feat(sessions): optional model for spawn tools and # model completion in composer

spawn_session and spawn_subsession accept an optional model parameter as
an exact provider/model-id (strict matching; unknown specs fail listing
available models; omitting it keeps the inherited model). The chat
composer opens a model completion menu on # and inserts a
#provider/model-id reference into the draft, which agents forward as
the model parameter.

.changeset/spawn-model-picker.md

@@ -0,0 +1,5 @@
+---
+"@jmfederico/pi-web": patch
+---
+
+Let agents pick a model when delegating work: `spawn_session` and `spawn_subsession` accept an optional `model` parameter as an exact `provider/model-id` (an unknown value is rejected; omitting it keeps the inherited model). In the chat composer, typing `#` opens a model completion menu that inserts a `#provider/model-id` reference into the draft, which agents forward as that parameter.

docs/config.html

@@ -103,6 +103,7 @@
             <a href="#catalog-refresh">Model catalog refresh</a>
             <a href="#session-tools">Session tools</a>
             <a href="#extension-dialogs">Extension dialogs</a>
+            <a href="#prompt-completions">Prompt completions</a>
             <a href="#completion-tools">Completion tools</a>
           </aside>
 
@@ -805,6 +806,15 @@
                 available when the child stops. Included output and transcripts follow a labeled marker and come last,
                 after PI WEB guidance.
               </p>
+              <p>
+                Both <code>spawn_session</code> and <code>spawn_subsession</code> accept an optional
+                <code>model</code> parameter, given as an exact <code>provider/model-id</code> such as
+                <code>anthropic/claude-sonnet-4-5</code>. When set, the new session starts on that model instead of
+                inheriting the dispatching session's model. The match is strict: an unknown or malformed value is
+                rejected with an error. A <code>#provider/model-id</code> reference in the prompt (see
+                <a href="#prompt-completions">Prompt completions</a>) is how users ask for a specific model; agents
+                forward that reference as this parameter.
+              </p>
               <p>
                 In <strong>Settings → Session daemon</strong>, these keys are saved on the selected machine. Restart the
                 session daemon on that machine after changing them.
@@ -879,6 +889,25 @@
             </section>
 
 
+            <section id="prompt-completions">
+              <h2>Prompt completions</h2>
+              <p>The chat composer opens completion menus on three trigger characters:</p>
+              <ul>
+                <li><code>/</code> at the very start of the draft completes session commands.</li>
+                <li>
+                  <code>@</code> completes file paths: <code>@</code> for tracked files, <code>@ </code> (at, then
+                  space) or <code>!@</code> for all files. Picking one inserts an <code>@path</code> reference into the
+                  draft, quoted automatically when the path contains spaces.
+                </li>
+                <li>
+                  <code>#</code> completes the models available to the session, filtered case-insensitively as you type
+                  (at most 12 entries). Picking one inserts a <code>#provider/model-id</code> reference into the draft,
+                  which tells agents the request should run on that model — for example as the <code>model</code>
+                  parameter of <code>spawn_session</code>.
+                </li>
+              </ul>
+            </section>
+
             <section id="completion-tools">
               <h2>Optional completion tools</h2>
               <p>

docs/config.md

@@ -273,6 +273,8 @@ A completion notice wakes an idle parent or queues behind in-flight work. Each n
 
 `list_subsessions`, `check_subsession`, and `read_subsession` never yield or change control flow. They are for deliberate inspection or recovery, not completion polling. While a child works, agent-facing `check_subsession` and `read_subsession` withhold partial output and direct the parent to continue independent work or yield at the join point. Output becomes available when the child stops. Included output and transcripts follow a labeled marker and come last, after PI WEB guidance.
 
+Both `spawn_session` and `spawn_subsession` accept an optional `model` parameter, given as an exact `provider/model-id` such as `anthropic/claude-sonnet-4-5`. When set, the new session starts on that model instead of inheriting the dispatching session's model. The match is strict: an unknown or malformed value is rejected with an error. A `#provider/model-id` reference in the prompt (see [Prompt completions](#prompt-completions)) is how users ask for a specific model; agents forward that reference as this parameter.
+
 In **Settings → Session daemon**, these keys are saved on the selected machine. Restart the session daemon on that machine after changing them.
 
 #### `askUser` and `ask_user`
@@ -331,6 +333,14 @@ Shortcut values are keyed by action id. Values are shortcut strings such as `mod
 
 Prefer Settings → Keyboard for editing shortcuts interactively.
 
+## Prompt completions
+
+The chat composer opens completion menus on three trigger characters:
+
+- `/` at the very start of the draft completes session commands.
+- `@` completes file paths: `@` for tracked files, `@ ` (at, then space) or `!@` for all files. Picking one inserts an `@path` reference into the draft, quoted automatically when the path contains spaces.
+- `#` completes the models available to the session, filtered case-insensitively as you type (at most 12 entries). Picking one inserts a `#provider/model-id` reference into the draft, which tells agents the request should run on that model — for example as the `model` parameter of `spawn_session`.
+
 ## Optional completion tools
 
 File and path `@` completions work without extra tools. If `fzf` is available on the PI WEB server's `PATH`, PI WEB uses it to improve completion filtering/ranking; otherwise it falls back to built-in ranking.

src/client/src/components/PromptEditor.modelCompletion.test.ts

@@ -0,0 +1,81 @@
+import { afterEach, describe, expect, it, vi } from "vitest";
+import { api, type SessionModel } from "../api";
+import { PromptEditor } from "./PromptEditor";
+
+const sonnet: SessionModel = { provider: "anthropic", id: "claude-sonnet-4-5", name: "Claude Sonnet 4.5" };
+const gpt: SessionModel = { provider: "openai", id: "gpt-5.2" };
+
+afterEach(() => {
+  vi.restoreAllMocks();
+});
+
+describe("PromptEditor model completions", () => {
+  it("requests models for the session and maps a # token to model completions", async () => {
+    const models = vi.spyOn(api, "models").mockResolvedValue({ models: [sonnet, gpt] });
+    const editor = new PromptEditor();
+    editor.sessionId = "session-1";
+    editor.cwd = "/repo";
+    editor.machineId = "remote-a";
+
+    await refreshCompletions(editor, "please use #cla");
+
+    expect(models).toHaveBeenCalledWith({ id: "session-1", cwd: "/repo" }, "remote-a");
+    expect(currentCompletions(editor)).toEqual([
+      { kind: "model", replaceFrom: 11, replaceTo: 15, insertText: "#anthropic/claude-sonnet-4-5", detail: "anthropic", description: "Claude Sonnet 4.5" },
+    ]);
+    expect(Reflect.get(editor, "selectedIndex")).toBe(0);
+  });
+
+  it("lists every model for a bare # token", async () => {
+    vi.spyOn(api, "models").mockResolvedValue({ models: [sonnet, gpt] });
+    const editor = new PromptEditor();
+    editor.sessionId = "session-1";
+    editor.cwd = "/repo";
+
+    await refreshCompletions(editor, "#");
+
+    expect(currentCompletions(editor)).toEqual([
+      { kind: "model", replaceFrom: 0, replaceTo: 1, insertText: "#anthropic/claude-sonnet-4-5", detail: "anthropic", description: "Claude Sonnet 4.5" },
+      { kind: "model", replaceFrom: 0, replaceTo: 1, insertText: "#openai/gpt-5.2", detail: "openai" },
+    ]);
+  });
+
+  it("clears completions when the models request fails", async () => {
+    vi.spyOn(api, "models").mockRejectedValue(new Error("models unavailable"));
+    const editor = new PromptEditor();
+    editor.sessionId = "session-1";
+    editor.cwd = "/repo";
+
+    await refreshCompletions(editor, "#cla");
+
+    expect(currentCompletions(editor)).toEqual([]);
+  });
+
+  it("does not request models without session context", async () => {
+    const models = vi.spyOn(api, "models").mockResolvedValue({ models: [sonnet] });
+    const editor = new PromptEditor();
+
+    await refreshCompletions(editor, "#cla");
+
+    expect(models).not.toHaveBeenCalled();
+    expect(currentCompletions(editor)).toEqual([]);
+  });
+});
+
+// refreshCompletions is private and driven by CodeMirror updates in production;
+// invoking it through Reflect mirrors the PromptEditor.draft.test.ts seam and
+// keeps the wiring test at the component boundary without a DOM harness.
+async function refreshCompletions(editor: PromptEditor, draft: string): Promise<void> {
+  Reflect.set(editor, "draft", draft);
+  const refresh: unknown = Reflect.get(editor, "refreshCompletions");
+  if (!isRefreshCompletions(refresh)) throw new Error("PromptEditor.refreshCompletions is not callable");
+  await refresh.call(editor);
+}
+
+function isRefreshCompletions(value: unknown): value is (this: PromptEditor) => Promise<void> {
+  return typeof value === "function";
+}
+
+function currentCompletions(editor: PromptEditor): unknown {
+  return Reflect.get(editor, "completions");
+}

