@vellumai/assistant 0.8.8 → 0.8.9-staging.1

package/src/config/bundled-skills/app-builder/TOOLS.json

@@ -74,6 +74,17 @@
           "change_summary": {
             "type": "string",
             "description": "Short summary of what changed, using git conventional commit format (e.g. 'feat: add coin flip app', 'fix: correct button alignment'). Used as the version history entry."
+          },
+          "icon": {
+            "type": "string",
+            "description": "Lenient alias. Prefer preview.icon. An emoji or image URL passed here is folded into preview.icon automatically."
+          },
+          "html": {
+            "type": "string",
+            "description": "Lenient alias. Prefer source_files. Raw HTML passed here is folded into source_files['src/index.html'] automatically."
+          },
+          "pages": {
+            "description": "Retired. Multi-page apps route inside the Preact app under src/components/. Passing this returns a guidance error."
           }
         },
         "required": ["name"]
@@ -150,6 +161,24 @@
       },
       "executor": "tools/app-generate-icon.ts",
       "execution_target": "host"
+    },
+    {
+      "name": "app_list",
+      "description": "List the persistent apps the user has built, returning each app's app_id, name, description, and timestamps (most recently updated first). Pass `query` to resolve an app the user mentions by name to its app_id before calling app_open. Use this whenever the user refers to an existing app and you don't already have its app_id in context.",
+      "category": "apps",
+      "risk": "low",
+      "input_schema": {
+        "type": "object",
+        "properties": {
+          "query": {
+            "type": "string",
+            "description": "Optional name (or partial name) the user mentioned. When provided, only matching apps are returned: an exact case-insensitive name match if one exists, otherwise substring matches in either direction (e.g. 'habit tracker' resolves 'Habit Tracker'). Omit to list all apps."
+          }
+        },
+        "required": []
+      },
+      "executor": "tools/app-list.ts",
+      "execution_target": "host"
     }
   ]
 }

package/src/config/bundled-skills/app-builder/references/DESIGN_SYSTEM.md

@@ -0,0 +1,48 @@
+# Design System
+
+A design system CSS is auto-injected inside a `@layer`, so your styles always take priority. It provides element defaults and automatic light/dark mode switching via `prefers-color-scheme`.
+
+**Use `--v-*` variables and `.v-*` classes** — they handle light/dark mode automatically. No manual dark mode CSS needed.
+
+---
+
+## Design tokens
+
+| Category        | Tokens                                                                                                                                                         |
+| --------------- | -------------------------------------------------------------------------------------------------------------------------------------------------------------- |
+| **Backgrounds** | `--v-bg`, `--v-surface`, `--v-surface-border`                                                                                                                  |
+| **Text**        | `--v-text`, `--v-text-secondary`, `--v-text-muted`                                                                                                             |
+| **Accent**      | `--v-accent`, `--v-accent-hover`                                                                                                                               |
+| **Status**      | `--v-success`, `--v-danger`, `--v-warning`                                                                                                                     |
+| **Spacing**     | `--v-spacing-xxs` (2px) / `-xs` (4px) / `-sm` (8px) / `-md` (12px) / `-lg` (16px) / `-xl` (24px) / `-xxl` (32px) / `-xxxl` (48px)                              |
+| **Radius**      | `--v-radius-xs` (2px) / `-sm` (4px) / `-md` (8px) / `-lg` (12px) / `-xl` (16px) / `-pill` (999px)                                                              |
+| **Shadows**     | `--v-shadow-sm`, `--v-shadow-md`, `--v-shadow-lg`                                                                                                              |
+| **Typography**  | `--v-font-family`, `--v-font-mono`, `--v-font-size-xs` (10px) / `-sm` (11px) / `-base` (14px) / `-lg` (17px) / `-xl` (22px) / `-2xl` (26px), `--v-line-height` |
+| **Animation**   | `--v-duration-fast` (0.15s) / `-standard` (0.25s) / `-slow` (0.4s)                                                                                             |
+| **Palettes**    | `--v-slate-{950..50}`, `--v-emerald-*`, `--v-violet-*`, `--v-indigo-*`, `--v-rose-*`, `--v-amber-*`                                                            |
+| **Constant**    | `--v-aux-white` (always `#FFFFFF` in both modes — use for text on filled/accent backgrounds)                                                                    |
+
+---
+
+## Utility classes
+
+`.v-button` (`.secondary` / `.danger` / `.ghost`), `.v-card`, `.v-list` / `.v-list-item`, `.v-badge` (`.success` / `.warning` / `.danger`), `.v-input-row`, `.v-empty-state`, `.v-toggle`.
+
+⚠️ **Never hardcode `color: white` or `color: #fff`.** Use `var(--v-aux-white)` for text on filled/accent backgrounds, or `var(--v-text)` / `var(--v-text-secondary)` for text on surface backgrounds. Hardcoded white causes invisible text on light surfaces.
+
+---
+
+## Custom themes
+
+When the user wants a specific branded look, write complete CSS with hardcoded colors and `@media (prefers-color-scheme: dark)` for dark variants. Don't mix `--v-*` auto-switching variables with hardcoded colors in the same element.
+
+---
+
+## Theme detection in JavaScript
+
+```javascript
+console.log(window.vellum.theme.mode); // 'light' or 'dark'
+window.addEventListener("vellum-theme-change", (e) => {
+  console.log("Theme:", e.detail.mode);
+});
+```

