@vellumai/assistant 0.6.3 → 0.6.5

package/src/messaging/providers/slack/render-transcript.ts

@@ -0,0 +1,443 @@
+/**
+ * Pure chronological thread-tag rendering for Slack transcripts.
+ *
+ * Given a list of stored messages (post-upgrade rows with structured metadata
+ * AND legacy pre-upgrade rows with `metadata === null`), produces a flat
+ * `{role, content}[]` chronologically ordered with compact thread tags so
+ * the model can reason across sibling threads in one channel.
+ *
+ * The function is pure: no I/O, no implicit clock reads. Time is taken from
+ * `opts.now` only when needed for relative formatting. Sort and tag rendering
+ * are deterministic.
+ *
+ * Consumers wire this into inbound history rendering and the compaction
+ * boundary.
+ */
+
+import { createHash } from "node:crypto";
+
+import type { ContentBlock, Message } from "../../../providers/types.js";
+import type { SlackMessageMetadata } from "./message-metadata.js";
+
+export interface RenderableSlackMessage {
+  role: "user" | "assistant";
+  content: string;
+  /** `null` indicates a legacy pre-upgrade row stored without Slack metadata. */
+  metadata: SlackMessageMetadata | null;
+  /**
+   * Sender display name to prepend to the tag line (e.g. `"@alice"`), or
+   * `null` to omit the label entirely. Callers should pass `null` when the
+   * label would be redundant with the `role` slot — i.e. assistant rows and
+   * user rows with no real Slack displayName. Reaction rows always need a
+   * subject, so they receive a role-derived fallback if `null` is passed.
+   */
+  senderLabel: string | null;
+  /** Fallback sort key for legacy rows; ignored when metadata.channelTs is set. */
+  createdAt: number;
+  /**
+   * Full structured content blocks parsed from the persisted row, when
+   * available. Optional so existing fixtures and callers that only need the
+   * flattened `content` string continue to compile. `renderSlackTranscript`
+   * consumes this field to preserve replayable Anthropic blocks
+   * (`tool_use`, `tool_result`, `thinking`, `redacted_thinking`, `image`,
+   * `file`) in their original order, emitting the tag line inline at the
+   * position of the first `text` block. Non-replayable blocks
+   * (`ui_surface`, `server_tool_use`, `web_search_tool_result`, unknown
+   * types) are stripped; when stripping empties the row entirely, a
+   * fallback tag-line text block is emitted so chronology is preserved.
+   */
+  readonly contentBlocks?: readonly ContentBlock[];
+}
+
+export interface RenderOptions {
+  /** Reserved for future relative-time rendering; currently unused. */
+  now?: Date;
+  /** Cap rendered reactions per parent message; default 5. */
+  maxReactionsPerMessage?: number;
+}
+
+const DEFAULT_MAX_REACTIONS = 5;
+
+/**
+ * Replayable Anthropic content-block types that we preserve verbatim from a
+ * persisted row when rendering the Slack chronological transcript.
+ *
+ * `text` is intentionally omitted — text content is subsumed into the tag
+ * line (e.g. `[11/14/23 14:25 @alice]: ...`) so callers reading the
+ * rendered output see one human-readable line per row rather than a raw
+ * text block stripped of thread context.
+ *
+ * Non-replayable types (`ui_surface`, `server_tool_use`,
+ * `web_search_tool_result`, unknown types) are dropped: `ui_surface` blocks
+ * are the assistant's ephemeral local UI scaffolding and not meaningful to
+ * replay; `server_tool_use` / `web_search_tool_result` carry provider-