src/client/src/components/PromptEditor.ts

@@ -5,12 +5,12 @@ import { EditorView, keymap, placeholder } from "@codemirror/view";
 import { defaultHighlightStyle, indentOnInput, indentUnit, syntaxHighlighting } from "@codemirror/language";
 import { LitElement, html, type PropertyValues } from "lit";
 import { customElement, property, query, state } from "lit/decorators.js";
-import { api, type FileSuggestion, type PromptAttachment, type SessionStatus, type SlashCommand } from "../api";
+import { api, type FileSuggestion, type PromptAttachment, type SessionModel, type SessionStatus, type SlashCommand } from "../api";
 import type { PromptAttachmentDelivery } from "../../../shared/apiTypes";
 import { capturePromptAttachments, effectivePromptAttachmentDelivery, isInlinePromptAttachment, promptAttachmentsCanUseInlineDelivery, type CapturedAttachment } from "../promptAttachmentCapture";
 import { inputModeForDraft, inputModesEqual, type InputMode } from "../inputModes";
 import { machineSessionKey } from "../machineKeys";
-import { detectPromptCompletionTrigger, fileCompletionInsertText, type PromptCompletionTrigger } from "../promptCompletions";
+import { detectPromptCompletionTrigger, fileCompletionInsertText, modelCompletionChoices, type PromptCompletionTrigger } from "../promptCompletions";
 import { clearDraft, loadDraft, saveDraft } from "../promptDraftStorage";
 import { loadAttachmentDelivery, saveAttachmentDelivery } from "../attachmentPreferences";
 import { createMobilePromptEnterMedia, readPromptEnterPreference, shouldSendPromptOnEnterShortcut, shouldUsePromptEnterShiftShortcut } from "../promptEnterBehavior";
@@ -283,7 +283,7 @@ export class PromptEditor extends LitElement {
             keyup: (event) => this.handleEditorKeyUp(event),
             blur: () => this.resetEditorModifierState(),
           }),
-          placeholder("Message pi... Use / for commands, @ for tracked files, @ space for all files"),
+          placeholder("Message pi... Use / for commands, @ for tracked files, @ space for all files, # for models"),
           this.editableCompartment.of(EditorView.editable.of(!this.disabled)),
           this.readOnlyCompartment.of(EditorState.readOnly.of(this.disabled)),
           EditorView.updateListener.of((update) => {
@@ -372,6 +372,15 @@ export class PromptEditor extends LitElement {
             ...(file.path.endsWith("/") && insertText.endsWith("\"") ? { cursorOffset: insertText.length - 1 } : {}),
           };
         });
+    } else if (trigger.kind === "model" && this.sessionId !== undefined && this.sessionId !== "" && this.cwd !== undefined && this.cwd !== "") {
+      const models = await api.models({ id: this.sessionId, cwd: this.cwd }, this.machineId).then((response) => response.models).catch(emptySessionModels);
+      if (version !== this.requestVersion) return;
+      this.completions = modelCompletionChoices(models, trigger.query).map((choice) => ({
+        kind: "model",
+        replaceFrom: trigger.from,
+        replaceTo: trigger.to,
+        ...choice,
+      }));
     }
   }
 
@@ -516,6 +525,10 @@ function emptyFileSuggestions(): FileSuggestion[] {
   return [];
 }
 
+function emptySessionModels(): SessionModel[] {
+  return [];
+}
+
 function filesFromDataTransfer(data: DataTransfer | null): File[] {
   if (data === null) return [];
   return Array.from(data.files);

src/client/src/components/shared.ts

@@ -72,7 +72,7 @@ export interface ChatLine {
 }
 
 export interface CompletionItem {
-  kind: "command" | "file";
+  kind: "command" | "file" | "model";
   replaceFrom: number;
   replaceTo: number;
   insertText: string;

src/client/src/inputModes.test.ts

@@ -31,4 +31,11 @@ describe("input mode helpers", () => {
     expect(inputModeForDraft("open !@\"vendor/file.ts")).toEqual({ kind: "file" });
     expect(inputModeForDraft("open \"src/main.ts")).toEqual({ kind: "normal" });
   });
+
+  it("collapses # model completion tokens to model mode", () => {
+    expect(inputModeForDraft("#anthropic/claude-opus")).toEqual({ kind: "model" });
+    expect(inputModeForDraft("use #gpt-5.2")).toEqual({ kind: "model" });
+    expect(inputModeForDraft("use #gpt-5.2 please")).toEqual({ kind: "normal" });
+    expect(inputModeForDraft("# ")).toEqual({ kind: "normal" });
+  });
 });

src/client/src/inputModes.ts

@@ -4,6 +4,7 @@ export type InputMode =
   | { kind: "normal" }
   | { kind: "command" }
   | { kind: "file" }
+  | { kind: "model" }
   | { kind: "shell"; excludeFromContext: boolean };
 
 export function inputModeForDraft(draft: string): InputMode {
@@ -11,7 +12,9 @@ export function inputModeForDraft(draft: string): InputMode {
   if (trimmed.startsWith("!@")) return { kind: "file" };
   if (trimmed.startsWith("!")) return { kind: "shell", excludeFromContext: trimmed.startsWith("!!") };
   if (currentToken(draft).startsWith("/")) return { kind: "command" };
-  if (detectPromptCompletionTrigger(draft)?.kind === "file") return { kind: "file" };
+  const trigger = detectPromptCompletionTrigger(draft);
+  if (trigger?.kind === "file") return { kind: "file" };
+  if (trigger?.kind === "model") return { kind: "model" };
   return { kind: "normal" };
 }
 

