@vellumai/assistant 0.8.7 → 0.8.8-dev.202606052332.17fc8ea

package/src/config/feature-flag-cache.ts

@@ -29,7 +29,7 @@
  */
 
 type FlagSlot = {
-  overrides: Record<string, boolean> | null;
+  overrides: Record<string, boolean | string> | null;
   fromGateway: boolean;
 };
 
@@ -44,7 +44,7 @@ function slot(): FlagSlot {
 }
 
 /** Read the current override cache. `null` means not yet populated. */
-export function getCachedOverrides(): Record<string, boolean> | null {
+export function getCachedOverrides(): Record<string, boolean | string> | null {
   return slot().overrides;
 }
 
@@ -66,7 +66,7 @@ export function isCachedFromGateway(): boolean {
  * `initFeatureFlagOverrides()` calls are no-ops.
  */
 export function setCachedOverrides(
-  overrides: Record<string, boolean>,
+  overrides: Record<string, boolean | string>,
   options: { fromGateway: boolean },
 ): void {
   const s = slot();

package/src/config/feature-flag-registry.json

@@ -26,11 +26,20 @@
       "defaultEnabled": false
     },
     {
-      "id": "prechat-onboarding-condensed-flow",
+      "id": "pre-chat-onboarding-experiment-2026-06-06",
       "scope": "client",
-      "key": "prechat-onboarding-condensed-flow",
-      "label": "Condensed Pre-chat Onboarding",
-      "description": "Enable the condensed pre-chat onboarding flow for a standard LaunchDarkly percentage rollout.",
+      "key": "pre-chat-onboarding-experiment-2026-06-06",
+      "label": "Pre-chat Onboarding Experiment 2026-06-06",
+      "description": "Pre-chat onboarding experiment with control and variant-a arms. Control shows the full funnel; variant-a shows the condensed pared-down flow.",
+      "defaultEnabled": "control",
+      "values": ["control", "variant-a"]
+    },
+    {
+      "id": "experiment-activation-flow-2026-06-03",
+      "scope": "client",
+      "key": "experiment-activation-flow-2026-06-03",
+      "label": "Activation Flow Experiment 2026-06-03",
+      "description": "Route allowlisted users to the activation-rail bootstrap template after pre-chat. Off by default and targeted through LaunchDarkly.",
       "defaultEnabled": false
     },
     {
@@ -57,6 +66,14 @@
       "description": "Control Developer nav visibility in macOS settings",
       "defaultEnabled": false
     },
+    {
+      "id": "system-schedule-toggles",
+      "scope": "assistant",
+      "key": "system-schedule-toggles",
+      "label": "System Schedule Toggles",
+      "description": "Show Heartbeat and Consolidation pause toggles on the schedules settings page. The toggles only pause automatic runs; manual Run now remains available.",
+      "defaultEnabled": false
+    },
     {
       "id": "developer-menu-items",
       "scope": "client",
@@ -363,7 +380,7 @@
     },
     {
       "id": "home-page",
-      "scope": "both",
+      "scope": "client",
       "key": "home-page",
       "label": "Home Page",
       "description": "Enable the Home page as the default landing view.",
@@ -450,11 +467,35 @@
       "defaultEnabled": false
     },
     {
-      "id": "self-intro-greeting",
+      "id": "empty-state-dynamic-greetings",
       "scope": "assistant",
+      "key": "empty-state-dynamic-greetings",
+      "label": "Dynamic Empty-State Greetings",
+      "description": "Enable the empty new-chat screen to refresh cached greeting options via the empty-state greeting LLM call site. When off, the assistant serves workspace-authored greetings or static fallbacks only.",
+      "defaultEnabled": false
+    },
+    {
+      "id": "self-intro-greeting",
+      "scope": "both",
       "key": "self-intro-greeting",
       "label": "Self-intro first message",
-      "description": "On the first conversation, send a natural self-introduction (e.g. \"Hi Vela, I'm alex. Nice to meet you.\") on the user's behalf and route it through real LLM inference, instead of serving the canned first greeting. Names come from the onboarding context; falls back to the canned greeting when no name is known. See assistant/src/daemon/first-greeting.ts (buildSelfIntroMessage).",
+      "description": "On the first conversation, send a natural self-introduction (e.g. \"Hi Vela, I'm alex. Nice to meet you.\") on the user's behalf and route it through real LLM inference, instead of serving the canned first greeting. Names come from the onboarding context; falls back to the canned greeting when no name is known. Exposed to clients so pre-hatch onboarding can compute the source-of-truth initial message, and to the assistant so older clients can still be gated server-side. See assistant/src/daemon/first-greeting.ts (buildSelfIntroMessage).",
+      "defaultEnabled": false
+    },
+    {
+      "id": "provider-first-profile-creation",
+      "scope": "client",
+      "key": "provider-first-profile-creation",
+      "label": "Provider-First Profile Creation",
+      "description": "New profile creation flow: pick (or inline-create) a provider first, then a model, with pre-filled provider and profile name/key, plus a quick-add \"+\" in the chat composer's Model Profile menu. When off, profile creation uses the previous field order and there is no composer quick-add.",
+      "defaultEnabled": false
+    },
+    {
+      "id": "acp",
+      "scope": "assistant",
+      "key": "acp",
+      "label": "ACP Coding Agents",
+      "description": "Enable spawning and steering external coding agents (Claude Code, Codex, Gemini) via the Agent Client Protocol. Alternative gate to the acp.enabled workspace config field; either enables the subsystem.",
       "defaultEnabled": false
     }
   ]