+ * specific `encrypted_content` that becomes stale and is rejected by the
+ * provider on re-send.
+ */
+const REPLAYABLE_BLOCK_TYPES = new Set<ContentBlock["type"]>([
+  "tool_use",
+  "tool_result",
+  "thinking",
+  "redacted_thinking",
+  "image",
+  "file",
+]);
+
+/**
+ * Compute a short, stable, deterministic alias for a Slack message ts.
+ *
+ * Used as the "parent label" inside thread-reply tags so the model can
+ * cross-reference children with their parent without leaking raw ts values.
+ * First 6 hex chars of sha256(channelTs) prefixed with `M`.
+ */
+export function parentAlias(channelTs: string): string {
+  const hash = createHash("sha256").update(channelTs).digest("hex");
+  return `M${hash.slice(0, 6)}`;
+}
+
+/**
+ * Format a Slack ts (`"1700000000.000100"`) as `MM/DD/YY HH:MM` (UTC).
+ *
+ * Slack ts is `<unix-seconds>.<microseconds>`; we treat it as a unix epoch
+ * second value for display purposes. Pure — derives only from the ts string.
+ */
+function formatSlackTs(channelTs: string): string {
+  const seconds = Number.parseFloat(channelTs);
+  if (!Number.isFinite(seconds)) return "??/??/?? ??:??";
+  return formatEpochMs(seconds * 1000);
+}
+
+/**
+ * Format an epoch millisecond timestamp as `MM/DD/YY HH:MM` (UTC).
+ */
+function formatEpochMs(ms: number): string {
+  if (!Number.isFinite(ms)) return "??/??/?? ??:??";
+  const d = new Date(ms);
+  const mo = String(d.getUTCMonth() + 1).padStart(2, "0");
+  const da = String(d.getUTCDate()).padStart(2, "0");
+  const yy = String(d.getUTCFullYear() % 100).padStart(2, "0");
+  const hh = String(d.getUTCHours()).padStart(2, "0");
+  const mm = String(d.getUTCMinutes()).padStart(2, "0");
+  return `${mo}/${da}/${yy} ${hh}:${mm}`;
+}
+
+/**
+ * Sort key for chronological ordering.
+ *
+ * Post-upgrade rows are ordered by their Slack `channelTs`; legacy rows
+ * (metadata === null) fall back to `createdAt`. Both produce a numeric
+ * seconds-since-epoch value so they intermix correctly.
+ */
+function sortKey(msg: RenderableSlackMessage): number {
+  if (msg.metadata) {
+    const n = Number.parseFloat(msg.metadata.channelTs);
+    if (Number.isFinite(n)) return n;
+  }
+  // createdAt is epoch ms; convert to seconds for like-with-like comparison.
+  return msg.createdAt / 1000;
+}
+
+/**
+ * Render a single non-reaction message (post-upgrade or legacy) as one
+ * tagged line.
+ */
+function renderMessage(msg: RenderableSlackMessage): string {
+  const meta = msg.metadata;
+  const senderPart = msg.senderLabel ? ` ${msg.senderLabel}` : "";
+  if (!meta) {
+    // Legacy pre-upgrade row: flat render, no thread tag.
+    const time = formatEpochMs(msg.createdAt);
+    return `[${time}${senderPart}]: ${msg.content}`;
+  }
+
+  const time = formatSlackTs(meta.channelTs);
+
+  if (meta.deletedAt !== undefined) {
+    const dtime = formatEpochMs(meta.deletedAt);
+    return `[${time}${senderPart} — deleted ${dtime}]`;
+  }
+
+  let head = `[${time}${senderPart}`;
+  if (meta.threadTs && meta.threadTs !== meta.channelTs) {
+    head += ` → ${parentAlias(meta.threadTs)}`;
+  }
+  if (meta.editedAt !== undefined) {
+    head += `, edited ${formatEpochMs(meta.editedAt)}`;
+  }
+  head += `]: ${msg.content}`;
+  return head;
+}
+
+/**
+ * Render a single reaction event as one tagged line.
+ *
+ * `[11/14/23 14:28 @bob reacted 👍 to M1a2b3c]` or
+ * `[11/14/23 14:28 @bob removed 👍 from M1a2b3c]`.
+ */
+function renderReaction(msg: RenderableSlackMessage): string | null {
