2358 lines
90 KiB
TypeScript
2358 lines
90 KiB
TypeScript
/**
|
|
* Shared pi SDK setup for fn engine agents.
|
|
*
|
|
* Uses Fusion auth for writes and legacy pi auth as a read-only fallback.
|
|
* Provides factory functions for creating triage and executor agent sessions.
|
|
*/
|
|
|
|
/* eslint-disable @typescript-eslint/no-explicit-any */
|
|
import { existsSync, readFileSync, realpathSync } from "node:fs";
|
|
import { exec, execFile } from "node:child_process";
|
|
import { promisify } from "node:util";
|
|
import { createRequire } from "node:module";
|
|
import { basename, dirname, join, relative, isAbsolute, resolve } from "node:path";
|
|
|
|
const execAsync = promisify(exec);
|
|
import {
|
|
createAgentSession,
|
|
createBashTool,
|
|
createCodingTools,
|
|
createEditTool,
|
|
createExtensionRuntime,
|
|
createFindTool,
|
|
createGrepTool,
|
|
createLsTool,
|
|
createReadOnlyTools,
|
|
createReadTool,
|
|
createWriteTool,
|
|
DefaultResourceLoader,
|
|
DefaultPackageManager,
|
|
discoverAndLoadExtensions,
|
|
ModelRegistry,
|
|
SessionManager,
|
|
SettingsManager,
|
|
type AgentSession,
|
|
type ToolDefinition,
|
|
} from "@earendil-works/pi-coding-agent";
|
|
import { customProviderRegistryKey, getEnabledPiExtensionPaths, getFusionAgentDir, getLegacyPiAgentDir, getProjectRootFromWorktree, reconcileClaudeCliPaths, reconcileDroidCliPaths, resolvePiExtensionProjectRoot } from "@fusion/core";
|
|
import type {
|
|
AgentPermissionPolicyActionCategory,
|
|
PermanentAgentActionCategory,
|
|
PermanentAgentGatingContext,
|
|
} from "@fusion/core";
|
|
import {
|
|
resolveSessionSkills,
|
|
createSkillsOverrideFromSelection,
|
|
type SkillSelectionContext,
|
|
} from "./skill-resolver.js";
|
|
import { isContextLimitError } from "./context-limit-detector.js";
|
|
import { createFusionAuthStorage, getModelRegistryModelsPath } from "./auth-storage.js";
|
|
import { piLog, extensionsLog } from "./logger.js";
|
|
import { readCustomProviders } from "./custom-providers.js";
|
|
import {
|
|
buildGateRejection,
|
|
evaluateAgentActionGate,
|
|
resolveGateOutcome,
|
|
type AgentActionGateContext,
|
|
} from "./agent-action-gate.js";
|
|
import { resolvePermanentAgentToolDecision } from "./permanent-agent-gating.js";
|
|
import type { SystemPromptLayers } from "./prompt-layers.js";
|
|
import { READONLY_ALLOWLIST, filterCustomToolsForReadonly, isReadonlyAllowed } from "./workflow-step-tool-policy.js";
|
|
import { createStreamingDeltaNormalizer } from "./streaming-delta.js";
|
|
import { isUnsupportedMessageRoleError } from "./transient-error-detector.js";
|
|
|
|
const RTK_ACCEPTED_REWRITE_EXIT_CODES = new Set([0, 3]);
|
|
const RTK_EXPECTED_PASSTHROUGH_EXIT_CODES = new Set([1, 2]);
|
|
const RTK_EXPECTED_FAIL_OPEN_ERROR_CODES = new Set(["ABORT_ERR", "ENOENT", "ETIMEDOUT"]);
|
|
const RTK_REWRITE_MAX_BUFFER_BYTES = 64 * 1024;
|
|
|
|
export type RtkRewriteMode = "off" | "rewrite";
|
|
|
|
export interface RtkRewriteOptions {
|
|
mode?: RtkRewriteMode;
|
|
timeoutMs?: number;
|
|
}
|
|
|
|
function normalizeRtkRewriteOptions(options?: RtkRewriteOptions): Required<RtkRewriteOptions> {
|
|
const modeEnv = process.env.FUSION_RTK_REWRITE?.toLowerCase();
|
|
const envMode: RtkRewriteMode = modeEnv === "1" || modeEnv === "true" || modeEnv === "rewrite" ? "rewrite" : "off";
|
|
const envTimeoutMs = Number.parseInt(process.env.FUSION_RTK_REWRITE_TIMEOUT_MS ?? "2000", 10);
|
|
const requestedTimeoutMs = options?.timeoutMs ?? envTimeoutMs;
|
|
return {
|
|
mode: options?.mode ?? envMode,
|
|
timeoutMs: Number.isFinite(requestedTimeoutMs) && requestedTimeoutMs > 0 ? requestedTimeoutMs : 2000,
|
|
};
|
|
}
|
|
|
|
function resolveRtkRewriteOptions(): Required<RtkRewriteOptions> {
|
|
return normalizeRtkRewriteOptions();
|
|
}
|
|
|
|
function getRtkErrorCode(error: Error | null): number | string | null {
|
|
if (!error) return 0;
|
|
const rawCode = (error as unknown as { code?: unknown }).code;
|
|
if (typeof rawCode === "number" || typeof rawCode === "string") return rawCode;
|
|
return null;
|
|
}
|
|
|
|
function shouldWarnForRtkFailure(code: number | string | null): boolean {
|
|
if (typeof code === "number") return !RTK_EXPECTED_PASSTHROUGH_EXIT_CODES.has(code);
|
|
if (typeof code === "string") return !RTK_EXPECTED_FAIL_OPEN_ERROR_CODES.has(code);
|
|
return true;
|
|
}
|
|
|
|
function rewriteCommandWithRtk(
|
|
command: string,
|
|
options: Required<RtkRewriteOptions>,
|
|
signal?: AbortSignal,
|
|
): Promise<string | null> {
|
|
if (signal?.aborted) {
|
|
return Promise.resolve(null);
|
|
}
|
|
|
|
return new Promise((resolve) => {
|
|
execFile(
|
|
"rtk",
|
|
["rewrite", command],
|
|
{
|
|
timeout: options.timeoutMs,
|
|
maxBuffer: RTK_REWRITE_MAX_BUFFER_BYTES,
|
|
signal,
|
|
windowsHide: true,
|
|
},
|
|
(error, stdout, stderr) => {
|
|
const code = getRtkErrorCode(error);
|
|
|
|
if (typeof code !== "number" || !RTK_ACCEPTED_REWRITE_EXIT_CODES.has(code)) {
|
|
if (shouldWarnForRtkFailure(code)) {
|
|
const reason = stderr?.toString().trim() || (error instanceof Error ? error.message : `exit ${String(code)}`);
|
|
piLog.warn(`[pi] rtk rewrite failed open: ${reason}`);
|
|
}
|
|
resolve(null);
|
|
return;
|
|
}
|
|
|
|
const rewritten = stdout.toString().trim();
|
|
resolve(rewritten && rewritten !== command ? rewritten : null);
|
|
},
|
|
);
|
|
});
|
|
}
|
|
|
|
export interface AgentResult {
|
|
session: AgentSession;
|
|
/** Path to the persisted session file (undefined for in-memory sessions). */
|
|
sessionFile?: string;
|
|
}
|
|
|
|
/**
|
|
* Process-global list of extension paths to inject into every createFnAgent
|
|
* session. Set once at startup by the host (cli's dashboard/daemon/serve)
|
|
* before any sessions are created. The paths are passed to pi's
|
|
* `DefaultResourceLoader` as `additionalExtensionPaths` so the cli's own
|
|
* `@runfusion/fusion` extension (registering `fn_*` tools) is loaded inside
|
|
* every agent session — including chat sessions that pass no `customTools`.
|
|
*/
|
|
let hostExtensionPaths: string[] = [];
|
|
|
|
export function setHostExtensionPaths(paths: readonly string[]): void {
|
|
hostExtensionPaths = [...paths];
|
|
}
|
|
|
|
export function getHostExtensionPaths(): readonly string[] {
|
|
return hostExtensionPaths;
|
|
}
|
|
|
|
export interface PromptableSession extends AgentSession {
|
|
promptWithFallback: (prompt: string, options?: unknown) => Promise<void>;
|
|
}
|
|
|
|
interface SessionManagerLike {
|
|
fileEntries?: Array<{ type?: string; message?: Record<string, unknown> }>;
|
|
appendMessage?: (message: Record<string, unknown>) => void;
|
|
_rewriteFile?: () => void;
|
|
}
|
|
|
|
interface ToolHookPayload {
|
|
toolCall: unknown;
|
|
args: unknown;
|
|
result: { content?: unknown; details?: unknown };
|
|
isError: boolean;
|
|
}
|
|
|
|
interface ToolHookResult {
|
|
content?: unknown;
|
|
details?: unknown;
|
|
isError?: boolean;
|
|
}
|
|
|
|
type AgentToolHookSession = AgentSession & {
|
|
agent?: {
|
|
afterToolCall?: (payload: ToolHookPayload) => Promise<ToolHookResult | undefined>;
|
|
state?: {
|
|
messages?: Array<Record<string, unknown>>;
|
|
};
|
|
};
|
|
__fusionToolResultGuardInstalled?: boolean;
|
|
__fusionMessageContentGuardInstalled?: boolean;
|
|
};
|
|
const FN_MEMORY_APPEND_TOOL_NAME = "fn_memory_append";
|
|
const FUSION_SHUTDOWN_WRAP_FLAG = "__fusionSessionShutdownDisposeWrapped";
|
|
|
|
type SessionShutdownEventShape = { type: "session_shutdown"; reason: "quit" | "reload" };
|
|
type ExtensionRunnerShutdownEmitter = {
|
|
hasHandlers: (event: "session_shutdown") => boolean;
|
|
emit: (event: SessionShutdownEventShape) => Promise<unknown>;
|
|
};
|
|
|
|
async function emitSessionShutdownEvent(
|
|
extensionRunner: ExtensionRunnerShutdownEmitter,
|
|
event: SessionShutdownEventShape,
|
|
): Promise<boolean> {
|
|
if (!extensionRunner.hasHandlers("session_shutdown")) {
|
|
return false;
|
|
}
|
|
await extensionRunner.emit(event);
|
|
return true;
|
|
}
|
|
|
|
/**
|
|
* Fusion creates raw pi `AgentSession` objects and many engine call sites
|
|
* invoke `session.dispose()` directly. Wrap dispose so we mirror
|
|
* `AgentSessionRuntime.dispose()` behavior and emit `session_shutdown` first.
|
|
*/
|
|
function wrapSessionDisposeWithShutdown(session: AgentSession): void {
|
|
const mutableSession = session as AgentSession & Record<string, unknown>;
|
|
if (mutableSession[FUSION_SHUTDOWN_WRAP_FLAG]) {
|
|
return;
|
|
}
|
|
mutableSession[FUSION_SHUTDOWN_WRAP_FLAG] = true;
|
|
|
|
const originalDispose =
|
|
typeof session.dispose === "function"
|
|
? session.dispose.bind(session)
|
|
: () => undefined;
|
|
let disposeStarted = false;
|
|
const wrappedDispose = async (): Promise<void> => {
|
|
if (disposeStarted) {
|
|
return;
|
|
}
|
|
disposeStarted = true;
|
|
|
|
const extensionRunner = (session as { extensionRunner?: unknown }).extensionRunner;
|
|
if (
|
|
extensionRunner &&
|
|
typeof (extensionRunner as ExtensionRunnerShutdownEmitter).hasHandlers === "function" &&
|
|
typeof (extensionRunner as ExtensionRunnerShutdownEmitter).emit === "function"
|
|
) {
|
|
try {
|
|
await emitSessionShutdownEvent(extensionRunner as ExtensionRunnerShutdownEmitter, {
|
|
type: "session_shutdown",
|
|
reason: "quit",
|
|
});
|
|
} catch (error: unknown) {
|
|
const message = error instanceof Error ? error.message : String(error);
|
|
piLog.warn(`Failed to emit session_shutdown during dispose: ${message}`);
|
|
}
|
|
}
|
|
|
|
try {
|
|
await Promise.resolve(originalDispose());
|
|
} catch (error: unknown) {
|
|
const message = error instanceof Error ? error.message : String(error);
|
|
piLog.warn(`Session dispose failed after session_shutdown emit: ${message}`);
|
|
}
|
|
};
|
|
|
|
(mutableSession as unknown as { dispose: () => void }).dispose = wrappedDispose as unknown as () => void;
|
|
}
|
|
|
|
export function _wrapSessionDisposeForTest(session: AgentSession): void {
|
|
wrapSessionDisposeWithShutdown(session);
|
|
}
|
|
|
|
function getSessionStateError(session: AgentSession): string {
|
|
const state = (session as any).state;
|
|
const error = state?.errorMessage ?? state?.error;
|
|
return typeof error === "string" ? error : "";
|
|
}
|
|
|
|
function clearSessionStateError(session: AgentSession): void {
|
|
const state = (session as any).state;
|
|
if (!state || typeof state !== "object") {
|
|
return;
|
|
}
|
|
|
|
// pi-coding-agent 0.70+ exposes `errorMessage` as readonly — writes are
|
|
// silently ignored. Pre-0.70 used mutable `state.error`. Best-effort clear
|
|
// both so transcripts carry forward to the next prompt cleanly.
