOLS.has(toolName)) return \"read\";\n\tif (WRITE_CATEGORY_TOOLS.has(toolName)) return \"write\";\n\tif (PROCESS_CATEGORY_TOOLS.has(toolName)) return \"process\";\n\tif (BROWSER_TOOL_PATTERN.test(toolName)) return \"browser\";\n\tif (toolName.startsWith(\"mcp__\")) return \"mcp\";\n\treturn undefined;\n}\n\n/**\n * Fill §6.3 category default timeouts for every active tool name that has no\n * explicit per-name entry. Explicit entries (user settings and built-in\n * defaults) always win and are all preserved; names without a category are\n * omitted so they fall through to the global `toolTimeoutMs` at runtime.\n */\nexport function applyCategoryTimeoutDefaults(\n\ttoolNames: Iterable<string>,\n\texplicit: Readonly<Record<string, number>>,\n): Record<string, number> {\n\tconst merged: Record<string, number> = Object.create(null);\n\tfor (const toolName of toolNames) {\n\t\tif (Object.hasOwn(explicit, toolName)) continue;\n\t\tconst category = resolveToolTimeoutCategory(toolName);\n\t\tif (category !== undefined) {\n\t\t\tmerged[toolName] = TOOL_CATEGORY_TIMEOUTS[category];\n\t\t}\n\t}\n\tfor (const [toolName, timeoutMs] of Object.entries(explicit)) {\n\t\tmerged[toolName] = timeoutMs;\n\t}\n\treturn merged;\n}\n\ntype ToolScheduler = NonNullable<AgentOptions[\"toolScheduler\"]>;\n\nexport interface ResolvedAgentToolSettings {\n\treadonly toolScheduler: ToolScheduler;\n\treadonly maxToolConcurrency: number | undefined;\n\treadonly strictExtensionClaims: boolean;\n\treadonly toolTimeoutMs: number;\n\treadonly toolTimeouts: Record<string, number>;\n}\n\nexport class AgentToolSettingsError extends Error {\n\treadonly settingName: string;\n\n\tconstructor(settingName: string) {\n\t\tsuper(`Invalid ${settingName} setting`);\n\t\tthis.name = \"AgentToolSettingsError\";\n\t\tthis.settingName = settingName;\n\t}\n}\n\nfunction parseScheduler(value: unknown, settingName: string): ToolScheduler | undefined {\n\tif (value === undefined) {\n\t\treturn undefined;\n\t}\n\tif (value === \"waves-v1\" || value === \"dag-v2\") {\n\t\treturn value;\n\t}\n\tthrow new AgentToolSettingsError(settingName);\n}\n\nfunction parseTimeout(value: unknown, settingName: string, defaultValue: number): number {\n\tif (value === undefined) {\n\t\treturn defaultValue;\n\t}\n\tif (typeof value !== \"number\" || !Number.isInteger(value) || value < 0 || value > MAX_TOOL_TIMEOUT_MS) {\n\t\tthrow new AgentToolSettingsError(settingName);\n\t}\n\treturn value;\n}\n\nfunction parseConcurrency(value: unknown): number | undefined {\n\tif (value === undefined) {\n\t\treturn DEFAULT_MAX_TOOL_CONCURRENCY;\n\t}\n\tif (typeof value !== \"number\" || !Number.isInteger(value) || value < 0) {\n\t\tthrow new AgentToolSettingsError(\"agent.maxToolConcurrency\");\n\t}\n\treturn value === 0 ? undefined : value;\n}\n\nfunction parseAgentSettings(value: AgentRuntimeSettings | undefined, settingName: string): AgentRuntimeSettings {\n\tif (value === undefined) {\n\t\treturn {};\n\t}\n\tif (typeof value !== \"object\" || value === null || Array.isArray(value)) {\n\t\tthrow new AgentToolSettingsError(settingName);\n\t}\n\treturn value;\n}\n\nfunction parseToolTimeoutEntries(\n\tvalue: AgentRuntimeSettings[\"toolTimeouts\"],\n\tsettingName: string,\n): readonly (readonly [string, number])[] {\n\tif (value === undefined) {\n\t\treturn [];\n\t}\n\tif (typeof value !== \"object\" || value === null || Array.isArray(value)) {\n\t\tthrow new AgentToolSettingsError(settingName);\n\t}\n\n\treturn Object.entries(value).map(([toolName, timeoutMs]) => {\n\t\tif (toolName.length === 0) {\n\t\t\tthrow new AgentToolSettingsError(settingName);\n\t\t}\n\t\treturn [toolName, parseTimeout(timeoutMs, `${settingName}.${toolName}`, DEFAULT_TOOL_TIMEOUT_MS)] as const;\n\t});\n}\n\nfunction resolveToolTimeouts(\n\tglobalSettings: AgentRuntimeSettings,\n\tprojectSettings: AgentRuntimeSettings,\n\teffectiveSettings: AgentRuntimeSettings,\n): Record<string, number> {\n\tconst timeouts: Record<string, number> = Object.create(null);\n\tconst entryGroups = [\n\t\tObject.entries(DEFAULT_BUILTIN_TOOL_TIMEOUTS),\n\t\tparseToolTimeoutEntries(globalSettings.toolTimeouts, \"agent.toolTimeouts\"),\n\t\tparseToolTimeoutEntries(projectSettings.toolTimeouts, \"agent.toolTimeouts\"),\n\t\tparseToolTimeoutEntries(effectiveSettings.toolTimeouts, \"agent.toolTimeouts\"),\n\t];\n\tfor (const entries of entryGroups) {\n\t\tfor (const [toolName, timeoutMs] of entries) {\n\t\t\ttimeouts[toolName] = timeoutMs;\n\t\t}\n\t}\n\treturn timeouts;\n}\n\nexport function resolveAgentToolSettings(\n\tsettingsManager: SettingsManager,\n\tenv: NodeJS.ProcessEnv | Readonly<Record<string, string | undefined>> = process.env,\n): ResolvedAgentToolSettings {\n\tconst globalSettings = parseAgentSettings(settingsManager.getGlobalSettings().agent, \"agent\");\n\tconst projectSettings = parseAgentSettings(settingsManager.getProjectSettings().agent, \"agent\");\n\tconst settings = parseAgentSettings(settingsManager.getAgentRuntimeSettings(), \"agent\");\n\tconst envScheduler = parseScheduler(env.OMK_TOOL_SCHEDULER, \"OMK_TOOL_SCHEDULER\");\n\tconst settingsScheduler = parseScheduler(settings.toolScheduler, \"agent.toolScheduler\");\n\n\treturn {\n\t\ttoolScheduler: envScheduler ?? settingsScheduler ?? DEFAULT_TOOL_SCHEDULER,\n\t\tmaxToolConcurrency: parseConcurrency(settings.maxToolConcurrency),\n\t\tstrictExtensionClaims: true,\n\t\ttoolTimeoutMs: parseTimeout(settings.toolTimeoutMs, \"agent.toolTimeoutMs\", DEFAULT_TOOL_TIMEOUT_MS),\n\t\ttoolTimeouts: resolveToolTimeouts(globalSettings, projectSettings, settings),\n\t};\n}\n"]}

package/dist/core/atomic-session-file.d.ts

@@ -0,0 +1,14 @@
+/**
+ * Atomically and durably replace the contents of `targetPath`.
+ *
+ * Writes a uniquely named temp file in the same directory, preserves the
+ * exact private mode 0600, fsyncs and closes the temp file, then renames it
+ * over the target so readers only ever observe the complete old or
+ * complete new content, never a truncated file. The parent directory is
+ * fsynced where supported. Windows and explicit unsupported-filesystem codes
+ * skip that final operation; real durability failures propagate. On any failure
+ * before the rename completes, the temp file is
+ * removed (best-effort) and the original target is left untouched.
+ */
+export declare function atomicRewriteFileSync(targetPath: string, data: string | Uint8Array): void;
+//# sourceMappingURL=atomic-session-file.d.ts.map

package/dist/core/atomic-session-file.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"atomic-session-file.d.ts","sourceRoot":"","sources":["../../src/core/atomic-session-file.ts"],"names":[],"mappings":"AAuBA;;;;;;;;;;;GAWG;AACH,wBAAgB,qBAAqB,CAAC,UAAU,EAAE,MAAM,EAAE,IAAI,EAAE,MAAM,GAAG,UAAU,GAAG,IAAI,CAkEzF","sourcesContent":["import { randomBytes } from \"crypto\";\nimport {\n\tcloseSync,\n\tfsyncSync,\n\tlstatSync,\n\topenSync,\n\trealpathSync,\n\trenameSync,\n\tstatSync,\n\tunlinkSync,\n\twriteSync,\n} from \"fs\";\nimport { basename, dirname, join } from \"path\";\nimport { enforcePrivateFileDescriptorModeSync } from \"./durable-file-mode.ts\";\n\nfunction assertRewriteTargetSafe(path: string): void {\n\ttry {\n\t\tif (statSync(path).nlink > 1) throw new Error(\"Atomic rewrite refused for a target with more than one hard link\");\n\t} catch (error) {\n\t\tif (!(error instanceof Error) || !(\"code\" in error) || Reflect.get(error, \"code\") !== \"ENOENT\") throw error;\n\t}\n}\n\n/**\n * Atomically and durably replace the contents of `targetPath`.\n *\n * Writes a uniquely named temp file in the same directory, preserves the\n * exact private mode 0600, fsyncs and closes the temp file, then renames it\n * over the target so readers only ever observe the complete old or\n * complete new content, never a truncated file. The parent directory is\n * fsynced where supported. Windows and explicit unsupported-filesystem codes\n * skip that final operation; real durability failures propagate. On any failure\n * before the rename completes, the temp file is\n * removed (best-effort) and the original target is left untouched.\n */\nexport function atomicRewriteFileSync(targetPath: string, data: string | Uint8Array): void {\n\tlet rewritePath: string;\n\ttry {\n\t\trewritePath = realpathSync.native(targetPath);\n\t} catch (error) {\n\t\tconst code = error instanceof Error && \"code\" in error ? Reflect.get(error, \"code\") : undefined;\n\t\tif (code !== \"ENOENT\") throw error;\n\t\tlet danglingLink = false;\n\t\ttry {\n\t\t\tdanglingLink = lstatSync(targetPath).isSymbolicLink();\n\t\t} catch (lstatError) {\n\t\t\tconst lstatCode =\n\t\t\t\tlstatError instanceof Error && \"code\" in lstatError ? Reflect.get(lstatError, \"code\") : undefined;\n\t\t\tif (lstatCode !== \"ENOENT\") throw lstatError;\n\t\t}\n\t\tif (danglingLink) throw error;\n\t\trewritePath = targetPath;\n\t}\n\tconst dir = dirname(rewritePath);\n\n\tassertRewriteTargetSafe(rewritePath);\n\tconst tempPath = join(dir, `.${basename(rewritePath)}.${process.pid}.${randomBytes(6).toString(\"hex\")}.tmp`);\n\tlet fd: number | undefined;\n\ttry {\n\t\tfd = openSync(tempPath, \"wx\", 0o600);\n\t\tenforcePrivateFileDescriptorModeSync(fd, rewritePath);\n\t\tconst buffer = typeof data === \"string\" ? Buffer.from(data, \"utf8\") : Buffer.from(data);\n\t\tlet offset = 0;\n\t\twhile (offset < buffer.length) {\n\t\t\toffset += writeSync(fd, buffer, offset);\n\t\t}\n\t\tfsyncSync(fd);\n\t\tcloseSync(fd);\n\t\tfd = undefined;\n\t\tassertRewriteTargetSafe(rewritePath);\n\t\trenameSync(tempPath, rewritePath);\n\t} catch (error) {\n\t\tif (fd !== undefined) {\n\t\t\ttry {\n\t\t\t\tcloseSync(fd);\n\t\t\t} catch (cleanupError) {\n\t\t\t\tif (!(cleanupError instanceof Error)) throw cleanupError;\n\t\t\t}\n\t\t}\n\t\ttry {\n\t\t\tunlinkSync(tempPath);\n\t\t} catch (cleanupError) {\n\t\t\tif (!(cleanupError instanceof Error)) throw cleanupError;\n\t\t}\n\t\tthrow error;\n\t}\n\n\tif (process.platform === \"win32\") return;\n\ttry {\n\t\tconst dirFd = openSync(dir, \"r\");\n\t\ttry {\n\t\t\tfsyncSync(dirFd);\n\t\t} finally {\n\t\t\tcloseSync(dirFd);\n\t\t}\n\t} catch (error) {\n\t\tconst code = error instanceof Error && \"code\" in error ? Reflect.get(error, \"code\") : undefined;\n\t\tif (typeof code !== \"string\" || !new Set([\"EINVAL\", \"ENOTSUP\", \"EOPNOTSUPP\", \"ENOSYS\"]).has(code)) {\n\t\t\tthrow error;\n\t\t}\n\t}\n}\n"]}

package/dist/core/atomic-session-file.js

@@ -0,0 +1,104 @@
+import { randomBytes } from "crypto";
+import { closeSync, fsyncSync, lstatSync, openSync, realpathSync, renameSync, statSync, unlinkSync, writeSync, } from "fs";
+import { basename, dirname, join } from "path";
+import { enforcePrivateFileDescriptorModeSync } from "./durable-file-mode.js";
+function assertRewriteTargetSafe(path) {
+    try {
+        if (statSync(path).nlink > 1)
+            throw new Error("Atomic rewrite refused for a target with more than one hard link");
+    }
+    catch (error) {
+        if (!(error instanceof Error) || !("code" in error) || Reflect.get(error, "code") !== "ENOENT")
+            throw error;
+    }
+}
+/**
+ * Atomically and durably replace the contents of `targetPath`.
+ *
+ * Writes a uniquely named temp file in the same directory, preserves the
+ * exact private mode 0600, fsyncs and closes the temp file, then renames it
+ * over the target so readers only ever observe the complete old or
+ * complete new content, never a truncated file. The parent directory is
+ * fsynced where supported. Windows and explicit unsupported-filesystem codes
+ * skip that final operation; real durability failures propagate. On any failure
+ * before the rename completes, the temp file is
+ * removed (best-effort) and the original target is left untouched.
