@vellumai/assistant 0.8.8 → 0.8.9-staging.2

package/src/plugins/defaults/overflow-reduce/middlewares/overflowReduce.ts

@@ -1,126 +0,0 @@
-import type { ContextWindowCompactOptions } from "../../../../context/window-manager.js";
-import {
-  createInitialReducerState,
-  reduceContextOverflow,
-  type ReducerState,
-} from "../../../../daemon/context-overflow-reducer.js";
-import type {
-  Middleware,
-  OverflowReduceArgs,
-  OverflowReduceResult,
-} from "../../../types.js";
-
-/**
- * Default middleware for the `overflowReduce` pipeline — implements the
- * historical tier-loop semantics (forced compaction, tool-result truncation,
- * media stubbing, injection downgrade) plus the post-step re-injection /
- * re-estimation dance.
- *
- * The middleware intentionally ignores `next`. Overflow reduction is a
- * *terminal* behavior: there is no downstream implementation to defer to when
- * a user-supplied middleware short-circuits. Later plugins may still wrap this
- * one (outer middleware can observe each reduction iteration via their own
- * `next` callback) but the default never delegates to a hypothetical base
- * handler — the inline loop was the base.
- */
-const defaultOverflowReduceMiddleware: Middleware<
-  OverflowReduceArgs,
-  OverflowReduceResult
-> = async function defaultOverflowReduceMiddleware(args, _next, _ctx) {
-  let messages = args.messages;
-  let runMessages = args.runMessages;
-  let injectionMode: "full" | "minimal" = "full";
-  let reducerState: ReducerState = createInitialReducerState();
-  let reducerCompacted = false;
-  let attempts = 0;
-
-  while (attempts < args.maxAttempts && !reducerState.exhausted) {
-    // Abort check at the top of every iteration. When the pipeline runner
-    // arms a timeout (or the caller aborts externally), `args.abortSignal`
-    // is linked to that trigger via `linkAbortSignal`, so this check lets
-    // us bail out BETWEEN iterations rather than letting another round of
-    // compaction / re-injection mutate `ctx.messages` after the turn has
-    // already failed. Individual `reduceContextOverflow` calls also honor
-    // the signal, but without this gate a fresh iteration could still
-    // start after the signal fires, since the previous one returned
-    // normally before the abort propagated.
-    args.abortSignal?.throwIfAborted();
-
-    attempts++;
-    args.emitActivityState();
-
-    const basisMessages = messages;
-    const step = await reduceContextOverflow(
-      basisMessages,
-      {
-        providerName: args.providerName,
-        systemPrompt: args.systemPrompt,
-        contextWindow: args.contextWindow,
-        targetTokens: args.preflightBudget,
-        toolTokenBudget: args.toolTokenBudget,
-      },
-      reducerState,
-      (msgs, signal, opts: ContextWindowCompactOptions) =>
-        args.compactFn(msgs, signal, opts),
-      args.abortSignal,
-    );
-
-    reducerState = step.state;
-    messages = step.messages;
-    injectionMode = step.state.injectionMode;
-
-    // Per-iteration compaction flag: whether THIS step just produced a
-    // fresh compaction. PKB / NOW re-injection is gated on this — see the
-    // reinjectForMode JSDoc for why the two signals differ.
-    const stepCompacted = step.compactionResult?.compacted === true;
-
-    // Let the orchestrator apply compaction side effects (circuit-breaker
-    // tracking, event emission, ctx mutation) before we re-inject.
-    if (step.compactionResult) {
-      await args.onCompactionResult(step.compactionResult, basisMessages);
-      if (stepCompacted) {
-        reducerCompacted = true;
-      }
-    }
-
-    // Second abort gate — if the side effects or the step itself took us
-    // past the deadline, don't rebuild runMessages or iterate again.
-    args.abortSignal?.throwIfAborted();
-
-    // Rebuild runMessages via the orchestrator-supplied helper (which
-    // re-runs `applyRuntimeInjections` with potentially downgraded mode
-    // and freshly re-hydrated PKB/NOW blocks after compaction). We pass
-    // the current reduced `messages` explicitly so the orchestrator never
-    // has to read from mutable shared state to rebuild runMessages — a
-    // tier that doesn't trigger compaction (tool-result truncation, media
-    // stubbing) won't update `ctx.messages` on its own.
-    //
-    // `stepCompacted` and `reducerCompacted` are both passed so the
-    // orchestrator can gate PKB / NOW re-injection per-iteration while
-    // keeping `slackChronologicalMessages` suppressed once any iteration
-    // has compacted.
-    runMessages = await args.reinjectForMode(
-      messages,
-      injectionMode,
-      stepCompacted,
-      reducerCompacted,
-    );
-
-    // Re-estimate with injections included — `step.estimatedTokens` was
-    // computed on bare history and doesn't account for tokens added by
-    // runtime injections.
-    const postInjectionTokens = args.estimatePostInjection(runMessages);
-    if (postInjectionTokens <= args.preflightBudget) break;
-  }
-
-  return {
-    messages,
-    runMessages,
-    injectionMode,
-    reducerState,
-    reducerCompacted,
-    attempts,
-  };
-};
-
-export default defaultOverflowReduceMiddleware;