package/src/config/bundled-skills/app-builder/references/RESPONSIVE.md

@@ -0,0 +1,57 @@
+# Responsive Baseline & Mobile-First Mode
+
+Every app must work across phone (~360px) to desktop (~1400px+).
+
+## Mode selection
+
+The conversation context's `<turn_context>` block carries an `interface:` field.
+
+**If `interface: ios`** (or any future mobile-web / android identifier):
+  → Mobile-first build. Design the narrow viewport first, enhance upward.
+
+**If `interface: macos` or `web`**:
+  → Desktop-first build. Design larger composition first; narrow fallback still meets the universal baseline below.
+
+**If field is absent or ambiguous**:
+  → Default to desktop-first unless the request implies phone use ("for my iPhone", "a tap-tracker I'll use on the go").
+
+---
+
+## Universal baseline (every build, regardless of interface)
+
+### Viewport & safe areas
+
+- Set viewport meta: `<meta name="viewport" content="width=device-width, initial-scale=1, viewport-fit=cover">`. Never set `user-scalable=no` — it blocks accessibility zoom.
+- Pad the root container with `env(safe-area-inset-*)` so content clears the notch: `padding-top: max(var(--v-spacing-lg), env(safe-area-inset-top))`, mirrored for `-bottom` / `-left` / `-right`.
+- Use `100dvh` (dynamic viewport height), not `100vh`, for full-height containers.
+
+### Form controls
+
+- `<input>`, `<textarea>`, `<select>` must be `font-size: 16px` or larger, or iOS Safari will zoom on focus and break the layout. This applies to every build.
+- Add `inputmode` to text fields with structured input: `numeric` for integers, `decimal` for amounts, `email`, `tel`, `url`.
+
+### Touch & hover
+
+- Interactive elements must be ≥44×44pt. `.v-button` already meets this; for custom controls, set `min-height: 44px` explicitly.
+- Gate hover affordances behind `@media (hover: hover)` so they don't stick on touch devices.
+- Disable text selection on app chrome with `user-select: none; -webkit-user-select: none`.
+
+### Layout fluidity
+
+- Use fluid widths only. No fixed-pixel layouts. Prefer `%`, `fr`, `minmax`, `clamp()` over `px` on container widths.
+- At narrow widths, collapse tables into stacked cards with labels and values arranged vertically.
+- Size `vellum.widgets.*` chart containers in `vw` / `%`, not fixed `px`.
+
+---
+
+## Mobile-first priorities (`interface: ios`)
+
+- Default body text to `--v-font-size-lg` (17px), not `--v-font-size-base` (14px).
+- Bump default vertical rhythm one step (e.g. `--v-spacing-md` → `--v-spacing-lg`).
+- One column as the default, not a fallback. Opt into multi-column only above `@media (min-width: 720px)`.
+- Bottom-anchor the primary action: `position: sticky; bottom: env(safe-area-inset-bottom)`.
+- Replace side modals and popovers with bottom sheets.
+
+## Desktop-first priorities (`interface: macos` / `web`)
+
+Multi-column composition, hover-rich affordances, denser information, side modals, inline primary actions. The universal baseline above is still the floor — narrow view must still work.