package/src/config/schemas/__tests__/memory-v2.test.ts

@@ -23,6 +23,7 @@ describe("MemoryV2ConfigSchema", () => {
       bm25_k1: 1.2,
       bm25_b: 0.4,
       consolidation_interval_hours: 4,
+      consolidation_enabled: true,
       consolidation_max_buffer_lines: 100,
       max_page_chars: 5000,
       consolidation_prompt_path: null,

package/src/config/schemas/__tests__/memory-v3.test.ts

@@ -0,0 +1,25 @@
+import { describe, expect, test } from "bun:test";
+
+import { MemoryV3ConfigSchema } from "../memory-v3.js";
+
+describe("MemoryV3ConfigSchema", () => {
+  test("parses an empty object to documented defaults", () => {
+    const parsed = MemoryV3ConfigSchema.parse({});
+    expect(parsed).toEqual({
+      workingSet: { maxPages: 150, evictWindow: 5 },
+      l2Concurrency: 16,
+    });
+  });
+
+  test("accepts an explicit l2Concurrency override", () => {
+    expect(MemoryV3ConfigSchema.parse({ l2Concurrency: 8 }).l2Concurrency).toBe(
+      8,
+    );
+  });
+
+  test("rejects a non-positive or non-integer l2Concurrency", () => {
+    expect(() => MemoryV3ConfigSchema.parse({ l2Concurrency: 0 })).toThrow();
+    expect(() => MemoryV3ConfigSchema.parse({ l2Concurrency: -4 })).toThrow();
+    expect(() => MemoryV3ConfigSchema.parse({ l2Concurrency: 1.5 })).toThrow();
+  });
+});

package/src/config/schemas/heartbeat.ts

@@ -56,6 +56,15 @@ export const HeartbeatConfigSchema = z
       .describe(
         "Maximum heartbeats that can run consecutively without a guardian message. Counter resets when the guardian sends a message. Set to null for unlimited.",
       ),
+    maxDailyRuns: z
+      .number({ error: "heartbeat.maxDailyRuns must be a number" })
+      .int("heartbeat.maxDailyRuns must be an integer")
+      .positive("heartbeat.maxDailyRuns must be a positive integer")
+      .nullable()
+      .default(2)
+      .describe(
+        "Maximum heartbeats that can run per calendar day. Resets at midnight local time. Set to null for unlimited.",
+      ),
     disposition: z
       .string({ error: "heartbeat.disposition must be a string" })
       .default(