+ */
+export function atomicRewriteFileSync(targetPath, data) {
+    let rewritePath;
+    try {
+        rewritePath = realpathSync.native(targetPath);
+    }
+    catch (error) {
+        const code = error instanceof Error && "code" in error ? Reflect.get(error, "code") : undefined;
+        if (code !== "ENOENT")
+            throw error;
+        let danglingLink = false;
+        try {
+            danglingLink = lstatSync(targetPath).isSymbolicLink();
+        }
+        catch (lstatError) {
+            const lstatCode = lstatError instanceof Error && "code" in lstatError ? Reflect.get(lstatError, "code") : undefined;
+            if (lstatCode !== "ENOENT")
+                throw lstatError;
+        }
+        if (danglingLink)
+            throw error;
+        rewritePath = targetPath;
+    }
+    const dir = dirname(rewritePath);
+    assertRewriteTargetSafe(rewritePath);
+    const tempPath = join(dir, `.${basename(rewritePath)}.${process.pid}.${randomBytes(6).toString("hex")}.tmp`);
+    let fd;
+    try {
+        fd = openSync(tempPath, "wx", 0o600);
+        enforcePrivateFileDescriptorModeSync(fd, rewritePath);
+        const buffer = typeof data === "string" ? Buffer.from(data, "utf8") : Buffer.from(data);
+        let offset = 0;
+        while (offset < buffer.length) {
+            offset += writeSync(fd, buffer, offset);
+        }
+        fsyncSync(fd);
+        closeSync(fd);
+        fd = undefined;
+        assertRewriteTargetSafe(rewritePath);
+        renameSync(tempPath, rewritePath);
+    }
+    catch (error) {
+        if (fd !== undefined) {
+            try {
+                closeSync(fd);
+            }
+            catch (cleanupError) {
+                if (!(cleanupError instanceof Error))
+                    throw cleanupError;
+            }
+        }
+        try {
+            unlinkSync(tempPath);
+        }
+        catch (cleanupError) {
+            if (!(cleanupError instanceof Error))
+                throw cleanupError;
+        }
+        throw error;
+    }
+    if (process.platform === "win32")
+        return;
+    try {
+        const dirFd = openSync(dir, "r");
+        try {
+            fsyncSync(dirFd);
+        }
+        finally {
+            closeSync(dirFd);
+        }
+    }
+    catch (error) {
+        const code = error instanceof Error && "code" in error ? Reflect.get(error, "code") : undefined;
+        if (typeof code !== "string" || !new Set(["EINVAL", "ENOTSUP", "EOPNOTSUPP", "ENOSYS"]).has(code)) {
+            throw error;
+        }
+    }
+}
+//# sourceMappingURL=atomic-session-file.js.map

package/dist/core/atomic-session-file.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"atomic-session-file.js","sourceRoot":"","sources":["../../src/core/atomic-session-file.ts"],"names":[],"mappings":"AAAA,OAAO,EAAE,WAAW,EAAE,MAAM,QAAQ,CAAC;AACrC,OAAO,EACN,SAAS,EACT,SAAS,EACT,SAAS,EACT,QAAQ,EACR,YAAY,EACZ,UAAU,EACV,QAAQ,EACR,UAAU,EACV,SAAS,GACT,MAAM,IAAI,CAAC;AACZ,OAAO,EAAE,QAAQ,EAAE,OAAO,EAAE,IAAI,EAAE,MAAM,MAAM,CAAC;AAC/C,OAAO,EAAE,oCAAoC,EAAE,MAAM,wBAAwB,CAAC;AAE9E,SAAS,uBAAuB,CAAC,IAAY,EAAQ;IACpD,IAAI,CAAC;QACJ,IAAI,QAAQ,CAAC,IAAI,CAAC,CAAC,KAAK,GAAG,CAAC;YAAE,MAAM,IAAI,KAAK,CAAC,kEAAkE,CAAC,CAAC;IACnH,CAAC;IAAC,OAAO,KAAK,EAAE,CAAC;QAChB,IAAI,CAAC,CAAC,KAAK,YAAY,KAAK,CAAC,IAAI,CAAC,CAAC,MAAM,IAAI,KAAK,CAAC,IAAI,OAAO,CAAC,GAAG,CAAC,KAAK,EAAE,MAAM,CAAC,KAAK,QAAQ;YAAE,MAAM,KAAK,CAAC;IAC7G,CAAC;AAAA,CACD;AAED;;;;;;;;;;;GAWG;AACH,MAAM,UAAU,qBAAqB,CAAC,UAAkB,EAAE,IAAyB,EAAQ;IAC1F,IAAI,WAAmB,CAAC;IACxB,IAAI,CAAC;QACJ,WAAW,GAAG,YAAY,CAAC,MAAM,CAAC,UAAU,CAAC,CAAC;IAC/C,CAAC;IAAC,OAAO,KAAK,EAAE,CAAC;QAChB,MAAM,IAAI,GAAG,KAAK,YAAY,KAAK,IAAI,MAAM,IAAI,KAAK,CAAC,CAAC,CAAC,OAAO,CAAC,GAAG,CAAC,KAAK,EAAE,MAAM,CAAC,CAAC,CAAC,CAAC,SAAS,CAAC;QAChG,IAAI,IAAI,KAAK,QAAQ;YAAE,MAAM,KAAK,CAAC;QACnC,IAAI,YAAY,GAAG,KAAK,CAAC;QACzB,IAAI,CAAC;YACJ,YAAY,GAAG,SAAS,CAAC,UAAU,CAAC,CAAC,cAAc,EAAE,CAAC;QACvD,CAAC;QAAC,OAAO,UAAU,EAAE,CAAC;YACrB,MAAM,SAAS,GACd,UAAU,YAAY,KAAK,IAAI,MAAM,IAAI,UAAU,CAAC,CAAC,CAAC,OAAO,CAAC,GAAG,CAAC,UAAU,EAAE,MAAM,CAAC,CAAC,CAAC,CAAC,SAAS,CAAC;YACnG,IAAI,SAAS,KAAK,QAAQ;gBAAE,MAAM,UAAU,CAAC;QAC9C,CAAC;QACD,IAAI,YAAY;YAAE,MAAM,KAAK,CAAC;QAC9B,WAAW,GAAG,UAAU,CAAC;IAC1B,CAAC;IACD,MAAM,GAAG,GAAG,OAAO,CAAC,WAAW,CAAC,CAAC;IAEjC,uBAAuB,CAAC,WAAW,CAAC,CAAC;IACrC,MAAM,QAAQ,GAAG,IAAI,CAAC,GAAG,EAAE,IAAI,QAAQ,CAAC,WAAW,CAAC,IAAI,OAAO,CAAC,GAAG,IAAI,WAAW,CAAC,CAAC,CAAC,CAAC,QAAQ,CAAC,KAAK,CAAC,MAAM,CAAC,CAAC;IAC7G,IAAI,EAAsB,CAAC;IAC3B,IAAI,CAAC;QACJ,EAAE,GAAG,QAAQ,CAAC,QAAQ,EAAE,IAAI,EAAE,KAAK,CAAC,CAAC;QACrC,oCAAoC,CAAC,EAAE,EAAE,WAAW,CAAC,CAAC;QACtD,MAAM,MAAM,GAAG,OAAO,IAAI,KAAK,QAAQ,CAAC,CAAC,CAAC,MAAM,CAAC,IAAI,CAAC,IAAI,EAAE,MAAM,CAAC,CAAC,CAAC,CAAC,MAAM,CAAC,IAAI,CAAC,IAAI,CAAC,CAAC;QACxF,IAAI,MAAM,GAAG,CAAC,CAAC;QACf,OAAO,MAAM,GAAG,MAAM,CAAC,MAAM,EAAE,CAAC;YAC/B,MAAM,IAAI,SAAS,CAAC,EAAE,EAAE,MAAM,EAAE,MAAM,CAAC,CAAC;QACzC,CAAC;QACD,SAAS,CAAC,EAAE,CAAC,CAAC;QACd,SAAS,CAAC,EAAE,CAAC,CAAC;QACd,EAAE,GAAG,SAAS,CAAC;QACf,uBAAuB,CAAC,WAAW,CAAC,CAAC;QACrC,UAAU,CAAC,QAAQ,EAAE,WAAW,CAAC,CAAC;IACnC,CAAC;IAAC,OAAO,KAAK,EAAE,CAAC;QAChB,IAAI,EAAE,KAAK,SAAS,EAAE,CAAC;YACtB,IAAI,CAAC;gBACJ,SAAS,CAAC,EAAE,CAAC,CAAC;YACf,CAAC;YAAC,OAAO,YAAY,EAAE,CAAC;gBACvB,IAAI,CAAC,CAAC,YAAY,YAAY,KAAK,CAAC;oBAAE,MAAM,YAAY,CAAC;YAC1D,CAAC;QACF,CAAC;QACD,IAAI,CAAC;YACJ,UAAU,CAAC,QAAQ,CAAC,CAAC;QACtB,CAAC;QAAC,OAAO,YAAY,EAAE,CAAC;YACvB,IAAI,CAAC,CAAC,YAAY,YAAY,KAAK,CAAC;gBAAE,MAAM,YAAY,CAAC;QAC1D,CAAC;QACD,MAAM,KAAK,CAAC;IACb,CAAC;IAED,IAAI,OAAO,CAAC,QAAQ,KAAK,OAAO;QAAE,OAAO;IACzC,IAAI,CAAC;QACJ,MAAM,KAAK,GAAG,QAAQ,CAAC,GAAG,EAAE,GAAG,CAAC,CAAC;QACjC,IAAI,CAAC;YACJ,SAAS,CAAC,KAAK,CAAC,CAAC;QAClB,CAAC;gBAAS,CAAC;YACV,SAAS,CAAC,KAAK,CAAC,CAAC;QAClB,CAAC;IACF,CAAC;IAAC,OAAO,KAAK,EAAE,CAAC;QAChB,MAAM,IAAI,GAAG,KAAK,YAAY,KAAK,IAAI,MAAM,IAAI,KAAK,CAAC,CAAC,CAAC,OAAO,CAAC,GAAG,CAAC,KAAK,EAAE,MAAM,CAAC,CAAC,CAAC,CAAC,SAAS,CAAC;QAChG,IAAI,OAAO,IAAI,KAAK,QAAQ,IAAI,CAAC,IAAI,GAAG,CAAC,CAAC,QAAQ,EAAE,SAAS,EAAE,YAAY,EAAE,QAAQ,CAAC,CAAC,CAAC,GAAG,CAAC,IAAI,CAAC,EAAE,CAAC;YACnG,MAAM,KAAK,CAAC;QACb,CAAC;IACF,CAAC;AAAA,CACD","sourcesContent":["import { randomBytes } from \"crypto\";\nimport {\n\tcloseSync,\n\tfsyncSync,\n\tlstatSync,\n\topenSync,\n\trealpathSync,\n\trenameSync,\n\tstatSync,\n\tunlinkSync,\n\twriteSync,\n} from \"fs\";\nimport { basename, dirname, join } from \"path\";\nimport { enforcePrivateFileDescriptorModeSync } from \"./durable-file-mode.ts\";\n\nfunction assertRewriteTargetSafe(path: string): void {\n\ttry {\n\t\tif (statSync(path).nlink > 1) throw new Error(\"Atomic rewrite refused for a target with more than one hard link\");\n\t} catch (error) {\n\t\tif (!(error instanceof Error) || !(\"code\" in error) || Reflect.get(error, \"code\") !== \"ENOENT\") throw error;\n\t}\n}\n\n/**\n * Atomically and durably replace the contents of `targetPath`.\n *\n * Writes a uniquely named temp file in the same directory, preserves the\n * exact private mode 0600, fsyncs and closes the temp file, then renames it\n * over the target so readers only ever observe the complete old or\n * complete new content, never a truncated file. The parent directory is\n * fsynced where supported. Windows and explicit unsupported-filesystem codes\n * skip that final operation; real durability failures propagate. On any failure\n * before the rename completes, the temp file is\n * removed (best-effort) and the original target is left untouched.\n */\nexport function atomicRewriteFileSync(targetPath: string, data: string | Uint8Array): void {\n\tlet rewritePath: string;\n\ttry {\n\t\trewritePath = realpathSync.native(targetPath);\n\t} catch (error) {\n\t\tconst code = error instanceof Error && \"code\" in error ? Reflect.get(error, \"code\") : undefined;\n\t\tif (code !== \"ENOENT\") throw error;\n\t\tlet danglingLink = false;\n\t\ttry {\n\t\t\tdanglingLink = lstatSync(targetPath).isSymbolicLink();\n\t\t} catch (lstatError) {\n\t\t\tconst lstatCode =\n\t\t\t\tlstatError instanceof Error && \"code\" in lstatError ? Reflect.get(lstatError, \"code\") : undefined;\n\t\t\tif (lstatCode !== \"ENOENT\") throw lstatError;\n\t\t}\n\t\tif (danglingLink) throw error;\n\t\trewritePath = targetPath;\n\t}\n\tconst dir = dirname(rewritePath);\n\n\tassertRewriteTargetSafe(rewritePath);\n\tconst tempPath = join(dir, `.${basename(rewritePath)}.${process.pid}.${randomBytes(6).toString(\"hex\")}.tmp`);\n\tlet fd: number | undefined;\n\ttry {\n\t\tfd = openSync(tempPath, \"wx\", 0o600);\n\t\tenforcePrivateFileDescriptorModeSync(fd, rewritePath);\n\t\tconst buffer = typeof data === \"string\" ? Buffer.from(data, \"utf8\") : Buffer.from(data);\n\t\tlet offset = 0;\n\t\twhile (offset < buffer.length) {\n\t\t\toffset += writeSync(fd, buffer, offset);\n\t\t}\n\t\tfsyncSync(fd);\n\t\tcloseSync(fd);\n\t\tfd = undefined;\n\t\tassertRewriteTargetSafe(rewritePath);\n\t\trenameSync(tempPath, rewritePath);\n\t} catch (error) {\n\t\tif (fd !== undefined) {\n\t\t\ttry {\n\t\t\t\tcloseSync(fd);\n\t\t\t} catch (cleanupError) {\n\t\t\t\tif (!(cleanupError instanceof Error)) throw cleanupError;\n\t\t\t}\n\t\t}\n\t\ttry {\n\t\t\tunlinkSync(tempPath);\n\t\t} catch (cleanupError) {\n\t\t\tif (!(cleanupError instanceof Error)) throw cleanupError;\n\t\t}\n\t\tthrow error;\n\t}\n\n\tif (process.platform === \"win32\") return;\n\ttry {\n\t\tconst dirFd = openSync(dir, \"r\");\n\t\ttry {\n\t\t\tfsyncSync(dirFd);\n\t\t} finally {\n\t\t\tcloseSync(dirFd);\n\t\t}\n\t} catch (error) {\n\t\tconst code = error instanceof Error && \"code\" in error ? Reflect.get(error, \"code\") : undefined;\n\t\tif (typeof code !== \"string\" || !new Set([\"EINVAL\", \"ENOTSUP\", \"EOPNOTSUPP\", \"ENOSYS\"]).has(code)) {\n\t\t\tthrow error;\n\t\t}\n\t}\n}\n"]}

package/dist/core/compaction/compaction.d.ts

@@ -24,9 +24,15 @@ export interface CompactionResult<T = unknown> {
 export interface CompactionSettings {
     enabled: boolean;
     reserveTokens: number;
+    reservedOutputTokens?: number;
+    reservedToolResultTokens?: number;
+    safetyMarginTokens?: number;
+    imageReserveTokens?: number;
     keepRecentTokens: number;
     /** Maximum fraction of the context window to use before compaction. Default: 0.9. */
     maxUsageRatio?: number;
+    rearmRatio?: number;
+    emergencyRatio?: number;
 }
 export declare const DEFAULT_COMPACTION_MAX_USAGE_RATIO = 0.9;
 export declare const DEFAULT_COMPACTION_SETTINGS: CompactionSettings;