package/src/config/bundled-skills/app-builder/references/SLIDES.md

@@ -0,0 +1,38 @@
+# Presentation Slide Design
+
+Slides are a different domain from apps. Skip app-specific patterns (contextual headers, search/filter, toast notifications, form validation, custom routes). Slides are static content — build navigation and layouts with custom HTML/CSS.
+
+## Key principles
+
+- One idea per slide — understood in 3 seconds
+- Layout variety — 3+ different types per deck, never consecutive same-type
+- 8 layout types: Title, Stats, Bullets, Quote, Comparison, Timeline, Visual/Immersive, Closing/CTA
+- Bold backgrounds — dark, gradient, or strongly tinted
+- Max 6 bullets per slide, max 3 sentences body text
+- Never go below 15px for any visible text
+
+## Navigation
+
+Build slide navigation as your own component. Common patterns:
+- Keyboard: `ArrowLeft` / `ArrowRight` / `Space` / `Escape`
+- Click affordances at left/right edges
+- Slide counter pill in a corner (e.g. `3 / 12`)
+- Optional progress bar at the top
+
+## Layout templates
+
+- **Title** — Centered headline, optional subtitle, no body. Full-bleed background or gradient.
+- **Stats** — One huge number, label below, supporting paragraph optional.
+- **Bullets** — Heading + 3–6 short bullets. Avoid wall-of-text.
+- **Quote** — Pull quote in large italic type, attribution below, contrasting background.
+- **Comparison** — Two columns (before/after, us/them, problem/solution). Visual symmetry matters.
+- **Timeline** — Horizontal or vertical sequence with dates and milestones.
+- **Visual/Immersive** — Full-bleed image, gradient, or generated graphic with minimal text overlay.
+- **Closing/CTA** — Headline + single call to action. Mirror the title slide aesthetic.
+
+## What to avoid
+
+- Generic Keynote / PowerPoint aesthetic (default white background, sans-serif body, bullet lists everywhere)
+- Tiny text below 15px — slides are read across rooms
+- Same layout type used 3+ times in a row — vary the rhythm
+- Body paragraphs longer than 3 sentences — split into multiple slides

package/src/config/bundled-skills/app-builder/tools/app-list.ts

@@ -0,0 +1,62 @@
+import * as appStore from "../../../../memory/app-store.js";
+import type {
+  ToolContext,
+  ToolExecutionResult,
+} from "../../../../tools/types.js";
+
+interface AppListEntry {
+  app_id: string;
+  name: string;
+  description?: string;
+  updated_at: number;
+  created_at: number;
+}
+
+function toEntry(app: appStore.AppDefinition): AppListEntry {
+  return {
+    app_id: app.id,
+    name: app.name,
+    ...(app.description ? { description: app.description } : {}),
+    updated_at: app.updatedAt,
+    created_at: app.createdAt,
+  };
+}
+
+export async function run(
+  input: Record<string, unknown>,
+  _context: ToolContext,
+): Promise<ToolExecutionResult> {
+  const apps = appStore.listApps();
+  const entries = apps.map(toEntry);
+
+  const query =
+    typeof input.query === "string" ? input.query.trim().toLowerCase() : "";
+
+  if (!query) {
+    return {
+      content: JSON.stringify({ count: entries.length, apps: entries }),
+      isError: false,
+    };
+  }
+
+  // Resolve the app the user mentioned by name. Prefer an exact (case-insensitive)
+  // name match, then fall back to substring matches in either direction so
+  // "habit tracker" resolves "Habit Tracker" and "my budget" resolves "Budget".
+  const exact = entries.filter((e) => e.name.toLowerCase() === query);
+  const matches =
+    exact.length > 0
+      ? exact
+      : entries.filter((e) => {
+          const name = e.name.toLowerCase();
+          return name.includes(query) || query.includes(name);
+        });
+
+  return {
+    content: JSON.stringify({
+      query: input.query,
+      match_count: matches.length,
+      matches,
+    }),
+    isError: false,
+  };
+}

