@piecemaker-legal/piecemaker 1.0.0

package/server/modules/websocket/services/chat-websocket.service.ts

@@ -0,0 +1,654 @@
+import path from 'node:path';
+
+import type { WebSocket } from 'ws';
+
+import { sessionsDb } from '@/modules/database/index.js';
+import { providerModelsService, sessionsService } from '@/modules/providers/index.js';
+import { chatRunRegistry } from '@/modules/websocket/services/chat-run-registry.service.js';
+import { connectedClients, WS_OPEN_STATE } from '@/modules/websocket/services/websocket-state.service.js';
+import {
+  getGlobalImageAssetsDir,
+  isImageAttachmentDescriptor,
+  normalizeAttachmentDescriptors,
+  type ChatAttachmentDescriptor,
+} from '@/shared/image-attachments.js';
+import type {
+  AnyRecord,
+  AuthenticatedWebSocketRequest,
+  LLMProvider,
+  ProviderPermissionDecision,
+  ProviderRuntimeWriter,
+} from '@/shared/types.js';
+import { parseIncomingJsonObject } from '@/shared/utils.js';
+
+/**
+ * Trust boundary for client-supplied image attachments: chat.send options come
+ * straight from the browser, and the provider runtimes read the referenced
+ * files off disk (Claude base64-encodes them into the prompt). Only images
+ * that live directly inside the global upload store (`~/.cloudcli/assets`,
+ * where POST /api/assets/images puts them) are allowed through — anything
+ * else (absolute paths elsewhere, traversal, subdirectories) is dropped.
+ *
+ * Exported for tests; `assetsRootOverride` exists only for them.
+ */
+export function filterAttachmentsToUploadStore(
+  attachments: unknown,
+  assetsRootOverride?: string,
+): ChatAttachmentDescriptor[] {
+  const assetsRoot = path.resolve(assetsRootOverride ?? getGlobalImageAssetsDir());
+
+  return normalizeAttachmentDescriptors(attachments).filter((descriptor) => {
+    // Relative paths are anchored in the store; absolute ones must already be in it.
+    const resolved = path.resolve(assetsRoot, descriptor.path);
+    const relative = path.relative(assetsRoot, resolved);
+    const isDirectChild =
+      relative.length > 0 &&
+      !relative.startsWith('..') &&
+      !path.isAbsolute(relative) &&
+      !relative.includes(path.sep) &&
+      !relative.includes('/');
+
+    if (!isDirectChild) {
+      console.warn(`[Chat] Dropping attachment outside the upload store: ${descriptor.path}`);
+    }
+    return isDirectChild;
+  });
+}
+
+/** Backward-compatible image filter consumed by existing websocket tests. */
+export function filterImagesToUploadStore(
+  images: unknown,
+  assetsRootOverride?: string,
+): ChatAttachmentDescriptor[] {
+  return filterAttachmentsToUploadStore(images, assetsRootOverride);
+}
+
+/** Application boundary for dispatching provider runs and approvals. */
+export type ProviderRuntimeGateway = {
+  hasRuntime(provider: string): boolean;
+  run(
+    provider: LLMProvider,
+    command: string,
+    options: AnyRecord,
+    writer: ProviderRuntimeWriter,
+  ): Promise<unknown>;
+  abort(provider: LLMProvider, sessionId: string): Promise<boolean>;
+  resolveToolApproval(requestId: string, payload: ProviderPermissionDecision): void;
+  getPendingApprovalsForSession(sessionId: string): unknown[];
+};
+
+type ChatWebSocketDependencies = {
+  /** Central dispatcher for every provider SDK/CLI runtime. */
+  runtime: ProviderRuntimeGateway;
+};
+
+/**
+ * Extracts the authenticated request user id in the formats currently produced
+ * by platform and OSS auth code paths.
+ */
+function readRequestUserId(
+  request: AuthenticatedWebSocketRequest | undefined
+): string | number | null {
+  const user = request?.user;
+  if (!user) {
+    return null;
+  }
+
+  if (typeof user.id === 'string' || typeof user.id === 'number') {
+    return user.id;
+  }
+
+  if (typeof user.userId === 'string' || typeof user.userId === 'number') {
+    return user.userId;
+  }
+
+  return null;
+}
+
+function sendJson(ws: WebSocket, payload: unknown): void {
+  if (ws.readyState === WS_OPEN_STATE) {
+    ws.send(JSON.stringify(payload));
+  }
+}
+
+/**
+ * Reports a protocol-level failure to the requesting client.