package/src/config/schemas/llm.ts

@@ -21,6 +21,7 @@ const LLMProvider = z.enum([
   "fireworks",
   "openrouter",
   "openai-compatible",
+  "minimax",
 ]);
 type LLMProvider = z.infer<typeof LLMProvider>;
 

package/src/config/schemas/memory-v2.ts

@@ -194,6 +194,14 @@ export const MemoryV2ConfigSchema = z
       .describe(
         "Hours between scheduled consolidation runs that synthesize buffered memories into concept pages",
       ),
+    consolidation_enabled: z
+      .boolean({
+        error: "memory.v2.consolidation_enabled must be a boolean",
+      })
+      .default(true)
+      .describe(
+        "Whether automatic scheduled consolidation runs are enabled. Manual run-now requests remain available while memory.v2.enabled is true.",
+      ),
     consolidation_max_buffer_lines: z
       .number({
         error: "memory.v2.consolidation_max_buffer_lines must be a number",

package/src/config/schemas/memory-v3.ts

@@ -33,6 +33,14 @@ export const MemoryV3ConfigSchema = z
     workingSet: MemoryV3WorkingSetSchema.default(
       MemoryV3WorkingSetSchema.parse({}),
     ),
+    l2Concurrency: z
+      .number({ error: "memory.v3.l2Concurrency must be a number" })
+      .int("memory.v3.l2Concurrency must be an integer")
+      .positive("memory.v3.l2Concurrency must be a positive integer")
+      .default(16)
+      .describe(
+        "Bounded fan-out for the per-leaf L2 selection (number of leaf selector calls in flight at once).",
+      ),
   })
   .describe("Memory v3 — topic-tree routing with a carry-forward working set");
 

package/src/config/schemas/platform.ts

@@ -87,6 +87,14 @@ export const DaemonConfigSchema = z
       .describe(
         "Whether the daemon records conversations even when no client is connected",
       ),
+    reapOrphanedSubprocesses: z
+      .boolean({
+        error: "daemon.reapOrphanedSubprocesses must be a boolean",
+      })
+      .default(false)
+      .describe(
+        "Whether the daemon, when running as PID 1 in a container, periodically reaps orphaned subprocesses that reparented to it. Off by default while the behavior is being validated.",
+      ),
   })
   .describe("Background daemon process configuration");
 

package/src/config/seed-inference-profiles.ts

@@ -58,7 +58,7 @@ const MANAGED_PROFILE_TEMPLATES: Record<string, ManagedProfileTemplate> = {
     label: "Quality",
     description: "Best results with the most capable model",
     maxTokens: 32000,
-    effort: "max",
+    effort: "high",
     thinking: { enabled: true, streamThinking: true },
     contextWindow: { maxInputTokens: DEFAULT_CONTEXT_WINDOW_MAX_INPUT_TOKENS },
   },
@@ -104,7 +104,7 @@ const USER_PROFILE_TEMPLATES: Record<string, ManagedProfileTemplate> = {
     label: "Quality",
     description: "Best results with the most capable model",
     maxTokens: 32000,
-    effort: "max",
+    effort: "high",
     thinking: { enabled: true, streamThinking: true },
     contextWindow: { maxInputTokens: DEFAULT_CONTEXT_WINDOW_MAX_INPUT_TOKENS },
   },

package/src/config/skills.ts

@@ -52,6 +52,7 @@ const VellumMetadataSchema = z
 
 const SkillMetadataSchema = z
   .object({
+    icon: z.string().optional(),
     emoji: z.string().optional(),
     vellum: VellumMetadataSchema.optional(),
   })
@@ -210,6 +211,7 @@ interface ParsedFrontmatter {
   displayName: string;
   description: string;
   body: string;
+  icon?: string;
   emoji?: string;
   includes?: string[];
   featureFlag?: string;
@@ -253,6 +255,7 @@ function parseFrontmatter(
   }
 