+  const meta = msg.metadata;
+  if (!meta || meta.eventKind !== "reaction" || !meta.reaction) return null;
+  const time = formatSlackTs(meta.channelTs);
+  const actor =
+    msg.senderLabel ?? (msg.role === "assistant" ? "@assistant" : "@user");
+  const verb = meta.reaction.op === "added" ? "reacted" : "removed";
+  const prep = meta.reaction.op === "added" ? "to" : "from";
+  const target = parentAlias(meta.reaction.targetChannelTs);
+  return `[${time} ${actor} ${verb} ${meta.reaction.emoji} ${prep} ${target}]`;
+}
+
+/**
+ * Build the content blocks for a single non-reaction message.
+ *
+ * Emits the tag line (`[MM/DD/YY HH:MM @sender ...]: body`) inline at the position of
+ * the first `text` block in `contentBlocks`, and preserves any replayable
+ * blocks (`tool_use`, `tool_result`, `thinking`, `redacted_thinking`,
+ * `image`, `file`) in their original order. Non-replayable blocks
+ * (`ui_surface`, `server_tool_use`, `web_search_tool_result`, unknown types)
+ * are dropped.
+ *
+ * Special cases:
+ * - **Deleted rows**: always emit a single tag-line block; replayable blocks
+ *   (if any) are discarded because the delete is a logical erasure.
+ * - **Legacy rows** (no structured `contentBlocks`, or empty array): fall
+ *   back to a single tag-line block to preserve pre-plumbing behaviour.
+ * - **Pure tool-only rows** (`contentBlocks` present but no `text` block):
+ *   emit only the replayable blocks — no tag line. Anthropic accepts role-
+ *   correct messages with only tool blocks.
+ * - **All-non-replayable rows** (`contentBlocks` present but every block is
+ *   filtered out — e.g. a row whose only blocks are `server_tool_use` or
+ *   `ui_surface`): emit a single fallback tag-line text block annotated
+ *   with the stripped block types/names. Dropping the row entirely would
+ *   silently alter chronology and can orphan adjacent tool_result context
+ *   in later repair/conversion steps.
+ */
+function buildMessageContentBlocks(
+  msg: RenderableSlackMessage,
+  tagLine: string,
+): ContentBlock[] {
+  const isDeleted = msg.metadata?.deletedAt !== undefined;
+  const blocks = msg.contentBlocks;
+
+  // Deleted rows: single tag line, drop any replayable content.
+  if (isDeleted) {
+    return [{ type: "text", text: tagLine }];
+  }
+
+  // Legacy / unplumbed rows: fall back to single tag line.
+  if (!blocks || blocks.length === 0) {
+    return [{ type: "text", text: tagLine }];
+  }
+
+  const out: ContentBlock[] = [];
+  let tagEmitted = false;
+  const strippedLabels: string[] = [];
+  for (const block of blocks) {
+    if (block.type === "text") {
+      if (!tagEmitted) {
+        out.push({ type: "text", text: tagLine });
+        tagEmitted = true;
+      }
+      // Subsequent text blocks are already subsumed into the tag line.
+      continue;
+    }
+    if (REPLAYABLE_BLOCK_TYPES.has(block.type)) {
+      out.push(block);
+      continue;
+    }
+    // Non-replayable (ui_surface, server_tool_use, web_search_tool_result,
+    // unknown) — drop, but remember what we saw in case we need a fallback.
+    if (block.type === "server_tool_use") {
+      strippedLabels.push(`server_tool_use(${block.name})`);
+    } else {
+      strippedLabels.push(block.type);
+    }
+  }
+
+  // Non-empty source fully filtered to nothing: emit a fallback tag line so
+  // the turn still appears in chronology. Annotate with the stripped block
+  // types/names so the model has a hint about what was there.