package/src/config/bundled-skills/document-editor/SKILL.md

@@ -1,44 +1,37 @@
 ---
 name: document-editor
-description: Rich text document editor with collaborative editing tools — create, read, update, and annotate documents
+description: Use whenever the user wants to write or draft an article, blog post, essay, report, or any long-form content. Creates the content in a rich text editor instead of dumping it in chat, so it can be streamed, reviewed, edited, and exported.
 compatibility: "Designed for Vellum personal assistants"
 metadata:
-  emoji: "📄"
+  emoji: "✍️"
   vellum:
-    display-name: "Document Editor"
+    display-name: "Document Writer"
     activation-hints:
-      - "User asks to write, draft, or collaborate on long-form content — use the document editor for a better editing experience"
-      - "When content will be iterated on, reviewed, or exported, prefer the document editor over inline markdown"
-      - "When a file attachment contains a draft or document the user wants to iterate on, open it in the editor"
+      - "User asks to write, draft, or compose an article, blog post, essay, report, story, or any long-form content — always create it in the document editor, not as a chat reply"
+      - "User wants written content they will iterate on, review, or export — use the editor instead of inline markdown"
+      - "A file attachment contains a draft or document the user wants to revise — open it in the editor"
+    avoid-when:
+      - "The user wants an interactive app, dashboard, calculator, game, or anything with state or data — use app-builder instead"
+      - "A one or two sentence answer is enough — just reply in chat"
 ---
 
-Create and edit long-form documents using the built-in rich text editor. Documents open in workspace mode with chat docked to the side.
+Write and edit long-form documents using the built-in rich text editor. Documents open in workspace mode with chat docked to the side. When a request is about writing prose (an article, blog post, report, essay, story, or similar), create it here rather than writing it into the chat response.
 
 ## Tools
 
-- **document_open** - Opens an existing document in the editor panel by `surface_id`. Use this when a document exists but isn't visible in the editor — for example after the user switches devices, refreshes the page, or when the editor panel was closed. Fetches the document from storage and sends it to the client.
-- **document_create** - Opens a new document editor with an optional title and initial Markdown content. Returns a `surface_id` for subsequent updates.
+- **document_create** - Opens a new document editor with an optional title and initial Markdown content. Returns a `surface_id` for subsequent updates. This is the entry point for any new piece of writing.
 - **document_update** - Updates content in an open document editor by `surface_id`. Supports `replace` (overwrite) and `append` (add to end) modes.
 - **document_read** - Reads the current content of a document by `surface_id` when it belongs to the current conversation, or when the current actor is the guardian/local user. Use to verify content before editing.
-- **document_list** - Lists documents. Without `query`, lists the current conversation's documents. With `query`, searches by title; guardian/local users can search across conversations, while other actors are scoped to the current conversation.
 - **document_find** - Searches a document for text or regex patterns. Returns matching lines with line numbers, match positions, and matched text.
 - **document_replace_text** - Targeted find-and-replace within a document. Supports literal and regex patterns (with backreferences). Optionally limit the number of replacements.
+- **document_list** - Lists documents. Without `query`, lists the current conversation's documents. With `query`, searches by title; guardian/local users can search across conversations, while other actors are scoped to the current conversation.
+- **document_open** - Opens an existing document in the editor panel by `surface_id`. Use this when a document exists but isn't visible in the editor — for example after the user switches devices, refreshes the page, or when the editor panel was closed. Fetches the document from storage and sends it to the client.
 - **document_delete** - Deletes a document by `surface_id`. Use to clean up unwanted documents.
 
-## Retrieving existing documents
-
-When the user asks to see, open, or pull up a document:
-
-1. Check the `<active_documents>` block in your context — it lists all documents in this conversation with their `surface_id` and title.
-2. If the document is NOT in `<active_documents>`, call `document_list` with a `query` matching the document title. For guardian/local users, this searches across previous conversations and sessions.
-3. Once you have the `surface_id`, call `document_open` to open the editor panel. This surfaces the editor on the client and returns document metadata (`surface_id`, `title`, `word_count`) — not the full content. If you need the actual document text, follow up with `document_read`.
-
-**Never** search the filesystem, conversation history, or archives to find a document. Always use `document_list` with a `query`.
-
-**If the user says they can't see a document you know exists** (e.g. after switching from macOS to web, or after a page refresh), call `document_open` with the `surface_id` to re-surface the editor panel on their current client.
-
 ## Creating a new document
 