src/client/src/promptCompletions.test.ts

@@ -1,5 +1,6 @@
 import { describe, expect, it } from "vitest";
-import { detectPromptCompletionTrigger, fileCompletionInsertText } from "./promptCompletions";
+import type { SessionModel } from "../../shared/apiTypes";
+import { detectPromptCompletionTrigger, fileCompletionInsertText, modelCompletionChoices } from "./promptCompletions";
 
 describe("detectPromptCompletionTrigger", () => {
   it("keeps all-file suggestions active when an @ space query contains spaces", () => {
@@ -54,6 +55,73 @@ describe("detectPromptCompletionTrigger", () => {
     });
     expect(detectPromptCompletionTrigger("/model")).toEqual({ kind: "command", query: "model", from: 0, to: 6 });
   });
+
+  it("detects model queries for tokens starting with #", () => {
+    expect(detectPromptCompletionTrigger("#")).toEqual({ kind: "model", query: "", from: 0, to: 1 });
+    expect(detectPromptCompletionTrigger("use #anthropic/claude-opus")).toEqual({ kind: "model", query: "anthropic/claude-opus", from: 4, to: 26 });
+  });
+
+  it("detects model queries at any cursor position but not inside other tokens or quotes", () => {
+    expect(detectPromptCompletionTrigger("use #claude now", 10)).toEqual({ kind: "model", query: "claud", from: 4, to: 10 });
+    expect(detectPromptCompletionTrigger("say hello#world")).toBeUndefined();
+    expect(detectPromptCompletionTrigger('say "#claude')).toBeUndefined();
+  });
+
+  it("shows model completion for a markdown header only on the bare # keystroke", () => {
+    expect(detectPromptCompletionTrigger("#")).toEqual({ kind: "model", query: "", from: 0, to: 1 });
+    expect(detectPromptCompletionTrigger("# ")).toBeUndefined();
+    expect(detectPromptCompletionTrigger("# Title")).toBeUndefined();
+  });
+});
+
+describe("modelCompletionChoices", () => {
+  const models: SessionModel[] = [
+    { provider: "anthropic", id: "claude-opus-4-5", name: "Claude Opus 4.5" },
+    { provider: "anthropic", id: "claude-sonnet-4-5", name: "Claude Sonnet 4.5" },
+    { provider: "openai", id: "gpt-5.2", name: "GPT-5.2" },
+    { provider: "google", id: "gemini-3-pro" },
+  ];
+
+  it("lists all models for an empty query, mapped to #provider/id insert texts", () => {
+    expect(modelCompletionChoices(models, "")).toEqual([
+      { insertText: "#anthropic/claude-opus-4-5", detail: "anthropic", description: "Claude Opus 4.5" },
+      { insertText: "#anthropic/claude-sonnet-4-5", detail: "anthropic", description: "Claude Sonnet 4.5" },
+      { insertText: "#openai/gpt-5.2", detail: "openai", description: "GPT-5.2" },
+      { insertText: "#google/gemini-3-pro", detail: "google" },
+    ]);
+  });
+
+  it("filters case-insensitively across provider/id, id, and display name", () => {
+    expect(modelCompletionChoices(models, "OPUS-4")).toEqual([
+      { insertText: "#anthropic/claude-opus-4-5", detail: "anthropic", description: "Claude Opus 4.5" },
+    ]);
+    expect(modelCompletionChoices(models, "openai")).toEqual([
+      { insertText: "#openai/gpt-5.2", detail: "openai", description: "GPT-5.2" },
+    ]);
+    expect(modelCompletionChoices(models, "sonnet 4.5")).toEqual([
+      { insertText: "#anthropic/claude-sonnet-4-5", detail: "anthropic", description: "Claude Sonnet 4.5" },
+    ]);
+  });
+
+  it("omits the description when the display name matches the id", () => {
+    expect(modelCompletionChoices([{ provider: "ollama", id: "qwen3", name: "qwen3" }], "")).toEqual([
+      { insertText: "#ollama/qwen3", detail: "ollama" },
+    ]);
+  });
+
+  it("skips models without a provider or id since they cannot form a #provider/id reference", () => {
+    expect(modelCompletionChoices([{ id: "orphan" }, { provider: "ghost" }, { provider: "openai", id: "gpt-5.2" }], "")).toEqual([
+      { insertText: "#openai/gpt-5.2", detail: "openai" },
+    ]);
+  });
+
+  it("caps the list at 12 models, preserving server order", () => {
+    const many = Array.from({ length: 20 }, (_, index) => ({ provider: "p", id: `m${String(index)}` }));
+    const choices = modelCompletionChoices(many, "");
+    expect(choices).toHaveLength(12);
+    expect(choices[0]?.insertText).toBe("#p/m0");
+    expect(choices[11]?.insertText).toBe("#p/m11");
+  });
 });
 
 describe("fileCompletionInsertText", () => {

src/client/src/promptCompletions.ts

@@ -1,6 +1,9 @@
+import type { SessionModel } from "../../shared/apiTypes";
+
 export type PromptCompletionTrigger =
   | { kind: "command"; query: string; from: number; to: number }
-  | { kind: "file"; query: string; from: number; to: number; fileScope?: "tracked" | "all" | undefined; allPrefix?: "@ " | "!@" | undefined; quoted?: boolean };
+  | { kind: "file"; query: string; from: number; to: number; fileScope?: "tracked" | "all" | undefined; allPrefix?: "@ " | "!@" | undefined; quoted?: boolean }
+  | { kind: "model"; query: string; from: number; to: number };
 
 export function detectPromptCompletionTrigger(draft: string, cursor = draft.length): PromptCompletionTrigger | undefined {
   const beforeCursor = draft.slice(0, cursor);
@@ -17,9 +20,46 @@ export function detectPromptCompletionTrigger(draft: string, cursor = draft.leng
   if (token.startsWith("/") && tokenStart === 0) return { kind: "command", query: token.slice(1), from: tokenStart, to: cursor };
   if (token.startsWith("!@")) return { kind: "file", query: token.slice(2), from: tokenStart, to: cursor, fileScope: "all", allPrefix: "!@" };
   if (token.startsWith("@")) return { kind: "file", query: token.slice(1), from: tokenStart, to: cursor, fileScope: "tracked" };
+  if (token.startsWith("#")) return { kind: "model", query: token.slice(1), from: tokenStart, to: cursor };
   return undefined;
 }
 