|
|
for (const key of ["errorMessage", "error"]) {
|
|
if (key in state) {
|
|
try {
|
|
state[key] = undefined;
|
|
} catch {
|
|
// readonly — no-op
|
|
}
|
|
}
|
|
}
|
|
}
|
|
|
|
function isThinkingReasoningConflictError(message: string): boolean {
|
|
return /cannot specify both\s+['"]?thinking['"]?\s+and\s+['"]?reasoning_effort['"]?/i.test(message);
|
|
}
|
|
|
|
function coercePreviewValue(value: unknown): string {
|
|
if (value === null) return "null";
|
|
if (value === undefined) return "undefined";
|
|
if (typeof value === "string") return value;
|
|
if (typeof value === "number" || typeof value === "boolean" || typeof value === "bigint") {
|
|
return String(value);
|
|
}
|
|
if (typeof value === "symbol") {
|
|
return value.description ? `symbol(${value.description})` : "symbol";
|
|
}
|
|
if (typeof value === "function") {
|
|
return `[function ${value.name || "anonymous"}]`;
|
|
}
|
|
return Object.prototype.toString.call(value);
|
|
}
|
|
|
|
function safePreviewJson(value: unknown): string {
|
|
const seen = new WeakSet<object>();
|
|
return JSON.stringify(value, (_key, candidate) => {
|
|
if (typeof candidate === "object" && candidate !== null) {
|
|
if (seen.has(candidate)) {
|
|
return "[Circular]";
|
|
}
|
|
seen.add(candidate);
|
|
}
|
|
return candidate;
|
|
});
|
|
}
|
|
|
|
export async function promptSessionAndCheck(session: AgentSession, prompt: string, options?: unknown): Promise<void> {
|
|
clearSessionStateError(session);
|
|
if (options === undefined) {
|
|
await session.prompt(prompt);
|
|
} else {
|
|
await (session.prompt as any)(prompt, options);
|
|
}
|
|
|
|
const stateError = getSessionStateError(session);
|
|
if (stateError) {
|
|
// pi-coding-agent swallows its own exceptions into state.errorMessage
|
|
// without preserving a stack. When the message looks like a generic
|
|
// TypeError (undefined/null property access), dump the session transcript
|
|
// shape so the malformed message can be identified next time.
|
|
if (/Cannot read propert(y|ies) of (undefined|null)/i.test(stateError)) {
|
|
try {
|
|
const messages = (session as any).agent?.state?.messages ?? (session as any).state?.messages;
|
|
if (Array.isArray(messages)) {
|
|
const recent = messages.slice(-6).map((m: Record<string, unknown>, idx: number) => {
|
|
const i = messages.length - 6 + idx;
|
|
const content = m?.content;
|
|
return {
|
|
index: i < 0 ? idx : i,
|
|
role: coercePreviewValue(m?.role),
|
|
contentType: Array.isArray(content) ? `array(len=${content.length})` : typeof content,
|
|
toolName: coercePreviewValue((m as { toolName?: unknown }).toolName),
|
|
stopReason: coercePreviewValue((m as { stopReason?: unknown }).stopReason),
|
|
};
|
|
});
|
|
piLog.error(`pi state error — transcript tail (${messages.length} msgs total): ${safePreviewJson(recent)}`);
|
|
} else {
|
|
piLog.error(`pi state error — state.messages is not an array: ${typeof messages}`);
|
|
}
|
|
} catch (inspectErr) {
|
|
piLog.warn(`pi state error — failed to inspect transcript: ${inspectErr instanceof Error ? inspectErr.message : String(inspectErr)}`);
|
|
}
|
|
}
|
|
// Some OpenAI-compatible providers (notably Moonshot/Kimi) end generation
|
|
// with a non-standard `finish_reason: repeat` when their server-side
|
|
// repetition detector trips. pi-ai surfaces this as a fatal state error,
|
|
// but for our purposes the assistant turn is already complete — treat it
|
|
// as a soft stop so the heartbeat keeps running.
|
|
if (/Provider finish_reason:\s*repeat\b/i.test(stateError)) {
|
|
piLog.warn(`pi state error — treating provider finish_reason=repeat as soft stop: ${stateError}`);
|
|
clearSessionStateError(session);
|
|
return;
|
|
}
|
|
// pi-ai's openai-codex-responses provider (Codex via ChatGPT-plan WebSocket)
|
|
// surfaces transport drops as bare "WebSocket error" / "WebSocket closed".
|
|
// The underlying ErrorEvent's `event.error` (cause/code) is dropped by
|
|
// pi-ai's `extractWebSocketError` (it only inspects `event.message`), so
|
|
// by the time we see the string the cause is gone. Tag the message with
|
|
// the model identity so retry/transient classification can at least tell
|
|
// which transport is unstable, and emit a structured warn for triage.
|
|
if (/^WebSocket (error|closed)\b/i.test(stateError) || /WebSocket stream closed before response\.completed/i.test(stateError)) {
|
|
const modelDesc = describeModel(session);
|
|
piLog.warn(`pi state error — Codex WebSocket transport drop (model=${modelDesc}): ${stateError}`);
|
|
throw new Error(`${stateError} (model=${modelDesc})`);
|
|
}
|
|
if (isUnsupportedMessageRoleError(stateError)) {
|
|
const modelDesc = describeModel(session);
|
|
const hint =
|
|
"Operator action required: this agent's configured model/provider rejected a message role. Check the agent model selection and provider compatibility (imported non-default 'company' agents may default to an incompatible model+provider combination).";
|
|
piLog.error(`pi state error — unsupported message role (model=${modelDesc}): ${stateError}`);
|
|
throw new Error(`${stateError} (model=${modelDesc}). ${hint}`);
|
|
}
|
|
throw new Error(stateError);
|
|
}
|
|
}
|
|
|
|
// Re-entry guard for the top-level dispatcher below. When `session.promptWithFallback`
|
|
// is attached (e.g. by `createFnAgent` at pi.ts:2012, where the rich model-swap +
|
|
// `isRetryableModelSelectionError` path lives), we want top-level callers like
|
|
// triage/executor to flow through it so missing-API-key, 401/403, rate-limit, etc.
|
|
// trigger the configured `fallbackModel`. The WeakSet prevents infinite recursion
|
|
// in case a session-attached `promptWithFallback` ever calls back into the
|
|
// top-level export with the same session.
|
|
const promptWithFallbackInFlight = new WeakSet<object>();
|
|
|
|
export async function promptWithFallback(session: AgentSession, prompt: string, options?: unknown): Promise<void> {
|
|
const sessionWithDispatch = session as AgentSession & {
|
|
promptWithFallback?: (prompt: string, options?: unknown) => Promise<void>;
|
|
};
|
|
if (
|
|
typeof sessionWithDispatch.promptWithFallback === "function" &&
|
|
!promptWithFallbackInFlight.has(session as unknown as object)
|
|
) {
|
|
promptWithFallbackInFlight.add(session as unknown as object);
|
|
try {
|
|
await sessionWithDispatch.promptWithFallback(prompt, options);
|
|
return;
|
|
} finally {
|
|
promptWithFallbackInFlight.delete(session as unknown as object);
|
|
}
|
|
}
|
|
|
|
piLog.log(`promptWithFallback: calling session.prompt (prompt length=${prompt.length})`);
|
|
try {
|
|
await promptSessionAndCheck(session, prompt, options);
|
|
piLog.log("promptWithFallback: prompt completed");
|
|
} catch (err: unknown) {
|
|
const errorMessage = err instanceof Error ? err.message : String(err);
|
|
if (!isContextLimitError(errorMessage)) {
|
|
piLog.error(`promptWithFallback: non-context error — propagating: ${errorMessage}`);
|
|
throw err;
|
|
}
|
|
|
|
// Context limit error — attempt auto-compaction and retry once
|
|
const promptMemoryRetry = await retryWithCompactedPromptMemory(session, prompt, options);
|
|
if (promptMemoryRetry.recovered) {
|
|
return;
|
|
}
|
|
if (promptMemoryRetry.error) {
|
|
const retryMessage = promptMemoryRetry.error instanceof Error ? promptMemoryRetry.error.message : String(promptMemoryRetry.error);
|
|
if (!isContextLimitError(retryMessage)) {
|
|
throw promptMemoryRetry.error;
|
|
}
|
|
}
|
|
|
|
const promptSectionRetry = await retryWithCompactedPromptSections(session, prompt, options);
|
|
if (promptSectionRetry.recovered) {
|
|
return;
|
|
}
|
|
if (promptSectionRetry.error) {
|
|
const retryMessage = promptSectionRetry.error instanceof Error ? promptSectionRetry.error.message : String(promptSectionRetry.error);
|
|
if (!isContextLimitError(retryMessage)) {
|
|
throw promptSectionRetry.error;
|
|
}
|
|
}
|
|
|
|
piLog.warn("promptWithFallback: context limit error — attempting auto-compaction");
|
|
await flushMemoryBeforeSessionCompaction(session);
|
|
const compactResult = await compactSessionContext(session);
|
|
if (!compactResult) {
|
|
piLog.error("promptWithFallback: compaction unavailable — propagating original error");
|
|
throw err;
|
|
}
|
|
|
|
piLog.log(`promptWithFallback: compaction succeeded (${compactResult.tokensBefore} tokens) — retrying prompt`);
|
|
try {
|
|
await promptSessionAndCheck(session, prompt, options);
|
|
piLog.log("promptWithFallback: prompt completed after auto-compaction");
|
|
} catch (retryErr: unknown) {
|
|
const retryErrorMessage = retryErr instanceof Error ? retryErr.message : String(retryErr);
|
|
piLog.error(`promptWithFallback: retry after auto-compaction failed: ${retryErrorMessage}`);
|
|
throw err; // Throw original error to preserve original context
|
|
}
|
|
}
|
|
}
|
|
|
|
/**
|
|
* Extract a human-readable model description from an AgentSession.
|
|
* Returns `"<provider>/<modelId>"` (e.g. `"anthropic/claude-sonnet-4-5"`)
|
|
* or `"unknown model"` when the session has no model set.
|
|
*/
|
|
export function describeModel(session: AgentSession): string {
|
|
const model = session.model;
|
|
if (!model) return "unknown model";
|
|
return `${model.provider}/${model.id}`;
|
|
}
|
|
|
|
/**
|
|
* Default instructions used when calling `session.compact()` for loop recovery.
|
|
* These guide the compaction summary to preserve essential context while
|
|
* freeing up the context window for continued work.