+  if (out.length === 0) {
+    const suffix =
+      strippedLabels.length > 0
+        ? ` [stripped non-replayable: ${strippedLabels.join(", ")}]`
+        : "";
+    return [{ type: "text", text: `${tagLine}${suffix}` }];
+  }
+  return out;
+}
+
+/**
+ * Render a chronological transcript with compact thread tags.
+ *
+ * Sort is stable: messages with identical sort keys preserve their input
+ * order so callers controlling input ordering can break ties deterministically.
+ *
+ * Reactions are rendered as their own lines (`[time @actor reacted ... to Mxxx]`),
+ * but capped per-target at `opts.maxReactionsPerMessage` (default 5). Excess
+ * reactions on the same target are collapsed into a single trailer line
+ * (`[…and N more reactions to Mxxx]`, singular `reaction` when N===1), emitted
+ * at the point the overflow window closes — i.e. immediately before the next
+ * event that is not an overflowing reaction for the same target — so trailers
+ * stay in chronological position rather than clustered at the end.
+ */
+export function renderSlackTranscript(
+  messages: RenderableSlackMessage[],
+  opts?: RenderOptions,
+): Message[] {
+  if (messages.length === 0) return [];
+
+  const maxReactions = Math.max(
+    1,
+    Math.floor(opts?.maxReactionsPerMessage ?? DEFAULT_MAX_REACTIONS),
+  );
+
+  // Stable sort: decorate-sort-undecorate so equal keys preserve input order.
+  const indexed = messages.map((m, i) => ({ m, i, k: sortKey(m) }));
+  indexed.sort((a, b) => {
+    if (a.k !== b.k) return a.k - b.k;
+    return a.i - b.i;
+  });
+  const sorted = indexed.map((x) => x.m);
+
+  // Per-target reaction counters used to enforce the cap.
+  const reactionCount = new Map<string, number>();
+  // Open per-target overflow windows. Excess reactions accumulate here; the
+  // window is closed (trailer emitted) as soon as the chronological walk
+  // reaches an event that is not an overflowing reaction for that target.
+  const overflowAccumulator = new Map<
+    string,
+    { excess: number; role: "user" | "assistant" }
+  >();
+
+  const trailerMessage = (
+    target: string,
+    acc: { excess: number; role: "user" | "assistant" },
+  ): Message => ({
+    role: acc.role,
+    content: [
+      {
+        type: "text" as const,
+        text: `[…and ${acc.excess} more ${acc.excess === 1 ? "reaction" : "reactions"} to ${parentAlias(target)}]`,
+      },
+    ],
+  });
+
+  const flushOverflowExcept = (out: Message[], keepTarget: string | null) => {
+    for (const target of Array.from(overflowAccumulator.keys())) {
+      if (target === keepTarget) continue;
+      const acc = overflowAccumulator.get(target)!;
+      out.push(trailerMessage(target, acc));
+      overflowAccumulator.delete(target);
+    }
+  };
+
+  const out: Message[] = [];
+  for (const m of sorted) {
+    const meta = m.metadata;
+    if (meta?.eventKind === "reaction" && meta.reaction) {
+      const target = meta.reaction.targetChannelTs;
+      const seen = reactionCount.get(target) ?? 0;
+      if (seen < maxReactions) {
+        // Reaction fits under the cap for `target`. Any open overflow windows
+        // for other targets are now behind us chronologically — close them.
+        flushOverflowExcept(out, null);
+        reactionCount.set(target, seen + 1);
+        const line = renderReaction(m);
+        if (line !== null) {
+          out.push({
+            role: m.role,
+            content: [{ type: "text" as const, text: line }],
+          });