package/dist/core/compaction/compaction.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"compaction.d.ts","sourceRoot":"","sources":["../../../src/core/compaction/compaction.ts"],"names":[],"mappings":"AAAA;;;;;GAKG;AAEH,OAAO,KAAK,EAAE,YAAY,EAAE,QAAQ,EAAE,aAAa,EAAE,MAAM,gBAAgB,CAAC;AAC5E,OAAO,KAAK,EAA6B,KAAK,EAAuB,KAAK,EAAE,MAAM,QAAQ,CAAC;AAQ3F,OAAO,EAA6C,KAAK,YAAY,EAAE,MAAM,uBAAuB,CAAC;AACrG,OAAO,EAIN,KAAK,cAAc,EAInB,MAAM,YAAY,CAAC;AAMpB,kEAAkE;AAClE,MAAM,WAAW,iBAAiB;IACjC,SAAS,EAAE,MAAM,EAAE,CAAC;IACpB,aAAa,EAAE,MAAM,EAAE,CAAC;CACxB;AAiED,8EAA8E;AAC9E,MAAM,WAAW,gBAAgB,CAAC,CAAC,GAAG,OAAO;IAC5C,OAAO,EAAE,MAAM,CAAC;IAChB,gBAAgB,EAAE,MAAM,CAAC;IACzB,YAAY,EAAE,MAAM,CAAC;IACrB,+FAA+F;IAC/F,OAAO,CAAC,EAAE,CAAC,CAAC;CACZ;AAMD,MAAM,WAAW,kBAAkB;IAClC,OAAO,EAAE,OAAO,CAAC;IACjB,aAAa,EAAE,MAAM,CAAC;IACtB,gBAAgB,EAAE,MAAM,CAAC;IACzB,qFAAqF;IACrF,aAAa,CAAC,EAAE,MAAM,CAAC;CACvB;AAED,eAAO,MAAM,kCAAkC,MAAM,CAAC;AAEtD,eAAO,MAAM,2BAA2B,EAAE,kBAKzC,CAAC;AAMF;;;GAGG;AACH,wBAAgB,sBAAsB,CAAC,KAAK,EAAE,KAAK,GAAG,MAAM,CAE3D;AAgBD;;GAEG;AACH,wBAAgB,qBAAqB,CAAC,OAAO,EAAE,YAAY,EAAE,GAAG,KAAK,GAAG,SAAS,CAShF;AAED,MAAM,WAAW,oBAAoB;IACpC,MAAM,EAAE,MAAM,CAAC;IACf,WAAW,EAAE,MAAM,CAAC;IACpB,cAAc,EAAE,MAAM,CAAC;IACvB,cAAc,EAAE,MAAM,GAAG,IAAI,CAAC;CAC9B;AAUD;;;GAGG;AACH,wBAAgB,qBAAqB,CAAC,QAAQ,EAAE,YAAY,EAAE,GAAG,oBAAoB,CA4BpF;AAED,iFAAiF;AACjF,wBAAgB,8BAA8B,CAC7C,QAAQ,EAAE,YAAY,EAAE,EACxB,eAAe,EAAE,YAAY,EAAE,GAC7B,oBAAoB,CAEtB;AAED,MAAM,MAAM,uBAAuB,GAAG,iBAAiB,GAAG,gBAAgB,CAAC;AAE3E,MAAM,WAAW,2BAA2B;IAC3C,aAAa,EAAE,MAAM,CAAC;IACtB,cAAc,EAAE,MAAM,CAAC;IACvB,mBAAmB,EAAE,MAAM,CAAC;IAC5B,qBAAqB,EAAE,MAAM,CAAC;IAC9B,SAAS,EAAE,uBAAuB,CAAC;CACnC;AAED,+FAA+F;AAC/F,wBAAgB,8BAA8B,CAC7C,aAAa,EAAE,MAAM,EACrB,QAAQ,EAAE,kBAAkB,GAC1B,2BAA2B,GAAG,SAAS,CAwBzC;AAED;;GAEG;AACH,wBAAgB,aAAa,CAAC,aAAa,EAAE,MAAM,EAAE,aAAa,EAAE,MAAM,EAAE,QAAQ,EAAE,kBAAkB,GAAG,OAAO,CAGjH;AAwBD;;;GAGG;AACH,wBAAgB,cAAc,CAAC,OAAO,EAAE,YAAY,GAAG,MAAM,CAwC5D;AAiDD;;;;GAIG;AACH,wBAAgB,kBAAkB,CAAC,OAAO,EAAE,YAAY,EAAE,EAAE,UAAU,EAAE,MAAM,EAAE,UAAU,EAAE,MAAM,GAAG,MAAM,CAe1G;AAED,MAAM,WAAW,cAAc;IAC9B,mCAAmC;IACnC,mBAAmB,EAAE,MAAM,CAAC;IAC5B,qFAAqF;IACrF,cAAc,EAAE,MAAM,CAAC;IACvB,uEAAuE;IACvE,WAAW,EAAE,OAAO,CAAC;CACrB;AAED;;;;;;;;;;;;;;;GAeG;AACH,wBAAgB,YAAY,CAC3B,OAAO,EAAE,YAAY,EAAE,EACvB,UAAU,EAAE,MAAM,EAClB,QAAQ,EAAE,MAAM,EAChB,gBAAgB,EAAE,MAAM,GACtB,cAAc,CAyDhB;AA0GD;;;GAGG;AACH,wBAAsB,eAAe,CACpC,eAAe,EAAE,YAAY,EAAE,EAC/B,KAAK,EAAE,KAAK,CAAC,GAAG,CAAC,EACjB,aAAa,EAAE,MAAM,EACrB,MAAM,EAAE,MAAM,GAAG,SAAS,EAC1B,OAAO,CAAC,EAAE,MAAM,CAAC,MAAM,EAAE,MAAM,CAAC,EAChC,MAAM,CAAC,EAAE,WAAW,EACpB,kBAAkB,CAAC,EAAE,MAAM,EAC3B,eAAe,CAAC,EAAE,MAAM,EACxB,aAAa,CAAC,EAAE,aAAa,EAC7B,QAAQ,CAAC,EAAE,QAAQ,GACjB,OAAO,CAAC,MAAM,CAAC,CAmDjB;AAMD,MAAM,WAAW,qBAAqB;IACrC,kCAAkC;IAClC,gBAAgB,EAAE,MAAM,CAAC;IACzB,qDAAqD;IACrD,mBAAmB,EAAE,YAAY,EAAE,CAAC;IACpC,2EAA2E;IAC3E,kBAAkB,EAAE,YAAY,EAAE,CAAC;IACnC,iEAAiE;IACjE,WAAW,EAAE,OAAO,CAAC;IACrB,YAAY,EAAE,MAAM,CAAC;IACrB,6DAA6D;IAC7D,eAAe,CAAC,EAAE,MAAM,CAAC;IACzB,yDAAyD;IACzD,OAAO,EAAE,cAAc,CAAC;IACxB,8CAA8C;IAC9C,QAAQ,EAAE,kBAAkB,CAAC;CAC7B;AAED,wBAAgB,iBAAiB,CAChC,WAAW,EAAE,YAAY,EAAE,EAC3B,QAAQ,EAAE,kBAAkB,GAC1B,qBAAqB,GAAG,SAAS,CAwEnC;AAqBD;;;;;;GAMG;AACH,wBAAsB,OAAO,CAC5B,WAAW,EAAE,qBAAqB,EAClC,KAAK,EAAE,KAAK,CAAC,GAAG,CAAC,EACjB,MAAM,EAAE,MAAM,GAAG,SAAS,EAC1B,OAAO,CAAC,EAAE,MAAM,CAAC,MAAM,EAAE,MAAM,CAAC,EAChC,kBAAkB,CAAC,EAAE,MAAM,EAC3B,MAAM,CAAC,EAAE,WAAW,EACpB,aAAa,CAAC,EAAE,aAAa,EAC7B,QAAQ,CAAC,EAAE,QAAQ,GACjB,OAAO,CAAC,gBAAgB,CAAC,CA2E3B","sourcesContent":["/**\n * Context compaction for long sessions.\n *\n * Pure functions for compaction logic. The session manager handles I/O,\n * and after compaction the session is reloaded.\n */\n\nimport type { AgentMessage, StreamFn, ThinkingLevel } from \"omk-agent-core\";\nimport type { AssistantMessage, Context, Model, SimpleStreamOptions, Usage } from \"omk-ai\";\nimport { completeSimple } from \"omk-ai\";\nimport {\n\tconvertToLlm,\n\tcreateBranchSummaryMessage,\n\tcreateCompactionSummaryMessage,\n\tcreateCustomMessage,\n} from \"../messages.ts\";\nimport { buildSessionContext, type CompactionEntry, type SessionEntry } from \"../session-manager.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// File Operation Tracking\n// ============================================================================\n\n/** Details stored in CompactionEntry.details for file tracking */\nexport interface CompactionDetails {\n\treadFiles: string[];\n\tmodifiedFiles: string[];\n}\n\n/**\n * Extract file operations from messages and previous compaction entries.\n */\nfunction extractFileOperations(\n\tmessages: AgentMessage[],\n\tentries: SessionEntry[],\n\tprevCompactionIndex: number,\n): FileOperations {\n\tconst fileOps = createFileOps();\n\n\tif (prevCompactionIndex >= 0) {\n\t\tconst prevCompaction = entries[prevCompactionIndex] as CompactionEntry;\n\t\tif (!prevCompaction.fromHook && prevCompaction.details) {\n\t\t\t// fromHook field kept for session file compatibility\n\t\t\tconst details = prevCompaction.details as CompactionDetails;\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\tfor (const f of details.modifiedFiles) fileOps.edited.add(f);\n\t\t\t}\n\t\t}\n\t}\n\n\t// Extract from tool calls in messages\n\tfor (const msg of messages) {\n\t\textractFileOpsFromMessage(msg, fileOps);\n\t}\n\n\treturn fileOps;\n}\n\n// ============================================================================\n// Message Extraction\n// ============================================================================\n\n/**\n * Extract AgentMessage from an entry if it produces one.\n * Returns undefined for entries that don't contribute to LLM context.\n */\nfunction getMessageFromEntry(entry: SessionEntry): AgentMessage | undefined {\n\tif (entry.type === \"message\") {\n\t\treturn entry.message;\n\t}\n\tif (entry.type === \"custom_message\") {\n\t\treturn createCustomMessage(entry.customType, entry.content, entry.display, entry.details, entry.timestamp);\n\t}\n\tif (entry.type === \"branch_summary\") {\n\t\treturn createBranchSummaryMessage(entry.summary, entry.fromId, entry.timestamp);\n\t}\n\tif (entry.type === \"compaction\") {\n\t\treturn createCompactionSummaryMessage(entry.summary, entry.tokensBefore, entry.timestamp);\n\t}\n\treturn undefined;\n}\n\nfunction getMessageFromEntryForCompaction(entry: SessionEntry): AgentMessage | undefined {\n\tif (entry.type === \"compaction\") {\n\t\treturn undefined;\n\t}\n\treturn getMessageFromEntry(entry);\n}\n\n/** Result from compact() - SessionManager adds uuid/parentUuid when saving */\nexport interface CompactionResult<T = unknown> {\n\tsummary: string;\n\tfirstKeptEntryId: string;\n\ttokensBefore: number;\n\t/** Extension-specific data (e.g., ArtifactIndex, version markers for structured compaction) */\n\tdetails?: T;\n}\n\n// ============================================================================\n// Types\n// ============================================================================\n\nexport interface CompactionSettings {\n\tenabled: boolean;\n\treserveTokens: number;\n\tkeepRecentTokens: number;\n\t/** Maximum fraction of the context window to use before compaction. Default: 0.9. */\n\tmaxUsageRatio?: number;\n}\n\nexport const DEFAULT_COMPACTION_MAX_USAGE_RATIO = 0.9;\n\nexport const DEFAULT_COMPACTION_SETTINGS: CompactionSettings = {\n\tenabled: true,\n\treserveTokens: 16384,\n\tkeepRecentTokens: 20000,\n\tmaxUsageRatio: DEFAULT_COMPACTION_MAX_USAGE_RATIO,\n};\n\n// ============================================================================\n// Token calculation\n// ============================================================================\n\n/**\n * Calculate total context tokens from usage.\n * Uses the native totalTokens field when available, falls back to computing from components.\n */\nexport function calculateContextTokens(usage: Usage): number {\n\treturn usage.totalTokens || usage.input + usage.output + usage.cacheRead + usage.cacheWrite;\n}\n\n/**\n * Get usage from an assistant message if available.\n * Skips aborted and error messages as they don't have valid usage data.\n */\nfunction getAssistantUsage(msg: AgentMessage): Usage | undefined {\n\tif (msg.role === \"assistant\" && \"usage\" in msg) {\n\t\tconst assistantMsg = msg as AssistantMessage;\n\t\tif (assistantMsg.stopReason !== \"aborted\" && assistantMsg.stopReason !== \"error\" && assistantMsg.usage) {\n\t\t\treturn assistantMsg.usage;\n\t\t}\n\t}\n\treturn undefined;\n}\n\n/**\n * Find the last non-aborted assistant message usage from session entries.\n */\nexport function getLastAssistantUsage(entries: SessionEntry[]): Usage | undefined {\n\tfor (let i = entries.length - 1; i >= 0; i--) {\n\t\tconst entry = entries[i];\n\t\tif (entry.type === \"message\") {\n\t\t\tconst usage = getAssistantUsage(entry.message);\n\t\t\tif (usage) return usage;\n\t\t}\n\t}\n\treturn undefined;\n}\n\nexport interface ContextUsageEstimate {\n\ttokens: number;\n\tusageTokens: number;\n\ttrailingTokens: number;\n\tlastUsageIndex: number | null;\n}\n\nfunction getLastAssistantUsageInfo(messages: AgentMessage[]): { usage: Usage; index: number } | undefined {\n\tfor (let i = messages.length - 1; i >= 0; i--) {\n\t\tconst usage = getAssistantUsage(messages[i]);\n\t\tif (usage) return { usage, index: i };\n\t}\n\treturn undefined;\n}\n\n/**\n * Estimate context tokens from messages, using the last assistant usage when available.\n * If there are messages after the last usage, estimate their tokens with estimateTokens.\n */\nexport function estimateContextTokens(messages: AgentMessage[]): ContextUsageEstimate {\n\tconst usageInfo = getLastAssistantUsageInfo(messages);\n\n\tif (!usageInfo) {\n\t\tlet estimated = 0;\n\t\tfor (const message of messages) {\n\t\t\testimated += estimateTokens(message);\n\t\t}\n\t\treturn {\n\t\t\ttokens: estimated,\n\t\t\tusageTokens: 0,\n\t\t\ttrailingTokens: estimated,\n\t\t\tlastUsageIndex: null,\n\t\t};\n\t}\n\n\tconst usageTokens = calculateContextTokens(usageInfo.usage);\n\tlet trailingTokens = 0;\n\tfor (let i = usageInfo.index + 1; i < messages.length; i++) {\n\t\ttrailingTokens += estimateTokens(messages[i]);\n\t}\n\n\treturn {\n\t\ttokens: usageTokens + trailingTokens,\n\t\tusageTokens,\n\t\ttrailingTokens,\n\t\tlastUsageIndex: usageInfo.index,\n\t};\n}\n\n/** Estimate context tokens after adding messages that have not been sent yet. */\nexport function estimateProjectedContextTokens(\n\tmessages: AgentMessage[],\n\tpendingMessages: AgentMessage[],\n): ContextUsageEstimate {\n\treturn estimateContextTokens([...messages, ...pendingMessages]);\n}\n\nexport type CompactionHeadroomLimit = \"max_usage_ratio\" | \"reserve_tokens\";\n\nexport interface CompactionHeadroomThreshold {\n\ttriggerTokens: number;\n\theadroomTokens: number;\n\tmaxUsageRatioTokens: number;\n\treserveBoundaryTokens: number;\n\tlimitedBy: CompactionHeadroomLimit;\n}\n\n/** Return the earliest compaction threshold from ratio-based headroom and absolute reserve. */\nexport function getCompactionHeadroomThreshold(\n\tcontextWindow: number,\n\tsettings: CompactionSettings,\n): CompactionHeadroomThreshold | undefined {\n\tif (!settings.enabled || !Number.isFinite(contextWindow) || contextWindow <= 0) {\n\t\treturn undefined;\n\t}\n\n\tconst windowTokens = Math.floor(contextWindow);\n\tconst reserveTokens = Number.isFinite(settings.reserveTokens) ? Math.max(0, Math.floor(settings.reserveTokens)) : 0;\n\tconst configuredMaxUsageRatio = settings.maxUsageRatio ?? DEFAULT_COMPACTION_MAX_USAGE_RATIO;\n\tconst maxUsageRatio =\n\t\tNumber.isFinite(configuredMaxUsageRatio) && configuredMaxUsageRatio > 0 && configuredMaxUsageRatio < 1\n\t\t\t? configuredMaxUsageRatio\n\t\t\t: DEFAULT_COMPACTION_MAX_USAGE_RATIO;\n\tconst maxUsageRatioTokens = Math.max(1, Math.floor(windowTokens * maxUsageRatio));\n\tconst reserveBoundaryTokens =\n\t\treserveTokens >= windowTokens ? maxUsageRatioTokens : Math.max(1, windowTokens - reserveTokens);\n\tconst triggerTokens = Math.min(maxUsageRatioTokens, reserveBoundaryTokens);\n\n\treturn {\n\t\ttriggerTokens,\n\t\theadroomTokens: windowTokens - triggerTokens,\n\t\tmaxUsageRatioTokens,\n\t\treserveBoundaryTokens,\n\t\tlimitedBy: reserveBoundaryTokens <= maxUsageRatioTokens ? \"reserve_tokens\" : \"max_usage_ratio\",\n\t};\n}\n\n/**\n * Check if compaction should trigger based on context usage.\n */\nexport function shouldCompact(contextTokens: number, contextWindow: number, settings: CompactionSettings): boolean {\n\tconst threshold = getCompactionHeadroomThreshold(contextWindow, settings);\n\treturn threshold !