|
|
*/
|
|
export const COMPACTION_FALLBACK_INSTRUCTIONS = [
|
|
"Summarize all completed steps concisely.",
|
|
"Preserve the current step number and any in-progress work details.",
|
|
"Keep references to key files, decisions, and error states.",
|
|
"Discard verbose tool output, repeated attempts, and exploration history.",
|
|
].join(" ");
|
|
|
|
const MAX_COMPACTED_PROMPT_MEMORY_CHARS = 8_000;
|
|
const MAX_COMPACTED_SUBTASK_GUIDANCE_CHARS = 1_200;
|
|
const MAX_COMPACTED_ATTACHMENTS_CHARS = 4_000;
|
|
const MAX_COMPACTED_EXISTING_SPEC_CHARS = 4_000;
|
|
const MAX_COMPACTED_TASK_PROMPT_CHARS = MAX_COMPACTED_EXISTING_SPEC_CHARS;
|
|
const MAX_COMPACTED_USER_COMMENTS_CHARS = 2_000;
|
|
|
|
function compactMarkdownMemorySection(sectionBody: string): string {
|
|
const lines = sectionBody.split("\n");
|
|
const kept: string[] = [];
|
|
let used = 0;
|
|
|
|
for (const line of lines) {
|
|
const trimmed = line.trimEnd();
|
|
const normalized = trimmed.trimStart();
|
|
const isUseful =
|
|
normalized.startsWith("##")
|
|
|| normalized.startsWith("- ")
|
|
|| normalized.startsWith("* ")
|
|
|| /^\d+\.\s/.test(normalized)
|
|
|| normalized.length === 0;
|
|
|
|
if (!isUseful) {
|
|
continue;
|
|
}
|
|
|
|
const nextLength = used + trimmed.length + 1;
|
|
if (nextLength > MAX_COMPACTED_PROMPT_MEMORY_CHARS) {
|
|
break;
|
|
}
|
|
|
|
kept.push(trimmed);
|
|
used = nextLength;
|
|
}
|
|
|
|
const compacted = kept.join("\n").trim();
|
|
if (compacted.length >= sectionBody.trim().length) {
|
|
return sectionBody.trim();
|
|
}
|
|
|
|
return [
|
|
compacted,
|
|
"",
|
|
`<!-- Memory compacted from ${sectionBody.length} characters to avoid context overflow. Use memory tools or the selected memory file later only if essential. -->`,
|
|
].join("\n").trim();
|
|
}
|
|
|
|
function compactPromptMemory(prompt: string): string | null {
|
|
const sectionPattern = /(^|\n)(## (?:Project Memory|Agent Memory|Memory)\n\n)([\s\S]*?)(?=\n## [^#]|\n# [^#]|$)/g;
|
|
let changed = false;
|
|
const compactedPrompt = prompt.replace(sectionPattern, (match, prefix: string, heading: string, body: string) => {
|
|
const trimmedBody = body.trim();
|
|
if (trimmedBody.length <= MAX_COMPACTED_PROMPT_MEMORY_CHARS) {
|
|
return match;
|
|
}
|
|
|
|
const compacted = compactMarkdownMemorySection(trimmedBody);
|
|
if (compacted.length >= trimmedBody.length) {
|
|
return match;
|
|
}
|
|
|
|
changed = true;
|
|
return `${prefix}${heading}${compacted}`;
|
|
});
|
|
|
|
return changed && compactedPrompt.length < prompt.length ? compactedPrompt : null;
|
|
}
|
|
|
|
function trimSubtaskSectionBody(body: string): string {
|
|
const paragraphs = body
|
|
.trim()
|
|
.split(/\n\s*\n/)
|
|
.map((p) => p.trim())
|
|
.filter(Boolean);
|
|
const rawFirstParagraph = paragraphs[0] ?? "Subtask guidance omitted for context limits.";
|
|
const maxFirstParagraphChars = Math.max(200, MAX_COMPACTED_SUBTASK_GUIDANCE_CHARS - 200);
|
|
const firstParagraph = rawFirstParagraph.length > maxFirstParagraphChars
|
|
? `${rawFirstParagraph.slice(0, maxFirstParagraphChars)}…`
|
|
: rawFirstParagraph;
|
|
return [
|
|
firstParagraph,
|
|
"",
|
|
"Follow the project's standard subtask split rules.",
|
|
].join("\n");
|
|
}
|
|
|
|
function compactAttachmentSectionBody(body: string): string {
|
|
if (body.length <= MAX_COMPACTED_ATTACHMENTS_CHARS) {
|
|
return body.trim();
|
|
}
|
|
|
|
const lines = body.split("\n");
|
|
const kept: string[] = [];
|
|
let inFence = false;
|
|
let fenceHasContent = false;
|
|
|
|
for (const line of lines) {
|
|
if (/^```/.test(line.trim())) {
|
|
if (!inFence) {
|
|
inFence = true;
|
|
fenceHasContent = false;
|
|
continue;
|
|
}
|
|
inFence = false;
|
|
if (fenceHasContent) {
|
|
kept.push("```", "_... attachment body trimmed ..._", "```");
|
|
}
|
|
continue;
|
|
}
|
|
|
|
if (inFence) {
|
|
fenceHasContent = true;
|
|
continue;
|
|
}
|
|
|
|
kept.push(line);
|
|
}
|
|
|
|
return kept.join("\n").trim();
|
|
}
|
|
|
|
function compactExistingSpecificationSectionBody(body: string): string {
|
|
const trimmed = body.trim();
|
|
if (trimmed.length <= MAX_COMPACTED_EXISTING_SPEC_CHARS) {
|
|
return trimmed;
|
|
}
|
|
|
|
const head = trimmed.slice(0, Math.floor(MAX_COMPACTED_EXISTING_SPEC_CHARS / 2));
|
|
const tail = trimmed.slice(-Math.floor(MAX_COMPACTED_EXISTING_SPEC_CHARS / 2));
|
|
return `${head}\n\n_... existing specification middle trimmed ..._\n\n${tail}`;
|
|
}
|
|
|
|
function extractMarkdownSection(document: string, headingName: string): string {
|
|
const heading = `## ${headingName}`;
|
|
const start = document.indexOf(heading);
|
|
if (start === -1) {
|
|
return "";
|
|
}
|
|
|
|
const afterHeading = start + heading.length;
|
|
const nextH2 = document.indexOf("\n## ", afterHeading);
|
|
const nextH1 = document.indexOf("\n# ", afterHeading);
|
|
const endCandidates = [nextH2, nextH1].filter((value) => value !== -1);
|
|
const end = endCandidates.length > 0 ? Math.min(...endCandidates) : document.length;
|
|
|
|
return document.slice(start, end).trim();
|
|
}
|
|
|
|
function compactTaskPromptStepsSection(section: string): string {
|
|
const stepTitles = Array.from(section.matchAll(/^### Step \d+:.*$/gm), (match) => match[0].trim());
|
|
if (stepTitles.length === 0) {
|
|
return section.trim();
|
|
}
|
|
|
|
return [
|
|
"## Steps",
|
|
...stepTitles,
|
|
"",
|
|
"_... step checklist details trimmed for context limits ..._",
|
|
].join("\n").trim();
|
|
}
|
|
|
|
function truncateCompactedSection(section: string, maxChars: number, label: string): string {
|
|
const trimmed = section.trim();
|
|
if (!trimmed || trimmed.length <= maxChars) {
|
|
return trimmed;
|
|
}
|
|
|
|
const marker = `_... ${label} trimmed for context limits ..._`;
|
|
const headBudget = Math.max(200, maxChars - marker.length - 2);
|
|
return [
|
|
`${trimmed.slice(0, headBudget).trimEnd()}…`,
|
|
"",
|
|
marker,
|
|
].join("\n").trim();
|
|
}
|
|
|
|
function compactTaskPromptSectionBody(body: string): string {
|
|
const trimmed = body.trim();
|
|
if (trimmed.length <= MAX_COMPACTED_TASK_PROMPT_CHARS) {
|
|
return trimmed;
|
|
}
|
|
|
|
const fencedMatch = /^```markdown\s*\n([\s\S]*?)\n```$/m.exec(trimmed);
|
|
const promptContent = fencedMatch ? fencedMatch[1].trim() : trimmed;
|
|
const firstSectionIndex = promptContent.indexOf("\n## ");
|
|
const preamble = (firstSectionIndex === -1 ? promptContent : promptContent.slice(0, firstSectionIndex)).trim();
|
|
const missionSection = extractMarkdownSection(promptContent, "Mission");
|
|
const dependenciesSection = extractMarkdownSection(promptContent, "Dependencies");
|
|
const fileScopeSection = extractMarkdownSection(promptContent, "File Scope");
|
|
const stepsSection = compactTaskPromptStepsSection(extractMarkdownSection(promptContent, "Steps"));
|
|
|
|
const compactedContent = [
|
|
truncateCompactedSection(preamble, 400, "task header"),
|
|
truncateCompactedSection(missionSection, 900, "mission"),
|
|
truncateCompactedSection(dependenciesSection, 500, "dependencies"),
|
|
truncateCompactedSection(fileScopeSection, 1_000, "file scope"),
|
|
truncateCompactedSection(stepsSection, 1_200, "steps outline"),
|
|
"_... remaining PROMPT.md sections trimmed for context limits ..._",
|
|
].filter(Boolean).join("\n\n").trim();
|
|
|
|
const narrowedContent = compactedContent.length <= MAX_COMPACTED_TASK_PROMPT_CHARS
|
|
? compactedContent
|
|
: compactExistingSpecificationSectionBody(compactedContent);
|
|
const finalContent = fencedMatch ? `\`\`\`markdown\n${narrowedContent}\n\`\`\`` : narrowedContent;
|
|
|
|
return finalContent.length < trimmed.length ? finalContent : compactExistingSpecificationSectionBody(trimmed);
|
|
}
|
|
|
|
function compactUserCommentsSectionBody(body: string): string {
|
|
const trimmed = body.trim();
|
|
if (trimmed.length <= MAX_COMPACTED_USER_COMMENTS_CHARS) {
|
|
return trimmed;
|
|
}
|
|
|
|
const lines = trimmed.split("\n");
|
|
const commentLines = lines.filter((line) => /^- \*\*\[[^\]]+\]\*\*/.test(line));
|
|
const staticLines = lines.filter((line) => !/^- \*\*\[[^\]]+\]\*\*/.test(line));
|
|
const sortedComments = [...commentLines].sort((a, b) => {
|
|
const aDate = a.match(/^- \*\*\[([^\]]+)\]\*\*/)?.[1] ?? "";
|
|
const bDate = b.match(/^- \*\*\[([^\]]+)\]\*\*/)?.[1] ?? "";
|
|
return bDate.localeCompare(aDate);
|
|
});
|
|
|
|
const kept: string[] = [];
|
|
let used = staticLines.join("\n").length;
|
|
for (const line of sortedComments) {
|
|
const next = used + line.length + 1;
|
|
if (next > MAX_COMPACTED_USER_COMMENTS_CHARS) {
|
|
break;
|
|
}
|
|
kept.push(line);
|
|
used = next;
|
|
}
|
|
|
|
const trimmedCount = Math.max(0, commentLines.length - kept.length);
|
|
return [
|
|
...staticLines,
|
|
"",
|
|
...kept,
|
|
...(trimmedCount > 0 ? ["", `_... ${trimmedCount} earlier comments trimmed ..._`] : []),
|
|
].join("\n").trim();
|
|
}
|
|
|
|
function compactLargePromptSections(prompt: string): string | null {
|
|
const sectionPattern = /(^|\n)(## (?:Subtask Consideration|Subtask Breakdown Requested|Attachments|Existing Specification|Task PROMPT\.md|User Comments)\n)((?:\n*```markdown[\s\S]*?\n```|[\s\S]*?))(?=\n## [^#]|\n# [^#]|$)/g;
|
|
let changed = false;
|
|
|
|
const compactedPrompt = prompt.replace(sectionPattern, (match, prefix: string, heading: string, body: string) => {
|
|
const headingName = heading.trim().replace(/^##\s+/, "");
|
|
const trimmedBody = body.trim();
|
|
|
|
const maxByHeading: Record<string, number> = {
|
|
"Subtask Consideration": MAX_COMPACTED_SUBTASK_GUIDANCE_CHARS,
|
|
"Subtask Breakdown Requested": MAX_COMPACTED_SUBTASK_GUIDANCE_CHARS,
|
|
Attachments: MAX_COMPACTED_ATTACHMENTS_CHARS,
|
|
"Existing Specification": MAX_COMPACTED_EXISTING_SPEC_CHARS,
|
|
"Task PROMPT.md": MAX_COMPACTED_TASK_PROMPT_CHARS,
|
|
"User Comments": MAX_COMPACTED_USER_COMMENTS_CHARS,
|
|
};
|
|
|
|
const maxChars = maxByHeading[headingName] ?? MAX_COMPACTED_PROMPT_MEMORY_CHARS;
|
|
if (trimmedBody.length <= maxChars) {
|
|
return match;
|
|
}
|
|
|
|
let compactedBody = trimmedBody;
|
|
if (headingName === "Subtask Consideration" || headingName === "Subtask Breakdown Requested") {
|
|
compactedBody = trimSubtaskSectionBody(trimmedBody);
|
|
} else if (headingName === "Attachments") {
|
|
compactedBody = compactAttachmentSectionBody(trimmedBody);
|
|
} else if (headingName === "Existing Specification") {
|
|
compactedBody = compactExistingSpecificationSectionBody(trimmedBody);
|
|
} else if (headingName === "Task PROMPT.md") {
|
|
compactedBody = compactTaskPromptSectionBody(trimmedBody);
|
|
} else if (headingName === "User Comments") {
|
|
compactedBody = compactUserCommentsSectionBody(trimmedBody);
|
|
}
|
|
|
|
const finalBody = [
|
|
compactedBody,
|
|
"",
|
|
`<!-- Section trimmed from ${trimmedBody.length} characters to fit context window. -->`,
|
|
].join("\n").trim();
|
|
|
|
if (finalBody.length >= trimmedBody.length) {
|
|
return match;
|
|
}
|
|
|
|
changed = true;
|
|
return `${prefix}${heading}${finalBody}`;
|
|
});
|
|
|
|
return changed && compactedPrompt.length < prompt.length ? compactedPrompt : null;
|
|
}
|
|
|
|
export const __testOnlyPromptCompaction = {
|
|
compactLargePromptSections,
|
|
};
|
|
|
|
async function retryWithCompactedPromptMemory(
|
|
session: AgentSession,
|
|
prompt: string,
|
|
options?: unknown,
|
|
): Promise<{ recovered: boolean; error?: unknown }> {
|
|
const compactedPrompt = compactPromptMemory(prompt);
|
|
if (!compactedPrompt) {
|
|
return { recovered: false };
|
|
}
|
|
|
|
piLog.log(
|
|
`promptWithFallback: retrying with compacted prompt memory (${prompt.length} → ${compactedPrompt.length} chars)`,
|
|
);
|
|
|
|
try {
|
|
await promptSessionAndCheck(session, compactedPrompt, options);
|
|
piLog.log("promptWithFallback: prompt completed after prompt-memory compaction");
|
|
return { recovered: true };
|
|
} catch (err: unknown) {
|
|
const errorMessage = err instanceof Error ? err.message : String(err);
|
|
piLog.error(`promptWithFallback: retry after prompt-memory compaction failed: ${errorMessage}`);
|
|
return { recovered: false, error: err };
|
|
}
|
|
}
|
|
|
|
async function retryWithCompactedPromptSections(
|
|
session: AgentSession,
|
|
prompt: string,
|
|
options?: unknown,
|
|
): Promise<{ recovered: boolean; error?: unknown }> {
|
|
const compactedPrompt = compactLargePromptSections(prompt);
|
|
if (!compactedPrompt) {
|
|
return { recovered: false };
|
|
}
|
|
|
|
piLog.log(
|
|
`promptWithFallback: retrying with compacted prompt sections (${prompt.length} → ${compactedPrompt.length} chars)`,
|
|
);
|
|
|
|
try {
|
|
await promptSessionAndCheck(session, compactedPrompt, options);
|
|
piLog.log("promptWithFallback: prompt completed after section compaction");
|
|
return { recovered: true };
|
|
} catch (err: unknown) {
|
|
const errorMessage = err instanceof Error ? err.message : String(err);
|
|
piLog.error(`promptWithFallback: retry after section compaction failed: ${errorMessage}`);
|
|
return { recovered: false, error: err };
|
|
}
|
|
}
|
|
|
|
async function flushMemoryBeforeSessionCompaction(session: AgentSession): Promise<void> {
|
|
if ((session as any).__fusionMemoryAppendAvailable !== true) {
|
|
return;
|
|
}
|
|
|
|
const flushPrompt = [
|
|
"Before context compaction, preserve only unresolved durable memory if needed.",
|
|
"If fn_memory_append is available and you learned reusable project decisions/conventions/pitfalls/open loops or private operating context that is not already saved, append it now.",
|
|
"Use scope=\"project\" for shared workspace knowledge and scope=\"agent\" for private operating context.",
|
|
"Use layer=\"long-term\" for durable facts and layer=\"daily\" for running notes/open loops.",
|
|
"If there is nothing durable to save, reply exactly: NONE.",
|
|
].join("\n");
|
|
|
|
try {
|
|
await promptSessionAndCheck(session, flushPrompt);
|
|
} catch (err: unknown) {
|
|
const errorMessage = err instanceof Error ? err.message : String(err);
|
|
piLog.warn(`promptWithFallback: memory flush before compaction skipped: ${errorMessage}`);
|
|
}
|
|
}
|
|
|
|
/**
|
|
* Compact an agent session's context to free up the context window.