   // metadata is already a parsed object from YAML — validate with Zod schema
+  let icon: string | undefined;
   let emoji: string | undefined;
   let parsedMeta: z.infer<typeof SkillMetadataSchema> | undefined;
   let vellum: z.infer<typeof VellumMetadataSchema> | undefined;
@@ -271,6 +274,7 @@ function parseFrontmatter(
       if (zodResult.success) {
         parsedMeta = zodResult.data;
         vellum = parsedMeta.vellum;
+        icon = parsedMeta.icon;
         emoji = vellum?.emoji ?? parsedMeta.emoji;
       } else {
         // Zod validation failed — fall back to raw parsed object so we don't
@@ -284,6 +288,9 @@ function parseFrontmatter(
         const raw = metadataRaw as Record<string, unknown>;
         parsedMeta = raw as z.infer<typeof SkillMetadataSchema>;
         vellum = raw?.vellum as z.infer<typeof VellumMetadataSchema>;
+        if (typeof parsedMeta?.icon === "string") {
+          icon = parsedMeta.icon;
+        }
         if (vellum && typeof vellum === "object") {
           emoji = typeof vellum.emoji === "string" ? vellum.emoji : undefined;
         }
@@ -337,6 +344,7 @@ function parseFrontmatter(
     displayName,
     description,
     body: strippedBody,
+    icon,
     emoji,
     includes,
     featureFlag,
@@ -489,6 +497,7 @@ function readSkillFromDirectory(
       directoryPath,
       skillFilePath,
       body: parsed.body,
+      icon: parsed.icon,
       emoji: parsed.emoji,
 
       source,
@@ -540,6 +549,7 @@ function readBundledSkillFromDirectory(
       skillFilePath,
       body: parsed.body,
       bundled: true,
+      icon: parsed.icon,
       emoji: parsed.emoji,
 
       source: "bundled",
@@ -599,6 +609,7 @@ function loadBundledSkills(): SkillSummary[] {
       directoryPath: skill.directoryPath,
       skillFilePath: skill.skillFilePath,
       bundled: true,
+      icon: skill.icon,
       emoji: skill.emoji,
 
       source: "bundled",
@@ -649,6 +660,7 @@ function skillSummaryFromDefinition(
     directoryPath: skill.directoryPath,
     skillFilePath: skill.skillFilePath,
     bundled: skill.bundled,
+    icon: skill.icon,
     emoji: skill.emoji,
     source,
     toolManifest: skill.toolManifest,
@@ -836,6 +848,7 @@ export function loadSkillCatalog(
           description: parsed.description,
           directoryPath: directory,
           skillFilePath,
+          icon: parsed.icon,
           emoji: parsed.emoji,
 
           source: "workspace",

package/src/context/compactor.ts

@@ -21,7 +21,6 @@
 import { optimizeImageForTransport } from "../agent/image-optimize.js";
 import type { CompactionConfig } from "../config/schemas/compaction.js";
 import type { LLMCallSite } from "../config/schemas/llm.js";
-import { stripInjectionsForCompaction } from "../daemon/conversation-runtime-assembly.js";
 import { filterMessagesForUntrustedActor } from "../daemon/message-provenance.js";
 import {
   getAttachmentContent,
@@ -42,6 +41,7 @@ import {
   type TrustClass,
 } from "../runtime/actor-trust-resolver.js";
 import { getLogger } from "../util/logger.js";
+import { stripInjectionsForCompaction } from "./strip-injections.js";
 import { estimatePromptTokens } from "./token-estimator.js";
 
 const log = getLogger("compactor");