== undefined && Number.isFinite(contextTokens) && contextTokens >= threshold.triggerTokens;\n}\n\n// ============================================================================\n// Cut point detection\n// ============================================================================\n\nconst ESTIMATED_IMAGE_CHARS = 4800;\n\nfunction estimateTextAndImageContentChars(content: string | Array<{ type: string; text?: string }>): number {\n\tif (typeof content === \"string\") {\n\t\treturn content.length;\n\t}\n\n\tlet chars = 0;\n\tfor (const block of content) {\n\t\tif (block.type === \"text\" && block.text) {\n\t\t\tchars += block.text.length;\n\t\t} else if (block.type === \"image\") {\n\t\t\tchars += ESTIMATED_IMAGE_CHARS;\n\t\t}\n\t}\n\treturn chars;\n}\n\n/**\n * Estimate token count for a message using chars/4 heuristic.\n * This is conservative (overestimates tokens).\n */\nexport function estimateTokens(message: AgentMessage): number {\n\tlet chars = 0;\n\n\tswitch (message.role) {\n\t\tcase \"user\": {\n\t\t\tchars = estimateTextAndImageContentChars(\n\t\t\t\t(message as { content: string | Array<{ type: string; text?: string }> }).content,\n\t\t\t);\n\t\t\treturn Math.ceil(chars / 4);\n\t\t}\n\t\tcase \"assistant\": {\n\t\t\tconst assistant = message as AssistantMessage;\n\t\t\tfor (const block of assistant.content) {\n\t\t\t\tif (block.type === \"text\") {\n\t\t\t\t\tchars += block.text.length;\n\t\t\t\t} else if (block.type === \"thinking\") {\n\t\t\t\t\tchars += block.thinking.length;\n\t\t\t\t} else if (block.type === \"toolCall\") {\n\t\t\t\t\tchars += block.name.length + JSON.stringify(block.arguments).length;\n\t\t\t\t}\n\t\t\t}\n\t\t\treturn Math.ceil(chars / 4);\n\t\t}\n\t\tcase \"custom\":\n\t\tcase \"toolResult\": {\n\t\t\tchars = estimateTextAndImageContentChars(message.content);\n\t\t\treturn Math.ceil(chars / 4);\n\t\t}\n\t\tcase \"bashExecution\": {\n\t\t\tchars = message.command.length + message.output.length;\n\t\t\treturn Math.ceil(chars / 4);\n\t\t}\n\t\tcase \"branchSummary\":\n\t\tcase \"compactionSummary\": {\n\t\t\tchars = message.summary.length;\n\t\t\treturn Math.ceil(chars / 4);\n\t\t}\n\t}\n\n\treturn 0;\n}\n\n/**\n * Find valid cut points: indices of user, assistant, custom, or bashExecution messages.\n * Never cut at tool results (they must follow their tool call).\n * When we cut at an assistant message with tool calls, its tool results follow it\n * and will be kept.\n * BashExecutionMessage is treated like a user message (user-initiated context).\n */\nfunction findValidCutPoints(entries: SessionEntry[], startIndex: number, endIndex: number): number[] {\n\tconst cutPoints: number[] = [];\n\tfor (let i = startIndex; i < endIndex; i++) {\n\t\tconst entry = entries[i];\n\t\tswitch (entry.type) {\n\t\t\tcase \"message\": {\n\t\t\t\tconst role = entry.message.role;\n\t\t\t\tswitch (role) {\n\t\t\t\t\tcase \"bashExecution\":\n\t\t\t\t\tcase \"custom\":\n\t\t\t\t\tcase \"branchSummary\":\n\t\t\t\t\tcase \"compactionSummary\":\n\t\t\t\t\tcase \"user\":\n\t\t\t\t\tcase \"assistant\":\n\t\t\t\t\t\tcutPoints.push(i);\n\t\t\t\t\t\tbreak;\n\t\t\t\t\tcase \"toolResult\":\n\t\t\t\t\t\tbreak;\n\t\t\t\t}\n\t\t\t\tbreak;\n\t\t\t}\n\t\t\tcase \"thinking_level_change\":\n\t\t\tcase \"model_change\":\n\t\t\tcase \"compaction\":\n\t\t\tcase \"branch_summary\":\n\t\t\tcase \"custom\":\n\t\t\tcase \"custom_message\":\n\t\t\tcase \"label\":\n\t\t\tcase \"session_info\":\n\t\t\t\tbreak;\n\t\t}\n\n\t\t// branch_summary and custom_message are user-role messages, valid cut points\n\t\tif (entry.type === \"branch_summary\" || entry.type === \"custom_message\") {\n\t\t\tcutPoints.push(i);\n\t\t}\n\t}\n\treturn cutPoints;\n}\n\n/**\n * Find the user message (or bashExecution) that starts the turn containing the given entry index.\n * Returns -1 if no turn start found before the index.\n * BashExecutionMessage is treated like a user message for turn boundaries.\n */\nexport function findTurnStartIndex(entries: SessionEntry[], entryIndex: number, startIndex: number): number {\n\tfor (let i = entryIndex; i >= startIndex; i--) {\n\t\tconst entry = entries[i];\n\t\t// branch_summary and custom_message are user-role messages, can start a turn\n\t\tif (entry.type === \"branch_summary\" || entry.type === \"custom_message\") {\n\t\t\treturn i;\n\t\t}\n\t\tif (entry.type === \"message\") {\n\t\t\tconst role = entry.message.role;\n\t\t\tif (role === \"user\" || role === \"bashExecution\") {\n\t\t\t\treturn i;\n\t\t\t}\n\t\t}\n\t}\n\treturn -1;\n}\n\nexport interface CutPointResult {\n\t/** Index of first entry to keep */\n\tfirstKeptEntryIndex: number;\n\t/** Index of user message that starts the turn being split, or -1 if not splitting */\n\tturnStartIndex: number;\n\t/** Whether this cut splits a turn (cut point is not a user message) */\n\tisSplitTurn: boolean;\n}\n\n/**\n * Find the cut point in session entries that keeps approximately `keepRecentTokens`.\n *\n * Algorithm: Walk backwards from newest, accumulating estimated message sizes.\n * Stop when we've accumulated >= keepRecentTokens. Cut at that point.\n *\n * Can cut at user OR assistant messages (never tool results). When cutting at an\n * assistant message with tool calls, its tool results come after and will be kept.\n *\n * Returns CutPointResult with:\n * - firstKeptEntryIndex: the entry index to start keeping from\n * - turnStartIndex: if cutting mid-turn, the user message that started that turn\n * - isSplitTurn: whether we're cutting in the middle of a turn\n *\n * Only considers entries between `startIndex` and `endIndex` (exclusive).\n */\nexport function findCutPoint(\n\tentries: SessionEntry[],\n\tstartIndex: number,\n\tendIndex: number,\n\tkeepRecentTokens: number,\n): CutPointResult {\n\tconst cutPoints = findValidCutPoints(entries, startIndex, endIndex);\n\n\tif (cutPoints.length === 0) {\n\t\treturn { firstKeptEntryIndex: startIndex, turnStartIndex: -1, isSplitTurn: false };\n\t}\n\n\t// Walk backwards from newest, accumulating estimated message sizes\n\tlet accumulatedTokens = 0;\n\tlet cutIndex = cutPoints[0]; // Default: keep from first message (not header)\n\n\tfor (let i = endIndex - 1; i >= startIndex; i--) {\n\t\tconst entry = entries[i];\n\t\tif (entry.type !== \"message\") continue;\n\n\t\t// Estimate this message's size\n\t\tconst messageTokens = estimateTokens(entry.message);\n\t\taccumulatedTokens += messageTokens;\n\n\t\t// Check if we've exceeded the budget\n\t\tif (accumulatedTokens >= keepRecentTokens) {\n\t\t\t// Find the closest valid cut point at or after this entry\n\t\t\tfor (let c = 0; c < cutPoints.length; c++) {\n\t\t\t\tif (cutPoints[c] >= i) {\n\t\t\t\t\tcutIndex = cutPoints[c];\n\t\t\t\t\tbreak;\n\t\t\t\t}\n\t\t\t}\n\t\t\tbreak;\n\t\t}\n\t}\n\n\t// Scan backwards from cutIndex to include any non-message entries (bash, settings, etc.)\n\twhile (cutIndex > startIndex) {\n\t\tconst prevEntry = entries[cutIndex - 1];\n\t\t// Stop at session header or compaction boundaries\n\t\tif (prevEntry.type === \"compaction\") {\n\t\t\tbreak;\n\t\t}\n\t\tif (prevEntry.type === \"message\") {\n\t\t\t// Stop if we hit any message\n\t\t\tbreak;\n\t\t}\n\t\t// Include this non-message entry (bash, settings change, etc.)\n\t\tcutIndex--;\n\t}\n\n\t// Determine if this is a split turn\n\tconst cutEntry = entries[cutIndex];\n\tconst isUserMessage = cutEntry.type === \"message\" && cutEntry.message.role === \"user\";\n\tconst turnStartIndex = isUserMessage ? -1 : findTurnStartIndex(entries, cutIndex, startIndex);\n\n\treturn {\n\t\tfirstKeptEntryIndex: cutIndex,\n\t\tturnStartIndex,\n\t\tisSplitTurn: !isUserMessage && turnStartIndex !== -1,\n\t};\n}\n\n// ============================================================================\n// Summarization\n// ============================================================================\n\nconst SUMMARIZATION_PROMPT = `The messages above are a conversation to summarize. Create a structured context checkpoint summary that another LLM will use to continue the work.\n\nUse this EXACT format:\n\n## Goal\n[What is the user trying to accomplish? Can be multiple items if the session covers different tasks.]\n\n## Constraints & Preferences\n- [Any constraints, preferences, or requirements mentioned by user]\n- [Or \"(none)\" if none were mentioned]\n\n## Progress\n### Done\n- [x] [Completed tasks/changes]\n\n### In Progress\n- [ ] [Current work]\n\n### Blocked\n- [Issues preventing progress, if any]\n\n## Key Decisions\n- **[Decision]**: [Brief rationale]\n\n## Next Steps\n1. [Ordered list of what should happen next]\n\n## Critical Context\n- [Any data, examples, or references needed to continue]\n- [Or \"(none)\" if not applicable]\n\nKeep each section concise. Preserve exact file paths, function names, and error messages.`;\n\nconst UPDATE_SUMMARIZATION_PROMPT = `The messages above are NEW conversation messages to incorporate into the existing summary provided in <previous-summary> tags.\n\nUpdate the existing structured summary with new information. RULES:\n- PRESERVE all existing information from the previous summary\n- ADD new progress, decisions, and context from the new messages\n- UPDATE the Progress section: move items from \"In Progress\" to \"Done\" when completed\n- UPDATE \"Next Steps\" based on what was accomplished\n- PRESERVE exact file paths, function names, and error messages\n- If something is no longer relevant, you may remove it\n\nUse this EXACT format:\n\n## Goal\n[Preserve existing goals, add new ones if the task expanded]\n\n## Constraints & Preferences\n- [Preserve existing, add new ones discovered]\n\n## Progress\n### Done\n- [x] [Include previously done items AND newly completed items]\n\n### In Progress\n- [ ] [Current work - update based on progress]\n\n### Blocked\n- [Current blockers - remove if resolved]\n\n## Key Decisions\n- **[Decision]**: [Brief rationale] (preserve all previous, add new)\n\n## Next Steps\n1. [Update based on current state]\n\n## Critical Context\n- [Preserve important context, add new if needed]\n\nKeep each section concise. Preserve exact file paths, function names, and error messages.`;\n\nfunction createSummarizationOptions(\n\tmodel: Model<any>,\n\tmaxTokens: number,\n\tapiKey: string | undefined,\n\theaders: Record<string, string> | undefined,\n\tsignal: AbortSignal | undefined,\n\tthinkingLevel: ThinkingLevel | undefined,\n): SimpleStreamOptions {\n\tconst options: SimpleStreamOptions = { maxTokens, signal, apiKey, headers };\n\tif (model.reasoning && thinkingLevel && thinkingLevel !== \"off\") {\n\t\toptions.reasoning = thinkingLevel;\n\t}\n\treturn options;\n}\n\nasync function completeSummarization(\n\tmodel: Model<any>,\n\tcontext: Context,\n\toptions: SimpleStreamOptions,\n\tstreamFn?: StreamFn,\n): Promise<AssistantMessage> {\n\tif (!streamFn) {\n\t\treturn completeSimple(model, context, options);\n\t}\n\tconst stream = await streamFn(model, context, options);\n\treturn stream.result();\n}\n\n/**\n * Generate a summary of the conversation using the LLM.\n * If previousSummary is provided, uses the update prompt to merge.\n */\nexport async function generateSummary(\n\tcurrentMessages: AgentMessage[],\n\tmodel: Model<any>,\n\treserveTokens: number,\n\tapiKey: string | undefined,\n\theaders?: Record<string, string>,\n\tsignal?: AbortSignal,\n\tcustomInstructions?: string,\n\tpreviousSummary?: string,\n\tthinkingLevel?: ThinkingLevel,\n\tstreamFn?: StreamFn,\n): Promise<string> {\n\tconst maxTokens = Math.min(\n\t\tMath.floor(0.8 * reserveTokens),\n\t\tmodel.maxTokens > 0 ? model.maxTokens : Number.POSITIVE_INFINITY,\n\t);\n\n\t// Use update prompt if we have a previous summary, otherwise initial prompt\n\tlet basePrompt = previousSummary ? UPDATE_SUMMARIZATION_PROMPT : SUMMARIZATION_PROMPT;\n\tif (customInstructions) {\n\t\tbasePrompt = `${basePrompt}\\n\\nAdditional focus: ${customInstructions}`;\n\t}\n\n\t// Serialize conversation to text so model doesn't try to continue it\n\t// Convert to LLM messages first (handles custom types like bashExecution, custom, etc.)