|
|
*
|
|
* Uses the SDK's native `session.compact()` method when available (the
|
|
* preferred path — it produces structured, LLM-generated summaries).
|
|
*
|
|
* @param session — The agent session to compact
|
|
* @param customInstructions — Optional instructions for the compaction summary.
|
|
* When not provided, uses COMPACTION_FALLBACK_INSTRUCTIONS.
|
|
* @returns The compaction result with summary and token metrics, or null if
|
|
* compaction was not available or failed.
|
|
*/
|
|
export async function compactSessionContext(
|
|
session: AgentSession,
|
|
customInstructions?: string,
|
|
): Promise<{ summary: string; tokensBefore: number } | null> {
|
|
const instructions = customInstructions ?? COMPACTION_FALLBACK_INSTRUCTIONS;
|
|
|
|
// Check if session.compact is available (runtime capability detection)
|
|
if (typeof (session as any).compact !== "function") {
|
|
return null;
|
|
}
|
|
|
|
try {
|
|
const result = await (session as any).compact(instructions);
|
|
if (result && typeof result === "object") {
|
|
return {
|
|
summary: result.summary ?? "",
|
|
tokensBefore: result.tokensBefore ?? 0,
|
|
};
|
|
}
|
|
return null;
|
|
} catch (err: unknown) {
|
|
const msg = err instanceof Error ? err.message : String(err);
|
|
piLog.warn(`Context compaction failed (will fall through to kill/requeue): ${msg}`);
|
|
return null;
|
|
}
|
|
}
|
|
|
|
export interface FallbackModelUsedPayload {
|
|
primaryModel: string;
|
|
fallbackModel: string;
|
|
triggerPoint: "session-creation" | "prompt-time";
|
|
taskId?: string;
|
|
taskTitle?: string;
|
|
timestamp?: string;
|
|
}
|
|
|
|
export type BuiltinWebToolName = "WebSearch" | "WebFetch";
|
|
|
|
export interface AgentOptions {
|
|
cwd: string;
|
|
systemPrompt: string;
|
|
/** Structured prompt layers for cross-session caching. When provided,
|
|
* the stable layer is used as systemPromptOverride and the dynamic
|
|
* layer as appendSystemPromptOverride. Falls back to systemPrompt
|
|
* when not provided. */
|
|
systemPromptLayers?: SystemPromptLayers;
|
|
tools?: "coding" | "readonly";
|
|
customTools?: ToolDefinition[];
|
|
/** Optional allowlist of builtin runtime web tools to keep enabled. */
|
|
builtinToolsAllowlist?: BuiltinWebToolName[];
|
|
onText?: (delta: string) => void;
|
|
onThinking?: (delta: string) => void;
|
|
onToolStart?: (name: string, args?: Record<string, unknown>) => void;
|
|
onToolEnd?: (name: string, isError: boolean, result?: unknown) => void;
|
|
/** Default model provider (e.g. "anthropic"). Used with `defaultModelId` to select a specific model. */
|
|
defaultProvider?: string;
|
|
/** Default model ID within the provider (e.g. "claude-sonnet-4-5"). Used with `defaultProvider`. */
|
|
defaultModelId?: string;
|
|
/** Optional fallback model provider used when the primary selected model hits
|
|
* a retryable provider-side failure such as rate limiting or overload. */
|
|
fallbackProvider?: string;
|
|
/** Optional fallback model ID used with `fallbackProvider`. */
|
|
fallbackModelId?: string;
|
|
/** Default thinking effort level (e.g. "medium", "high"). When provided, sets the session's thinking level after creation. */
|
|
defaultThinkingLevel?: string;
|
|
/** Optional pre-configured SessionManager. When provided, the agent session
|
|
* uses this instead of creating an in-memory session. Pass a file-based
|
|
* SessionManager to enable session persistence and pause/resume. */
|
|
sessionManager?: SessionManager;
|
|
/** Optional skill selection context. When provided, the agent session's
|
|
* skills are filtered according to project execution settings and any
|
|
* caller-requested skill names. Omit to use default skill discovery
|
|
* (all discovered skills included). */
|
|
skillSelection?: SkillSelectionContext;
|
|
/** Convenience: skill names to include in the session. When provided
|
|
* (and `skillSelection` is not), auto-constructs a SkillSelectionContext
|
|
* from the cwd and these names. Ignored when `skillSelection` is set. */
|
|
skills?: string[];
|
|
/** Extra directories to scan for skills (each holding `<id>/SKILL.md`), in
|
|
* addition to the default cwd/agent-dir roots. Forwarded to the resource
|
|
* loader so callers (e.g. plugins that install skills to a private dir) can
|
|
* make `skills`/`skillSelection` names discoverable in the live session. */
|
|
additionalSkillPaths?: string[];
|
|
/** Optional task-scoped env injected into this session's subprocess tools only. */
|
|
taskEnv?: NodeJS.ProcessEnv;
|
|
/** Last-chance abort hook fired immediately before `createAgentSession`.
|
|
* See `AgentRuntimeOptions.beforeSpawnSession`. */
|
|
beforeSpawnSession?: () => Promise<void> | void;
|
|
/** Callback fired when runtime falls back from primary model to fallback model. */
|
|
onFallbackModelUsed?: (payload: FallbackModelUsedPayload) => Promise<void> | void;
|
|
/** Optional task context for fallback notifications. */
|
|
taskId?: string;
|
|
taskTitle?: string;
|
|
actionGateContext?: AgentActionGateContext;
|
|
/** Permanent-agent action gating context forwarded by runtime/session helpers. */
|
|
permanentAgentGating?: PermanentAgentGatingContext;
|
|
}
|
|
|
|
function resolveCustomProviderApiType(apiType: string): "anthropic" | "openai-responses" | "openai-completions" {
|
|
if (apiType === "anthropic-compatible") {
|
|
return "anthropic";
|
|
}
|
|
if (apiType === "openai-responses") {
|
|
return "openai-responses";
|
|
}
|
|
return "openai-completions";
|
|
}
|
|
|
|
function resolveConfiguredModel(
|
|
modelRegistry: ModelRegistry,
|
|
kind: "primary" | "fallback",
|
|
provider?: string,
|
|
modelId?: string,
|
|
) {
|
|
if (!provider || !modelId) {
|
|
return undefined;
|
|
}
|
|
|
|
const model = modelRegistry.find(provider, modelId);
|
|
if (model) {
|
|
return model;
|
|
}
|
|
|
|
// Fall back to constructing a model on-the-fly if the provider is known.
|
|
// This mirrors the pi CLI's buildFallbackModel behaviour, which accepts any
|
|
// model ID for a configured provider (e.g. any OpenRouter model string) even
|
|
// when it isn't in the built-in or custom model list.
|
|
const providerModels = modelRegistry.getAll().filter((m) => m.provider === provider);
|
|
if (providerModels.length > 0) {
|
|
const baseModel = providerModels[0]!;
|
|
piLog.warn(`${kind} model ${provider}/${modelId} not in registry; using provider base model as template`);
|
|
return { ...baseModel, id: modelId, name: modelId };
|
|
}
|
|
|
|
throw new Error(
|
|
`Configured model ${provider}/${modelId} (${kind} selection) was not found in the pi model registry. `
|
|
+ "If this model comes from a custom provider, verify Settings → Custom Providers (stored in ~/.fusion/settings.json) includes this provider/model, "
|
|
+ "or choose an available model from /api/models.",
|
|
);
|
|
}
|
|
|
|
export function isRetryableModelSelectionError(message: string): boolean {
|
|
// An unsupported message-role rejection (e.g. a reasoning model sending the
|
|
// "developer" system role to a provider that only accepts
|
|
// system/user/assistant/tool) is fundamentally a model+provider
|
|
// compatibility problem. Treat it as a model-selection error so a configured
|
|
// fallback model is tried once before the task is marked failed. The
|
|
// `usingFallback` guard upstream keeps this to a single swap, so an
|
|
// incompatible fallback fails terminally rather than looping.
|
|
if (isUnsupportedMessageRoleError(message)) {
|
|
return true;
|
|
}
|
|
const normalized = message.toLowerCase();
|
|
return normalized.includes("rate limit")
|
|
|| normalized.includes("too many requests")
|
|
|| normalized.includes("429")
|
|
|| normalized.includes("401")
|
|
|| normalized.includes("403")
|
|
|| normalized.includes("unauthorized")
|
|
|| normalized.includes("forbidden")
|
|
|| normalized.includes("authentication")
|
|
|| normalized.includes("invalid api key")
|
|
|| normalized.includes("invalid key")
|
|
|| normalized.includes("api key")
|
|
|| normalized.includes("overloaded")
|
|
|| normalized.includes("quota")
|
|
|| normalized.includes("capacity")
|
|
|| normalized.includes("temporarily unavailable")
|
|
|| normalized.includes("invalid temperature");
|
|
}
|
|
|
|
interface PackageManagerSettingsView {
|
|
getGlobalSettings(): Record<string, any>;
|
|
getProjectSettings(): Record<string, any>;
|
|
getNpmCommand(): string[] | undefined;
|
|
}
|
|
|
|
function readJsonObject(path: string): Record<string, any> {
|
|
if (!existsSync(path)) {
|
|
return {};
|
|
}
|
|
|
|
try {
|
|
const parsed = JSON.parse(readFileSync(path, "utf-8"));
|
|
return parsed && typeof parsed === "object" ? parsed as Record<string, any> : {};
|
|
} catch {
|
|
return {};
|
|
}
|
|
}
|
|
|
|
function normalizeSessionHistoryEntries(sessionManager: SessionManagerLike): void {
|
|
const entries = sessionManager.fileEntries;
|
|
if (!Array.isArray(entries) || entries.length === 0) {
|
|
return;
|
|
}
|
|
|
|
let changed = false;
|
|
for (const entry of entries) {
|
|
if (entry?.type !== "message" || !entry.message || typeof entry.message !== "object") {
|
|
continue;
|
|
}
|
|
const role = entry.message.role;
|
|
if (role !== "assistant" && role !== "toolResult") {
|
|
continue;
|
|
}
|
|
if (!("content" in entry.message)) {
|
|
entry.message.content = [];
|
|
changed = true;
|
|
}
|
|
}
|
|
|
|
if (changed) {
|
|
sessionManager._rewriteFile?.();
|
|
}
|
|
}
|
|
|
|
function normalizeAssistantOrToolResultMessage(message: unknown): message is Record<string, unknown> {
|
|
if (!message || typeof message !== "object") {
|
|
return false;
|
|
}
|
|
|
|
const role = (message as Record<string, unknown>).role;
|
|
if (role !== "assistant" && role !== "toolResult" && role !== "user") {
|
|
return false;
|
|
}
|
|
|
|
const obj = message as Record<string, unknown>;
|
|
// `user` messages may carry content as a string (plain prompt) — leave those alone.
|
|
// For any other shape (undefined, null, object, etc.) coerce to an empty array so
|
|
// pi-coding-agent's _getUserMessageText (content.filter(...)) can't crash.
|
|
if (role === "user") {
|
|
if (typeof obj.content !== "string" && !Array.isArray(obj.content)) {
|
|
obj.content = [];
|
|
}
|
|
return true;
|
|
}
|
|
|
|
if (!Array.isArray(obj.content)) {
|
|
obj.content = [];
|
|
}
|
|
return true;
|
|
}
|
|
|
|
function syncNormalizedMessageIntoAgentState(session: AgentToolHookSession, message: Record<string, unknown>): void {
|
|
const messages = session.agent?.state?.messages;
|
|
if (!Array.isArray(messages) || messages.length === 0) {
|
|
return;
|
|
}
|
|
|
|
for (let i = messages.length - 1; i >= 0; i--) {
|
|
const candidate = messages[i];
|
|
if (!candidate || typeof candidate !== "object") {
|
|
continue;
|
|
}
|
|
if (candidate === message) {
|
|
normalizeAssistantOrToolResultMessage(candidate);
|
|
return;
|
|
}
|
|
|
|
if (candidate.role !== message.role) {
|
|
continue;
|
|
}
|
|
|
|
if (candidate.role === "toolResult") {
|
|
if (candidate.toolCallId === message.toolCallId && candidate.toolName === message.toolName) {
|
|
normalizeAssistantOrToolResultMessage(candidate);
|
|
return;
|
|
}
|
|
continue;
|
|
}
|
|
|
|
if (candidate.timestamp === message.timestamp) {
|
|
normalizeAssistantOrToolResultMessage(candidate);
|
|
return;
|
|
}
|
|
}
|
|
}
|
|
|
|
function installToolResultContentGuard(session: AgentToolHookSession): void {
|
|
if (session.__fusionToolResultGuardInstalled || !session.agent?.afterToolCall) {
|
|
return;
|
|
}
|
|
|
|
const originalAfterToolCall = session.agent.afterToolCall.bind(session.agent) as any;
|
|
(session.agent as any).afterToolCall = async (payload: ToolHookPayload) => {
|
|
const hookResult = await originalAfterToolCall(payload);
|
|
if (!hookResult || typeof hookResult !== "object") {
|
|
return hookResult;
|
|
}
|
|
|
|
const content = hookResult.content ?? payload.result?.content ?? [];
|
|
return {
|
|
content: Array.isArray(content) ? content : [],
|
|
details: hookResult.details ?? payload.result?.details,
|
|
isError: hookResult.isError ?? payload.isError,
|
|
};
|
|
};
|
|
session.__fusionToolResultGuardInstalled = true;
|
|
}
|
|
|
|
function installMessageContentGuard(session: AgentToolHookSession, sessionManager: SessionManagerLike): void {
|
|
if (session.__fusionMessageContentGuardInstalled) {
|
|
return;
|
|
}
|
|
|
|
// Sweep any pre-existing state.messages (e.g. restored from a session file)
|
|
// so messages with malformed content can't crash pi-coding-agent's
|
|
// _getUserMessageText / similar array traversals before our event hooks fire.