+This is the default path when the user asks you to write something.
+
 1. **Create the document**: Call `document_create` with a title (inferred from the request). Call the tool immediately, not after conversational preamble. Capture the `surface_id` from the response — every subsequent `document_update` call must reference it.
 2. **Write content in Markdown**: Use proper structure (`#` for titles, `##` for sections), **bold**, _italic_, code blocks, tables, lists, blockquotes as appropriate.
 3. **CRITICAL - Stream content in chunks**: Call `document_update` MULTIPLE times, not just once. Break content into logical chunks (paragraphs, sections, or every 200-300 words). Call `document_update` with `mode: "append"` for EACH chunk separately. The user experiences real-time content appearing as you write.
@@ -59,6 +52,18 @@ When the user requests changes to a document:
    - `document_find` + `document_replace_text` — **for everything else**. Fixing typos, renaming terms, swapping sections, reordering content, adjusting formatting, or any edit that touches only part of the document. This is the default choice for edits. It avoids rewriting the entire document and eliminates the risk of accidentally dropping content.
 4. **Do NOT use `document_update` with `mode: "replace"` for targeted edits.** Rewriting the entire document to change a few words or rearrange sections is wasteful and error-prone.
 
+## Retrieving existing documents
+
+When the user asks to see, open, or pull up a document:
+
+1. Check the `<active_documents>` block in your context — it lists all documents in this conversation with their `surface_id` and title.
+2. If the document is NOT in `<active_documents>`, call `document_list` with a `query` matching the document title. For guardian/local users, this searches across previous conversations and sessions.
+3. Once you have the `surface_id`, call `document_open` to open the editor panel. This surfaces the editor on the client and returns document metadata (`surface_id`, `title`, `word_count`) — not the full content. If you need the actual document text, follow up with `document_read`.
+
+**Never** search the filesystem, conversation history, or archives to find a document. Always use `document_list` with a `query`.
+
+**If the user says they can't see a document you know exists** (e.g. after switching from macOS to web, or after a page refresh), call `document_open` with the `surface_id` to re-surface the editor panel on their current client.
+
 ## Find & Replace
 
 Use `document_find` and `document_replace_text` for surgical edits that target specific text patterns without rewriting the entire document.
@@ -117,7 +122,7 @@ Users can leave inline comments on documents. Open comments are surfaced in a `<
 ## Anti-Patterns
 
 - **Don't use `app_create` for blog posts, articles, or written content.** Use `document_create` — apps are for interactive content with state/data.
-- **Don't output the full content in chat.** The content goes in the document editor, not in the chat response. Acknowledge what you're doing and stream to the editor.
+- **Don't write the article into the chat response.** Long-form prose goes in the document editor via `document_create`, not in chat and not into a `.md` file in the workspace. Acknowledge what you're doing and stream to the editor.
 - **Don't wait to generate everything before sending.** Stream content in chunks via `document_update` with `mode: "append"` so users see progress in real time.
 
 ## Usage Notes

package/src/config/bundled-skills/document-editor/TOOLS.json

@@ -42,7 +42,7 @@
     },
     {
       "name": "document_update",
-      "description": "Update content in an open document editor. Use this to stream generated content or apply edits.",
+      "description": "Append to or fully replace a document's content. Takes the full Markdown in `content` plus a `mode`: \"append\" adds it to the end (use this to stream generated content in chunks), \"replace\" overwrites the entire document. This tool does NOT do targeted edits — for changing specific words, phrases, or sections, use document_replace_text instead. There is no old/new or find/replace parameter here.",
       "category": "document-editor",
       "risk": "low",
       "input_schema": {

package/src/config/bundled-tool-registry.ts

@@ -23,6 +23,7 @@ import * as acpSteer from "./bundled-skills/acp/tools/acp-steer.js";
 import * as appCreate from "./bundled-skills/app-builder/tools/app-create.js";
 import * as appDelete from "./bundled-skills/app-builder/tools/app-delete.js";
 import * as appGenerateIcon from "./bundled-skills/app-builder/tools/app-generate-icon.js";
+import * as appList from "./bundled-skills/app-builder/tools/app-list.js";
 import * as appRefresh from "./bundled-skills/app-builder/tools/app-refresh.js";
 // ── app-control ────────────────────────────────────────────────────────────────
 import * as appControlClick from "./bundled-skills/app-control/tools/app-control-click.js";
@@ -142,6 +143,7 @@ export const bundledToolRegistry = new Map<string, SkillToolScript>([
   ["app-builder:tools/app-delete.ts", appDelete],
   ["app-builder:tools/app-refresh.ts", appRefresh],
   ["app-builder:tools/app-generate-icon.ts", appGenerateIcon],
+  ["app-builder:tools/app-list.ts", appList],
 