\n\tconst llmMessages = convertToLlm(currentMessages);\n\tconst conversationText = serializeConversation(llmMessages);\n\n\t// Build the prompt with conversation wrapped in tags\n\tlet promptText = `<conversation>\\n${conversationText}\\n</conversation>\\n\\n`;\n\tif (previousSummary) {\n\t\tpromptText += `<previous-summary>\\n${previousSummary}\\n</previous-summary>\\n\\n`;\n\t}\n\tpromptText += basePrompt;\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\tconst completionOptions = createSummarizationOptions(model, maxTokens, apiKey, headers, signal, thinkingLevel);\n\n\tconst response = await completeSummarization(\n\t\tmodel,\n\t\t{ systemPrompt: SUMMARIZATION_SYSTEM_PROMPT, messages: summarizationMessages },\n\t\tcompletionOptions,\n\t\tstreamFn,\n\t);\n\n\tif (response.stopReason === \"error\") {\n\t\tthrow new Error(`Summarization failed: ${response.errorMessage || \"Unknown error\"}`);\n\t}\n\n\tconst textContent = response.content\n\t\t.filter((c): c is { type: \"text\"; text: string } => c.type === \"text\")\n\t\t.map((c) => c.text)\n\t\t.join(\"\\n\");\n\n\treturn textContent;\n}\n\n// ============================================================================\n// Compaction Preparation (for extensions)\n// ============================================================================\n\nexport interface CompactionPreparation {\n\t/** UUID of first entry to keep */\n\tfirstKeptEntryId: string;\n\t/** Messages that will be summarized and discarded */\n\tmessagesToSummarize: AgentMessage[];\n\t/** Messages that will be turned into turn prefix summary (if splitting) */\n\tturnPrefixMessages: AgentMessage[];\n\t/** Whether this is a split turn (cut point in middle of turn) */\n\tisSplitTurn: boolean;\n\ttokensBefore: number;\n\t/** Summary from previous compaction, for iterative update */\n\tpreviousSummary?: string;\n\t/** File operations extracted from messagesToSummarize */\n\tfileOps: FileOperations;\n\t/** Compaction settions from settings.jsonl\t*/\n\tsettings: CompactionSettings;\n}\n\nexport function prepareCompaction(\n\tpathEntries: SessionEntry[],\n\tsettings: CompactionSettings,\n): CompactionPreparation | undefined {\n\tif (pathEntries.length > 0 && pathEntries[pathEntries.length - 1].type === \"compaction\") {\n\t\treturn undefined;\n\t}\n\n\tlet prevCompactionIndex = -1;\n\tfor (let i = pathEntries.length - 1; i >= 0; i--) {\n\t\tif (pathEntries[i].type === \"compaction\") {\n\t\t\tprevCompactionIndex = i;\n\t\t\tbreak;\n\t\t}\n\t}\n\n\tlet previousSummary: string | undefined;\n\tlet boundaryStart = 0;\n\tif (prevCompactionIndex >= 0) {\n\t\tconst prevCompaction = pathEntries[prevCompactionIndex] as CompactionEntry;\n\t\tpreviousSummary = prevCompaction.summary;\n\t\tconst firstKeptEntryIndex = pathEntries.findIndex((entry) => entry.id === prevCompaction.firstKeptEntryId);\n\t\tboundaryStart = firstKeptEntryIndex >= 0 ? firstKeptEntryIndex : prevCompactionIndex + 1;\n\t}\n\tconst boundaryEnd = pathEntries.length;\n\n\tconst tokensBefore = estimateContextTokens(buildSessionContext(pathEntries).messages).tokens;\n\n\tconst cutPoint = findCutPoint(pathEntries, boundaryStart, boundaryEnd, settings.keepRecentTokens);\n\n\t// Get UUID of first kept entry\n\tconst firstKeptEntry = pathEntries[cutPoint.firstKeptEntryIndex];\n\tif (!firstKeptEntry?.id) {\n\t\treturn undefined; // Session needs migration\n\t}\n\tconst firstKeptEntryId = firstKeptEntry.id;\n\n\tconst historyEnd = cutPoint.isSplitTurn ? cutPoint.turnStartIndex : cutPoint.firstKeptEntryIndex;\n\n\t// Messages to summarize (will be discarded after summary)\n\tconst messagesToSummarize: AgentMessage[] = [];\n\tfor (let i = boundaryStart; i < historyEnd; i++) {\n\t\tconst msg = getMessageFromEntryForCompaction(pathEntries[i]);\n\t\tif (msg) messagesToSummarize.push(msg);\n\t}\n\n\t// Messages for turn prefix summary (if splitting a turn)\n\tconst turnPrefixMessages: AgentMessage[] = [];\n\tif (cutPoint.isSplitTurn) {\n\t\tfor (let i = cutPoint.turnStartIndex; i < cutPoint.firstKeptEntryIndex; i++) {\n\t\t\tconst msg = getMessageFromEntryForCompaction(pathEntries[i]);\n\t\t\tif (msg) turnPrefixMessages.push(msg);\n\t\t}\n\t}\n\n\t// Extract file operations from messages and previous compaction\n\tconst fileOps = extractFileOperations(messagesToSummarize, pathEntries, prevCompactionIndex);\n\n\t// Also extract file ops from turn prefix if splitting\n\tif (cutPoint.isSplitTurn) {\n\t\tfor (const msg of turnPrefixMessages) {\n\t\t\textractFileOpsFromMessage(msg, fileOps);\n\t\t}\n\t}\n\n\treturn {\n\t\tfirstKeptEntryId,\n\t\tmessagesToSummarize,\n\t\tturnPrefixMessages,\n\t\tisSplitTurn: cutPoint.isSplitTurn,\n\t\ttokensBefore,\n\t\tpreviousSummary,\n\t\tfileOps,\n\t\tsettings,\n\t};\n}\n\n// ============================================================================\n// Main compaction function\n// ============================================================================\n\nconst TURN_PREFIX_SUMMARIZATION_PROMPT = `This is the PREFIX of a turn that was too large to keep. The SUFFIX (recent work) is retained.\n\nSummarize the prefix to provide context for the retained suffix:\n\n## Original Request\n[What did the user ask for in this turn?]\n\n## Early Progress\n- [Key decisions and work done in the prefix]\n\n## Context for Suffix\n- [Information needed to understand the retained recent work]\n\nBe concise. Focus on what's needed to understand the kept suffix.`;\n\n/**\n * Generate summaries for compaction using prepared data.\n * Returns CompactionResult - SessionManager adds uuid/parentUuid when saving.\n *\n * @param preparation - Pre-calculated preparation from prepareCompaction()\n * @param customInstructions - Optional custom focus for the summary\n */\nexport async function compact(\n\tpreparation: CompactionPreparation,\n\tmodel: Model<any>,\n\tapiKey: string | undefined,\n\theaders?: Record<string, string>,\n\tcustomInstructions?: string,\n\tsignal?: AbortSignal,\n\tthinkingLevel?: ThinkingLevel,\n\tstreamFn?: StreamFn,\n): Promise<CompactionResult> {\n\tconst {\n\t\tfirstKeptEntryId,\n\t\tmessagesToSummarize,\n\t\tturnPrefixMessages,\n\t\tisSplitTurn,\n\t\ttokensBefore,\n\t\tpreviousSummary,\n\t\tfileOps,\n\t\tsettings,\n\t} = preparation;\n\n\t// Generate summaries (can be parallel if both needed) and merge into one\n\tlet summary: string;\n\n\tif (isSplitTurn && turnPrefixMessages.length > 0) {\n\t\t// Generate both summaries in parallel\n\t\tconst [historyResult, turnPrefixResult] = await Promise.all([\n\t\t\tmessagesToSummarize.length > 0\n\t\t\t\t? generateSummary(\n\t\t\t\t\t\tmessagesToSummarize,\n\t\t\t\t\t\tmodel,\n\t\t\t\t\t\tsettings.reserveTokens,\n\t\t\t\t\t\tapiKey,\n\t\t\t\t\t\theaders,\n\t\t\t\t\t\tsignal,\n\t\t\t\t\t\tcustomInstructions,\n\t\t\t\t\t\tpreviousSummary,\n\t\t\t\t\t\tthinkingLevel,\n\t\t\t\t\t\tstreamFn,\n\t\t\t\t\t)\n\t\t\t\t: Promise.resolve(\"No prior history.\"),\n\t\t\tgenerateTurnPrefixSummary(\n\t\t\t\tturnPrefixMessages,\n\t\t\t\tmodel,\n\t\t\t\tsettings.reserveTokens,\n\t\t\t\tapiKey,\n\t\t\t\theaders,\n\t\t\t\tsignal,\n\t\t\t\tthinkingLevel,\n\t\t\t\tstreamFn,\n\t\t\t),\n\t\t]);\n\t\t// Merge into single summary\n\t\tsummary = `${historyResult}\\n\\n---\\n\\n**Turn Context (split turn):**\\n\\n${turnPrefixResult}`;\n\t} else {\n\t\t// Just generate history summary\n\t\tsummary = await generateSummary(\n\t\t\tmessagesToSummarize,\n\t\t\tmodel,\n\t\t\tsettings.reserveTokens,\n\t\t\tapiKey,\n\t\t\theaders,\n\t\t\tsignal,\n\t\t\tcustomInstructions,\n\t\t\tpreviousSummary,\n\t\t\tthinkingLevel,\n\t\t\tstreamFn,\n\t\t);\n\t}\n\n\t// Compute file lists and append to summary\n\tconst { readFiles, modifiedFiles } = computeFileLists(fileOps);\n\tsummary += formatFileOperations(readFiles, modifiedFiles);\n\n\tif (!firstKeptEntryId) {\n\t\tthrow new Error(\"First kept entry has no UUID - session may need migration\");\n\t}\n\n\treturn {\n\t\tsummary,\n\t\tfirstKeptEntryId,\n\t\ttokensBefore,\n\t\tdetails: { readFiles, modifiedFiles } as CompactionDetails,\n\t};\n}\n\n/**\n * Generate a summary for a turn prefix (when splitting a turn).\n */\nasync function generateTurnPrefixSummary(\n\tmessages: AgentMessage[],\n\tmodel: Model<any>,\n\treserveTokens: number,\n\tapiKey: string | undefined,\n\theaders?: Record<string, string>,\n\tsignal?: AbortSignal,\n\tthinkingLevel?: ThinkingLevel,\n\tstreamFn?: StreamFn,\n): Promise<string> {\n\tconst maxTokens = Math.min(\n\t\tMath.floor(0.5 * reserveTokens),\n\t\tmodel.maxTokens > 0 ? model.maxTokens : Number.POSITIVE_INFINITY,\n\t); // Smaller budget for turn prefix\n\tconst llmMessages = convertToLlm(messages);\n\tconst conversationText = serializeConversation(llmMessages);\n\tconst promptText = `<conversation>\\n${conversationText}\\n</conversation>\\n\\n${TURN_PREFIX_SUMMARIZATION_PROMPT}`;\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\tconst response = await completeSummarization(\n\t\tmodel,\n\t\t{ systemPrompt: SUMMARIZATION_SYSTEM_PROMPT, messages: summarizationMessages },\n\t\tcreateSummarizationOptions(model, maxTokens, apiKey, headers, signal, thinkingLevel),\n\t\tstreamFn,\n\t);\n\n\tif (response.stopReason === \"error\") {\n\t\tthrow new Error(`Turn prefix summarization failed: ${response.errorMessage || \"Unknown error\"}`);\n\t}\n\n\treturn response.content\n\t\t.filter((c): c is { type: \"text\"; text: string } => c.type === \"text\")\n\t\t.map((c) => c.text)\n\t\t.join(\"\\n\");\n}\n"]}
+{"version":3,"file":"compaction.d.ts","sourceRoot":"","sources":["../../../src/core/compaction/compaction.ts"],"names":[],"mappings":"AAAA;;;;;GAKG;AAEH,OAAO,KAAK,EAAE,YAAY,EAAE,QAAQ,EAAE,aAAa,EAAE,MAAM,gBAAgB,CAAC;AAC5E,OAAO,KAAK,EAA6B,KAAK,EAAuB,KAAK,EAAE,MAAM,QAAQ,CAAC;AAS3F,OAAO,EAA6C,KAAK,YAAY,EAAE,MAAM,uBAAuB,CAAC;AACrG,OAAO,EAIN,KAAK,cAAc,EAInB,MAAM,YAAY,CAAC;AAMpB,kEAAkE;AAClE,MAAM,WAAW,iBAAiB;IACjC,SAAS,EAAE,MAAM,EAAE,CAAC;IACpB,aAAa,EAAE,MAAM,EAAE,CAAC;CACxB;AAiED,8EAA8E;AAC9E,MAAM,WAAW,gBAAgB,CAAC,CAAC,GAAG,OAAO;IAC5C,OAAO,EAAE,MAAM,CAAC;IAChB,gBAAgB,EAAE,MAAM,CAAC;IACzB,YAAY,EAAE,MAAM,CAAC;IACrB,+FAA+F;IAC/F,OAAO,CAAC,EAAE,CAAC,CAAC;CACZ;AAMD,MAAM,WAAW,kBAAkB;IAClC,OAAO,EAAE,OAAO,CAAC;IACjB,aAAa,EAAE,MAAM,CAAC;IACtB,oBAAoB,CAAC,EAAE,MAAM,CAAC;IAC9B,wBAAwB,CAAC,EAAE,MAAM,CAAC;IAClC,kBAAkB,CAAC,EAAE,MAAM,CAAC;IAC5B,kBAAkB,CAAC,EAAE,MAAM,CAAC;IAC5B,gBAAgB,EAAE,MAAM,CAAC;IACzB,qFAAqF;IACrF,aAAa,CAAC,EAAE,MAAM,CAAC;IACvB,UAAU,CAAC,EAAE,MAAM,CAAC;IACpB,cAAc,CAAC,EAAE,MAAM,CAAC;CACxB;AAED,eAAO,MAAM,kCAAkC,MAAM,CAAC;AAEtD,eAAO,MAAM,2BAA2B,EAAE,kBAKzC,CAAC;AAMF;;;GAGG;AACH,wBAAgB,sBAAsB,CAAC,KAAK,EAAE,KAAK,GAAG,MAAM,CAE3D;AAgBD;;GAEG;AACH,wBAAgB,qBAAqB,CAAC,OAAO,EAAE,YAAY,EAAE,GAAG,KAAK,GAAG,SAAS,CAShF;AAED,MAAM,WAAW,oBAAoB;IACpC,MAAM,EAAE,MAAM,CAAC;IACf,WAAW,EAAE,MAAM,CAAC;IACpB,cAAc,EAAE,MAAM,CAAC;IACvB,cAAc,EAAE,MAAM,GAAG,IAAI,CAAC;CAC9B;AAUD;;;GAGG;AACH,wBAAgB,qBAAqB,CAAC,QAAQ,EAAE,YAAY,EAAE,GAAG,oBAAoB,CA4BpF;AAED,iFAAiF;AACjF,wBAAgB,8BAA8B,CAC7C,QAAQ,EAAE,YAAY,EAAE,EACxB,eAAe,EAAE,YAAY,EAAE,GAC7B,oBAAoB,CAEtB;AAED,MAAM,MAAM,uBAAuB,GAAG,iBAAiB,GAAG,gBAAgB,CAAC;AAE3E,MAAM,WAAW,2BAA2B;IAC3C,aAAa,EAAE,MAAM,CAAC;IACtB,cAAc,EAAE,MAAM,CAAC;IACvB,mBAAmB,EAAE,MAAM,CAAC;IAC5B,qBAAqB,EAAE,MAAM,CAAC;IAC9B,SAAS,EAAE,uBAAuB,CAAC;CACnC;AAED,+FAA+F;AAC/F,wBAAgB,8BAA8B,CAC7C,aAAa,EAAE,MAAM,EACrB,QAAQ,EAAE,kBAAkB,GAC1B,2BAA2B,GAAG,SAAS,CAkCzC;AAED;;GAEG;AACH,wBAAgB,aAAa,CAAC,aAAa,EAAE,MAAM,EAAE,aAAa,EAAE,MAAM,EAAE,QAAQ,EAAE,kBAAkB,GAAG,OAAO,CAGjH;AAwBD;;;GAGG;AACH,wBAAgB,cAAc,CAAC,OAAO,EAAE,YAAY,GAAG,MAAM,CAwC5D;AAiDD;;;;GAIG;AACH,wBAAgB,kBAAkB,CAAC,OAAO,EAAE,YAAY,EAAE,EAAE,UAAU,EAAE,MAAM,EAAE,UAAU,EAAE,MAAM,GAAG,MAAM,CAe1G;AAED,MAAM,WAAW,cAAc;IAC9B,mCAAmC;IACnC,mBAAmB,EAAE,MAAM,CAAC;IAC5B,qFAAqF;IACrF,cAAc,EAAE,MAAM,CAAC;IACvB,uEAAuE;IACvE,WAAW,EAAE,OAAO,CAAC;CACrB;AAED;;;;;;;;;;;;;;;GAeG;AACH,wBAAgB,YAAY,CAC3B,OAAO,EAAE,YAAY,EAAE,EACvB,UAAU,EAAE,MAAM,EAClB,QAAQ,EAAE,MAAM,EAChB,gBAAgB,EAAE,MAAM,GACtB,cAAc,CAyDhB;AA0GD;;;GAGG;AACH,wBAAsB,eAAe,CACpC,eAAe,EAAE,YAAY,EAAE,EAC/B,KAAK,EAAE,KAAK,CAAC,GAAG,CAAC,EACjB,aAAa,EAAE,MAAM,EACrB,MAAM,EAAE,MAAM,GAAG,SAAS,EAC1B,OAAO,CAAC,EAAE,MAAM,CAAC,MAAM,EAAE,MAAM,CAAC,EAChC,MAAM,CAAC,EAAE,WAAW,EACpB,kBAAkB,CAAC,EAAE,MAAM,EAC3B,eAAe,CAAC,EAAE,MAAM,EACxB,aAAa,CAAC,EAAE,aAAa,EAC7B,QAAQ,CAAC,EAAE,QAAQ,GACjB,OAAO,CAAC,MAAM,CAAC,CAmDjB;AAMD,MAAM,WAAW,qBAAqB;IACrC,kCAAkC;IAClC,gBAAgB,EAAE,MAAM,CAAC;IACzB,qDAAqD;IACrD,mBAAmB,EAAE,YAAY,EAAE,CAAC;IACpC,2EAA2E;IAC3E,kBAAkB,EAAE,YAAY,EAAE,CAAC;IACnC,iEAAiE;IACjE,WAAW,EAAE,OAAO,CAAC;IACrB,YAAY,EAAE,MAAM,CAAC;IACrB,6DAA6D;IAC7D,eAAe,CAAC,EAAE,MAAM,CAAC;IACzB,yDAAyD;IACzD,OAAO,EAAE,cAAc,CAAC;IACxB,8CAA8C;IAC9C,QAAQ,EAAE,kBAAkB,CAAC;CAC7B;AAED,wBAAgB,iBAAiB,CAChC,WAAW,EAAE,YAAY,EAAE,EAC3B,QAAQ,EAAE,kBAAkB,GAC1B,qBAAqB,GAAG,SAAS,CAwEnC;AAqBD;;;;;;GAMG;AACH,wBAAsB,OAAO,CAC5B,WAAW,EAAE,qBAAqB,EAClC,KAAK,EAAE,KAAK,CAAC,GAAG,CAAC,EACjB,MAAM,EAAE,MAAM,GAAG,SAAS,EAC1B,OAAO,CAAC,EAAE,MAAM,CAAC,MAAM,EAAE,MAAM,CAAC,EAChC,kBAAkB,CAAC,EAAE,MAAM,EAC3B,MAAM,CAAC,EAAE,WAAW,EACpB,aAAa,CAAC,EAAE,aAAa,EAC7B,QAAQ,CAAC,EAAE,QAAQ,GACjB,OAAO,CAAC,gBAAgB,CAAC,CA2E3B","sourcesContent":["/**\n * Context compaction for long sessions.\n *\n * Pure functions for compaction logic. The session manager handles I/O,\n * and after compaction the session is reloaded.\n */\n\nimport type { AgentMessage, StreamFn, ThinkingLevel } from \"omk-agent-core\";\nimport type { AssistantMessage, Context, Model, SimpleStreamOptions, Usage } from \"omk-ai\";\nimport { completeSimple } from \"omk-ai\";\nimport { computeReservedTokenBudget } from \"../context-budget-reserved-tokens.ts\";\nimport {\n\tconvertToLlm,\n\tcreateBranchSummaryMessage,\n\tcreateCompactionSummaryMessage,\n\tcreateCustomMessage,\n} from \"../messages.ts\";\nimport { buildSessionContext, type CompactionEntry, type SessionEntry } from \"../session-manager.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// File Operation Tracking\n// ============================================================================\n\n/** Details stored in CompactionEntry.details for file tracking */\nexport interface CompactionDetails {\n\treadFiles: string[];\n\tmodifiedFiles: string[];\n}\n\n/**\n * Extract file operations from messages and previous compaction entries.\n */\nfunction extractFileOperations(\n\tmessages: AgentMessage[],\n\tentries: SessionEntry[],\n\tprevCompactionIndex: number,\n): FileOperations {\n\tconst fileOps = createFileOps();\n\n\tif (prevCompactionIndex >= 0) {\n\t\tconst prevCompaction = entries[prevCompactionIndex] as CompactionEntry;\n\t\tif (!prevCompaction.fromHook && prevCompaction.details) {\n\t\t\t// fromHook field kept for session file compatibility\n\t\t\tconst details = prevCompaction.details as CompactionDetails;\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\tfor (const f of details.modifiedFiles) fileOps.edited.add(f);\n\t\t\t}\n\t\t}\n\t}\n\n\t// Extract from tool calls in messages\n\tfor (const msg of messages) {\n\t\textractFileOpsFromMessage(msg, fileOps);\n\t}\n\n\treturn fileOps;\n}\n\n// ============================================================================\n// Message Extraction\n// ============================================================================\n\n/**\n * Extract AgentMessage from an entry if it produces one.\n * Returns undefined for entries that don't contribute to LLM context.\n */\nfunction getMessageFromEntry(entry: SessionEntry): AgentMessage | undefined {\n\tif (entry.type === \"message\") {\n\t\treturn entry.message;\n\t}\n\tif (entry.type === \"custom_message\") {\n\t\treturn createCustomMessage(entry.customType, entry.content, entry.display, entry.details, entry.timestamp);\n\t}\n\tif (entry.type === \"branch_summary\") {\n\t\treturn createBranchSummaryMessage(entry.summary, entry.fromId, entry.timestamp);\n\t}\n\tif (entry.type === \"compaction\") {\n\t\treturn createCompactionSummaryMessage(entry.summary, entry.tokensBefore, entry.timestamp);\n\t}\n\treturn undefined;\n}\n\nfunction getMessageFromEntryForCompaction(entry: SessionEntry): AgentMessage | undefined {\n\tif (entry.type === \"compaction\") {\n\t\treturn undefined;\n\t}\n\treturn getMessageFromEntry(entry);\n}\n\n/** Result from compact() - SessionManager adds uuid/parentUuid when saving */\nexport interface CompactionResult<T = unknown> {\n\tsummary: string;\n\tfirstKeptEntryId: string;\n\ttokensBefore: number;\n\t/** Extension-specific data (e.g., ArtifactIndex, version markers for structured compaction) */\n\tdetails?: T;\n}\n\n// ============================================================================\n// Types\n// ============================================================================\n\nexport interface CompactionSettings {\n\tenabled: boolean;\n\treserveTokens: number;\n\treservedOutputTokens?: number;\n\treservedToolResultTokens?: number;\n\tsafetyMarginTokens?: number;\n\timageReserveTokens?: number;\n\tkeepRecentTokens: number;\n\t/** Maximum fraction of the context window to use before compaction. Default: 0.9. */\n\tmaxUsageRatio?: number;\n\trearmRatio?: number;\n\temergencyRatio?: number;\n}\n\nexport const DEFAULT_COMPACTION_MAX_USAGE_RATIO = 0.9;\n\nexport const DEFAULT_COMPACTION_SETTINGS: CompactionSettings = {\n\tenabled: true,\n\treserveTokens: 16384,\n\tkeepRecentTokens: 20000,\n\tmaxUsageRatio: DEFAULT_COMPACTION_MAX_USAGE_RATIO,\n};\n\n// ============================================================================\n// Token calculation\n// ============================================================================\n\n/**\n * Calculate total context tokens from usage.\n * Uses the native totalTokens field when available, falls back to computing from components.\n */\nexport function calculateContextTokens(usage: Usage): number {\n\treturn usage.totalTokens || usage.input + usage.output + usage.cacheRead + usage.cacheWrite;\n}\n\n/**\n * Get usage from an assistant message if available.\n * Skips aborted and error messages as they don't have valid usage data.\n */\nfunction getAssistantUsage(msg: AgentMessage): Usage | undefined {\n\tif (msg.role === \"assistant\" && \"usage\" in msg) {\n\t\tconst assistantMsg = msg as AssistantMessage;\n\t\tif (assistantMsg.stopReason !== \"aborted\" && assistantMsg.stopReason !== \"error\" && assistantMsg.usage) {\n\t\t\treturn assistantMsg.usage;\n\t\t}\n\t}\n\treturn undefined;\n}\n\n/**\n * Find the last non-aborted assistant message usage from session entries.\n */\nexport function getLastAssistantUsage(entries: SessionEntry[]): Usage | undefined {\n\tfor (let i = entries.length - 1; i >= 0; i--) {\n\t\tconst entry = entries[i];\n\t\tif (entry.type === \"message\") {\n\t\t\tconst usage = getAssistantUsage(entry.message);\n\t\t\tif (usage) return usage;\n\t\t}\n\t}\n\treturn undefined;\n}\n\nexport interface ContextUsageEstimate {\n\ttokens: number;\n\tusageTokens: number;\n\ttrailingTokens: number;\n\tlastUsageIndex: number | null;\n}\n\nfunction getLastAssistantUsageInfo(messages: AgentMessage[]): { usage: Usage; index: number } | undefined {\n\tfor (let i = messages.length - 1; i >= 0; i--) {\n\t\tconst usage = getAssistantUsage(messages[i]);\n\t\tif (usage) return { usage, index: i };\n\t}\n\treturn undefined;\n}\n\n/**\n * Estimate context tokens from messages, using the last assistant usage when available.\n * If there are messages after the last usage, estimate their tokens with estimateTokens.\n */\nexport function estimateContextTokens(messages: AgentMessage[]): ContextUsageEstimate {\n\tconst usageInfo = getLastAssistantUsageInfo(messages);\n\n\tif (!usageInfo) {\n\t\tlet estimated = 0;\n\t\tfor (const message of messages) {\n\t\t\testimated += estimateTokens(message);\n\t\t}\n\t\treturn {\n\t\t\ttokens: estimated,\n\t\t\tusageTokens: 0,\n\t\t\ttrailingTokens: estimated,\n\t\t\tlastUsageIndex: null,\n\t\t};\n\t}\n\n\tconst usageTokens = calculateContextTokens(usageInfo.usage);\n\tlet trailingTokens = 0;\n\tfor (let i = usageInfo.index + 1; i < messages.length; i++) {\n\t\ttrailingTokens += estimateTokens(messages[i]);\n\t}\n\n\treturn {\n\t\ttokens: usageTokens + trailingTokens,\n\t\tusageTokens,\n\t\ttrailingTokens,\n\t\tlastUsageIndex: usageInfo.index,\n\t};\n}\n\n/** Estimate context tokens after adding messages that have not been sent yet. */\nexport function estimateProjectedContextTokens(\n\tmessages: AgentMessage[],\n\tpendingMessages: AgentMessage[],\n): ContextUsageEstimate {\n\treturn estimateContextTokens([...messages, ...pendingMessages]);\n}\n\nexport type CompactionHeadroomLimit = \"max_usage_ratio\" | \"reserve_tokens\";\n\nexport interface CompactionHeadroomThreshold {\n\ttriggerTokens: number;\n\theadroomTokens: number;\n\tmaxUsageRatioTokens: number;\n\treserveBoundaryTokens: number;\n\tlimitedBy: CompactionHeadroomLimit;\n}\n\n/** Return the earliest compaction threshold from ratio-based headroom and absolute reserve. */\nexport function getCompactionHeadroomThreshold(\n\tcontextWindow: number,\n\tsettings: CompactionSettings,\n): CompactionHeadroomThreshold | undefined {\n\tif (!settings.enabled || !Number.isFinite(contextWindow) || contextWindow <= 0) {\n\t\treturn undefined;\n\t}\n\n\tconst windowTokens = Math.floor(contextWindow);\n\tconst reserveTokens = Number.isFinite(settings.reserveTokens) ? Math.max(0, Math.floor(settings.reserveTokens)) : 0;\n\tconst configuredMaxUsageRatio = settings.maxUsageRatio ?? DEFAULT_COMPACTION_MAX_USAGE_RATIO;\n\tconst maxUsageRatio =\n\t\tNumber.isFinite(configuredMaxUsageRatio) && configuredMaxUsageRatio > 0 && configuredMaxUsageRatio < 1\n\t\t\t? configuredMaxUsageRatio\n\t\t\t: DEFAULT_COMPACTION_MAX_USAGE_RATIO;\n\tconst maxUsageRatioTokens = Math.max(1, Math.floor(windowTokens * maxUsageRatio));\n\tconst reservedBudget = computeReservedTokenBudget({\n\t\tmodelContextWindow: windowTokens,\n\t\tsystemPromptTokens: 0,\n\t\treservedOutputTokens: settings.reservedOutputTokens ?? reserveTokens,\n\t\treservedToolResultTokens: settings.reservedToolResultTokens ?? 0,\n\t\tsafetyMarginTokens: settings.safetyMarginTokens ?? 0,\n\t\timageReserveTokens: settings.imageReserveTokens ?? 0,\n\t});\n\tconst reserveBoundaryTokens =\n\t\treservedBudget.overflow || reservedBudget.effectiveBudget === 0\n\t\t\t? maxUsageRatioTokens\n\t\t\t: Math.max(1, reservedBudget.effectiveBudget);\n\tconst triggerTokens = Math.min(maxUsageRatioTokens, reserveBoundaryTokens);\n\n\treturn {\n\t\ttriggerTokens,\n\t\theadroomTokens: windowTokens - triggerTokens,\n\t\tmaxUsageRatioTokens,\n\t\treserveBoundaryTokens,\n\t\tlimitedBy: reserveBoundaryTokens <= maxUsageRatioTokens ? \"reserve_tokens\" : \"max_usage_ratio\",\n\t};\n}\n\n/**\n * Check if compaction should trigger based on context usage.\n */\nexport function shouldCompact(contextTokens: number, contextWindow: number, settings: CompactionSettings): boolean {\n\tconst threshold = getCompactionHeadroomThreshold(contextWindow, settings);\n\treturn threshold !