|
|
const existingMessages = session.agent?.state?.messages;
|
|
if (Array.isArray(existingMessages)) {
|
|
for (const candidate of existingMessages) {
|
|
normalizeAssistantOrToolResultMessage(candidate);
|
|
}
|
|
}
|
|
|
|
if (typeof session.subscribe === "function") {
|
|
session.subscribe((event: unknown) => {
|
|
if (!event || typeof event !== "object" || (event as { type?: string }).type !== "message_end") {
|
|
return;
|
|
}
|
|
|
|
const message = (event as { message?: unknown }).message;
|
|
if (!normalizeAssistantOrToolResultMessage(message)) {
|
|
return;
|
|
}
|
|
|
|
syncNormalizedMessageIntoAgentState(session, message);
|
|
});
|
|
}
|
|
|
|
if (typeof sessionManager.appendMessage === "function") {
|
|
const originalAppendMessage = sessionManager.appendMessage.bind(sessionManager);
|
|
sessionManager.appendMessage = (message: Record<string, unknown>) => {
|
|
normalizeAssistantOrToolResultMessage(message);
|
|
syncNormalizedMessageIntoAgentState(session, message);
|
|
return originalAppendMessage(message);
|
|
};
|
|
}
|
|
|
|
session.__fusionMessageContentGuardInstalled = true;
|
|
}
|
|
|
|
function hasPackageManagerSettings(settings: Record<string, any>): boolean {
|
|
return Array.isArray(settings.packages) || Array.isArray(settings.npmCommand);
|
|
}
|
|
|
|
function siblingAgentDir(agentDir: string, siblingRoot: ".fusion" | ".pi"): string | undefined {
|
|
if (basename(agentDir) !== "agent") {
|
|
return undefined;
|
|
}
|
|
return join(dirname(dirname(agentDir)), siblingRoot, "agent");
|
|
}
|
|
|
|
function createReadOnlyPiSettingsView(cwd: string, agentDir: string): PackageManagerSettingsView {
|
|
const projectRoot = resolvePiExtensionProjectRoot(cwd);
|
|
const fusionAgentDir = agentDir.includes(`${join(".fusion", "agent")}`)
|
|
? agentDir
|
|
: siblingAgentDir(agentDir, ".fusion");
|
|
const legacyAgentDir = agentDir.includes(`${join(".pi", "agent")}`)
|
|
? agentDir
|
|
: siblingAgentDir(agentDir, ".pi");
|
|
const legacyGlobalSettings = legacyAgentDir ? readJsonObject(join(legacyAgentDir, "settings.json")) : {};
|
|
const fusionGlobalSettings = fusionAgentDir ? readJsonObject(join(fusionAgentDir, "settings.json")) : {};
|
|
const directGlobalSettings = readJsonObject(join(agentDir, "settings.json"));
|
|
const globalSettings = { ...legacyGlobalSettings, ...directGlobalSettings, ...fusionGlobalSettings };
|
|
const fusionProjectSettings = readJsonObject(join(projectRoot, ".fusion", "settings.json"));
|
|
const mergedSettings = { ...globalSettings, ...fusionProjectSettings };
|
|
|
|
return {
|
|
getGlobalSettings: () => structuredClone(globalSettings),
|
|
getProjectSettings: () => structuredClone(fusionProjectSettings),
|
|
getNpmCommand: () => Array.isArray(mergedSettings.npmCommand)
|
|
? [...mergedSettings.npmCommand]
|
|
: undefined,
|
|
};
|
|
}
|
|
|
|
function getPackageManagerAgentDir(): string {
|
|
const fusionAgentDir = getFusionAgentDir();
|
|
const legacyAgentDir = getLegacyPiAgentDir();
|
|
const fusionSettings = readJsonObject(join(fusionAgentDir, "settings.json"));
|
|
const legacySettings = readJsonObject(join(legacyAgentDir, "settings.json"));
|
|
|
|
if (hasPackageManagerSettings(fusionSettings) || !existsSync(legacyAgentDir)) {
|
|
return fusionAgentDir;
|
|
}
|
|
if (hasPackageManagerSettings(legacySettings)) {
|
|
return legacyAgentDir;
|
|
}
|
|
return existsSync(fusionAgentDir) ? fusionAgentDir : legacyAgentDir;
|
|
}
|
|
|
|
/**
|
|
* Resolve the absolute path to Fusion's vendored `@fusion/pi-claude-cli`
|
|
* extension entry. Used by `registerExtensionProviders` to ensure the fork
|
|
* always wins over any externally-installed `pi-claude-cli`.
|
|
*
|
|
* Returns null when the vendored package isn't available (e.g. someone
|
|
* embedded `@fusion/engine` standalone without bundling the fork) — callers
|
|
* should treat that as "no override needed, leave external paths alone".
|
|
*/
|
|
function resolveVendoredClaudeCliEntry(): string | null {
|
|
try {
|
|
const require_ = createRequire(import.meta.url);
|
|
const pkgJsonPath = require_.resolve("@fusion/pi-claude-cli/package.json");
|
|
const pkgJson = JSON.parse(readFileSync(pkgJsonPath, "utf-8")) as {
|
|
pi?: { extensions?: unknown };
|
|
};
|
|
const extensions = pkgJson.pi?.extensions;
|
|
if (!Array.isArray(extensions) || extensions.length === 0) return null;
|
|
const entry = extensions[0];
|
|
if (typeof entry !== "string" || entry.length === 0) return null;
|
|
const path = resolve(dirname(pkgJsonPath), entry);
|
|
return existsSync(path) ? path : null;
|
|
} catch {
|
|
return null;
|
|
}
|
|
}
|
|
|
|
/**
|
|
* Resolve the absolute path to Fusion's vendored `@fusion/droid-cli`
|
|
* extension entry. Used by `registerExtensionProviders` to ensure the
|
|
* vendored extension always wins over any externally-installed `droid-cli`.
|
|
*/
|
|
function resolveVendoredDroidCliEntry(): string | null {
|
|
try {
|
|
const require_ = createRequire(import.meta.url);
|
|
const pkgJsonPath = require_.resolve("@fusion/droid-cli/package.json");
|
|
const pkgJson = JSON.parse(readFileSync(pkgJsonPath, "utf-8")) as {
|
|
pi?: { extensions?: unknown };
|
|
};
|
|
const extensions = pkgJson.pi?.extensions;
|
|
if (!Array.isArray(extensions) || extensions.length === 0) return null;
|
|
const entry = extensions[0];
|
|
if (typeof entry !== "string" || entry.length === 0) return null;
|
|
const path = resolve(dirname(pkgJsonPath), entry);
|
|
return existsSync(path) ? path : null;
|
|
} catch {
|
|
return null;
|
|
}
|
|
}
|
|
|
|
async function registerExtensionProviders(cwd: string, modelRegistry: ModelRegistry): Promise<void> {
|
|
try {
|
|
const agentDir = getPackageManagerAgentDir();
|
|
const packageManager = new DefaultPackageManager({
|
|
cwd,
|
|
agentDir,
|
|
settingsManager: createReadOnlyPiSettingsView(cwd, agentDir) as any,
|
|
});
|
|
const resolvedPaths = await packageManager.resolve();
|
|
const packageExtensionPaths = resolvedPaths.extensions
|
|
.filter((resource) => resource.enabled)
|
|
.map((resource) => resource.path);
|
|
|
|
// Always prefer Fusion's vendored `@fusion/pi-claude-cli` over any external
|
|
// `pi-claude-cli` install (e.g. a global `npm install -g pi-claude-cli`,
|
|
// or `npm:pi-claude-cli` in agent settings). Upstream has known timing
|
|
// and once-and-lock MCP-config bugs that we fix in the fork; loading both
|
|
// also produces unpredictable provider-registration winners.
|
|
const vendoredClaudeCli = resolveVendoredClaudeCliEntry();
|
|
const reconciledPaths = reconcileClaudeCliPaths(
|
|
[...getEnabledPiExtensionPaths(cwd), ...packageExtensionPaths],
|
|
vendoredClaudeCli,
|
|
);
|
|
|
|
// Prefer Fusion's vendored `@fusion/droid-cli` over any external
|
|
// `droid-cli` install. Side-by-side loading of two extensions that
|
|
// register the same provider name produces unpredictable winners.
|
|
const vendoredDroidCli = resolveVendoredDroidCliEntry();
|
|
const doubleReconciledPaths = reconcileDroidCliPaths(
|
|
reconciledPaths,
|
|
vendoredDroidCli,
|
|
);
|
|
|
|
const extensionsResult = await discoverAndLoadExtensions(
|
|
doubleReconciledPaths,
|
|
cwd,
|
|
join(resolvePiExtensionProjectRoot(cwd), ".fusion", "disabled-auto-extension-discovery"),
|
|
);
|
|
|
|
for (const { path, error } of extensionsResult.errors) {
|
|
extensionsLog.warn(`Failed to load ${path}: ${error}`);
|
|
}
|
|
|
|
for (const { name, config, extensionPath } of extensionsResult.runtime.pendingProviderRegistrations) {
|
|
try {
|
|
modelRegistry.registerProvider(name, config);
|
|
} catch (error) {
|
|
const message = error instanceof Error ? error.message : String(error);
|
|
extensionsLog.warn(`Failed to register provider from ${extensionPath}: ${message}`);
|
|
}
|
|
}
|
|
|
|
extensionsResult.runtime.pendingProviderRegistrations = [];
|
|
modelRegistry.refresh();
|
|
} catch (error) {
|
|
const message = error instanceof Error ? error.message : String(error);
|
|
extensionsLog.error(`Failed to discover extensions: ${message}`);
|
|
createExtensionRuntime();
|
|
modelRegistry.refresh();
|
|
}
|
|
}
|
|
|
|
// ── Worktree Path Boundary Helpers ──────────────────────────────────────────
|
|
|
|
export { getProjectRootFromWorktree };
|
|
|
|
async function isRegisteredGitWorktree(projectRoot: string, worktreePath: string): Promise<boolean> {
|
|
try {
|
|
const { stdout } = await execAsync("git worktree list --porcelain", {
|
|
cwd: projectRoot,
|
|
encoding: "utf-8",
|
|
});
|
|
const resolvedWorktree = normalizeExistingPathForGitComparison(worktreePath);
|
|
return stdout.split("\n").some((line) =>
|
|
line.startsWith("worktree ") && normalizeExistingPathForGitComparison(line.slice("worktree ".length)) === resolvedWorktree
|
|
);
|
|
} catch {
|
|
return false;
|
|
}
|
|
}
|
|
|
|
function normalizeExistingPathForGitComparison(path: string): string {
|
|
try {
|
|
return realpathSync.native(path);
|
|
} catch {
|
|
return resolve(path);
|
|
}
|
|
}
|
|
|
|
async function isCompleteGitWorktree(worktreePath: string): Promise<boolean> {
|
|
try {
|
|
const { stdout } = await execAsync("git rev-parse --show-toplevel", {
|
|
cwd: worktreePath,
|
|
encoding: "utf-8",
|
|
});
|
|
return normalizeExistingPathForGitComparison(stdout.trim()) === normalizeExistingPathForGitComparison(worktreePath);
|
|
} catch {
|
|
return false;
|
|
}
|
|
}
|
|
|
|
async function assertValidWorktreeSession(cwd: string, projectRoot: string): Promise<void> {
|
|
if (!existsSync(cwd)) {
|
|
throw new Error(`Refusing to start coding agent in missing worktree: ${cwd}`);
|
|
}
|
|
if (!existsSync(join(cwd, ".git")) || !await isCompleteGitWorktree(cwd)) {
|
|
throw new Error(`Refusing to start coding agent in incomplete worktree: ${cwd}`);
|
|
}
|
|
if (!await isRegisteredGitWorktree(projectRoot, cwd)) {
|
|
throw new Error(`Refusing to start coding agent in unregistered git worktree: ${cwd}`);
|
|
}
|
|
}
|
|
|
|
/**
|
|
* Check if a path is allowed to be accessed from a worktree session.
|
|
* Rules:
|
|
* - Paths inside the worktree are always allowed
|
|
* - Project root .fusion/memory/ files are allowed (for durable project learnings)
|
|
* - Task attachments under .fusion/tasks/N/attachments/ are allowed (for reading context files)
|
|
* - Sibling task specs (.fusion/tasks/N/PROMPT.md and task.json) are allowed for
|
|
* read-only tools (read/glob/grep) so agents can consult dependency specs.