   // app-control
   ["app-control:tools/app-control-start.ts", appControlStart],

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

@@ -26,12 +26,13 @@
       "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.",
-      "defaultEnabled": false
+      "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",
@@ -488,6 +489,14 @@
       "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/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/context/strip-injections.ts

@@ -87,8 +87,14 @@ const RUNTIME_INJECTION_PREFIXES: InjectionMatcher[] = [
   { prefix: "<info>\n", suffix: "\n</info>" },
   "<voice_call_control>",
   "<workspace_top_level>", // backward-compat: strip legacy workspace blocks
-  // NOTE: <workspace> is intentionally NOT stripped — workspace context
-  // persists in history so the assistant retains workspace grounding.
+  // 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>",

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 {
@@ -198,8 +199,7 @@ export class ContextWindowManager {
    * `compactedMessages` so `compactedPersistedMessages` only reflects DB
    * rows. Decremented after a successful compaction.
    */
-  nonPersistedPrefixCount = 0;
-  summaryIsInjected = false;
+  private _nonPersistedPrefixCount = 0;
   private _resolvedSystemPrompt: string | undefined;
 
   constructor(options: ContextWindowManagerOptions) {
@@ -215,6 +215,21 @@ export class ContextWindowManager {
     this.config = config;
   }
 
+  /** Leading non-persisted inherited-context messages the compactor preserves. */
+  get nonPersistedPrefixCount(): number {
+    return this._nonPersistedPrefixCount;
+  }
+
+  /**
+   * Seed the non-persisted inherited-context prefix when a forked/sub
+   * conversation inherits its parent's in-memory context. The compactor folds
+   * this prefix into the summary and {@link consumeCompactionState} decrements
+   * it as the leading messages are compacted away.
+   */
+  seedNonPersistedPrefix(count: number): void {
+    this._nonPersistedPrefixCount = count;
+  }
+
   private get estimationProviderName(): string {
     return this.provider.tokenEstimationProvider ?? this.provider.name;
   }
@@ -352,7 +367,7 @@ export class ContextWindowManager {
       signal,
       overrideProfile: options?.overrideProfile ?? null,
       actorTrustClass: options?.actorTrustClass,
-      nonPersistedPrefixCount: this.nonPersistedPrefixCount,
+      nonPersistedPrefixCount: this._nonPersistedPrefixCount,
     };
 
     // Retry budget for the compactor itself. Lives here (not in the
@@ -390,7 +405,7 @@ export class ContextWindowManager {
         ...baseArgs,
         messages: result.messages,
         previousEstimatedInputTokens: previousEstimate,
-        nonPersistedPrefixCount: this.nonPersistedPrefixCount,
+        nonPersistedPrefixCount: this._nonPersistedPrefixCount,
       });
       if (!nextResult.compacted) break;
       const nextEstimate = this.recomputePostCompactionEstimate(
@@ -434,21 +449,20 @@ export class ContextWindowManager {
   }
 
   /**
-   * Decrement the non-persisted prefix bookkeeping and clear the
-   * injected-summary flag after a productive compaction. Called once per
-   * successful internal attempt so multi-attempt runs keep the count
-   * honest as the leading injected messages get folded into the summary.
+   * Decrement the non-persisted prefix bookkeeping after a productive
+   * compaction. Called once per successful internal attempt so multi-attempt
+   * runs keep the count honest as the leading injected messages get folded
+   * into the summary.
    */
   private consumeCompactionState(compactedMessages: number): void {
     const compactedAway = Math.min(
-      this.nonPersistedPrefixCount,
+      this._nonPersistedPrefixCount,
       compactedMessages,
     );
-    this.nonPersistedPrefixCount = Math.max(
+    this._nonPersistedPrefixCount = Math.max(
       0,
-      this.nonPersistedPrefixCount - compactedAway,
+      this._nonPersistedPrefixCount - compactedAway,
     );
-    this.summaryIsInjected = false;
   }
 }
 