== undefined && Number.isFinite(contextTokens) && contextTokens >= threshold.triggerTokens;\n}\n\n// ============================================================================\n// Cut point detection\n// ============================================================================\n\nconst ESTIMATED_IMAGE_CHARS = 4800;\n\nfunction estimateTextAndImageContentChars(content: string | Array<{ type: string; text?: string }>): number {\n\tif (typeof content === \"string\") {\n\t\treturn content.length;\n\t}\n\n\tlet chars = 0;\n\tfor (const block of content) {\n\t\tif (block.type === \"text\" && block.text) {\n\t\t\tchars += block.text.length;\n\t\t} else if (block.type === \"image\") {\n\t\t\tchars += ESTIMATED_IMAGE_CHARS;\n\t\t}\n\t}\n\treturn chars;\n}\n\n/**\n * Estimate token count for a message using chars/4 heuristic.\n * This is conservative (overestimates tokens).\n */\nexport function estimateTokens(message: AgentMessage): number {\n\tlet chars = 0;\n\n\tswitch (message.role) {\n\t\tcase \"user\": {\n\t\t\tchars = estimateTextAndImageContentChars(\n\t\t\t\t(message as { content: string | Array<{ type: string; text?: string }> }).content,\n\t\t\t);\n\t\t\treturn Math.ceil(chars / 4);\n\t\t}\n\t\tcase \"assistant\": {\n\t\t\tconst assistant = message as AssistantMessage;\n\t\t\tfor (const block of assistant.content) {\n\t\t\t\tif (block.type === \"text\") {\n\t\t\t\t\tchars += block.text.length;\n\t\t\t\t} else if (block.type === \"thinking\") {\n\t\t\t\t\tchars += block.thinking.length;\n\t\t\t\t} else if (block.type === \"toolCall\") {\n\t\t\t\t\tchars += block.name.length + JSON.stringify(block.arguments).length;\n\t\t\t\t}\n\t\t\t}\n\t\t\treturn Math.ceil(chars / 4);\n\t\t}\n\t\tcase \"custom\":\n\t\tcase \"toolResult\": {\n\t\t\tchars = estimateTextAndImageContentChars(message.content);\n\t\t\treturn Math.ceil(chars / 4);\n\t\t}\n\t\tcase \"bashExecution\": {\n\t\t\tchars = message.command.length + message.output.length;\n\t\t\treturn Math.ceil(chars / 4);\n\t\t}\n\t\tcase \"branchSummary\":\n\t\tcase \"compactionSummary\": {\n\t\t\tchars = message.summary.length;\n\t\t\treturn Math.ceil(chars / 4);\n\t\t}\n\t}\n\n\treturn 0;\n}\n\n/**\n * Find valid cut points: indices of user, assistant, custom, or bashExecution messages.\n * Never cut at tool results (they must follow their tool call).\n * When we cut at an assistant message with tool calls, its tool results follow it\n * and will be kept.\n * BashExecutionMessage is treated like a user message (user-initiated context).\n */\nfunction findValidCutPoints(entries: SessionEntry[], startIndex: number, endIndex: number): number[] {\n\tconst cutPoints: number[] = [];\n\tfor (let i = startIndex; i < endIndex; i++) {\n\t\tconst entry = entries[i];\n\t\tswitch (entry.type) {\n\t\t\tcase \"message\": {\n\t\t\t\tconst role = entry.message.role;\n\t\t\t\tswitch (role) {\n\t\t\t\t\tcase \"bashExecution\":\n\t\t\t\t\tcase \"custom\":\n\t\t\t\t\tcase \"branchSummary\":\n\t\t\t\t\tcase \"compactionSummary\":\n\t\t\t\t\tcase \"user\":\n\t\t\t\t\tcase \"assistant\":\n\t\t\t\t\t\tcutPoints.push(i);\n\t\t\t\t\t\tbreak;\n\t\t\t\t\tcase \"toolResult\":\n\t\t\t\t\t\tbreak;\n\t\t\t\t}\n\t\t\t\tbreak;\n\t\t\t}\n\t\t\tcase \"thinking_level_change\":\n\t\t\tcase \"model_change\":\n\t\t\tcase \"compaction\":\n\t\t\tcase \"branch_summary\":\n\t\t\tcase \"custom\":\n\t\t\tcase \"custom_message\":\n\t\t\tcase \"label\":\n\t\t\tcase \"session_info\":\n\t\t\t\tbreak;\n\t\t}\n\n\t\t// branch_summary and custom_message are user-role messages, valid cut points\n\t\tif (entry.type === \"branch_summary\" || entry.type === \"custom_message\") {\n\t\t\tcutPoints.push(i);\n\t\t}\n\t}\n\treturn cutPoints;\n}\n\n/**\n * Find the user message (or bashExecution) that starts the turn containing the given entry index.\n * Returns -1 if no turn start found before the index.\n * BashExecutionMessage is treated like a user message for turn boundaries.\n */\nexport function findTurnStartIndex(entries: SessionEntry[], entryIndex: number, startIndex: number): number {\n\tfor (let i = entryIndex; i >= startIndex; i--) {\n\t\tconst entry = entries[i];\n\t\t// branch_summary and custom_message are user-role messages, can start a turn\n\t\tif (entry.type === \"branch_summary\" || entry.type === \"custom_message\") {\n\t\t\treturn i;\n\t\t}\n\t\tif (entry.type === \"message\") {\n\t\t\tconst role = entry.message.role;\n\t\t\tif (role === \"user\" || role === \"bashExecution\") {\n\t\t\t\treturn i;\n\t\t\t}\n\t\t}\n\t}\n\treturn -1;\n}\n\nexport interface CutPointResult {\n\t/** Index of first entry to keep */\n\tfirstKeptEntryIndex: number;\n\t/** Index of user message that starts the turn being split, or -1 if not splitting */\n\tturnStartIndex: number;\n\t/** Whether this cut splits a turn (cut point is not a user message) */\n\tisSplitTurn: boolean;\n}\n\n/**\n * Find the cut point in session entries that keeps approximately `keepRecentTokens`.\n *\n * Algorithm: Walk backwards from newest, accumulating estimated message sizes.\n * Stop when we've accumulated >= keepRecentTokens. Cut at that point.\n *\n * Can cut at user OR assistant messages (never tool results). When cutting at an\n * assistant message with tool calls, its tool results come after and will be kept.\n *\n * Returns CutPointResult with:\n * - firstKeptEntryIndex: the entry index to start keeping from\n * - turnStartIndex: if cutting mid-turn, the user message that started that turn\n * - isSplitTurn: whether we're cutting in the middle of a turn\n *\n * Only considers entries between `startIndex` and `endIndex` (exclusive).\n */\nexport function findCutPoint(\n\tentries: SessionEntry[],\n\tstartIndex: number,\n\tendIndex: number,\n\tkeepRecentTokens: number,\n): CutPointResult {\n\tconst cutPoints = findValidCutPoints(entries, startIndex, endIndex);\n\n\tif (cutPoints.length === 0) {\n\t\treturn { firstKeptEntryIndex: startIndex, turnStartIndex: -1, isSplitTurn: false };\n\t}\n\n\t// Walk backwards from newest, accumulating estimated message sizes\n\tlet accumulatedTokens = 0;\n\tlet cutIndex = cutPoints[0]; // Default: keep from first message (not header)\n\n\tfor (let i = endIndex - 1; i >= startIndex; i--) {\n\t\tconst entry = entries[i];\n\t\tif (entry.type !== \"message\") continue;\n\n\t\t// Estimate this message's size\n\t\tconst messageTokens = estimateTokens(entry.message);\n\t\taccumulatedTokens += messageTokens;\n\n\t\t// Check if we've exceeded the budget\n\t\tif (accumulatedTokens >= keepRecentTokens) {\n\t\t\t// Find the closest valid cut point at or after this entry\n\t\t\tfor (let c = 0; c < cutPoints.length; c++) {\n\t\t\t\tif (cutPoints[c] >= i) {\n\t\t\t\t\tcutIndex = cutPoints[c];\n\t\t\t\t\tbreak;\n\t\t\t\t}\n\t\t\t}\n\t\t\tbreak;\n\t\t}\n\t}\n\n\t// Scan backwards from cutIndex to include any non-message entries (bash, settings, etc.)\n\twhile (cutIndex > startIndex) {\n\t\tconst prevEntry = entries[cutIndex - 1];\n\t\t// Stop at session header or compaction boundaries\n\t\tif (prevEntry.type === \"compaction\") {\n\t\t\tbreak;\n\t\t}\n\t\tif (prevEntry.type === \"message\") {\n\t\t\t// Stop if we hit any message\n\t\t\tbreak;\n\t\t}\n\t\t// Include this non-message entry (bash, settings change, etc.)\n\t\tcutIndex--;\n\t}\n\n\t// Determine if this is a split turn\n\tconst cutEntry = entries[cutIndex];\n\tconst isUserMessage = cutEntry.type === \"message\" && cutEntry.message.role === \"user\";\n\tconst turnStartIndex = isUserMessage ? -1 : findTurnStartIndex(entries, cutIndex, startIndex);\n\n\treturn {\n\t\tfirstKeptEntryIndex: cutIndex,\n\t\tturnStartIndex,\n\t\tisSplitTurn: !isUserMessage && turnStartIndex !== -1,\n\t};\n}\n\n// ============================================================================\n// Summarization\n// ============================================================================\n\nconst SUMMARIZATION_PROMPT = `The messages above are a conversation to summarize. Create a structured context checkpoint summary that another LLM will use to continue the work.\n\nUse this EXACT format:\n\n## Goal\n[What is the user trying to accomplish? Can be multiple items if the session covers different tasks.]\n\n## Constraints & Preferences\n- [Any constraints, preferences, or requirements mentioned by user]\n- [Or \"(none)\" if none were mentioned]\n\n## Progress\n### Done\n- [x] [Completed tasks/changes]\n\n### In Progress\n- [ ] [Current work]\n\n### Blocked\n- [Issues preventing progress, if any]\n\n## Key Decisions\n- **[Decision]**: [Brief rationale]\n\n## Next Steps\n1. [Ordered list of what should happen next]\n\n## Critical Context\n- [Any data, examples, or references needed to continue]\n- [Or \"(none)\" if not applicable]\n\nKeep each section concise. Preserve exact file paths, function names, and error messages.`;\n\nconst UPDATE_SUMMARIZATION_PROMPT = `The messages above are NEW conversation messages to incorporate into the existing summary provided in <previous-summary> tags.\n\nUpdate the existing structured summary with new information. RULES:\n- PRESERVE all existing information from the previous summary\n- ADD new progress, decisions, and context from the new messages\n- UPDATE the Progress section: move items from \"In Progress\" to \"Done\" when completed\n- UPDATE \"Next Steps\" based on what was accomplished\n- PRESERVE exact file paths, function names, and error messages\n- If something is no longer relevant, you may remove it\n\nUse this EXACT format:\n\n## Goal\n[Preserve existing goals, add new ones if the task expanded]\n\n## Constraints & Preferences\n- [Preserve existing, add new ones discovered]\n\n## Progress\n### Done\n- [x] [Include previously done items AND newly completed items]\n\n### In Progress\n- [ ] [Current work - update based on progress]\n\n### Blocked\n- [Current blockers - remove if resolved]\n\n## Key Decisions\n- **[Decision]**: [Brief rationale] (preserve all previous, add new)\n\n## Next Steps\n1. [Update based on current state]\n\n## Critical Context\n- [Preserve important context, add new if needed]\n\nKeep each section concise. Preserve exact file paths, function names, and error messages.`;\n\nfunction createSummarizationOptions(\n\tmodel: Model<any>,\n\tmaxTokens: number,\n\tapiKey: string | undefined,\n\theaders: Record<string, string> | undefined,\n\tsignal: AbortSignal | undefined,\n\tthinkingLevel: ThinkingLevel | undefined,\n): SimpleStreamOptions {\n\tconst options: SimpleStreamOptions = { maxTokens, signal, apiKey, headers };\n\tif (model.reasoning && thinkingLevel && thinkingLevel !== \"off\") {\n\t\toptions.reasoning = thinkingLevel;\n\t}\n\treturn options;\n}\n\nasync function completeSummarization(\n\tmodel: Model<any>,\n\tcontext: Context,\n\toptions: SimpleStreamOptions,\n\tstreamFn?: StreamFn,\n): Promise<AssistantMessage> {\n\tif (!streamFn) {\n\t\treturn completeSimple(model, context, options);\n\t}\n\tconst stream = await streamFn(model, context, options);\n\treturn stream.result();\n}\n\n/**\n * Generate a summary of the conversation using the LLM.\n * If previousSummary is provided, uses the update prompt to merge.\n */\nexport async function generateSummary(\n\tcurrentMessages: AgentMessage[],\n\tmodel: Model<any>,\n\treserveTokens: number,\n\tapiKey: string | undefined,\n\theaders?: Record<string, string>,\n\tsignal?: AbortSignal,\n\tcustomInstructions?: string,\n\tpreviousSummary?: string,\n\tthinkingLevel?: ThinkingLevel,\n\tstreamFn?: StreamFn,\n): Promise<string> {\n\tconst maxTokens = Math.min(\n\t\tMath.floor(0.8 * reserveTokens),\n\t\tmodel.maxTokens > 0 ? model.maxTokens : Number.POSITIVE_INFINITY,\n\t);\n\n\t// Use update prompt if we have a previous summary, otherwise initial prompt\n\tlet basePrompt = previousSummary ? UPDATE_SUMMARIZATION_PROMPT : SUMMARIZATION_PROMPT;\n\tif (customInstructions) {\n\t\tbasePrompt = `${basePrompt}\\n\\nAdditional focus: ${customInstructions}`;\n\t}\n\n\t// Serialize conversation to text so model doesn't try to continue it\n\t// Convert to LLM messages first (handles custom types like bashExecution, custom, etc.)