package/src/plugins/defaults/overflow-reduce/package.json

@@ -1,15 +0,0 @@
-{
-  "name": "default-overflow-reduce",
-  "version": "1.0.0",
-  "description": "First-party default plugin wrapping the assistant's built-in overflow-reduce pipeline with a passthrough implementation.",
-  "private": true,
-  "license": "MIT",
-  "type": "module",
-  "main": "./register.ts",
-  "engines": {
-    "node": ">=20.12.0"
-  },
-  "peerDependencies": {
-    "@vellumai/plugin-api": "^0.8.0"
-  }
-}

package/src/plugins/defaults/overflow-reduce/register.ts

@@ -1,42 +0,0 @@
-/**
- * Default `overflowReduce` plugin — extracted verbatim from the inline
- * preflight reducer loop that previously lived in
- * `daemon/conversation-agent-loop.ts` (the `while (preflightAttempts < …)`
- * block around lines 1045–1156 before PR 23).
- *
- * The plugin owns the reducer tier-loop (forced compaction, tool-result
- * truncation, media stubbing, injection downgrade) and the post-step
- * re-injection / re-estimation dance. Orchestrator-specific coupling
- * (activity emission, circuit-breaker tracking, compaction-result
- * application, runtime injection reassembly) is threaded in through the
- * callbacks carried on {@link OverflowReduceArgs}; the plugin itself has no
- * access to the agent-loop context object.
- *
- * The forced-compaction tier runs through the orchestrator-supplied
- * `compactFn`, which routes into the `compaction` plugin pipeline so
- * registered compaction middleware observes reducer-initiated invocations
- * alongside the orchestrator-owned call sites. Non-compaction tiers
- * (tool-result truncation, media stubbing, injection downgrade) remain
- * in-process: they mutate message arrays directly without crossing a
- * pipeline boundary. The reducer itself runs under the `overflowReduce`
- * pipeline, so the full layering is `overflowReduce` → reducer tier loop
- * → (for the forced-compaction tier only) nested `compaction` pipeline.
- */
-
-import { type Plugin } from "../../types.js";
-import defaultOverflowReduceMiddleware from "./middlewares/overflowReduce.js";
-import pkg from "./package.json" with { type: "json" };
-
-/**
- * The default plugin registered at bootstrap. No `init`/`onShutdown` —
- * registering the middleware is the only behavior.
- */
-export const defaultOverflowReducePlugin: Plugin = {
-  manifest: {
-    name: pkg.name,
-    version: pkg.version,
-  },
-  middleware: {
-    overflowReduce: defaultOverflowReduceMiddleware,
-  },
-};

package/src/plugins/defaults/title-generate/register.ts

@@ -1,35 +0,0 @@
-/**
- * Default `title-generate` plugin.
- *
- * Contributes two hooks, both pure triggers that delegate the title work to
- * the service (`memory/conversation-title-service.ts`):
- *
- * - `user-prompt-submit` (`./hooks/user-prompt-submit.ts`) — first-pass title
- *   generation from the submitted prompt.
- * - `stop` (`./hooks/stop.ts`) — second-pass regeneration once the
- *   conversation reaches its third user turn, for a title that reflects the
- *   established topic.
- *
- * Persistence and the resulting `conversation_title_updated` / `sync_changed`
- * broadcast are owned by the title service.
- *
- * Registered via a side-effect import from
- * `daemon/external-plugins-bootstrap.ts` so it is present in the registry
- * by the time {@link bootstrapPlugins} runs.
- */
-
-import { type Plugin } from "../../types.js";
-import stop from "./hooks/stop.js";
-import userPromptSubmit from "./hooks/user-prompt-submit.js";
-import pkg from "./package.json" with { type: "json" };
-
-export const defaultTitleGeneratePlugin: Plugin = {
-  manifest: {
-    name: pkg.name,
-    version: pkg.version,
-  },
-  hooks: {
-    "user-prompt-submit": userPromptSubmit,
-    stop,
-  },
-};