+export interface ModelCompletionChoice {
+  insertText: string;
+  detail: string;
+  description?: string;
+}
+
+const MODEL_COMPLETION_LIMIT = 12;
+
+export function modelCompletionChoices(models: readonly SessionModel[], query: string): ModelCompletionChoice[] {
+  const needle = query.toLowerCase();
+  const choices: ModelCompletionChoice[] = [];
+  for (const model of models) {
+    // A completion must produce a strict provider/model-id reference, so models
+    // missing either half of the identity can never be inserted.
+    if (!hasQualifiedModelId(model)) continue;
+    if (!modelMatchesQuery(model, needle)) continue;
+    choices.push({
+      insertText: `#${model.provider}/${model.id}`,
+      detail: model.provider,
+      ...(model.name !== undefined && model.name !== "" && model.name !== model.id ? { description: model.name } : {}),
+    });
+    if (choices.length >= MODEL_COMPLETION_LIMIT) break;
+  }
+  return choices;
+}
+
+function hasQualifiedModelId(model: SessionModel): model is SessionModel & { provider: string; id: string } {
+  return typeof model.provider === "string" && model.provider !== "" && typeof model.id === "string" && model.id !== "";
+}
+
+function modelMatchesQuery(model: SessionModel & { provider: string; id: string }, needle: string): boolean {
+  return `${model.provider}/${model.id}`.toLowerCase().includes(needle)
+    || model.id.toLowerCase().includes(needle)
+    || (model.name?.toLowerCase().includes(needle) ?? false);
+}
+
 export function fileCompletionInsertText(path: string, quoted: boolean, allPrefix?: "@ " | "!@"): string {
   const prefix = allPrefix ?? "@";
   if (!quoted && !path.includes(" ")) return `${prefix}${path}`;

src/server/sessions/piSessionService.delegationTools.test.ts

@@ -1,6 +1,7 @@
 import { mkdtemp, rm, writeFile } from "node:fs/promises";
 import { tmpdir } from "node:os";
 import { join } from "node:path";
+import type { ExtensionContext } from "@earendil-works/pi-coding-agent";
 import { afterEach, describe, expect, it, vi } from "vitest";
 import { createPiWebCustomToolDefinitions, sessionAllowsDelegationTools, type PiSessionManager } from "./piSessionService.js";
 import type { SubsessionToolDeps } from "./spawnSubsessionTool.js";
@@ -14,13 +15,14 @@ afterEach(async () => {
 
 function delegationDeps() {
   const spawn = vi.fn(() => Promise.resolve({ sessionId: "independent-1", cwd: "/workspace" }));
+  const subsessionSpawn = vi.fn(() => Promise.resolve({ sessionId: "child-1", cwd: "/workspace" }));
   const subsessions: SubsessionToolDeps = {
-    spawn: vi.fn(() => Promise.resolve({ sessionId: "child-1", cwd: "/workspace" })),
+    spawn: subsessionSpawn,
     list: vi.fn(() => Promise.resolve([])),
     check: vi.fn(() => Promise.resolve({ sessionId: "child-1", cwd: "/workspace", status: "idle" as const, finalText: "", messageCount: 0 })),
     read: vi.fn(() => Promise.resolve({ sessionId: "child-1", cwd: "/workspace", status: "idle" as const, entries: [], total: 0, matched: 0, start: 0, hasMore: false })),
   };
-  return { spawn, subsessions };
+  return { spawn, subsessions, subsessionSpawn };
 }
 
 function toolNames(definitions: ReturnType<typeof createPiWebCustomToolDefinitions>): string[] {
@@ -35,6 +37,20 @@ function manager(id: string, file: string | undefined, entries: readonly unknown
   });
 }
 
+const dispatchModel = { provider: "anthropic", id: "claude-sonnet" };
+
+function ctxFor(sessionId: string, sessionFile: string | undefined, model?: unknown): ExtensionContext {
+  // The delegation tools only read sessionManager and model from the context.
+  // eslint-disable-next-line @typescript-eslint/consistent-type-assertions -- test stub with the minimal surface the tools read.
+  return { sessionManager: manager(sessionId, sessionFile), ...(model === undefined ? {} : { model }) } as unknown as ExtensionContext;
+}
+
+function findTool(definitions: ReturnType<typeof createPiWebCustomToolDefinitions>, name: string) {
+  const tool = definitions.find((definition) => definition.name === name);
+  if (tool === undefined) throw new Error(`missing tool ${name}`);
+  return tool;
+}
+
 describe("delegation tool capability boundary", () => {
   it("provides every globally enabled delegation tool to unrestricted sessions", () => {
     const { spawn, subsessions } = delegationDeps();
@@ -63,6 +79,39 @@ describe("delegation tool capability boundary", () => {
     expect(toolNames(createPiWebCustomToolDefinitions("/workspace", false, spawn, subsessions))).toEqual(["edit"]);
   });
 
+  it("wires the dispatching session identity, inherited model, and model spec into spawn_session", async () => {
+    const { spawn, subsessions } = delegationDeps();
+    const spawnTool = findTool(createPiWebCustomToolDefinitions("/workspace", true, spawn, subsessions), "spawn_session");
+
+    await spawnTool.execute("call-1", { prompt: "go", model: "openai/gpt-5" }, undefined, undefined, ctxFor("spawner-7", "/sessions/spawner-7.jsonl", dispatchModel));
+
+    expect(spawn).toHaveBeenCalledWith({
+      spawningCwd: "/workspace",
+      spawningSessionId: "spawner-7",
+      prompt: "go",
+      cwd: undefined,
+      model: dispatchModel,
+      modelSpec: "openai/gpt-5",
+    });
+  });
+
+  it("wires the parent identity, inherited model, and model spec into spawn_subsession", async () => {
+    const { spawn, subsessions, subsessionSpawn } = delegationDeps();
+    const spawnTool = findTool(createPiWebCustomToolDefinitions("/workspace", true, spawn, subsessions), "spawn_subsession");
+
+    await spawnTool.execute("call-2", { prompt: "go", model: "openai/gpt-5" }, undefined, undefined, ctxFor("parent-9", "/sessions/parent-9.jsonl", dispatchModel));
+
+    expect(subsessionSpawn).toHaveBeenCalledWith({
+      spawningCwd: "/workspace",
+      parentSessionId: "parent-9",
+      parentSessionFile: "/sessions/parent-9.jsonl",
+      prompt: "go",
+      cwd: undefined,
+      model: dispatchModel,
+      modelSpec: "openai/gpt-5",
+    });
+  });
+
   it.each(["human-created", "spawn_session-created"])("allows delegation for a %s session without tracked-child provenance", async () => {
     const sessionManager = manager("session-1", undefined);
     const open = vi.fn(() => { throw new Error("no parent session should be opened"); });

src/server/sessions/piSessionService.spawnSession.test.ts

@@ -4,6 +4,7 @@ import type { SpawnTargetDecision } from "./spawnTargetResolver.js";
 import { CapturingSessionEventHub, fakeRuntime, runtimeCreator, sessionGateway, testModel, testModelRuntime, type RuntimeCreator } from "./piSessionService.testSupport.js";
 
 const TEST_AGENT_DIR = "/tmp/pi-web-test-agent";
+const TEST_MODEL_SPEC = "anthropic/claude-sonnet-4-5-20250929";
 
 describe("PiSessionService", () => {
   describe("spawnSession", () => {
@@ -25,7 +26,7 @@ describe("PiSessionService", () => {
     it("starts a session at the resolved target, delivers the prompt, and logs the spawn", async () => {
       const { fake, service, log } = spawnService({ allowed: true, cwd: "/workspace-feature" });
 