|
|
* - All other paths outside the worktree are rejected
|
|
*
|
|
* @param worktreePath - Absolute path to the worktree directory
|
|
* @param projectRoot - Absolute path to the project root (derived from worktree)
|
|
* @param requestedPath - The path being accessed
|
|
* @param toolName - Tool making the request (controls read-only exceptions)
|
|
* @returns true if allowed, false if rejected
|
|
*/
|
|
function isWorktreeAllowedPath(
|
|
worktreePath: string,
|
|
projectRoot: string,
|
|
requestedPath: string,
|
|
toolName?: string,
|
|
): boolean {
|
|
// Normalize paths
|
|
const worktreeResolved = resolve(worktreePath);
|
|
const projectRootResolved = resolve(projectRoot);
|
|
const requestedResolved = isAbsolute(requestedPath) ? resolve(requestedPath) : resolve(worktreeResolved, requestedPath);
|
|
|
|
// Check if path is inside the worktree
|
|
const relToWorktree = relative(worktreeResolved, requestedResolved);
|
|
if (!relToWorktree.startsWith("..") && !isAbsolute(relToWorktree)) {
|
|
return true; // Path is inside the worktree
|
|
}
|
|
|
|
// Exception: project root `.fusion/memory/` files for durable project learnings
|
|
const relToProjectRoot = relative(projectRootResolved, requestedResolved).replace(/\\/g, "/");
|
|
if (
|
|
relToProjectRoot === ".fusion/memory" ||
|
|
relToProjectRoot === ".fusion/memory/" ||
|
|
relToProjectRoot.startsWith(".fusion/memory/")
|
|
) {
|
|
return true;
|
|
}
|
|
|
|
// Exception: task attachments under `.fusion/tasks/*/attachments/*`
|
|
if (relToProjectRoot.match(/^\.fusion\/tasks\/[^/]+\/attachments\//)) {
|
|
return true;
|
|
}
|
|
|
|
// Exception (read-only): sibling task specs so the agent can consult the
|
|
// PROMPT.md / task.json of dependency tasks without needing them copied
|
|
// into the worktree. `glob`/`grep` are narrow enough to allow as well so
|
|
// the agent can discover them; writes and bash remain restricted.
|
|
const readOnlyTools = new Set(["read", "glob", "grep"]);
|
|
if (toolName && readOnlyTools.has(toolName) &&
|
|
/^\.fusion\/tasks\/[^/]+\/(PROMPT\.md|task\.json)$/.test(relToProjectRoot)) {
|
|
return true;
|
|
}
|
|
|
|
// All other paths outside the worktree are rejected
|
|
return false;
|
|
}
|
|
|
|
/**
|
|
* Wrap tools with worktree boundary validation.
|
|
* When cwd is a worktree path, file operations are validated against worktree boundaries.
|
|
*
|
|
* @param tools - Array of tool definitions to wrap
|
|
* @param worktreePath - Absolute path to the worktree directory (if applicable)
|
|
* @param projectRoot - Absolute path to the project root (if applicable)
|
|
* @returns Wrapped tools with boundary validation
|
|
*/
|
|
/**
|
|
* Build a tool result payload in the shape pi-coding-agent / pi-ai expect
|
|
* (content as an array of typed blocks, isError=true) rather than a bare
|
|
* `{ok:false,error}` object. Returning the bare object leaves the toolResult
|
|
* message with `content: undefined`, which pi's downstream handling later
|
|
* crashes on with "Cannot read properties of undefined (reading 'filter')".
|
|
*/
|
|
function boundaryRejection(message: string, details?: Record<string, unknown>) {
|
|
return {
|
|
content: [{ type: "text", text: message }],
|
|
isError: true,
|
|
ok: false,
|
|
error: message,
|
|
...(details ? { details } : {}),
|
|
};
|
|
}
|
|
|
|
function normalizeApprovalRequestCategory(
|
|
category: PermanentAgentActionCategory,
|
|
): AgentPermissionPolicyActionCategory {
|
|
if (category === "none") {
|
|
return "command_execution";
|
|
}
|
|
return category;
|
|
}
|
|
|
|
function buildPermanentAgentApprovalDedupeKey(input: {
|
|
requesterActorId?: string;
|
|
taskId?: string;
|
|
toolName: string;
|
|
category: PermanentAgentActionCategory;
|
|
}): string {
|
|
return [
|
|
input.requesterActorId ?? "",
|
|
input.taskId ?? "",
|
|
input.toolName,
|
|
input.category,
|
|
].join("|");
|
|
}
|
|
|
|
const GATE_BYPASS_TOOL_NAMES = new Set([
|
|
"fn_heartbeat_done",
|
|
"fn_send_message",
|
|
"fn_post_room_message",
|
|
]);
|
|
|
|
export function wrapToolsWithBoundary(
|
|
tools: ToolDefinition[],
|
|
worktreePath: string | null,
|
|
projectRoot: string | null,
|
|
): ToolDefinition[] {
|
|
if (!worktreePath || !projectRoot) {
|
|
return tools; // Not a worktree session, no wrapping needed
|
|
}
|
|
|
|
return tools.map((tool) => {
|
|
// Only wrap tools that access the filesystem
|
|
const fileToolNames = new Set(["read", "write", "edit", "glob", "grep", "bash"]);
|
|
if (!fileToolNames.has(tool.name)) {
|
|
return tool;
|
|
}
|
|
|
|
// Store the original execute function
|
|
const originalExecute = tool.execute as any;
|
|
|
|
|
|
return {
|
|
...tool,
|
|
|
|
execute: async (...args: any[]) => {
|
|
const _toolCallId = args[0] as string;
|
|
const params = args[1] as Record<string, unknown>;
|
|
const _signal = args[2] as AbortSignal | undefined;
|
|
|
|
// Check path argument for file operations
|
|
const pathArg = params.path as string | undefined;
|
|
if (pathArg && !isWorktreeAllowedPath(worktreePath, projectRoot, pathArg, tool.name)) {
|
|
const relToProject = relative(projectRoot, pathArg);
|
|
return boundaryRejection(
|
|
`Path "${relToProject}" is outside the worktree boundary. ` +
|
|
`Coding agents can only modify files inside the current worktree. ` +
|
|
`Exceptions (read-only): .fusion/memory/, .fusion/tasks/*/attachments/, ` +
|
|
`and .fusion/tasks/*/{PROMPT.md,task.json} for dependency context.`,
|
|
);
|
|
}
|
|
|
|
// For bash, also check the working directory if specified
|
|
const cwdArg = params.cwd as string | undefined;
|
|
if (tool.name === "bash" && cwdArg && !isWorktreeAllowedPath(worktreePath, projectRoot, cwdArg, tool.name)) {
|
|
return boundaryRejection(
|
|
`Working directory is outside the worktree boundary. ` +
|
|
`Commands must run inside the worktree.`,
|
|
);
|
|
}
|
|
|
|
// Call the original tool implementation with all arguments passed through
|
|
return originalExecute(...args);
|
|
},
|
|
};
|
|
});
|
|
}
|
|
|
|
export function wrapToolsWithRtkRewrite(
|
|
tools: ToolDefinition[],
|
|
options: RtkRewriteOptions = resolveRtkRewriteOptions(),
|
|
): ToolDefinition[] {
|
|
const resolvedOptions = normalizeRtkRewriteOptions(options);
|
|
|
|
if (resolvedOptions.mode !== "rewrite") {
|
|
return tools;
|
|
}
|
|
|
|
return tools.map((tool) => {
|
|
if (tool.name !== "bash") {
|
|
return tool;
|
|
}
|
|
|
|
const originalExecute = tool.execute as any;
|
|
return {
|
|
...tool,
|
|
execute: async (...args: any[]) => {
|
|
const params = args[1] as Record<string, unknown> | undefined;
|
|
const command = params?.command;
|
|
if (typeof command !== "string" || !command.trim()) {
|
|
return originalExecute(...args);
|
|
}
|
|
|
|
const signal = args[2] as AbortSignal | undefined;
|
|
const rewrittenCommand = await rewriteCommandWithRtk(command, resolvedOptions, signal);
|
|
if (!rewrittenCommand) {
|
|
return originalExecute(...args);
|
|
}
|
|
|
|
const rewrittenArgs = [...args];
|
|
rewrittenArgs[1] = { ...(params ?? {}), command: rewrittenCommand };
|
|
return originalExecute(...rewrittenArgs);
|
|
},
|
|
};
|
|
});
|
|
}
|
|
|
|
export function wrapToolsWithPermanentAgentGating(
|
|
tools: ToolDefinition[],
|
|
gating: PermanentAgentGatingContext | undefined,
|
|
): ToolDefinition[] {
|
|
if (!gating) {
|
|
return tools;
|
|
}
|
|
|
|
return tools.map((tool) => {
|
|
// FN-3852/FN-3855: terminal completion and send-message coordination
|
|
// primitives must never be approval-gated, or open sessions can deadlock.
|
|
if (GATE_BYPASS_TOOL_NAMES.has(tool.name)) {
|
|
return tool;
|
|
}
|
|
|
|
const originalExecute = tool.execute as any;
|
|
return {
|
|
...tool,
|
|
execute: async (...args: any[]) => {
|
|
const params = (args[1] ?? {}) as Record<string, unknown>;
|
|
const decision = resolvePermanentAgentToolDecision({
|
|
toolName: tool.name,
|
|
args: params,
|
|
gating,
|
|
});
|
|
|
|
if (decision.disposition === "allow") {
|
|
return originalExecute(...args);
|
|
}
|
|
|
|
const details: Record<string, unknown> = {
|
|
disposition: decision.disposition,
|
|
category: decision.category,
|
|
toolName: decision.toolName,
|
|
...(decision.disposition === "require-approval" ? { requiresApproval: true } : {}),
|
|
};
|
|
|
|
if (decision.disposition === "require-approval") {
|
|
const dedupeKey = buildPermanentAgentApprovalDedupeKey({
|
|
requesterActorId: gating.requester?.actorId,
|
|
taskId: gating.taskId,
|
|
toolName: decision.toolName,
|
|
category: decision.category,
|
|
});
|
|
details.approvalDedupeKey = dedupeKey;
|
|
|
|
let approvalRequest = await gating.findPendingApprovalRequest?.(dedupeKey);
|
|
if (!approvalRequest && gating.createApprovalRequest) {
|
|
approvalRequest = await gating.createApprovalRequest({
|
|
category: normalizeApprovalRequestCategory(decision.category),
|
|
toolName: decision.toolName,
|
|
args: params,
|
|
});
|
|
}
|
|
|
|
if (approvalRequest?.id) {
|
|
details.approvalRequestId = approvalRequest.id;
|
|
}
|
|
}
|
|
|
|
const reason = decision.disposition === "block"
|
|
? `Action blocked by permanent-agent policy (${decision.category}) for tool ${decision.toolName}`
|
|
: `Action requires approval (${decision.category}) before tool ${decision.toolName} can run`;
|
|
|
|
return boundaryRejection(reason, details);
|
|
},
|
|
};
|
|
});
|
|
}
|
|
|
|
export function wrapToolsWithActionGate(
|
|
tools: ToolDefinition[],
|
|
gateContext: AgentActionGateContext | undefined,
|
|
): ToolDefinition[] {
|
|
if (!gateContext || gateContext.isEphemeral) {
|
|
return tools;
|
|
}
|
|
|
|
return tools.map((tool) => {
|
|
// FN-3852/FN-3855: terminal completion and send-message coordination
|
|
// primitives must never be approval-gated, or open sessions can deadlock.
|
|
if (GATE_BYPASS_TOOL_NAMES.has(tool.name)) {
|
|
return tool;
|
|
}
|
|
|
|
const originalExecute = tool.execute as any;
|
|
return {
|
|
...tool,
|
|
execute: async (...args: any[]) => {
|
|
const params = (args[1] ?? {}) as Record<string, unknown>;
|
|
const decision = evaluateAgentActionGate({
|
|
agentId: gateContext.agentId,
|
|
taskId: gateContext.taskId,
|
|
toolName: tool.name,
|
|
args: params,
|
|
permissionPolicy: gateContext.permissionPolicy,
|
|
});
|
|
|
|
const latestApproval = gateContext.findApprovalByDedupeKey
|
|
? await gateContext.findApprovalByDedupeKey(decision.approvalDedupeKey)
|
|
: await gateContext.findPendingApprovalByDedupeKey?.(decision.approvalDedupeKey).then((request) =>
|
|
request ? { id: request.id, status: "pending" as const } : null
|
|
);
|
|
|
|
const gateOutcome = resolveGateOutcome(decision, latestApproval ?? null);
|
|
|
|
if (gateOutcome.outcome === "allow") {
|
|
return originalExecute(...args);
|
|
}
|
|
|
|
if (gateOutcome.outcome === "execute-once-then-complete") {
|
|
const result = await originalExecute(...args);
|
|
if (gateOutcome.approvalRequestId) {
|
|
await gateContext.markApprovalCompleted?.(gateOutcome.approvalRequestId);
|
|
}
|
|
return result;
|
|
}
|
|
|
|
if (gateOutcome.outcome === "block") {
|
|
if (latestApproval?.status === "denied") {
|
|
return buildGateRejection(
|
|
{
|
|
...decision,
|
|
metadata: {
|
|
...decision.metadata,
|
|
approvalRequestId: latestApproval.id,
|
|
dedupeKey: decision.approvalDedupeKey,
|
|
},
|
|
},
|
|
"Action was denied by approver. The agent must not retry this action.",
|
|
);
|
|
}
|
|
|
|
return buildGateRejection(
|
|
decision,
|
|
`Action blocked by permission policy (${decision.category}) for ${gateContext.agentName}`,
|
|
);
|
|
}
|
|
|
|
let approvalRequestId = gateOutcome.approvalRequestId;
|
|
if (!approvalRequestId) {
|
|
const created = await gateContext.createApprovalRequest(decision, params) as { id?: string } | null;
|
|
approvalRequestId = created?.id;
|
|
if (approvalRequestId) {
|
|
await gateContext.pauseForApproval?.({ approvalRequestId, decision });
|
|
}
|
|
}
|
|
|
|
return buildGateRejection(
|
|
{
|
|
...decision,
|
|
metadata: {
|
|
...decision.metadata,
|
|
...(approvalRequestId ? { approvalRequestId } : {}),
|
|
dedupeKey: decision.approvalDedupeKey,
|
|
},
|
|
},
|
|
`Action requires approval (request ${approvalRequestId ?? "pending"}). Agent and task have been paused; will resume once a decision is made.`,
|
|
);
|
|
},
|
|
};
|
|
});
|
|
}
|
|
|
|
/**
|
|
* Create a pi agent session configured for fn.
|
|
* Reuses the user's existing pi auth and model configuration.