package/src/context/strip-injections.ts

@@ -0,0 +1,128 @@
+/**
+ * Runtime-injection stripping for compaction and overflow recovery.
+ *
+ * Runtime injections (turn context, memory, NOW.md, workspace, Slack
+ * chronological, etc.) persist in message history to keep the conversation
+ * prefix stable for Anthropic's prefix caching. They only need to be removed
+ * when compaction rewrites the message array, so the compactor summarizes the
+ * raw persistent messages rather than the ephemeral injected blocks.
+ *
+ * This module is the compaction-layer home for that strip so both the agent
+ * loop (which drives the compaction pipeline) and the compactor can call it
+ * without reaching up into the daemon orchestrator.
+ */
+import type { Message } from "../providers/types.js";
+
+/**
+ * A matcher for an injected text block. A plain string matches by prefix
+ * (`startsWith`). A `{ prefix, suffix }` wrapper requires BOTH the opening
+ * prefix and the closing suffix, so user-authored content that merely begins
+ * with an injection-like opening tag (e.g. a message discussing `<info>`
+ * markup) is not mistaken for an injected block and dropped. This mirrors
+ * `countMemoryPrefixBlocks`, which only treats `<memory>…</memory>` /
+ * `<info>…</info>` blocks as injected when the full wrapper is present.
+ */
+export type InjectionMatcher = string | { prefix: string; suffix: string };
+
+/**
+ * Remove text blocks from user messages that match any of the given matchers.
+ * If stripping removes all content blocks from a message, the message itself
+ * is dropped.
+ *
+ * This is the shared primitive behind the individual strip* functions and
+ * the `stripInjectionsForCompaction` pipeline.
+ */
+export function stripUserTextBlocksByPrefix(
+  messages: Message[],
+  matchers: InjectionMatcher[],
+): Message[] {
+  return messages
+    .map((message) => {
+      if (message.role !== "user") return message;
+      const nextContent = message.content.filter((block) => {
+        if (block.type !== "text") return true;
+        return !matchers.some((m) =>
+          typeof m === "string"
+            ? block.text.startsWith(m)
+            : block.text.startsWith(m.prefix) && block.text.endsWith(m.suffix),
+        );
+      });
+      if (nextContent.length === message.content.length) return message;
+      if (nextContent.length === 0) return null;
+      return { ...message, content: nextContent };
+    })
+    .filter(
+      (message): message is NonNullable<typeof message> => message != null,
+    );
+}
+
+/** Matchers stripped by the pipeline (order doesn't matter — single pass). */
+const RUNTIME_INJECTION_PREFIXES: InjectionMatcher[] = [
+  "<channel_capabilities>",
+  "<channel_command_context>",
+  "<disk_pressure_warning>",
+  "<channel_turn_context>", // backward-compat: strip legacy separate channel blocks
+  "<guardian_context>",
+  "<inbound_actor_context>", // backward-compat: strip legacy separate actor blocks
+  "<interface_turn_context>", // backward-compat: strip legacy separate interface blocks
+  // NOTE: <turn_context> is intentionally NOT stripped — unified turn context
+  // blocks persist in history so the assistant retains temporal/actor grounding.
+  "<background_turn>",
+  "<memory_context __injected>",
+  "<memory_context>", // backward-compat: strip legacy blocks from pre-__injected history
+  // The static `memory-v2-static` block (`<info>\n…</info>`) and the
+  // dynamic activation block (`<memory>\n…</memory>`, plus legacy
+  // `<memory __injected>…`) are both stripped so each compaction
+  // re-injects the freshest essentials/threads/recent/buffer view and
+  // re-runs the activation pipeline, matching the `<knowledge_base>`
+  // cadence. The activation pipeline dedupes via `everInjected`, and
+  // compaction handles aggregate growth, so accumulation does not cause
+  // unbounded context growth. Both wrappers may appear in persisted rows.
+  //
+  // These two use the full `{ prefix, suffix }` wrapper shape (not a bare
+  // prefix) so that user-authored text merely starting with `<memory>\n` or
+  // `<info>\n` is never silently dropped during compaction/`/clean`. This
+  // matches the full-wrapper requirement in `countMemoryPrefixBlocks`.
+  { prefix: "<memory>\n", suffix: "\n</memory>" },
+  { prefix: "<info>\n", suffix: "\n</info>" },
+  "<voice_call_control>",
+  "<workspace_top_level>", // backward-compat: strip legacy workspace blocks
+  // The `<workspace>` top-level block is stripped so each compaction re-injects
+  // a fresh directory snapshot rather than carrying a stale listing into the
+  // summary — matching the `<knowledge_base>`/`<NOW.md>` cadence and keeping the
+  // `workspace-context` injector's presence detection in lockstep (the block is
+  // present exactly when compaction would strip it). The full `{ prefix, suffix }`
+  // wrapper shape ensures user-authored text merely starting with `<workspace>\n`
+  // is never mistaken for an injected block.
+  { prefix: "<workspace>\n", suffix: "\n</workspace>" },
+  "<temporal_context>\nToday:", // backward-compat: strip legacy temporal blocks
+  "<active_subagents>",
+  "<active_workspace>",
+  "<active_dynamic_page>",
+  "<non_interactive_context>",
+  // Shared prefix catches both the current NOW.md tag and any pre-line-limit
+  // variant that may linger in in-flight histories during a rolling deploy.
+  "<NOW.md Always keep this up to date",
+  "<now_scratchpad>", // backward-compat: strip legacy blocks from pre-rename history
+  "<knowledge_base>",
+  "<pkb>", // backward-compat: strip legacy tag from pre-rename history
+  "<system_reminder>",
+  "<transport_hints>",
+  // The Slack active-thread focus block is non-persisted and injected on
+  // the FINAL user turn only. Strip it here so re-assembly during compaction
+  // and overflow recovery does not duplicate it across turns.
+  "<active_thread>",
+  "<system_notice>One or more tool calls returned an error.",
+];
+
+/**
+ * Strip all runtime-injected context from message history in a single pass.
+ *
+ * Used only during compaction and overflow recovery — not on normal turns.
+ * Runtime injections persist in history to keep the conversation prefix
+ * stable for Anthropic's prefix caching. Stripping is only needed when
+ * compaction rewrites the message array (cache miss is expected anyway).
+ */
+export function stripInjectionsForCompaction(messages: Message[]): Message[] {
+  return stripUserTextBlocksByPrefix(messages, RUNTIME_INJECTION_PREFIXES);
+}