-      const result = await service.spawnSession({ spawningCwd: "/workspace", prompt: "continue the plan", cwd: "/workspace-feature" });
+      const result = await service.spawnSession({ spawningCwd: "/workspace", spawningSessionId: "spawner-1", prompt: "continue the plan", cwd: "/workspace-feature" });
 
       expect(result).toEqual({ sessionId: "spawned-1", cwd: "/workspace-feature" });
       expect(fake.calls.prompt).toEqual([{ text: "continue the plan", options: undefined }]);
@@ -53,17 +54,34 @@ describe("PiSessionService", () => {
         heartbeatIntervalMs: 60_000,
       });
 
-      await service.spawnSession({ spawningCwd: "/workspace", prompt: "continue", cwd: "/workspace-feature", model });
+      await service.spawnSession({ spawningCwd: "/workspace", spawningSessionId: "spawner-1", prompt: "continue", cwd: "/workspace-feature", model });
 
       expect(initialModel).toBe(model);
       expect(delegationToolsEnabled).toBe(true);
       await service.dispose();
     });
 
+    it("names the spawned session's model in the result", async () => {
+      const spawned = fakeRuntime("spawned-1", { sessionFile: "/tmp/spawned-1.jsonl", model: testModel() });
+      const service = new PiSessionService(new CapturingSessionEventHub(), {
+        agentDir: TEST_AGENT_DIR,
+      modelRuntime: testModelRuntime,
+        createAgentRuntime: runtimeCreator(spawned.runtime),
+        sessionManager: sessionGateway([]),
+        spawnTargets: { resolveSpawnTarget: () => Promise.resolve({ allowed: true, cwd: "/workspace-feature" }) },
+        heartbeatIntervalMs: 60_000,
+      });
+
+      const result = await service.spawnSession({ spawningCwd: "/workspace", spawningSessionId: "spawner-1", prompt: "continue", cwd: "/workspace-feature" });
+
+      expect(result).toEqual({ sessionId: "spawned-1", cwd: "/workspace-feature", model: TEST_MODEL_SPEC });
+      await service.dispose();
+    });
+
     it("rejects an out-of-project target without starting a session", async () => {
       const { fake, service } = spawnService({ allowed: false, reason: "out-of-project", allowedCwds: ["/workspace"] });
 
-      await expect(service.spawnSession({ spawningCwd: "/workspace", prompt: "go", cwd: "/elsewhere" }))
+      await expect(service.spawnSession({ spawningCwd: "/workspace", spawningSessionId: "spawner-1", prompt: "go", cwd: "/elsewhere" }))
         .rejects.toThrow("cwd must be a workspace of this project. Allowed: /workspace");
       expect(fake.calls.prompt).toEqual([]);
       expect(service.activeCount()).toBe(0);
@@ -73,7 +91,7 @@ describe("PiSessionService", () => {
     it("rejects when the spawning session is not in a registered project", async () => {
       const { service } = spawnService({ allowed: false, reason: "not-registered" });
 
-      await expect(service.spawnSession({ spawningCwd: "/workspace", prompt: "go", cwd: undefined }))
+      await expect(service.spawnSession({ spawningCwd: "/workspace", spawningSessionId: "spawner-1", prompt: "go", cwd: undefined }))
         .rejects.toThrow("Spawning session is not in a registered project");
       await service.dispose();
     });
@@ -88,9 +106,120 @@ describe("PiSessionService", () => {
         heartbeatIntervalMs: 60_000,
       });
 
-      await expect(service.spawnSession({ spawningCwd: "/workspace", prompt: "go", cwd: undefined }))
+      await expect(service.spawnSession({ spawningCwd: "/workspace", spawningSessionId: "spawner-1", prompt: "go", cwd: undefined }))
         .rejects.toThrow("Spawning sessions is disabled");
       await service.dispose();
     });
+
+    describe("model spec resolution", () => {
+      /**
+       * Harness: the spawner comes online via `service.start`, then the spawn
+       * creates the next queued runtime. `initialModels` records every
+       * creation-time model so tests can see exactly what the spawned session
+       * was started with.
+       */
+      function specService(spawnerPatch: Parameters<typeof fakeRuntime>[1] = {}) {
+        const spawner = fakeRuntime("spawner-1", { sessionFile: "/tmp/spawner-1.jsonl", ...spawnerPatch });
+        const spawned = fakeRuntime("spawned-2", { sessionFile: "/tmp/spawned-2.jsonl", model: testModel() });
+        const initialModels: PiAgentSession["model"][] = [];
+        const runtimes = [spawner.runtime, spawned.runtime];
+        let index = 0;
+        const createAgentRuntime: RuntimeCreator = async (_createRuntime, options) => {
+          await Promise.resolve();
+          initialModels.push(options.initialModel);
+          const runtime = runtimes[index] ?? spawned.runtime;
+          index += 1;
+          return runtime;
+        };
+        const service = new PiSessionService(new CapturingSessionEventHub(), {
+          agentDir: TEST_AGENT_DIR,
+      modelRuntime: testModelRuntime,
+          createAgentRuntime,
+          sessionManager: sessionGateway([]),
+          spawnTargets: { resolveSpawnTarget: () => Promise.resolve({ allowed: true, cwd: "/workspace-feature" }) },
+          heartbeatIntervalMs: 60_000,
+        });
+        return { service, spawner, spawned, initialModels };
+      }
+
+      it("resolves the spec against the spawning session's scoped models and names it in the result", async () => {
+        const scoped = testModel();
+        const { service, initialModels } = specService({ scopedModels: [{ model: scoped }] });
+        await service.start("/workspace");
+
+        const result = await service.spawnSession({ spawningCwd: "/workspace", spawningSessionId: "spawner-1", prompt: "go", cwd: "/workspace-feature", modelSpec: TEST_MODEL_SPEC });
+
+        expect(initialModels).toHaveLength(2);
+        expect(initialModels[0]).toBeUndefined();
+        expect(initialModels[1]).toBe(scoped);
+        expect(result).toEqual({ sessionId: "spawned-2", cwd: "/workspace-feature", model: TEST_MODEL_SPEC });
+        await service.dispose();
+      });
+
+      it("falls back to a direct runtime lookup when the spec is not among the available candidates", async () => {
+        // The shared test runtime has no configured auth, so its available
+        // snapshot is empty; only the getModel fallback can resolve the spec.
+        const { service, initialModels } = specService();
+        await service.start("/workspace");
+
+        const result = await service.spawnSession({ spawningCwd: "/workspace", spawningSessionId: "spawner-1", prompt: "go", cwd: "/workspace-feature", modelSpec: TEST_MODEL_SPEC });
+
+        expect(initialModels[1]).toMatchObject({ provider: "anthropic", id: "claude-sonnet-4-5-20250929" });
+        expect(result).toEqual({ sessionId: "spawned-2", cwd: "/workspace-feature", model: TEST_MODEL_SPEC });
+        await service.dispose();
+      });
+
+      it.each(["no-slash", "anthropic/", "/id"])("rejects the malformed spec %s without starting a session", async (modelSpec) => {
+        const { service, spawned, initialModels } = specService({ scopedModels: [{ model: testModel() }] });
+        await service.start("/workspace");
+
+        await expect(service.spawnSession({ spawningCwd: "/workspace", spawningSessionId: "spawner-1", prompt: "go", cwd: "/workspace-feature", modelSpec }))
+          .rejects.toThrow(`Unknown model "${modelSpec}". Pass an exact "provider/model-id".`);
+        expect(initialModels).toEqual([undefined]);
+        expect(spawned.calls.prompt).toEqual([]);
+        expect(service.activeCount()).toBe(1);
+        await service.dispose();
+      });
+
+      it("rejects an unknown spec without starting a session", async () => {
+        const { service, spawned, initialModels } = specService({ scopedModels: [{ model: testModel() }] });
+        await service.start("/workspace");
+
+        await expect(service.spawnSession({ spawningCwd: "/workspace", spawningSessionId: "spawner-1", prompt: "go", cwd: "/workspace-feature", modelSpec: "anthropic/does-not-exist" }))
+          .rejects.toThrow('Unknown model "anthropic/does-not-exist". Pass an exact "provider/model-id".');
+        expect(initialModels).toEqual([undefined]);
+        expect(spawned.calls.prompt).toEqual([]);
+        expect(service.activeCount()).toBe(1);
+        await service.dispose();
+      });
+
+      it("rejects an unknown spec even when the spawning session has no available models", async () => {
+        const { service } = specService();
+        await service.start("/workspace");
+
+        await expect(service.spawnSession({ spawningCwd: "/workspace", spawningSessionId: "spawner-1", prompt: "go", cwd: "/workspace-feature", modelSpec: "ghost/model" }))
+          .rejects.toThrow('Unknown model "ghost/model". Pass an exact "provider/model-id".');
+        await service.dispose();
+      });
+
+      it("does not look up the spawning session when no model spec is given", async () => {
+        const spawned = fakeRuntime("spawned-2", { sessionFile: "/tmp/spawned-2.jsonl", model: testModel() });
+        const service = new PiSessionService(new CapturingSessionEventHub(), {
+          agentDir: TEST_AGENT_DIR,
+      modelRuntime: testModelRuntime,
+          createAgentRuntime: runtimeCreator(spawned.runtime),
+          sessionManager: sessionGateway([]),
+          spawnTargets: { resolveSpawnTarget: () => Promise.resolve({ allowed: true, cwd: "/workspace-feature" }) },
+          heartbeatIntervalMs: 60_000,
+        });
+
+        // "ghost" is not a resolvable session, and the default path must not care.
+        const result = await service.spawnSession({ spawningCwd: "/workspace", spawningSessionId: "ghost", prompt: "go", cwd: "/workspace-feature" });
+
+        expect(result).toEqual({ sessionId: "spawned-2", cwd: "/workspace-feature", model: TEST_MODEL_SPEC });
+        expect(spawned.calls.prompt).toEqual([{ text: "go", options: undefined }]);
+        await service.dispose();
+      });
+    });
   });
 });