|
|
*
|
|
* Returned sessions are wrapped so `session.dispose()` emits pi's
|
|
* `session_shutdown` extension event before teardown.
|
|
*/
|
|
export async function createFnAgent(options: AgentOptions): Promise<AgentResult> {
|
|
piLog.log(`createFnAgent called (tools=${options.tools}, provider=${options.defaultProvider}, model=${options.defaultModelId})`);
|
|
const authStorage = createFusionAuthStorage();
|
|
const modelRegistry = ModelRegistry.create(authStorage, getModelRegistryModelsPath());
|
|
|
|
// Resolve the project root early so extension providers, skill discovery,
|
|
// and resource loading all use the correct root when cwd is a worktree,
|
|
// subdirectory, or any path other than the project root itself.
|
|
const resolvedProjectRoot = getProjectRootFromWorktree(options.cwd) ?? resolvePiExtensionProjectRoot(options.cwd);
|
|
await registerExtensionProviders(resolvedProjectRoot, modelRegistry);
|
|
|
|
const customProviders = readCustomProviders();
|
|
for (const provider of customProviders) {
|
|
try {
|
|
const registryKey = customProviderRegistryKey(provider, customProviders);
|
|
modelRegistry.registerProvider(registryKey, {
|
|
baseUrl: provider.baseUrl,
|
|
api: resolveCustomProviderApiType(provider.apiType),
|
|
apiKey: provider.apiKey,
|
|
models: (provider.models ?? []).map((model) => ({
|
|
id: model.id,
|
|
name: model.name,
|
|
reasoning: false,
|
|
input: ["text" as const],
|
|
cost: {
|
|
input: 0,
|
|
output: 0,
|
|
cacheRead: 0,
|
|
cacheWrite: 0,
|
|
},
|
|
contextWindow: 128000,
|
|
maxTokens: 16384,
|
|
})),
|
|
});
|
|
piLog.log(`Registered custom provider "${provider.name}" (key=${registryKey}, id=${provider.id})`);
|
|
} catch (error) {
|
|
const message = error instanceof Error ? error.message : String(error);
|
|
const registryKey = customProviderRegistryKey(provider, customProviders);
|
|
piLog.warn(`Failed to register custom provider "${provider.name}" (key=${registryKey}, id=${provider.id}, apiType=${provider.apiType}, baseUrl=${provider.baseUrl}): ${message}`);
|
|
}
|
|
}
|
|
modelRegistry.refresh();
|
|
|
|
// Build the pi built-in tool set. We deliberately do NOT use the bundled
|
|
// `createCodingTools` / `createReadOnlyTools` presets — they're missing
|
|
// tools that pi-claude-cli's Claude→pi name mapping depends on (Glob→find,
|
|
// Grep→grep). When a coding session ran via Claude CLI tried `Glob`, pi
|
|
// returned "Tool find not found" and the agent looped. Compose explicitly
|
|
// so every tool referenced by tool-mapping.ts is registered.
|
|
const bashToolOptions = options.taskEnv
|
|
? {
|
|
spawnHook: ({ command, cwd, env }: { command: string; cwd: string; env: NodeJS.ProcessEnv }) => ({
|
|
command,
|
|
cwd,
|
|
env: {
|
|
...env,
|
|
...options.taskEnv,
|
|
},
|
|
}),
|
|
}
|
|
: undefined;
|
|
|
|
const isReadonly = options.tools === "readonly";
|
|
const builtins = [
|
|
createReadTool(options.cwd),
|
|
createBashTool(options.cwd, bashToolOptions),
|
|
createEditTool(options.cwd),
|
|
createWriteTool(options.cwd),
|
|
createGrepTool(options.cwd),
|
|
createFindTool(options.cwd),
|
|
createLsTool(options.cwd),
|
|
] as ToolDefinition[];
|
|
const tools = isReadonly
|
|
? builtins.filter((tool) => isReadonlyAllowed(tool.name))
|
|
: builtins;
|
|
// Suppress lint about unused presets — kept in scope for incremental migration.
|
|
void createCodingTools;
|
|
void createReadOnlyTools;
|
|
|
|
// Detect if this is a worktree session and apply path boundaries
|
|
const worktreePath = options.cwd;
|
|
const worktreeProjectRoot = getProjectRootFromWorktree(worktreePath);
|
|
if (worktreeProjectRoot) {
|
|
await assertValidWorktreeSession(worktreePath, worktreeProjectRoot);
|
|
}
|
|
const boundaryContext = { worktreePath, worktreeProjectRoot };
|
|
|
|
// resolvedProjectRoot was computed above (before registerExtensionProviders)
|
|
// and is reused here for resource loader and skill discovery.
|
|
|
|
// Compaction is explicitly enabled to prevent context-window overflow during
|
|
// long-running agent conversations (triage, execution, review, merge).
|
|
// When the context fills up, pi auto-compacts the conversation history to
|
|
// keep the session alive without manual intervention. This must remain enabled
|
|
// as a reliability safeguard — disabling it would cause overflow failures.
|
|
const settingsManager = SettingsManager.inMemory({
|
|
compaction: { enabled: true },
|
|
retry: { enabled: true, maxRetries: 3 },
|
|
});
|
|
|
|
// Resolve explicit model selection if provider and model ID are specified.
|
|
// If the primary configured model cannot be resolved but a fallback model is
|
|
// configured, prefer the fallback as the initial model selection.
|
|
let selectedModel;
|
|
let fallbackModel;
|
|
try {
|
|
selectedModel = resolveConfiguredModel(
|
|
modelRegistry,
|
|
"primary",
|
|
options.defaultProvider,
|
|
options.defaultModelId,
|
|
);
|
|
} catch (primaryResolutionError) {
|
|
if (!options.fallbackProvider || !options.fallbackModelId) {
|
|
throw primaryResolutionError;
|
|
}
|
|
fallbackModel = resolveConfiguredModel(
|
|
modelRegistry,
|
|
"fallback",
|
|
options.fallbackProvider,
|
|
options.fallbackModelId,
|
|
);
|
|
selectedModel = fallbackModel;
|
|
}
|
|
|
|
if (!fallbackModel) {
|
|
fallbackModel = resolveConfiguredModel(
|
|
modelRegistry,
|
|
"fallback",
|
|
options.fallbackProvider,
|
|
options.fallbackModelId,
|
|
);
|
|
}
|
|
|
|
// Resolve skill selection: explicit skillSelection wins over convenience `skills`
|
|
let effectiveSkillSelection: SkillSelectionContext | undefined = options.skillSelection;
|
|
if (!effectiveSkillSelection && options.skills && options.skills.length > 0) {
|
|
piLog.log(`Using skills from convenience parameter: [${options.skills.join(", ")}]`);
|
|
effectiveSkillSelection = {
|
|
projectRootDir: resolvedProjectRoot,
|
|
requestedSkillNames: options.skills,
|
|
sessionPurpose: "executor",
|
|
};
|
|
}
|
|
|
|
// Resolve skill selection if provided
|
|
let skillsOverrideFn: ReturnType<typeof createSkillsOverrideFromSelection> | undefined;
|
|
if (effectiveSkillSelection) {
|
|
const selectionResult = resolveSessionSkills(effectiveSkillSelection);
|
|
if (selectionResult.diagnostics.length > 0) {
|
|
const purpose = effectiveSkillSelection.sessionPurpose ?? "skills";
|
|
for (const diag of selectionResult.diagnostics) {
|
|
const msg = `[skills] [${purpose}] ${diag.type}: ${diag.message}`;
|
|
if (diag.type === "error") piLog.error(msg);
|
|
else if (diag.type === "warning") piLog.warn(msg);
|
|
else piLog.log(msg);
|
|
}
|
|
}
|
|
skillsOverrideFn = createSkillsOverrideFromSelection(selectionResult, {
|
|
requestedSkillNames: effectiveSkillSelection.requestedSkillNames,
|
|
sessionPurpose: effectiveSkillSelection.sessionPurpose,
|
|
});
|
|
}
|
|
|
|
// `tools: "readonly"` MUST mean a hermetically sealed read-only session with
|
|
// respect to host extension injection. Host extensions (`@runfusion/fusion`)
|
|
// can register write tools like `fn_task_create`, so they are deliberately
|
|
// EXCLUDED in readonly mode. Caller-supplied `customTools` are preserved,
|
|
// since heartbeat/reviewer flows explicitly provide engine-owned tools.
|
|
// This keeps summarizer/compaction sessions safe while retaining intended
|
|
// delegation/memory tools for readonly engine sessions.
|
|
const effectiveExtensionPaths = isReadonly ? [] : hostExtensionPaths;
|
|
if (isReadonly && hostExtensionPaths.length > 0) {
|
|
piLog.log(`readonly session — host extensions (${hostExtensionPaths.length}) skipped`);
|
|
}
|
|
|
|
const resourceLoader = new DefaultResourceLoader({
|
|
cwd: resolvedProjectRoot,
|
|
agentDir: getFusionAgentDir(),
|
|
settingsManager,
|
|
systemPromptOverride: () => options.systemPromptLayers?.stable ?? options.systemPrompt,
|
|
appendSystemPromptOverride: () =>
|
|
options.systemPromptLayers?.dynamic
|
|
? [options.systemPromptLayers.dynamic]
|
|
: [],
|
|
...(effectiveExtensionPaths.length > 0 ? { additionalExtensionPaths: [...effectiveExtensionPaths] } : {}),
|
|
...(options.additionalSkillPaths && options.additionalSkillPaths.length > 0
|
|
? { additionalSkillPaths: [...options.additionalSkillPaths] }
|
|
: {}),
|
|
...(skillsOverrideFn ? { skillsOverride: skillsOverrideFn } : {}),
|
|
});
|
|
await resourceLoader.reload();
|
|
|
|
const sessionManager = options.sessionManager ?? SessionManager.inMemory();
|
|
normalizeSessionHistoryEntries(sessionManager as unknown as SessionManagerLike);
|
|
|
|
const createSessionWithModel = async (modelOverride?: typeof selectedModel) => {
|
|
// pi-coding-agent 0.68+: `tools` is a string[] allowlist of tool names, not
|
|
// Tool instances. We need boundary-wrapped versions of the built-ins, so we
|
|
// suppress the defaults with `noTools: "builtin"` and register our wrapped
|
|
// tools through `customTools` instead. The wrapped tools preserve the same
|
|
// names (`read`, `bash`, ...) as the built-ins they replace.
|
|
const readonlyFilteredCustomTools = isReadonly
|
|
? filterCustomToolsForReadonly(options.customTools ?? [])
|
|
: { allowed: options.customTools ?? [], denied: [] };
|
|
if (isReadonly && readonlyFilteredCustomTools.denied.length > 0) {
|
|
piLog.warn(
|
|
`[pi] readonly mode: dropped ${readonlyFilteredCustomTools.denied.length} denied custom tool(s): ${readonlyFilteredCustomTools.denied.join(", ")}`,
|
|
);
|
|
}
|
|
|
|
const toolChainStart: ToolDefinition[] = [
|
|
...(tools as ToolDefinition[]),
|
|
...readonlyFilteredCustomTools.allowed,
|
|
];
|
|
const toolsWithRtkRewrite = wrapToolsWithRtkRewrite(toolChainStart);
|
|
const toolsWithPermanentGating = wrapToolsWithPermanentAgentGating(
|
|
toolsWithRtkRewrite,
|
|
options.permanentAgentGating,
|
|
);
|
|
const toolsWithActionGate = wrapToolsWithActionGate(
|
|
toolsWithPermanentGating,
|
|
options.actionGateContext,
|
|
);
|
|
const customToolList: ToolDefinition[] = wrapToolsWithBoundary(
|
|
toolsWithActionGate,
|
|
boundaryContext.worktreePath,
|
|
boundaryContext.worktreeProjectRoot,
|
|
);
|
|
// Sort tools alphabetically by name for deterministic ordering.
|
|
// Prompt caching requires the tool list to be byte-identical across
|
|
// sessions — reordering breaks cache prefix matching.
|
|
// Exception: fn_heartbeat_done must remain last (stable terminal signal
|
|
// required by the heartbeat executor — see agent-heartbeat.ts).
|
|
customToolList.sort((a, b) => a.name.localeCompare(b.name));
|
|
const heartbeatDoneIdx = customToolList.findIndex((t) => t.name === "fn_heartbeat_done");
|
|
if (heartbeatDoneIdx >= 0 && heartbeatDoneIdx < customToolList.length - 1) {
|
|
const [doneTool] = customToolList.splice(heartbeatDoneIdx, 1);
|
|
customToolList.push(doneTool);
|
|
}
|
|
// Last-chance abort hook. Fires *here* — after every awaited setup step
|
|
// in createFnAgent (provider registration, worktree validation, resource
|
|
// loader reload) and immediately before the actual LLM session spawn.
|
|
// This is the latest synchronous decision point where the engine can
|
|
// honor a pause that flipped during this function's setup window.