package/src/context/token-estimator.ts

@@ -359,6 +359,29 @@ export function estimatePromptTokens(
   return correction === 1.0 ? raw : Math.ceil(raw * correction);
 }
 
+/**
+ * Calibrated prompt-token estimate including the tool-definition budget.
+ *
+ * Combines the per-tool budget ({@link estimateToolsTokens}) with the
+ * message/system estimate ({@link estimatePromptTokens}) under the EWMA
+ * calibration correction. This is the estimate the overflow gate consumes;
+ * the pre-send calibration capture in `agent/loop.ts` deliberately stays on
+ * `estimatePromptTokensRaw` so the calibrator trains against the uncorrected
+ * value rather than chasing its own output.
+ */
+export function estimatePromptTokensWithTools(
+  history: Message[],
+  systemPrompt: string | undefined,
+  tools: ToolDefinition[],
+  providerName: string,
+): number {
+  const toolTokenBudget = tools.length > 0 ? estimateToolsTokens(tools) : 0;
+  return estimatePromptTokens(history, systemPrompt, {
+    providerName,
+    toolTokenBudget,
+  });
+}
+
 function stableJson(value: unknown): string {
   try {
     return JSON.stringify(value);

package/src/context/tool-result-truncation.ts

@@ -5,29 +5,6 @@ import type {
   ToolResultContent,
 } from "../providers/types.js";
 