+ *
+ * Protocol errors deliberately use their own `kind` (instead of the provider
+ * `error` message kind) so the frontend can distinguish "your request was
+ * invalid" from "the model run produced an error" without inspecting text.
+ */
+function sendProtocolError(
+  ws: WebSocket,
+  code: string,
+  error: string,
+  sessionId?: string
+): void {
+  sendJson(ws, {
+    kind: 'protocol_error',
+    code,
+    error,
+    sessionId: sessionId ?? null,
+    timestamp: new Date().toISOString(),
+  });
+}
+
+function readRequiredSessionId(data: AnyRecord): string | null {
+  const sessionId = typeof data.sessionId === 'string' ? data.sessionId.trim() : '';
+  return sessionId.length > 0 ? sessionId : null;
+}
+
+/**
+ * Handles `chat.send`: resolves the session row (provider, project path, and
+ * provider-native id all come from the database — never from the client),
+ * registers the run, and dispatches to the provider runtime.
+ */
+async function handleChatSend(
+  ws: WebSocket,
+  userId: string | number | null,
+  data: AnyRecord,
+  dependencies: ChatWebSocketDependencies
+): Promise<void> {
+  const resolved = resolveSendTarget(ws, data, dependencies, 'chat.send');
+  if (!resolved) {
+    return;
+  }
+
+  await dispatchRun(ws, userId, resolved.sessionId, resolved.session, data, dependencies);
+}
+
+type ResolvedSendTarget = {
+  sessionId: string;
+  session: NonNullable<ReturnType<typeof sessionsDb.getSessionById>>;
+  provider: LLMProvider;
+};
+
+/**
+ * Shared front half of `chat.send` and `chat.edit-send`: the session row and
+ * provider come from the database, never from the client.
+ */
+function resolveSendTarget(
+  ws: WebSocket,
+  data: AnyRecord,
+  dependencies: ChatWebSocketDependencies,
+  frameName: string,
+): ResolvedSendTarget | null {
+  const sessionId = readRequiredSessionId(data);
+  if (!sessionId) {
+    sendProtocolError(ws, 'SESSION_ID_REQUIRED', `${frameName} requires a sessionId.`);
+    return null;
+  }
+
+  const session = sessionsDb.getSessionById(sessionId);
+  if (!session) {
+    sendProtocolError(
+      ws,
+      'SESSION_NOT_FOUND',
+      `Session "${sessionId}" was not found. Create it via POST /api/providers/sessions first.`,
+      sessionId
+    );
+    return null;
+  }
+
+  const provider = session.provider as LLMProvider;
+  if (!dependencies.runtime.hasRuntime(provider)) {
+    sendProtocolError(ws, 'UNSUPPORTED_PROVIDER', `Provider "${provider}" is not available.`, sessionId);
+    return null;
+  }
+
+  return { sessionId, session, provider };
+}
+
+/**
+ * Registers the run and hands the turn to the provider runtime.
+ *
+ * `extraRuntimeOptions` is how an edited message asks the provider to resume
+ * partway instead of continuing from the tip; a normal send passes nothing.
+ */
+async function dispatchRun(
+  ws: WebSocket | null,
+  userId: string | number | null,
+  sessionId: string,
+  session: NonNullable<ReturnType<typeof sessionsDb.getSessionById>>,
+  data: AnyRecord,
+  dependencies: ChatWebSocketDependencies,
+  extraRuntimeOptions: AnyRecord = {},
+  beforeRun?: (run: NonNullable<ReturnType<typeof chatRunRegistry.startRun>>) => void | Promise<void>,
+): Promise<{ started: boolean; error: string | null }> {
+  const provider = session.provider as LLMProvider;
+
+  const run = chatRunRegistry.startRun({
+    appSessionId: sessionId,
+    provider,
+    providerSessionId: session.provider_session_id,
+    connection: ws,
+    userId,
+  });
+
+  if (!run) {
+    if (ws) {
+      sendProtocolError(
+        ws,
+        'RUN_IN_PROGRESS',
+        `Session "${sessionId}" already has a run in progress.`,
+        sessionId
+      );
+    }
+    return { started: false, error: 'A run is already in progress for this session.' };
+  }
+
+  const clientOptions = (data.options ?? {}) as AnyRecord;
+  const command = typeof data.content === 'string' ? data.content : '';
+
+  // Record what this turn runs with so reopening the session later restores the
+  // same model and reasoning effort, and so the resume path has a
+  // session-scoped model answer to use.