|
|
if (options.beforeSpawnSession) {
|
|
await options.beforeSpawnSession();
|
|
}
|
|
const createSessionOptions: Parameters<typeof createAgentSession>[0] = {
|
|
cwd: options.cwd,
|
|
authStorage,
|
|
modelRegistry,
|
|
resourceLoader,
|
|
noTools: "builtin",
|
|
customTools: customToolList,
|
|
sessionManager,
|
|
settingsManager,
|
|
...(modelOverride ? { model: modelOverride } : {}),
|
|
};
|
|
|
|
if (options.builtinToolsAllowlist && options.builtinToolsAllowlist.length > 0) {
|
|
const safeBuiltinAllowlist = isReadonly
|
|
? options.builtinToolsAllowlist.filter((name) => READONLY_ALLOWLIST.includes(name as (typeof READONLY_ALLOWLIST)[number]))
|
|
: options.builtinToolsAllowlist;
|
|
createSessionOptions.tools = [
|
|
...new Set([
|
|
...customToolList.map((tool) => tool.name),
|
|
...safeBuiltinAllowlist,
|
|
]),
|
|
].sort();
|
|
}
|
|
|
|
return createAgentSession(createSessionOptions);
|
|
};
|
|
|
|
const emitFallbackUsed = async (triggerPoint: "session-creation" | "prompt-time"): Promise<void> => {
|
|
if (!options.onFallbackModelUsed || !selectedModel || !fallbackModel) {
|
|
return;
|
|
}
|
|
await options.onFallbackModelUsed({
|
|
primaryModel: `${selectedModel.provider}/${selectedModel.id}`,
|
|
fallbackModel: `${fallbackModel.provider}/${fallbackModel.id}`,
|
|
triggerPoint,
|
|
taskId: options.taskId,
|
|
taskTitle: options.taskTitle,
|
|
timestamp: new Date().toISOString(),
|
|
});
|
|
};
|
|
|
|
let sessionResult;
|
|
let usingFallback = false;
|
|
try {
|
|
sessionResult = await createSessionWithModel(selectedModel);
|
|
piLog.log(`Session created successfully (model=${selectedModel ? `${selectedModel.provider}/${selectedModel.id}` : "default"})`);
|
|
} catch (err: any) {
|
|
if (!fallbackModel || !selectedModel || !isRetryableModelSelectionError(err?.message || "")) {
|
|
piLog.error(`Session creation failed: ${err.message}`);
|
|
throw err;
|
|
}
|
|
piLog.warn(`Primary model failed (${err.message}), trying fallback`);
|
|
usingFallback = true;
|
|
sessionResult = await createSessionWithModel(fallbackModel);
|
|
await emitFallbackUsed("session-creation");
|
|
piLog.log("Fallback session created successfully");
|
|
}
|
|
|
|
let activeSession = sessionResult.session;
|
|
wrapSessionDisposeWithShutdown(activeSession);
|
|
installToolResultContentGuard(activeSession as AgentToolHookSession);
|
|
installMessageContentGuard(activeSession as AgentToolHookSession, sessionManager as unknown as SessionManagerLike);
|
|
(activeSession as any).__fusionMemoryAppendAvailable = options.customTools?.some((tool) => tool.name === FN_MEMORY_APPEND_TOOL_NAME) === true;
|
|
const promptableSession = activeSession as PromptableSession;
|
|
|
|
let thinkingCompatibilityDisabled = false;
|
|
const applyThinkingLevelIfSupported = (targetSession: AgentSession, sourceModel: string): void => {
|
|
if (!options.defaultThinkingLevel || thinkingCompatibilityDisabled) {
|
|
return;
|
|
}
|
|
try {
|
|
(targetSession as PromptableSession).setThinkingLevel(options.defaultThinkingLevel as any);
|
|
} catch (err: unknown) {
|
|
const message = err instanceof Error ? err.message : String(err);
|
|
if (!isThinkingReasoningConflictError(message)) {
|
|
throw err;
|
|
}
|
|
thinkingCompatibilityDisabled = true;
|
|
piLog.warn(`Disabling explicit thinking level for model ${sourceModel}: ${message}`);
|
|
}
|
|
};
|
|
|
|
const wireFallbackHooks = (targetSession: PromptableSession): void => {
|
|
installToolResultContentGuard(targetSession as unknown as AgentToolHookSession);
|
|
installMessageContentGuard(
|
|
targetSession as unknown as AgentToolHookSession,
|
|
sessionManager as unknown as SessionManagerLike,
|
|
);
|
|
(targetSession as any).__fusionMemoryAppendAvailable = options.customTools?.some((tool) => tool.name === FN_MEMORY_APPEND_TOOL_NAME) === true;
|
|
const deltaNormalizer = createStreamingDeltaNormalizer();
|
|
targetSession.subscribe((event) => {
|
|
if (event.type === "message_update") {
|
|
const msgEvent = event.assistantMessageEvent;
|
|
if (msgEvent.type === "text_delta") {
|
|
// Repair dropped sentence-boundary spaces at the shared engine delta chokepoint,
|
|
// including tool-call cross-message boundaries (see streaming-delta.ts).
|
|
options.onText?.(deltaNormalizer.normalize(msgEvent.partial, msgEvent.contentIndex, msgEvent.delta, "text"));
|
|
} else if (msgEvent.type === "thinking_delta") {
|
|
// Repair dropped sentence-boundary spaces at the shared engine delta chokepoint,
|
|
// including tool-call cross-message boundaries (see streaming-delta.ts).
|
|
options.onThinking?.(deltaNormalizer.normalize(msgEvent.partial, msgEvent.contentIndex, msgEvent.delta, "thinking"));
|
|
}
|
|
}
|
|
if (event.type === "tool_execution_start") {
|
|
options.onToolStart?.(event.toolName, event.args as Record<string, unknown> | undefined);
|
|
}
|
|
if (event.type === "tool_execution_end") {
|
|
options.onToolEnd?.(event.toolName, event.isError, event.result);
|
|
}
|
|
});
|
|
};
|
|
|
|
const swapPromptSession = async (modelToUse: typeof selectedModel): Promise<PromptableSession> => {
|
|
if (!modelToUse) {
|
|
throw new Error("Cannot swap session without a resolved model");
|
|
}
|
|
wrapSessionDisposeWithShutdown(activeSession);
|
|
try {
|
|
activeSession.dispose();
|
|
} catch (err: unknown) {
|
|
const msg = err instanceof Error ? err.message : String(err);
|
|
piLog.warn(`Failed to dispose session during swap: ${msg}`);
|
|
}
|
|
const next = (await createSessionWithModel(modelToUse)).session as PromptableSession;
|
|
wireFallbackHooks(next);
|
|
wrapSessionDisposeWithShutdown(next);
|
|
applyThinkingLevelIfSupported(next, `${modelToUse.provider}/${modelToUse.id}`);
|
|
Object.setPrototypeOf(promptableSession, Object.getPrototypeOf(next));
|
|
Object.assign(promptableSession, next);
|
|
promptableSession.promptWithFallback = next.promptWithFallback ?? promptableSession.promptWithFallback;
|
|
activeSession = next;
|
|
return next;
|
|
};
|
|
|
|
promptableSession.promptWithFallback = async (prompt: string, promptOptions?: unknown) => {
|
|
try {
|
|
await promptSessionAndCheck(activeSession, prompt, promptOptions);
|
|
return;
|
|
} catch (err: any) {
|
|
const errorMessage = err?.message || "";
|
|
if (isContextLimitError(errorMessage)) {
|
|
// Context limit error — attempt auto-compaction and retry once
|
|
const promptMemoryRetry = await retryWithCompactedPromptMemory(activeSession, prompt, promptOptions);
|
|
if (promptMemoryRetry.recovered) {
|
|
return;
|
|
}
|
|
if (promptMemoryRetry.error) {
|
|
const retryMessage = promptMemoryRetry.error instanceof Error ? promptMemoryRetry.error.message : String(promptMemoryRetry.error);
|
|
if (!isContextLimitError(retryMessage)) {
|
|
throw promptMemoryRetry.error;
|
|
}
|
|
}
|
|
|
|
const promptSectionRetry = await retryWithCompactedPromptSections(activeSession, prompt, promptOptions);
|
|
if (promptSectionRetry.recovered) {
|
|
return;
|
|
}
|
|
if (promptSectionRetry.error) {
|
|
const retryMessage = promptSectionRetry.error instanceof Error ? promptSectionRetry.error.message : String(promptSectionRetry.error);
|
|
if (!isContextLimitError(retryMessage)) {
|
|
throw promptSectionRetry.error;
|
|
}
|
|
}
|
|
|
|
piLog.warn("promptWithFallback: context limit error — attempting auto-compaction");
|
|
await flushMemoryBeforeSessionCompaction(activeSession);
|
|
const compactResult = await compactSessionContext(activeSession);
|
|
if (compactResult) {
|
|
piLog.log(`promptWithFallback: compaction succeeded (${compactResult.tokensBefore} tokens) — retrying prompt`);
|
|
try {
|
|
await promptSessionAndCheck(activeSession, prompt, promptOptions);
|
|
return;
|
|
} catch (retryErr: any) {
|
|
const retryErrorMessage = retryErr?.message || "";
|
|
piLog.error(`promptWithFallback: retry after auto-compaction failed: ${retryErrorMessage}`);
|
|
// Throw original error to preserve original context
|
|
throw err;
|
|
}
|
|
} else {
|
|
piLog.error("promptWithFallback: compaction unavailable — propagating original error");
|
|
throw err;
|
|
}
|
|
}
|
|
|
|
if (!usingFallback && options.defaultThinkingLevel && !thinkingCompatibilityDisabled && isThinkingReasoningConflictError(errorMessage)) {
|
|
thinkingCompatibilityDisabled = true;
|
|
piLog.warn(`Prompt failed with thinking/reasoning conflict; retrying without explicit thinking level: ${errorMessage}`);
|
|
const recoveredSession = await swapPromptSession(selectedModel);
|
|
await promptSessionAndCheck(recoveredSession, prompt, promptOptions);
|
|
return;
|
|
}
|
|
|
|
if (!fallbackModel || usingFallback || !isRetryableModelSelectionError(errorMessage)) {
|
|
throw err;
|
|
}
|
|
|
|
usingFallback = true;
|
|
const fallbackSession = await swapPromptSession(fallbackModel);
|
|
await emitFallbackUsed("prompt-time");
|
|
|
|
// Retry with fallback model, also with auto-compaction support
|
|
try {
|
|
await promptSessionAndCheck(fallbackSession, prompt, promptOptions);
|
|
return;
|
|
} catch (fallbackErr: any) {
|
|
const fallbackErrorMessage = fallbackErr?.message || "";
|
|
if (isContextLimitError(fallbackErrorMessage)) {
|
|
const promptMemoryRetry = await retryWithCompactedPromptMemory(fallbackSession, prompt, promptOptions);
|
|
if (promptMemoryRetry.recovered) {
|
|
return;
|
|
}
|
|
if (promptMemoryRetry.error) {
|
|
const retryMessage = promptMemoryRetry.error instanceof Error ? promptMemoryRetry.error.message : String(promptMemoryRetry.error);
|
|
if (!isContextLimitError(retryMessage)) {
|
|
throw promptMemoryRetry.error;
|
|
}
|
|
}
|
|
|
|
const promptSectionRetry = await retryWithCompactedPromptSections(fallbackSession, prompt, promptOptions);
|
|
if (promptSectionRetry.recovered) {
|
|
return;
|
|
}
|
|
if (promptSectionRetry.error) {
|
|
const retryMessage = promptSectionRetry.error instanceof Error ? promptSectionRetry.error.message : String(promptSectionRetry.error);
|
|
if (!isContextLimitError(retryMessage)) {
|
|
throw promptSectionRetry.error;
|
|
}
|
|
}
|
|
|
|
piLog.warn("promptWithFallback: fallback session context limit error — attempting auto-compaction");
|
|
await flushMemoryBeforeSessionCompaction(fallbackSession);
|
|
const compactResult = await compactSessionContext(fallbackSession);
|
|
if (compactResult) {
|
|
piLog.log(`promptWithFallback: fallback compaction succeeded (${compactResult.tokensBefore} tokens) — retrying`);
|
|
try {
|
|
await promptSessionAndCheck(fallbackSession, prompt, promptOptions);
|
|
return;
|
|
} catch (retryErr: any) {
|
|
const retryErrorMessage = retryErr?.message || "";
|
|
piLog.error(`promptWithFallback: fallback retry after auto-compaction failed: ${retryErrorMessage}`);
|
|
throw fallbackErr; // Throw original fallback error
|
|
}
|
|
} else {
|
|
piLog.error("promptWithFallback: fallback compaction unavailable — propagating original error");
|
|
throw fallbackErr;
|
|
}
|
|
}
|
|
throw fallbackErr;
|
|
}
|
|
}
|
|
};
|
|
|
|
// Apply thinking level if specified (with compatibility fallback).
|
|
applyThinkingLevelIfSupported(
|
|
promptableSession,
|
|
selectedModel ? `${selectedModel.provider}/${selectedModel.id}` : describeModel(promptableSession),
|
|
);
|
|
|
|
// Wire up event listeners
|
|
const deltaNormalizer = createStreamingDeltaNormalizer();
|
|
promptableSession.subscribe((event) => {
|
|
if (event.type === "message_update") {
|
|
const msgEvent = event.assistantMessageEvent;
|
|
if (msgEvent.type === "text_delta") {
|
|
// Repair dropped sentence-boundary spaces at the shared engine delta chokepoint,
|
|
// including tool-call cross-message boundaries (see streaming-delta.ts).
|
|
options.onText?.(deltaNormalizer.normalize(msgEvent.partial, msgEvent.contentIndex, msgEvent.delta, "text"));
|
|
} else if (msgEvent.type === "thinking_delta") {
|
|
// Repair dropped sentence-boundary spaces at the shared engine delta chokepoint,
|
|
// including tool-call cross-message boundaries (see streaming-delta.ts).
|
|
options.onThinking?.(deltaNormalizer.normalize(msgEvent.partial, msgEvent.contentIndex, msgEvent.delta, "thinking"));
|
|
}
|
|
}
|
|
if (event.type === "tool_execution_start") {
|
|
options.onToolStart?.(event.toolName, event.args as Record<string, unknown> | undefined);
|
|
}
|
|
if (event.type === "tool_execution_end") {
|
|
options.onToolEnd?.(event.toolName, event.isError, event.result);
|
|
}
|
|
});
|
|
|
|
return { session: promptableSession, sessionFile: promptableSession.sessionFile };
|
|
}
|