-/**
- * Maximum share of the context window that a single tool result may occupy.
- */
-const MAX_TOOL_RESULT_CONTEXT_SHARE = 0.3;
-
-/**
- * Absolute cap on tool-result characters (~100K tokens).
- */
-export const HARD_MAX_TOOL_RESULT_CHARS = 400_000;
-
-/**
- * Calculate the maximum allowed characters for a tool result based on the
- * context window size. Uses ~4 chars per token as a rough heuristic.
- */
-export function calculateMaxToolResultChars(
-  contextWindowTokens: number,
-): number {
-  return Math.min(
-    HARD_MAX_TOOL_RESULT_CHARS,
-    Math.floor(contextWindowTokens * MAX_TOOL_RESULT_CONTEXT_SHARE * 4),
-  );
-}
-
 /**
  * Aggressively truncate all tool-result text across an entire message history.
  *

package/src/context/window-manager.ts

@@ -40,7 +40,8 @@ const INTERNAL_CONTEXT_SUMMARY_MESSAGES = new WeakSet<Message>();
 
 // ---------------------------------------------------------------------------
 // Public types — preserved for downstream consumers (agent loop, conversation,
-// plugin pipeline, applyCompactionResult, routes/playground/force-compact).
+// default compaction plugin, applyCompactionResult,
+// routes/playground/force-compact).
 // ---------------------------------------------------------------------------
 
 export interface ContextWindowResult {
@@ -90,7 +91,6 @@ export interface ShouldCompactResult {
 }
 
 export interface ContextWindowCompactOptions {
-  lastCompactedAt?: number;
   /** Skip the auto-threshold check (used for /compact and recovery). */
   force?: boolean;
   /**
@@ -104,14 +104,12 @@ export interface ContextWindowCompactOptions {
    */
   precomputedEstimate?: number;
   /**
-   * Legacy fields retained for backwards compatibility with existing
-   * callers. The new assistant-driven compactor does not consume them —
-   * the model decides where to cut and what to keep — but accepting them
+   * Legacy field retained for backwards compatibility with existing
+   * callers. The new assistant-driven compactor does not consume it —
+   * the model decides where to cut and what to keep — but accepting it
    * here lets callers keep their existing call sites unchanged.
    */
   minKeepRecentUserTurns?: number;
-  conversationOriginChannel?: string;
-  targetInputTokensOverride?: number;
   /**
    * Trust class of the actor whose turn triggered compaction. Forwarded to
    * the compactor so the image manifest excludes guardian-only attachments

package/src/credential-execution/executable-discovery.ts

@@ -17,6 +17,7 @@
  */
 
 import { existsSync } from "node:fs";
+import { createRequire } from "node:module";
 import { dirname, join } from "node:path";
 
 import { getIsContainerized } from "../config/env-registry.js";
@@ -160,6 +161,21 @@ export function discoverLocalCes():
     return { mode: "local-source", sourcePath: sourceEntry };
   }
 
+  // npm-layout fallback: resolve via node_modules when installed as a package
+  try {
+    const _require = createRequire(import.meta.url);
+    const pkgPath = _require.resolve(
+      "@vellumai/credential-executor/package.json",
+    );
+    const npmSourceEntry = join(dirname(pkgPath), "src", "main.ts");
+    if (existsSync(npmSourceEntry)) {
+      log.info({ path: npmSourceEntry }, "Found CES source via npm package");
+      return { mode: "local-source", sourcePath: npmSourceEntry };
+    }
+  } catch {
+    // Package not installed — fall through to unavailable
+  }
+
   const reason = `CES executable not found. Searched: ${searchPaths.join(", ")}; also checked source at ${sourceEntry}`;
   log.warn(reason);
   return { mode: "unavailable", reason };

package/src/daemon/__tests__/conversation-lifecycle-auto-analyze.test.ts

@@ -165,6 +165,12 @@ function makeDisposeContext(
   const ctx = {
     conversationId: overrides.conversationId ?? "conv-1",
     processing: false,
+    isProcessing(this: { processing: boolean }) {
+      return this.processing;
+    },
+    setProcessing(this: { processing: boolean }, value: boolean) {
+      this.processing = value;
+    },
     abortController,
     prompter,
     secretPrompter,