+        }
+      } else {
+        // Reaction overflows for `target`. Close any other open windows, then
+        // extend this target's open window.
+        flushOverflowExcept(out, target);
+        const acc = overflowAccumulator.get(target) ?? {
+          excess: 0,
+          role: m.role,
+        };
+        acc.excess += 1;
+        overflowAccumulator.set(target, acc);
+      }
+      continue;
+    }
+    // Non-reaction event: every open overflow window closes here.
+    flushOverflowExcept(out, null);
+    const tagLine = renderMessage(m);
+    const blocks = buildMessageContentBlocks(m, tagLine);
+    if (blocks.length === 0) continue;
+    out.push({
+      role: m.role,
+      content: blocks,
+    });
+  }
+
+  // End of the walk: flush any still-open overflow windows.
+  flushOverflowExcept(out, null);
+
+  return filterOrphanToolPairs(out);
+}
+
+/**
+ * Final safety pass that drops unpaired `tool_use` ↔ `tool_result` blocks.
+ *
+ * Anthropic's API requires every `tool_use` in an assistant turn to be
+ * matched by a `tool_result` in the following user turn (and vice versa).
+ * In normal operation `renderSlackTranscript` emits fully-paired turns
+ * because the persisted transcript reflects completed tool exchanges, but
+ * edge cases (mid-turn compaction, partial failures, a race between
+ * producer persistence and consumer persistence) can leave an orphan in
+ * the rendered output. Sending an orphan to the provider hard-fails the
+ * entire request, so we defensively prune any unpaired block here.
+ *
+ * Server-side block types (`server_tool_use`, `web_search_tool_result`) are
+ * stripped earlier by `buildMessageContentBlocks` — they carry stale
+ * provider-specific `encrypted_content` and are never replayed — so they
+ * cannot reach this filter.
+ *
+ * A message that becomes empty after filtering (e.g. an assistant row that
+ * carried only an orphaned `tool_use`) is dropped entirely rather than
+ * emitted as `{role, content: []}` — empty-content messages are also
+ * rejected by the provider.
+ */
+function filterOrphanToolPairs(messages: Message[]): Message[] {
+  const produced = new Set<string>();
+  const consumed = new Set<string>();
+  for (const msg of messages) {
+    for (const b of msg.content) {
+      if (b.type === "tool_use") produced.add(b.id);
+      else if (b.type === "tool_result" || b.type === "web_search_tool_result")
+        consumed.add(b.tool_use_id);
+    }
+  }
+  const out: Message[] = [];
+  for (const msg of messages) {
+    const kept: ContentBlock[] = [];
+    for (const b of msg.content) {
+      if (b.type === "tool_use" && !consumed.has(b.id)) continue;
+      if (
+        (b.type === "tool_result" || b.type === "web_search_tool_result") &&
+        !produced.has(b.tool_use_id)
+      )
+        continue;
+      kept.push(b);
+    }
+    if (kept.length > 0) out.push({ role: msg.role, content: kept });
+  }
+  return out;
+}
+
+/**
+ * Extract the first text-block text from each rendered message.
+ *
+ * Used by callers (e.g. the active-thread focus block) that need a flat
+ * `string[]` of rendered tag lines rather than the structured `Message[]`
+ * output. Messages with no text block yield an empty string.
+ */
+export function extractTagLineTexts(rendered: Message[]): string[] {
+  return rendered.map((msg) => {
+    const first = msg.content.find((b) => b.type === "text");
+    return first && first.type === "text" ? first.text : "";
+  });
+}