package/src/plugins/defaults/tool-error/register.ts

@@ -1,23 +0,0 @@
-/**
- * Default `tool-error` plugin.
- *
- * Contributes a `post-tool-use` hook that coaches the model to retry or report
- * a failed tool call, bounded per tool so an unrecoverable error doesn't churn.
- * The coaching is surfaced via `additionalContext` (a separate provider-only
- * block), leaving the tool result's own content untouched. The decision logic
- * lives in `./hooks/post-tool-use.ts`.
- */
-
-import { type Plugin } from "../../types.js";
-import postToolUse from "./hooks/post-tool-use.js";
-import pkg from "./package.json" with { type: "json" };
-
-export const defaultToolErrorPlugin: Plugin = {
-  manifest: {
-    name: pkg.name,
-    version: pkg.version,
-  },
-  hooks: {
-    "post-tool-use": postToolUse,
-  },
-};

package/src/plugins/defaults/tool-result-truncate/register.ts

@@ -1,24 +0,0 @@
-/**
- * Default `tool-result-truncate` plugin.
- *
- * Contributes a `post-tool-use` hook that tail-drops an oversized tool result
- * down to a character budget derived from the model's context window before
- * the result is sent to the provider. The truncation implementation lives in
- * `./terminal.ts`; the hook in `./hooks/post-tool-use.ts` wires it into the
- * lifecycle. Defaults register before user plugins, so this runs at the front
- * of the hook chain.
- */
-
-import { type Plugin } from "../../types.js";
-import postToolUse from "./hooks/post-tool-use.js";
-import pkg from "./package.json" with { type: "json" };
-
-export const defaultToolResultTruncatePlugin: Plugin = {
-  manifest: {
-    name: pkg.name,
-    version: pkg.version,
-  },
-  hooks: {
-    "post-tool-use": postToolUse,
-  },
-};

package/src/plugins/external-api.ts

@@ -1,104 +0,0 @@
-/**
- * External plugin runtime — `globalThis.__vellumPluginRuntime` bridge.
- *
- * Workspace-local plugins (`<workspaceDir>/plugins/*`) get dynamic-imported by
- * {@link loadUserPlugins}. They need to call `registerPlugin()` from their
- * module body and they typically also want to read secrets, subscribe to
- * runtime events, etc.
- *
- * Importing those symbols by absolute path
- * (`/abs/path/to/assistant/src/plugins/registry.js`) works when the daemon is
- * running from source — both the daemon and the dynamic-imported plugin
- * resolve the same on-disk file and share module identity. It DOES NOT work
- * when the daemon is a `bun --compile` binary: the daemon's modules are baked
- * into the executable, and any absolute-path import re-loads a fresh disk
- * copy. `registerPlugin()` then writes into a disjoint registry instance and
- * the daemon never sees the plugin.
- *
- * The fix is to expose a single, stable handle on `globalThis` that plugins
- * read at module-load time. The daemon's bundled modules attach themselves
- * here once at startup; plugins consume the same instance regardless of how
- * the daemon was built.
- *
- * Plugins use the bridge like this:
- *
- *   const runtime = (globalThis as { __vellumPluginRuntime?: ... })
- *     .__vellumPluginRuntime;
- *   if (!runtime || runtime.version !== 1) throw new Error("...");
- *   const { registerPlugin, assistantEventHub, getSecureKeyAsync } = runtime;
- *
- * Type-only imports (`import type { Plugin } from "..."`) remain free to use
- * absolute paths or workspace-local copies — the TypeScript compiler erases
- * them and they have no module-identity effect at runtime.
- *
- * See `assistant/docs/plugins.md` for the full authoring contract and
- * `assistant/examples/plugins/echo/register.ts` for a worked example.
- */
-
-import { assistantEventHub } from "../runtime/assistant-event-hub.js";
-import { getSecureKeyAsync } from "../security/secure-keys.js";
-import { registerPlugin } from "./registry.js";
-
-/**
- * The handle plugins read from `globalThis.__vellumPluginRuntime`.
- *
- * `version` is the contract version. Plugins should assert against the
- * version they were authored against and refuse to register if they see a
- * different one — bumping it is a breaking change to the plugin surface.
- *
- * The field set is intentionally small: the most-commonly-needed symbols
- * across both static (module-load-time) and dynamic (init-time) plugin
- * lifecycle. Plugins that need richer runtime state can still receive it
- * through {@link PluginInitContext} during `init()`.
- */
-export interface VellumPluginRuntime {
-  readonly version: 1;
-  readonly registerPlugin: typeof registerPlugin;
-  readonly assistantEventHub: typeof assistantEventHub;
-  readonly getSecureKeyAsync: typeof getSecureKeyAsync;
-}
-
-/** Stable globalThis key. Don't rename — plugins reference it by string. */
-const RUNTIME_GLOBAL_KEY = "__vellumPluginRuntime" as const;
-
-interface GlobalWithRuntime {
-  [RUNTIME_GLOBAL_KEY]?: VellumPluginRuntime;
-}
-
-/**
- * Install the plugin runtime bridge on `globalThis`. Idempotent — repeat
- * calls are no-ops, so it's safe to invoke from tests that also touch the
- * lifecycle path.
- *
- * Must be called BEFORE {@link loadUserPlugins} runs, otherwise plugins that
- * touch `globalThis.__vellumPluginRuntime` in their module body will throw.
- */
-export function installPluginRuntime(): void {
-  const g = globalThis as GlobalWithRuntime;
-  if (g[RUNTIME_GLOBAL_KEY]) return;
-  g[RUNTIME_GLOBAL_KEY] = {
-    version: 1,
-    registerPlugin,
-    assistantEventHub,
-    getSecureKeyAsync,
-  };
-}
-
-/**
- * Read the installed runtime. Returns `undefined` if {@link installPluginRuntime}
- * hasn't been called yet — plugins should treat this as a fatal error.
- *
- * Exposed mainly for tests and for the rare in-process consumer that wants
- * to read the bridge from the same module graph that installed it.
- */
-export function getPluginRuntime(): VellumPluginRuntime | undefined {
-  return (globalThis as GlobalWithRuntime)[RUNTIME_GLOBAL_KEY];
-}
-
-/**
- * Tear down the runtime handle. Test-only — production code never uninstalls
- * because the runtime lifetime is the daemon lifetime.
- */
-export function uninstallPluginRuntimeForTests(): void {
-  delete (globalThis as GlobalWithRuntime)[RUNTIME_GLOBAL_KEY];
-}