\n\tconst llmMessages = convertToLlm(currentMessages);\n\tconst conversationText = serializeConversation(llmMessages);\n\n\t// Build the prompt with conversation wrapped in tags\n\tlet promptText = `<conversation>\\n${conversationText}\\n</conversation>\\n\\n`;\n\tif (previousSummary) {\n\t\tpromptText += `<previous-summary>\\n${previousSummary}\\n</previous-summary>\\n\\n`;\n\t}\n\tpromptText += basePrompt;\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\tconst completionOptions = createSummarizationOptions(model, maxTokens, apiKey, headers, signal, thinkingLevel);\n\n\tconst response = await completeSummarization(\n\t\tmodel,\n\t\t{ systemPrompt: SUMMARIZATION_SYSTEM_PROMPT, messages: summarizationMessages },\n\t\tcompletionOptions,\n\t\tstreamFn,\n\t);\n\n\tif (response.stopReason === \"error\") {\n\t\tthrow new Error(`Summarization failed: ${response.errorMessage || \"Unknown error\"}`);\n\t}\n\n\tconst textContent = response.content\n\t\t.filter((c): c is { type: \"text\"; text: string } => c.type === \"text\")\n\t\t.map((c) => c.text)\n\t\t.join(\"\\n\");\n\n\treturn textContent;\n}\n\n// ============================================================================\n// Compaction Preparation (for extensions)\n// ============================================================================\n\nexport interface CompactionPreparation {\n\t/** UUID of first entry to keep */\n\tfirstKeptEntryId: string;\n\t/** Messages that will be summarized and discarded */\n\tmessagesToSummarize: AgentMessage[];\n\t/** Messages that will be turned into turn prefix summary (if splitting) */\n\tturnPrefixMessages: AgentMessage[];\n\t/** Whether this is a split turn (cut point in middle of turn) */\n\tisSplitTurn: boolean;\n\ttokensBefore: number;\n\t/** Summary from previous compaction, for iterative update */\n\tpreviousSummary?: string;\n\t/** File operations extracted from messagesToSummarize */\n\tfileOps: FileOperations;\n\t/** Compaction settions from settings.jsonl\t*/\n\tsettings: CompactionSettings;\n}\n\nexport function prepareCompaction(\n\tpathEntries: SessionEntry[],\n\tsettings: CompactionSettings,\n): CompactionPreparation | undefined {\n\tif (pathEntries.length > 0 && pathEntries[pathEntries.length - 1].type === \"compaction\") {\n\t\treturn undefined;\n\t}\n\n\tlet prevCompactionIndex = -1;\n\tfor (let i = pathEntries.length - 1; i >= 0; i--) {\n\t\tif (pathEntries[i].type === \"compaction\") {\n\t\t\tprevCompactionIndex = i;\n\t\t\tbreak;\n\t\t}\n\t}\n\n\tlet previousSummary: string | undefined;\n\tlet boundaryStart = 0;\n\tif (prevCompactionIndex >= 0) {\n\t\tconst prevCompaction = pathEntries[prevCompactionIndex] as CompactionEntry;\n\t\tpreviousSummary = prevCompaction.summary;\n\t\tconst firstKeptEntryIndex = pathEntries.findIndex((entry) => entry.id === prevCompaction.firstKeptEntryId);\n\t\tboundaryStart = firstKeptEntryIndex >= 0 ? firstKeptEntryIndex : prevCompactionIndex + 1;\n\t}\n\tconst boundaryEnd = pathEntries.length;\n\n\tconst tokensBefore = estimateContextTokens(buildSessionContext(pathEntries).messages).tokens;\n\n\tconst cutPoint = findCutPoint(pathEntries, boundaryStart, boundaryEnd, settings.keepRecentTokens);\n\n\t// Get UUID of first kept entry\n\tconst firstKeptEntry = pathEntries[cutPoint.firstKeptEntryIndex];\n\tif (!firstKeptEntry?.id) {\n\t\treturn undefined; // Session needs migration\n\t}\n\tconst firstKeptEntryId = firstKeptEntry.id;\n\n\tconst historyEnd = cutPoint.isSplitTurn ? cutPoint.turnStartIndex : cutPoint.firstKeptEntryIndex;\n\n\t// Messages to summarize (will be discarded after summary)\n\tconst messagesToSummarize: AgentMessage[] = [];\n\tfor (let i = boundaryStart; i < historyEnd; i++) {\n\t\tconst msg = getMessageFromEntryForCompaction(pathEntries[i]);\n\t\tif (msg) messagesToSummarize.push(msg);\n\t}\n\n\t// Messages for turn prefix summary (if splitting a turn)\n\tconst turnPrefixMessages: AgentMessage[] = [];\n\tif (cutPoint.isSplitTurn) {\n\t\tfor (let i = cutPoint.turnStartIndex; i < cutPoint.firstKeptEntryIndex; i++) {\n\t\t\tconst msg = getMessageFromEntryForCompaction(pathEntries[i]);\n\t\t\tif (msg) turnPrefixMessages.push(msg);\n\t\t}\n\t}\n\n\t// Extract file operations from messages and previous compaction\n\tconst fileOps = extractFileOperations(messagesToSummarize, pathEntries, prevCompactionIndex);\n\n\t// Also extract file ops from turn prefix if splitting\n\tif (cutPoint.isSplitTurn) {\n\t\tfor (const msg of turnPrefixMessages) {\n\t\t\textractFileOpsFromMessage(msg, fileOps);\n\t\t}\n\t}\n\n\treturn {\n\t\tfirstKeptEntryId,\n\t\tmessagesToSummarize,\n\t\tturnPrefixMessages,\n\t\tisSplitTurn: cutPoint.isSplitTurn,\n\t\ttokensBefore,\n\t\tpreviousSummary,\n\t\tfileOps,\n\t\tsettings,\n\t};\n}\n\n// ============================================================================\n// Main compaction function\n// ============================================================================\n\nconst TURN_PREFIX_SUMMARIZATION_PROMPT = `This is the PREFIX of a turn that was too large to keep. The SUFFIX (recent work) is retained.\n\nSummarize the prefix to provide context for the retained suffix:\n\n## Original Request\n[What did the user ask for in this turn?]\n\n## Early Progress\n- [Key decisions and work done in the prefix]\n\n## Context for Suffix\n- [Information needed to understand the retained recent work]\n\nBe concise. Focus on what's needed to understand the kept suffix.`;\n\n/**\n * Generate summaries for compaction using prepared data.\n * Returns CompactionResult - SessionManager adds uuid/parentUuid when saving.\n *\n * @param preparation - Pre-calculated preparation from prepareCompaction()\n * @param customInstructions - Optional custom focus for the summary\n */\nexport async function compact(\n\tpreparation: CompactionPreparation,\n\tmodel: Model<any>,\n\tapiKey: string | undefined,\n\theaders?: Record<string, string>,\n\tcustomInstructions?: string,\n\tsignal?: AbortSignal,\n\tthinkingLevel?: ThinkingLevel,\n\tstreamFn?: StreamFn,\n): Promise<CompactionResult> {\n\tconst {\n\t\tfirstKeptEntryId,\n\t\tmessagesToSummarize,\n\t\tturnPrefixMessages,\n\t\tisSplitTurn,\n\t\ttokensBefore,\n\t\tpreviousSummary,\n\t\tfileOps,\n\t\tsettings,\n\t} = preparation;\n\n\t// Generate summaries (can be parallel if both needed) and merge into one\n\tlet summary: string;\n\n\tif (isSplitTurn && turnPrefixMessages.length > 0) {\n\t\t// Generate both summaries in parallel\n\t\tconst [historyResult, turnPrefixResult] = await Promise.all([\n\t\t\tmessagesToSummarize.length > 0\n\t\t\t\t? generateSummary(\n\t\t\t\t\t\tmessagesToSummarize,\n\t\t\t\t\t\tmodel,\n\t\t\t\t\t\tsettings.reserveTokens,\n\t\t\t\t\t\tapiKey,\n\t\t\t\t\t\theaders,\n\t\t\t\t\t\tsignal,\n\t\t\t\t\t\tcustomInstructions,\n\t\t\t\t\t\tpreviousSummary,\n\t\t\t\t\t\tthinkingLevel,\n\t\t\t\t\t\tstreamFn,\n\t\t\t\t\t)\n\t\t\t\t: Promise.resolve(\"No prior history.\"),\n\t\t\tgenerateTurnPrefixSummary(\n\t\t\t\tturnPrefixMessages,\n\t\t\t\tmodel,\n\t\t\t\tsettings.reserveTokens,\n\t\t\t\tapiKey,\n\t\t\t\theaders,\n\t\t\t\tsignal,\n\t\t\t\tthinkingLevel,\n\t\t\t\tstreamFn,\n\t\t\t),\n\t\t]);\n\t\t// Merge into single summary\n\t\tsummary = `${historyResult}\\n\\n---\\n\\n**Turn Context (split turn):**\\n\\n${turnPrefixResult}`;\n\t} else {\n\t\t// Just generate history summary\n\t\tsummary = await generateSummary(\n\t\t\tmessagesToSummarize,\n\t\t\tmodel,\n\t\t\tsettings.reserveTokens,\n\t\t\tapiKey,\n\t\t\theaders,\n\t\t\tsignal,\n\t\t\tcustomInstructions,\n\t\t\tpreviousSummary,\n\t\t\tthinkingLevel,\n\t\t\tstreamFn,\n\t\t);\n\t}\n\n\t// Compute file lists and append to summary\n\tconst { readFiles, modifiedFiles } = computeFileLists(fileOps);\n\tsummary += formatFileOperations(readFiles, modifiedFiles);\n\n\tif (!firstKeptEntryId) {\n\t\tthrow new Error(\"First kept entry has no UUID - session may need migration\");\n\t}\n\n\treturn {\n\t\tsummary,\n\t\tfirstKeptEntryId,\n\t\ttokensBefore,\n\t\tdetails: { readFiles, modifiedFiles } as CompactionDetails,\n\t};\n}\n\n/**\n * Generate a summary for a turn prefix (when splitting a turn).\n */\nasync function generateTurnPrefixSummary(\n\tmessages: AgentMessage[],\n\tmodel: Model<any>,\n\treserveTokens: number,\n\tapiKey: string | undefined,\n\theaders?: Record<string, string>,\n\tsignal?: AbortSignal,\n\tthinkingLevel?: ThinkingLevel,\n\tstreamFn?: StreamFn,\n): Promise<string> {\n\tconst maxTokens = Math.min(\n\t\tMath.floor(0.5 * reserveTokens),\n\t\tmodel.maxTokens > 0 ? model.maxTokens : Number.POSITIVE_INFINITY,\n\t); // Smaller budget for turn prefix\n\tconst llmMessages = convertToLlm(messages);\n\tconst conversationText = serializeConversation(llmMessages);\n\tconst promptText = `<conversation>\\n${conversationText}\\n</conversation>\\n\\n${TURN_PREFIX_SUMMARIZATION_PROMPT}`;\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\tconst response = await completeSummarization(\n\t\tmodel,\n\t\t{ systemPrompt: SUMMARIZATION_SYSTEM_PROMPT, messages: summarizationMessages },\n\t\tcreateSummarizationOptions(model, maxTokens, apiKey, headers, signal, thinkingLevel),\n\t\tstreamFn,\n\t);\n\n\tif (response.stopReason === \"error\") {\n\t\tthrow new Error(`Turn prefix summarization failed: ${response.errorMessage || \"Unknown error\"}`);\n\t}\n\n\treturn response.content\n\t\t.filter((c): c is { type: \"text\"; text: string } => c.type === \"text\")\n\t\t.map((c) => c.text)\n\t\t.join(\"\\n\");\n}\n"]}

package/dist/core/compaction/compaction.js

@@ -5,6 +5,7 @@
  * and after compaction the session is reloaded.
  */
 import { completeSimple } from "omk-ai";
+import { computeReservedTokenBudget } from "../context-budget-reserved-tokens.js";
 import { convertToLlm, createBranchSummaryMessage, createCompactionSummaryMessage, createCustomMessage, } from "../messages.js";
 import { buildSessionContext } from "../session-manager.js";
 import { computeFileLists, createFileOps, extractFileOpsFromMessage, formatFileOperations, SUMMARIZATION_SYSTEM_PROMPT, serializeConversation, } from "./utils.js";
@@ -160,7 +161,17 @@ export function getCompactionHeadroomThreshold(contextWindow, settings) {
         ? configuredMaxUsageRatio
         : DEFAULT_COMPACTION_MAX_USAGE_RATIO;
     const maxUsageRatioTokens = Math.max(1, Math.floor(windowTokens * maxUsageRatio));
-    const reserveBoundaryTokens = reserveTokens >= windowTokens ? maxUsageRatioTokens : Math.max(1, windowTokens - reserveTokens);
+    const reservedBudget = computeReservedTokenBudget({
+        modelContextWindow: windowTokens,
+        systemPromptTokens: 0,
+        reservedOutputTokens: settings.reservedOutputTokens ?? reserveTokens,
+        reservedToolResultTokens: settings.reservedToolResultTokens ?? 0,
+        safetyMarginTokens: settings.safetyMarginTokens ?? 0,
+        imageReserveTokens: settings.imageReserveTokens ?? 0,
+    });
+    const reserveBoundaryTokens = reservedBudget.overflow || reservedBudget.effectiveBudget === 0
+        ? maxUsageRatioTokens
+        : Math.max(1, reservedBudget.effectiveBudget);
     const triggerTokens = Math.min(maxUsageRatioTokens, reserveBoundaryTokens);
     return {
         triggerTokens,