+  if (typeof clientOptions.model === 'string' && clientOptions.model.trim()) {
+    providerModelsService.setSessionModel(provider, sessionId, clientOptions.model);
+  }
+  if (typeof clientOptions.effort === 'string' && clientOptions.effort.trim()) {
+    providerModelsService.setSessionEffort(provider, sessionId, clientOptions.effort);
+  }
+
+  const attachmentCandidates = [
+    ...normalizeAttachmentDescriptors(clientOptions.images),
+    ...normalizeAttachmentDescriptors(clientOptions.files),
+    ...normalizeAttachmentDescriptors(clientOptions.attachments),
+  ];
+  const verifiedAttachments = filterAttachmentsToUploadStore(attachmentCandidates);
+  const uniqueAttachments = verifiedAttachments.filter(
+    (descriptor, index, all) => all.findIndex((candidate) => candidate.path === descriptor.path) === index,
+  );
+
+  // The provider runtimes receive the stable app session id. When their
+  // CLI/SDK needs the provider-native id for resume, they resolve it from the
+  // session row themselves (sessionsService.resolveProviderSessionId).
+  // Brand-new sessions have no provider id yet, so the runtime starts fresh
+  // and announces one, which the gateway writer captures and maps back to the
+  // app session id.
+  const runtimeOptions: AnyRecord = {
+    ...clientOptions,
+    ...extraRuntimeOptions,
+    // Attachments are re-validated server-side: only direct children of the
+    // global upload store may reach provider runtimes or their file tools.
+    attachments: uniqueAttachments,
+    images: uniqueAttachments.filter(isImageAttachmentDescriptor),
+    files: uniqueAttachments.filter((descriptor) => !isImageAttachmentDescriptor(descriptor)),
+    sessionId,
+    cwd: clientOptions.cwd ?? session.project_path ?? undefined,
+    projectPath: session.project_path ?? clientOptions.projectPath,
+  };
+
+  let failure: string | null = null;
+  try {
+    // Runs only now that the session is reserved, because an edit rewinds the
+    // conversation here and a rewind for a run that was never admitted cannot
+    // be taken back. Inside the try so a rewind that throws still releases the
+    // run instead of leaving the session processing forever.
+    await beforeRun?.(run);
+    await dependencies.runtime.run(provider, command, runtimeOptions, run.writer);
+  } catch (error) {
+    failure = error instanceof Error ? error.message : String(error);
+    console.error(`[Chat] Provider runtime "${provider}" failed`, { sessionId, error: failure });
+  } finally {
+    // Safety net: a runtime that crashed (or resolved) without emitting its
+    // terminal `complete` would otherwise leave the session stuck in
+    // "processing" forever on every connected client. Scoped to THIS run —
+    // a queued message can start the session's next run before this promise
+    // settles, and the session-keyed completeRun would kill that new run.
+    chatRunRegistry.completeRunIfCurrent(run, { exitCode: 1 });
+  }
+
+  return { started: true, error: failure };
+}
+
+/**
+ * Handles `chat.edit-send`: replaces an already-sent message and everything
+ * after it with a new turn.
+ *
+ * Nothing is deleted. The provider resumes the conversation partway and
+ * appends the replacement, so the abandoned attempt stays in the transcript
+ * file and is simply no longer part of the live conversation — the same shape
+ * Claude Code's rewind and Codex's fork-with-cut-point produce.