package/src/messaging/providers/slack/types.ts

@@ -119,6 +119,10 @@ export interface SlackConversationsOpenResponse extends SlackApiResponse {
 
 export type SlackReactionAddResponse = SlackApiResponse;
 
+export type SlackConversationJoinResponse = SlackApiResponse & {
+  channel?: SlackConversation;
+};
+
 export type SlackConversationLeaveResponse = SlackApiResponse;
 
 export interface SlackChatDeleteResponse extends SlackApiResponse {

package/src/messaging/style-analyzer.ts

@@ -127,7 +127,7 @@ export async function extractStylePatterns(
     .map((e, i) => `--- Message ${i + 1} ---\n${e}`)
     .join("\n\n");
 
-  const provider = await getConfiguredProvider();
+  const provider = await getConfiguredProvider("styleAnalyzer");
   if (!provider) {
     return { stylePatterns: [], contactObservations: [] };
   }
@@ -147,7 +147,10 @@ export async function extractStylePatterns(
     promptMessages,
     [storeStyleAnalysisTool],
     STYLE_EXTRACTION_SYSTEM_PROMPT,
-    { signal: AbortSignal.timeout(30_000) },
+    {
+      signal: AbortSignal.timeout(30_000),
+      config: { callSite: "styleAnalyzer" },
+    },
   );
 
   const toolBlock = response.content.find((b) => b.type === "tool_use");

package/src/notifications/README.md

@@ -31,7 +31,7 @@ The candidate set is serialized into a compact `<conversation-candidates>` block
 
 ### 3. Decision
 
-The decision engine (`decision-engine.ts`) sends the signal to an LLM (configured via `notifications.decisionModelIntent`) along with available channels, the user's preference summary, and the conversation candidate set. The LLM responds with a structured decision: whether to notify, which channels, rendered copy per channel, a deduplication key, and **per-channel conversation actions**.
+The decision engine (`decision-engine.ts`) sends the signal to an LLM (configured via `llm.callSites.notificationDecision`) along with available channels, the user's preference summary, and the conversation candidate set. The LLM responds with a structured decision: whether to notify, which channels, rendered copy per channel, a deduplication key, and **per-channel conversation actions**.
 
 **Conversation actions:** For each selected channel, the LLM decides:
 
@@ -538,10 +538,14 @@ Preferences are sanitized against prompt injection (angle brackets replaced with
 
 ## Configuration
 
-All settings live under the `notifications` key in `config.json`:
+The decision engine and preference extractor pick their per-call LLM config
+from the unified `llm` block. Override defaults by setting either of:
 
-| Key                                 | Type   | Default               | Description                                                              |
-| ----------------------------------- | ------ | --------------------- | ------------------------------------------------------------------------ |
-| `notifications.decisionModelIntent` | string | `"latency-optimized"` | Model intent used for both the decision engine and preference extraction |
+| Key                                  | Type    | Default        | Description                                                                 |
+| ------------------------------------ | ------- | -------------- | --------------------------------------------------------------------------- |
+| `llm.callSites.notificationDecision` | object  | _(unset)_      | Provider/model/effort/etc. override for the decision engine call site       |
+| `llm.callSites.preferenceExtraction` | object  | _(unset)_      | Provider/model/effort/etc. override for the preference extractor call site  |
+
+When a call site override is unset, the resolver falls back to `llm.default`.
 
 The notification pipeline is always active -- signals are processed and dispatched as soon as the daemon is running. The audit trail (events, decisions, deliveries) is written for every signal.

package/src/notifications/decision-engine.ts

@@ -12,7 +12,6 @@
 import { v4 as uuid } from "uuid";
 
 import { getDeliverableChannels } from "../channels/config.js";
-import { getConfig } from "../config/loader.js";
 import { listGuardianChannels } from "../contacts/contact-store.js";
 import { resolveGuardianPersona } from "../prompts/persona-resolver.js";
 import { buildCoreIdentityContext } from "../prompts/system-prompt.js";
@@ -22,7 +21,7 @@ import {
   getConfiguredProvider,
   userMessage,
 } from "../providers/provider-send-message.js";
-import type { ModelIntent, Provider } from "../providers/types.js";
+import type { Provider } from "../providers/types.js";
 import { getLogger } from "../util/logger.js";
 import { truncate } from "../util/truncate.js";
 import {
@@ -61,9 +60,9 @@ const PROMPT_VERSION = "v4";
 /**
  * Maximum character budget for identity context injected into the notification
  * decision prompt. We truncate to prevent oversized prompts when SOUL.md /
- * IDENTITY.md / USER.md are large — exceeding the provider context window
- * would cause the LLM call to fail and silently degrade to deterministic
- * fallback for all notifications.
+ * IDENTITY.md / users/<slug>.md are large — exceeding the provider context
+ * window would cause the LLM call to fail and silently degrade to
+ * deterministic fallback for all notifications.
  */
 const MAX_IDENTITY_CONTEXT_CHARS = 2000;
 