src/server/sessions/piSessionService.spawnSubsession.test.ts

@@ -98,6 +98,38 @@ describe("PiSessionService", () => {
       await service.dispose();
     });
 
+    it("resolves a model spec against the parent's models and names it in the result", async () => {
+      const scoped = testModel();
+      const parent = fakeRuntime("parent-1", { sessionFile: "/tmp/parent-1.jsonl", scopedModels: [{ model: scoped }] });
+      const child = fakeRuntime("child-1", { sessionFile: "/tmp/child-1.jsonl", sessionManager: fakeSessionManager("/workspace-feature"), model: scoped });
+      const initialModels: PiAgentSession["model"][] = [];
+      const runtimes = [parent.runtime, child.runtime];
+      let index = 0;
+      const createAgentRuntime: RuntimeCreator = async (_createRuntime, options) => {
+        await Promise.resolve();
+        initialModels.push(options.initialModel);
+        const runtime = runtimes[index] ?? child.runtime;
+        index += 1;
+        return runtime;
+      };
+      const service = new PiSessionService(new CapturingSessionEventHub(), {
+        agentDir: TEST_AGENT_DIR,
+      modelRuntime: testModelRuntime,
+        createAgentRuntime,
+        sessionManager: sessionGateway([]),
+        archiveStore: emptyArchiveStore(),
+        spawnTargets: { resolveSpawnTarget: () => Promise.resolve({ allowed: true, cwd: "/workspace-feature" }) },
+        heartbeatIntervalMs: 60_000,
+      });
+
+      await service.start("/workspace");
+      const result = await service.spawnSubsession({ spawningCwd: "/workspace", parentSessionId: "parent-1", parentSessionFile: "/tmp/parent-1.jsonl", prompt: "do the slice", cwd: "/workspace-feature", modelSpec: "anthropic/claude-sonnet-4-5-20250929" });
+
+      expect(initialModels[1]).toBe(scoped);
+      expect(result).toEqual({ sessionId: "child-1", cwd: "/workspace-feature", model: "anthropic/claude-sonnet-4-5-20250929" });
+      await service.dispose();
+    });
+
     it("persists tracked child links in the parent and child sessions", async () => {
       const parentPersisted: { customType: string; data?: unknown }[] = [];
       const childPersisted: { customType: string; data?: unknown }[] = [];

src/server/sessions/piSessionService.ts

@@ -120,6 +120,30 @@ function spawnTargetError(decision: Extract<SpawnTargetDecision, { allowed: fals
   return new Error(`cwd must be a workspace of this project. Allowed: ${decision.allowedCwds.join(", ")}`);
 }
 
+function modelSpecOf(model: { provider: string; id: string }): string {
+  return `${model.provider}/${model.id}`;
+}
+
+/**
+ * Parse a strict `provider/model-id` spec: split on the first `/` (model ids
+ * may themselves contain `/`) and require both parts to be non-empty.
+ */
+function parseModelSpec(spec: string): { provider: string; modelId: string } | undefined {
+  const slash = spec.indexOf("/");
+  if (slash <= 0 || slash === spec.length - 1) return undefined;
+  return { provider: spec.slice(0, slash), modelId: spec.slice(slash + 1) };
+}
+
+/**
+ * Error for a spawn-tool model spec that matched nothing. States the facts —
+ * the bad spec and the required format — with deliberately no model list
+ * (a list would invite guesses). The agent loop turns the throw into an
+ * error tool result; how to recover is the agent's call.
+ */
+function unknownSpawnModelError(modelSpec: string): Error {
+  return new Error(`Unknown model "${modelSpec}". Pass an exact "provider/model-id".`);
+}
+
 function authLossWarningKey(sessionId: string, provider: string, modelId: string): string {
   return `${sessionId}:${provider}/${modelId}`;
 }
@@ -1178,13 +1202,23 @@ export class PiSessionService implements SessionRouteService {
     if (this.spawnTargets === undefined) throw new Error("Spawning sessions is disabled");
     const decision = await this.spawnTargets.resolveSpawnTarget(input.spawningCwd, input.cwd);
     if (!decision.allowed) throw spawnTargetError(decision);
-    const created = await this.start(decision.cwd, input.model === undefined ? {} : { initialModel: input.model });
+    // A model spec overrides the inherited model. Only a spec triggers a
+    // spawning-session lookup; the default path must not depend on it.
+    const model = input.modelSpec === undefined
+      ? input.model
+      : await this.resolveSpawnModel(input.spawningSessionId, input.modelSpec);
+    const created = await this.start(decision.cwd, model === undefined ? {} : { initialModel: model });
+    const modelUsed = this.active.get(created.id)?.runtime.session.model;
     await this.prompt(created.id, input.prompt);
     this.logger.info(
       { spawningCwd: input.spawningCwd, sessionId: created.id, cwd: decision.cwd, promptLength: input.prompt.length },
       "spawn_session started a new session",
     );
-    return { sessionId: created.id, cwd: decision.cwd };
+    return {
+      sessionId: created.id,
+      cwd: decision.cwd,
+      ...(modelUsed === undefined ? {} : { model: modelSpecOf(modelUsed) }),
+    };
   }
 