+ */
+async function handleChatEditSend(
+  ws: WebSocket,
+  userId: string | number | null,
+  data: AnyRecord,
+  dependencies: ChatWebSocketDependencies
+): Promise<void> {
+  const resolved = resolveSendTarget(ws, data, dependencies, 'chat.edit-send');
+  if (!resolved) {
+    return;
+  }
+
+  const { sessionId, session, provider } = resolved;
+  const anchorId = typeof data.anchorId === 'string' ? data.anchorId.trim() : '';
+  if (!anchorId) {
+    sendProtocolError(ws, 'ANCHOR_REQUIRED', 'chat.edit-send requires the anchorId of the message being replaced.', sessionId);
+    return;
+  }
+
+  let resumeThroughId: string | null;
+  try {
+    const anchor = await sessionsService.resolveEditAnchor(sessionId, anchorId);
+    if (!anchor) {
+      sendProtocolError(
+        ws,
+        'EDIT_NOT_SUPPORTED',
+        `Provider "${provider}" cannot replace an already-sent message.`,
+        sessionId
+      );
+      return;
+    }
+    if (!anchor.found) {
+      sendProtocolError(ws, 'ANCHOR_NOT_FOUND', 'That message is no longer in the transcript.', sessionId);
+      return;
+    }
+    resumeThroughId = anchor.resumeThroughId;
+  } catch (error) {
+    const message = error instanceof Error ? error.message : String(error);
+    sendProtocolError(ws, 'ANCHOR_LOOKUP_FAILED', `Could not read the transcript: ${message}`, sessionId);
+    return;
+  }
+
+  // Providers split here on what their runtime can do. Claude resumes its
+  // transcript partway, so the anchor rides along as a run option. Codex
+  // cannot — a thread only grows — so the conversation is rewound on disk and
+  // the run that follows is an ordinary resume of whatever the session then
+  // points at. Which of the two applies is decided here; the rewind itself
+  // waits until the run has actually been admitted.
+  const rewinds = sessionsService.providerRewindsForEdit(sessionId);
+
+  await dispatchRun(
+    ws,
+    userId,
+    sessionId,
+    session,
+    data,
+    dependencies,
+    // `null` is meaningful: the edited turn was the first prompt, so the
+    // conversation starts over instead of resuming.
+    rewinds
+      ? {}
+      : { resumeAnchorId: resumeThroughId ?? undefined, resumeFromScratch: resumeThroughId === null },
+    async (run) => {
+      // Emitted through the run's writer so it is sequenced and replayed like
+      // any other event — a second tab watching this session has to truncate
+      // too.
+      //
+      // Before the rewind, not after it. A rewind that has to branch spawns a
+      // process and waits on a JSON-RPC round trip, and holding the frame
+      // until that came back left the message the user had just edited away
+      // sitting on screen for about a second — the very flicker this feature
+      // exists to avoid. Announcing first is safe because a rewind that fails
+      // still ends the run, and the terminal `complete` makes every client
+      // re-read the transcript, which puts back anything that turned out not
+      // to have been replaced after all.
+      run.writer.send({
+        kind: 'history_truncated',
+        provider,
+        sessionId,
+        anchorId,
+      });
+
+      if (rewinds) {
+        try {
+          await sessionsService.rewindSessionForEdit(sessionId, resumeThroughId);
+        } catch (error) {
+          const message = error instanceof Error ? error.message : String(error);
+          sendProtocolError(ws, 'EDIT_REWIND_FAILED', `Could not rewind the conversation: ${message}`, sessionId);
+          // Ends the run before the provider is asked to continue a
+          // conversation that was not rewound after all.
+          throw error;
+        }
+      }
+    },
+  );
+}
+
+/**
+ * Handles `chat.abort`: cancels the run for one app session and emits the
+ * terminal `complete` on its behalf (runtimes skip their own complete for
+ * aborted runs, and the registry drops any duplicate).
+ */
+async function handleChatAbort(
+  ws: WebSocket,
+  data: AnyRecord,
+  dependencies: ChatWebSocketDependencies
+): Promise<void> {
+  const sessionId = readRequiredSessionId(data);
+  if (!sessionId) {
+    sendProtocolError(ws, 'SESSION_ID_REQUIRED', 'chat.abort requires a sessionId.');
+    return;
+  }
+
+  const run = chatRunRegistry.getRun(sessionId);
+  if (!run || run.status !== 'running') {
+    sendProtocolError(ws, 'NO_ACTIVE_RUN', `Session "${sessionId}" has no active run.`, sessionId);
+    return;
+  }
+
+  const success = await dependencies.runtime.abort(run.provider, sessionId);
+
+  chatRunRegistry.completeRun(sessionId, {
+    exitCode: success ? 0 : 1,
+    aborted: true,
+  });
+}
+
+/**
+ * Handles `chat.subscribe`: for each requested session, reports whether a run
+ * is processing, re-attaches the live stream to this socket, replays missed
+ * events (seq > lastSeq), and includes pending permission requests.