@@ -717,9 +716,6 @@ export async function evaluateSignal(
   availableChannels: NotificationChannel[],
   preferenceContext?: string,
 ): Promise<NotificationDecision> {
-  const config = getConfig();
-  const decisionModelIntent = config.notifications.decisionModelIntent;
-
   // When no explicit preference context is provided, load the user's
   // stored notification preferences from the memory-backed store.
   // Wrapped in try/catch so a DB failure doesn't break the decision path.
@@ -750,7 +746,7 @@ export async function evaluateSignal(
     );
   }
 
-  const provider = await getConfiguredProvider();
+  const provider = await getConfiguredProvider("notificationDecision");
   if (!provider) {
     log.warn(
       "Configured provider unavailable for notification decision, using fallback",
@@ -774,7 +770,6 @@ export async function evaluateSignal(
       signal,
       availableChannels,
       resolvedPreferenceContext,
-      decisionModelIntent,
       candidateSet,
     );
   } catch (err) {
@@ -805,7 +800,6 @@ async function classifyWithLLM(
   signal: NotificationSignal,
   availableChannels: NotificationChannel[],
   preferenceContext: string | undefined,
-  modelIntent: ModelIntent,
   candidateSet?: ConversationCandidateSet,
 ): Promise<NotificationDecision> {
   const { signal: abortSignal, cleanup } = createTimeout(DECISION_TIMEOUT_MS);
@@ -858,7 +852,7 @@ async function classifyWithLLM(
       systemPrompt,
       {
         config: {
-          modelIntent,
+          callSite: "notificationDecision",
           max_tokens: 2048,
           tool_choice: {
             type: "tool" as const,

package/src/notifications/preference-extractor.ts

@@ -11,7 +11,6 @@
  * - "Weeknights after 10pm: only critical notifications"
  */
 
-import { getConfig } from "../config/loader.js";
 import {
   createTimeout,
   extractToolUse,
@@ -140,15 +139,12 @@ const EXTRACTION_TOOL = {
 export async function extractPreferences(
   message: string,
 ): Promise<ExtractionResult> {
-  const provider = await getConfiguredProvider();
+  const provider = await getConfiguredProvider("preferenceExtraction");
   if (!provider) {
     log.debug("No provider available for preference extraction");
     return { detected: false, preferences: [] };
   }
 
-  const config = getConfig();
-  const modelIntent = config.notifications.decisionModelIntent;
-
   const { signal, cleanup } = createTimeout(EXTRACTION_TIMEOUT_MS);
 
   try {
@@ -158,7 +154,7 @@ export async function extractPreferences(
       SYSTEM_PROMPT,
       {
         config: {
-          modelIntent,
+          callSite: "preferenceExtraction",
           max_tokens: 1024,
           tool_choice: {
             type: "tool" as const,

package/src/notifications/signal.ts

@@ -97,6 +97,11 @@ export const NOTIFICATION_SOURCE_EVENT_NAMES = [
     id: "voice.response_ready",
     description: "Voice response ready for playback",
   },
+  {
+    id: "credential.health_alert",
+    description:
+      "OAuth credential health issue detected (expired, revoked, missing scopes)",
+  },
 ] as const;
 
 export type NotificationSourceEventName =

package/src/oauth/__tests__/identity-verifier.test.ts

@@ -34,6 +34,7 @@ function makeProviderRow(
     tokenExchangeUrl: "https://example.com/token",
     refreshUrl: null,
     tokenEndpointAuthMethod: "client_secret_post",
+    tokenExchangeBodyFormat: "form",
     userinfoUrl: null,
     baseUrl: null,
     defaultScopes: "[]",