   /**
@@ -1197,11 +1231,17 @@ export class PiSessionService implements SessionRouteService {
     if (this.spawnTargets === undefined) throw new Error("Spawning sessions is disabled");
     const decision = await this.spawnTargets.resolveSpawnTarget(input.spawningCwd, input.cwd);
     if (!decision.allowed) throw spawnTargetError(decision);
+    // A model spec overrides the inherited model and is resolved against the
+    // parent's model runtime; only a spec triggers that lookup.
+    const model = input.modelSpec === undefined
+      ? input.model
+      : await this.resolveSpawnModel(input.parentSessionId, input.modelSpec);
     const created = await this.startSession(decision.cwd, {
       ...(input.parentSessionFile === undefined ? {} : { parentSession: input.parentSessionFile }),
-      ...(input.model === undefined ? {} : { initialModel: input.model }),
+      ...(model === undefined ? {} : { initialModel: model }),
       creationProvenance: "tracked-subsession",
     });
+    const modelUsed = this.active.get(created.id)?.runtime.session.model;
     const parentSessionFile = nonEmptyString(input.parentSessionFile);
     const link: TrackedSubsessionLink = {
       parentSessionId: input.parentSessionId,
@@ -1218,7 +1258,42 @@ export class PiSessionService implements SessionRouteService {
       { parentSessionId: input.parentSessionId, sessionId: created.id, cwd: decision.cwd, promptLength: input.prompt.length },
       "spawn_subsession started a tracked child session",
     );
-    return { sessionId: created.id, cwd: decision.cwd };
+    return {
+      sessionId: created.id,
+      cwd: decision.cwd,
+      ...(modelUsed === undefined ? {} : { model: modelSpecOf(modelUsed) }),
+    };
+  }
+
+  /**
+   * The models a session may pick from: its scoped set when model-scoped,
+   * otherwise the runtime's available snapshot. Refreshes the runtime catalog
+   * first so callers see newly configured providers and models.
+   */
+  private async sessionModelCandidates(session: PiAgentSession): Promise<readonly AgentModel[]> {
+    await session.modelRuntime.refresh();
+    return session.scopedModels.length > 0
+      ? session.scopedModels.map((scoped) => scoped.model)
+      : session.modelRuntime.getAvailableSnapshot();
+  }
+
+  /**
+   * Resolve a strict `provider/model-id` spec from a spawn tool against the
+   * *spawning* session's model runtime, using the same candidates
+   * {@link setModel} offers plus a direct runtime lookup as fallback. Unknown
+   * or malformed specs throw; the agent loop turns that into an error tool
+   * result the spawning agent can retry from.
+   */
+  private async resolveSpawnModel(spawningSessionId: string, modelSpec: string): Promise<AgentModel> {
+    const session = await this.getOrOpen(spawningSessionId);
+    const parsed = parseModelSpec(modelSpec);
+    const candidates = await this.sessionModelCandidates(session);
+    const model = parsed === undefined
+      ? undefined
+      : candidates.find((candidate) => candidate.provider === parsed.provider && candidate.id === parsed.modelId)
+        ?? session.modelRuntime.getModel(parsed.provider, parsed.modelId);
+    if (model === undefined) throw unknownSpawnModelError(modelSpec);
+    return model;
   }
 
   /**
@@ -1822,10 +1897,7 @@ export class PiSessionService implements SessionRouteService {
 
   async availableModels(ref: PiSessionLookup): Promise<ClientSessionModel[]> {
     const session = await this.getOrOpen(ref);
-    await session.modelRuntime.refresh();
-    const models = session.scopedModels.length > 0
-      ? session.scopedModels.map((scoped) => scoped.model)
-      : session.modelRuntime.getAvailableSnapshot();
+    const models = await this.sessionModelCandidates(session);
     return models.map(modelToClientModel);
   }
 
@@ -1833,11 +1905,8 @@ export class PiSessionService implements SessionRouteService {
     await this.assertWritable(ref);
     const session = await this.getOrOpen(ref);
     this.assertTreeNavigationInactive(session, "change models");
-    await session.modelRuntime.refresh();
+    const candidates = await this.sessionModelCandidates(session);
     this.assertTreeNavigationInactive(session, "change models");
-    const candidates = session.scopedModels.length > 0
-      ? session.scopedModels.map((scoped) => scoped.model)
-      : session.modelRuntime.getAvailableSnapshot();
     const model = candidates.find((candidate) => candidate.provider === provider && candidate.id === modelId)
       ?? session.modelRuntime.getModel(provider, modelId);
     if (model === undefined) throw new Error(`Model not found: ${provider}/${modelId}`);

src/server/sessions/spawnSessionTool.test.ts

@@ -2,20 +2,23 @@ import type { ExtensionContext } from "@earendil-works/pi-coding-agent";
 import { describe, expect, it, vi } from "vitest";
 import { createSpawnSessionToolDefinition } from "./spawnSessionTool.js";
 
-// eslint-disable-next-line @typescript-eslint/consistent-type-assertions -- test stub with the minimal surface the tool reads.
-const ctx = {} as ExtensionContext;
 const dispatchModel = { provider: "anthropic", id: "claude-sonnet" };
+
+function ctxFor(sessionId: string, model?: unknown): ExtensionContext {
+  const sessionManager = { getSessionId: () => sessionId };
+  // The spawn tool only reads sessionManager.getSessionId and model.
   // eslint-disable-next-line @typescript-eslint/consistent-type-assertions -- test stub with the minimal surface the tool reads.
-const ctxWithModel = { model: dispatchModel } as ExtensionContext;
+  return { sessionManager, ...(model === undefined ? {} : { model }) } as unknown as ExtensionContext;
+}
 
 describe("createSpawnSessionToolDefinition", () => {
-  it("passes the spawning cwd, explicit cwd, dispatching model, and prompt to spawn callback", async () => {
+  it("passes the spawning identity, explicit cwd, dispatching model, and prompt to spawn callback", async () => {
     const spawn = vi.fn(() => Promise.resolve({ sessionId: "new-1", cwd: "/repos/a-feature" }));
     const tool = createSpawnSessionToolDefinition("/repos/a", { spawn });
 
-    const result = await tool.execute("call-1", { prompt: "do the thing", cwd: "/repos/a-feature" }, undefined, undefined, ctxWithModel);
+    const result = await tool.execute("call-1", { prompt: "do the thing", cwd: "/repos/a-feature" }, undefined, undefined, ctxFor("spawner-1", dispatchModel));
 