+ *
+ * This single message replaces the old `check-session-status`,
+ * `get-pending-permissions`, and Claude-only writer reconnect flows.
+ */
+function handleChatSubscribe(
+  ws: WebSocket,
+  data: AnyRecord,
+  dependencies: ChatWebSocketDependencies
+): void {
+  const targets = Array.isArray(data.sessions) ? data.sessions : [];
+
+  for (const target of targets) {
+    if (!target || typeof target !== 'object') {
+      continue;
+    }
+
+    const sessionId = typeof (target as AnyRecord).sessionId === 'string'
+      ? ((target as AnyRecord).sessionId as string).trim()
+      : '';
+    if (!sessionId) {
+      continue;
+    }
+
+    const lastSeqRaw = (target as AnyRecord).lastSeq;
+    const lastSeq = typeof lastSeqRaw === 'number' && Number.isFinite(lastSeqRaw)
+      ? Math.max(0, Math.floor(lastSeqRaw))
+      : 0;
+
+    const run = chatRunRegistry.getRun(sessionId);
+    const isProcessing = chatRunRegistry.isProcessing(sessionId);
+
+    // Future live events for this run should land on the socket that asked —
+    // this is what makes mid-stream page refreshes work for all providers.
+    if (isProcessing) {
+      chatRunRegistry.attachConnection(sessionId, ws);
+    }
+
+    // Pending approvals are tracked under the app session id inside the
+    // Claude runtime, so they can be looked up directly.
+    const pendingPermissions = dependencies.runtime.getPendingApprovalsForSession(sessionId);
+
+    sendJson(ws, {
+      kind: 'chat_subscribed',
+      sessionId,
+      isProcessing,
+      lastSeq: run?.lastSeq ?? 0,
+      pendingPermissions,
+      timestamp: new Date().toISOString(),
+    });
+
+    // Replay only for RUNNING runs, strictly after the ack. Completed runs
+    // are fully persisted to the provider transcript and served over REST —
+    // replaying them (e.g. after a page reload where the client's lastSeq is
+    // 0) would duplicate messages the history fetch already returned.
+    if (isProcessing) {
+      for (const event of chatRunRegistry.replayEvents(sessionId, lastSeq)) {
+        sendJson(ws, event);
+      }
+    }
+  }
+}
+
+/**
+ * Handles `chat.permission-response`: forwards a tool-approval decision to the
+ * pending approval resolver (Claude is the only provider with interactive
+ * approvals today, but the message is intentionally provider-neutral).
+ */
+function handlePermissionResponse(data: AnyRecord, dependencies: ChatWebSocketDependencies): void {
+  if (typeof data.requestId !== 'string' || data.requestId.length === 0) {
+    return;
+  }
+
+  dependencies.runtime.resolveToolApproval(data.requestId, {
+    allow: Boolean(data.allow),
+    updatedInput: data.updatedInput,
+    message: typeof data.message === 'string' ? data.message : undefined,
+    rememberEntry: data.rememberEntry,
+  });
+}
+
+/**
+ * Handles authenticated chat websocket messages used by the main chat panel.
+ *
+ * Inbound protocol (client to server):
+ * - `chat.send`                { sessionId, content, options? }
+ * - `chat.abort`               { sessionId }
+ * - `chat.subscribe`           { sessions: [{ sessionId, lastSeq? }] }
+ * - `chat.permission-response` { requestId, allow, updatedInput?, message?, rememberEntry? }
+ *
+ * Outbound protocol (server to client): every frame is `kind`-based — either
+ * a provider `NormalizedMessage` (with `seq`) or a gateway event
+ * (`chat_subscribed`, `session_upserted`, `loading_progress`,
+ * `protocol_error`).