package/src/daemon/conversation-agent-loop-handlers.ts

@@ -76,10 +76,8 @@ import {
   cleanAssistantContent,
   drainDirectiveDisplayBuffer,
 } from "./assistant-attachments.js";
-import type {
-  AgentLoopConversationContext,
-  AssistantSurface,
-} from "./conversation-agent-loop.js";
+import type { Conversation } from "./conversation.js";
+import type { AssistantSurface } from "./conversation-agent-loop.js";
 import {
   buildConversationErrorMessage,
   classifyConversationError,
@@ -295,29 +293,11 @@ export interface EventHandlerState {
    * never claims content a flush has not yet written.
    */
   lastPersistedContentSeq: number | undefined;
-  /**
-   * Whether the workspace top-level block should be (re)injected on this
-   * turn. Compaction's prepare phase strips the workspace / NOW.md / PKB
-   * blocks off the tail, so it is set after any successful compaction to
-   * force the workspace overview back in. On an ordinary turn the block is
-   * already present in history, so it defaults `false` to avoid burning
-   * tokens re-injecting it redundantly.
-   */
-  shouldInjectWorkspace: boolean;
-  /**
-   * Whether the reducer has compacted `ctx.messages`, gating the Slack
-   * chronological-transcript override on re-injection. The captured
-   * transcript is the full persisted history; blindly replaying it after
-   * compaction would overwrite the reduced messages and undo compaction, so
-   * once this is `true` the override falls back to the reduced
-   * `ctx.messages`.
-   */
-  reducerCompacted: boolean;
 }
 
 /** Immutable context shared across event handlers within a single agent loop run. */
 export interface EventHandlerDeps {
-  readonly ctx: AgentLoopConversationContext;
+  readonly ctx: Conversation;
   readonly onEvent: (msg: ServerMessage) => void;
   readonly reqId: string;
   readonly isFirstMessage: boolean;
@@ -392,8 +372,6 @@ export function createEventHandlerState(): EventHandlerState {
     pendingPartialFlushPromise: undefined,
     currentMessageContent: [],
     lastPersistedContentSeq: undefined,
-    shouldInjectWorkspace: false,
-    reducerCompacted: false,
   };
 }
 
@@ -655,9 +633,7 @@ function computeToolUseStatusText(
   return `Running ${friendlyToolName(name)}`;
 }
 
-function resolveAssistantReplyTimestampTimezone(
-  ctx: AgentLoopConversationContext,
-): string {
+function resolveAssistantReplyTimestampTimezone(ctx: Conversation): string {
   const config = getConfig();
   return resolveTurnTimezoneContext({
     configuredUserTimeZone: config.ui?.userTimezone ?? null,
@@ -1659,6 +1635,8 @@ export async function handleMessageComplete(
       deps.ctx.conversationId,
       state.pendingNotifiedInferenceProfile,
     );
+    deps.ctx.lastNotifiedInferenceProfile =
+      state.pendingNotifiedInferenceProfile;
     state.pendingNotifiedInferenceProfile = null;
   }
 
@@ -2302,17 +2280,14 @@ export async function dispatchAgentEvent(
         // so re-injection re-applies onto the stripped history even when the
         // pipeline ran but did not compact. When it did compact, commit the
         // durable result (DB-record fields, Slack provenance, SSE) — which
-        // overwrites `ctx.messages` with the compacted history — and flip the
-        // per-turn re-injection guards the orchestrator reads. This runs
+        // overwrites `ctx.messages` with the compacted history. This runs
         // before the loop's `reinject` hook (the loop awaits this dispatch),
-        // so the guards are set in time. A failed durable commit re-throws
-        // below to abort the turn rather than re-injecting against
-        // half-applied state.
+        // so the committed history is in place in time. A failed durable
+        // commit re-throws below to abort the turn rather than re-injecting
+        // against half-applied state.
         deps.ctx.messages = event.basis;
         if (event.result.compacted) {
           await deps.applyCompaction(event.result, event.basis);
-          state.reducerCompacted = true;
-          state.shouldInjectWorkspace = true;
         }
         break;
       case "history_stripped":