package/src/proactive-artifact/aux-message-injector.ts

@@ -1,97 +0,0 @@
-/**
- * Auxiliary message injection for proactive artifacts.
- *
- * Injects an assistant message into a conversation without going through
- * the normal agent loop. Defers injection while the conversation is
- * actively processing to preserve chronological message ordering.
- */
-
-import { createAssistantMessage } from "../agent/message-types.js";
-import { findConversation } from "../daemon/conversation-store.js";
-import {
-  conversationMessagesSyncTag,
-  SYNC_TAGS,
-} from "../daemon/message-types/sync.js";
-import { addMessage } from "../memory/conversation-crud.js";
-import type { BroadcastFn } from "../notifications/adapters/macos.js";
-import { getLogger } from "../util/logger.js";
-
-const log = getLogger("aux-message-injector");
-const IDLE_POLL_MS = 200;
-const IDLE_TIMEOUT_MS = 60_000;
-
-async function waitForIdle(conversationId: string): Promise<boolean> {
-  const start = Date.now();
-  while (Date.now() - start < IDLE_TIMEOUT_MS) {
-    const conv = findConversation(conversationId);
-    if (!conv || !conv.isProcessing()) return true;
-    await new Promise((resolve) => setTimeout(resolve, IDLE_POLL_MS));
-  }
-  return false;
-}
-
-export async function injectAuxAssistantMessage(params: {
-  conversationId: string;
-  text: string;
-  broadcastMessage: BroadcastFn;
-}): Promise<void> {
-  const conv = findConversation(params.conversationId);
-  if (conv?.isProcessing()) {
-    const reachedIdle = await waitForIdle(params.conversationId);
-    if (!reachedIdle) {
-      log.warn(
-        { conversationId: params.conversationId },
-        "Timed out waiting for conversation idle; injecting anyway",
-      );
-    }
-  }
-
-  const msg = await addMessage(
-    params.conversationId,
-    "assistant",
-    JSON.stringify([{ type: "text", text: params.text }]),
-    { skipIndexing: true },
-  );
-
-  const current = findConversation(params.conversationId);
-  if (current && !current.isProcessing()) {
-    current.getMessages().push(createAssistantMessage(params.text));
-
-    params.broadcastMessage({
-      type: "assistant_text_delta",
-      text: params.text,
-      conversationId: params.conversationId,
-    });
-    params.broadcastMessage({
-      type: "message_complete",
-      conversationId: params.conversationId,
-      messageId: msg.id,
-      source: "aux",
-    });
-  }
-
-  // The injected aux message bumps `lastMessageAt`, which can move the
-  // row in the paginated sidebar window — that's a shape change. macOS
-  // refreshes its full list on `conversation_list_invalidated`; web
-  // refreshes via the paired `sync_changed` `conversationsList` tag.
-  //
-  // TODO(electron-cutover): drop the `conversation_list_invalidated`
-  // emission once macOS migrates to the Electron client and consumes
-  // `sync_changed` directly. See `runtime/sync/resource-sync-events.ts`
-  // for the symmetric helper used by route handlers.
-  params.broadcastMessage(
-    {
-      type: "conversation_list_invalidated",
-      reason: "reordered",
-    },
-    undefined,
-    { targetInterfaceId: "macos" },
-  );
-  params.broadcastMessage({
-    type: "sync_changed",
-    tags: [
-      SYNC_TAGS.conversationsList,
-      conversationMessagesSyncTag(params.conversationId),
-    ],
-  });
-}