+ */
+/**
+ * Runs a turn for a session with no client attached.
+ *
+ * Used by scheduled messages, which fire from a timer: there is no socket to
+ * report errors to and no audience to stream to. The run is registered exactly
+ * like an interactive one, so anyone who opens the session while it is going
+ * subscribes and replays it from the start, and the session shows as busy
+ * everywhere in the meantime.
+ *
+ * Resolves when the provider run settles. Returns false when the session has
+ * gone away or is busy without `interruptActiveRun`, which the caller reports
+ * on the schedule.
+ */
+export async function runDetachedChatTurn(
+  input: {
+    sessionId: string;
+    userId: string | number | null;
+    content: string;
+    options?: AnyRecord;
+    /**
+     * Aborts a run already in progress instead of refusing to start. A
+     * scheduled message sets this: the user picked the time knowing it might
+     * land mid-run, so the timer outranks whatever is running.
+     */
+    interruptActiveRun?: boolean;
+  },
+  dependencies: ChatWebSocketDependencies,
+): Promise<{ started: boolean; error: string | null }> {
+  const session = sessionsDb.getSessionById(input.sessionId);
+  if (!session) {
+    return { started: false, error: 'The session no longer exists.' };
+  }
+
+  const provider = session.provider as LLMProvider;
+  if (!dependencies.runtime.hasRuntime(provider)) {
+    return { started: false, error: `Provider "${provider}" is not available.` };
+  }
+
+  const activeRun = chatRunRegistry.getRun(input.sessionId);
+  if (activeRun && activeRun.status === 'running') {
+    if (!input.interruptActiveRun) {
+      return { started: false, error: 'A run was already in progress for this session.' };
+    }
+    // Same shape as `chat.abort`: cancel the provider run and emit the
+    // terminal `complete` on its behalf, so every watching client sees the
+    // interrupted run end before this turn's stream begins. The interrupted
+    // run's own dispatch settles later through completeRunIfCurrent, which is
+    // scoped to that run and cannot touch the one started here.
+    const aborted = await dependencies.runtime.abort(activeRun.provider, input.sessionId);
+    chatRunRegistry.completeRun(input.sessionId, {
+      exitCode: aborted ? 0 : 1,
+      aborted: true,
+    });
+  }
+
+  return dispatchRun(
+    null,
+    input.userId,
+    input.sessionId,
+    session,
+    { sessionId: input.sessionId, content: input.content, options: input.options ?? {} },
+    dependencies,
+  );
+}
+
+export function handleChatConnection(
+  ws: WebSocket,
+  request: AuthenticatedWebSocketRequest,
+  dependencies: ChatWebSocketDependencies
+): void {
+  console.log('[INFO] Chat WebSocket connected');
+  connectedClients.add(ws);
+
+  const userId = readRequestUserId(request);
+
+  ws.on('message', async (rawMessage) => {
+    try {
+      const parsed = parseIncomingJsonObject(rawMessage);
+      if (!parsed) {
+        throw new Error('Invalid websocket payload');
+      }
+
+      const data = parsed as AnyRecord;
+      const messageType = typeof data.type === 'string' ? data.type : '';
+
+      switch (messageType) {
+        case 'chat.edit-send':
+          await handleChatEditSend(ws, userId, data, dependencies);
+          return;
+        case 'chat.send':
+          await handleChatSend(ws, userId, data, dependencies);
+          return;
+        case 'chat.abort':
+          await handleChatAbort(ws, data, dependencies);
+          return;
+        case 'chat.subscribe':
+          handleChatSubscribe(ws, data, dependencies);
+          return;
+        case 'chat.permission-response':
+          handlePermissionResponse(data, dependencies);
+          return;
+        default:
+          sendProtocolError(ws, 'UNKNOWN_MESSAGE_TYPE', `Unknown message type "${messageType}".`);
+          return;
+      }
+    } catch (error) {
+      const message = error instanceof Error ? error.message : String(error);
+      console.error('[ERROR] Chat WebSocket error:', message);
+      sendProtocolError(ws, 'INTERNAL_ERROR', message);
+    }
+  });
+
+  ws.on('close', () => {
+    console.log('[INFO] Chat client disconnected');
+    connectedClients.delete(ws);
+  });
+}