-    expect(spawn).toHaveBeenCalledWith({ spawningCwd: "/repos/a", prompt: "do the thing", cwd: "/repos/a-feature", model: dispatchModel });
+    expect(spawn).toHaveBeenCalledWith({ spawningCwd: "/repos/a", spawningSessionId: "spawner-1", prompt: "do the thing", cwd: "/repos/a-feature", model: dispatchModel });
     expect(result.details).toEqual({ sessionId: "new-1", cwd: "/repos/a-feature" });
     expect(result.content[0]).toMatchObject({ type: "text", text: "Started independent session new-1 in /repos/a-feature." });
   });
@@ -32,16 +35,45 @@ describe("createSpawnSessionToolDefinition", () => {
     const spawn = vi.fn(() => Promise.resolve({ sessionId: "new-2", cwd: "/repos/a" }));
     const tool = createSpawnSessionToolDefinition("/repos/a", { spawn });
 
-    await tool.execute("call-2", { prompt: "continue" }, undefined, undefined, ctx);
+    await tool.execute("call-2", { prompt: "continue" }, undefined, undefined, ctxFor("spawner-1"));
 
-    expect(spawn).toHaveBeenCalledWith({ spawningCwd: "/repos/a", prompt: "continue", cwd: undefined });
+    expect(spawn).toHaveBeenCalledWith({ spawningCwd: "/repos/a", spawningSessionId: "spawner-1", prompt: "continue", cwd: undefined });
+  });
+
+  it("forwards an explicit model as a model spec alongside the inherited model", async () => {
+    const spawn = vi.fn(() => Promise.resolve({ sessionId: "new-3", cwd: "/repos/a", model: "openai/gpt-5" }));
+    const tool = createSpawnSessionToolDefinition("/repos/a", { spawn });
+
+    const result = await tool.execute("call-3", { prompt: "continue", model: "openai/gpt-5" }, undefined, undefined, ctxFor("spawner-1", dispatchModel));
+
+    expect(spawn).toHaveBeenCalledWith({
+      spawningCwd: "/repos/a",
+      spawningSessionId: "spawner-1",
+      prompt: "continue",
+      cwd: undefined,
+      model: dispatchModel,
+      modelSpec: "openai/gpt-5",
+    });
+    expect(result.details).toEqual({ sessionId: "new-3", cwd: "/repos/a", model: "openai/gpt-5" });
+    expect(result.content[0]).toMatchObject({ type: "text", text: "Started independent session new-3 in /repos/a using model openai/gpt-5." });
+  });
+
+  it("teaches the model parameter format and the #provider/model-id reference convention", () => {
+    const tool = createSpawnSessionToolDefinition("/repos/a", { spawn: vi.fn() });
+
+    expect(tool.parameters).toMatchObject({
+      properties: {
+        // eslint-disable-next-line @typescript-eslint/no-unsafe-assignment -- stringMatching yields `any` against the loosely typed tool schema.
+        model: { description: expect.stringMatching(/provider\/model-id.*#provider\/model-id.*Omit to inherit/s) },
+      },
+    });
   });
 
   it("propagates the spawn callback error so the agent loop reports it", async () => {
     const spawn = vi.fn(() => Promise.reject(new Error("cwd must be a workspace of this project. Allowed: /repos/a")));
     const tool = createSpawnSessionToolDefinition("/repos/a", { spawn });
 
-    await expect(tool.execute("call-4", { prompt: "x", cwd: "/elsewhere" }, undefined, undefined, ctx))
+    await expect(tool.execute("call-4", { prompt: "x", cwd: "/elsewhere" }, undefined, undefined, ctxFor("spawner-1")))
       .rejects.toThrow("cwd must be a workspace of this project. Allowed: /repos/a");
   });
 });

src/server/sessions/spawnSessionTool.ts

@@ -4,16 +4,22 @@ import { defineTool, type ExtensionContext } from "@earendil-works/pi-coding-age
 export interface SpawnSessionResult {
   sessionId: string;
   cwd: string;
+  /** Model the spawned session runs with, as `provider/id`; absent when unknown. */
+  model?: string;
 }
 
 export type SpawnSessionModel = NonNullable<ExtensionContext["model"]>;
 
 export interface SpawnSessionInvocation {
   spawningCwd: string;
+  /** Id of the dispatching session; used to resolve {@link modelSpec} against its model runtime. */
+  spawningSessionId: string;
   prompt: string;
   cwd: string | undefined;
   /** Current model from the dispatching session, used as the spawned session's default. */
   model?: SpawnSessionModel;
+  /** Strict `provider/model-id` requested by the dispatcher; overrides {@link model} when set. */
+  modelSpec?: string;
 }
 
 export interface SpawnSessionToolDeps {
@@ -29,6 +35,9 @@ const SpawnSessionParams = Type.Object({
   cwd: Type.Optional(Type.String({
     description: "Working directory for the new session. Must be a workspace (worktree, or root) of the same project as this session. Defaults to this session's working directory.",
   })),
+  model: Type.Optional(Type.String({
+    description: 'Model for the new session, as an exact "provider/model-id" such as "anthropic/claude-sonnet-4-5". When the user references a model as #provider/model-id in their request, forward it here. An unknown value is rejected. Omit to inherit this session\'s model.',
+  })),
 });
 
 /**
@@ -50,12 +59,15 @@ export function createSpawnSessionToolDefinition(spawningCwd: string, deps: Spaw
       // valid workspace) rather than crash.
       const result = await deps.spawn({
         spawningCwd,
+        spawningSessionId: ctx.sessionManager.getSessionId(),
         prompt: params.prompt,
         cwd: params.cwd,
         ...(ctx.model === undefined ? {} : { model: ctx.model }),
+        ...(params.model === undefined ? {} : { modelSpec: params.model }),
       });
+      const modelNote = result.model === undefined ? "" : ` using model ${result.model}`;
       return {
-        content: [{ type: "text", text: `Started independent session ${result.sessionId} in ${result.cwd}.` }],
+        content: [{ type: "text", text: `Started independent session ${result.sessionId} in ${result.cwd}${modelNote}.` }],
         details: result,
       };
     },

src/server/sessions/spawnSubsessionTool.test.ts

@@ -112,6 +112,36 @@ describe("createSubsessionToolDefinitions", () => {
     });
   });
 
+  it("spawn_subsession forwards an explicit model as a model spec and names the model used", async () => {
+    const spawn = vi.fn(() => Promise.resolve({ sessionId: "child-3", cwd: "/repos/a", model: "openai/gpt-5" }));
+    const { spawn: spawnTool } = tools({ spawn });
+
+    const result = await spawnTool.execute("call-model", { prompt: "do it", model: "openai/gpt-5" }, undefined, undefined, ctxFor("parent-1", "/sessions/parent-1.jsonl", dispatchModel));
+
+    expect(spawn).toHaveBeenCalledWith({
+      spawningCwd: "/repos/a",
+      parentSessionId: "parent-1",
+      parentSessionFile: "/sessions/parent-1.jsonl",
+      prompt: "do it",
+      cwd: undefined,
+      model: dispatchModel,
+      modelSpec: "openai/gpt-5",
+    });
+    expect(result.details).toEqual({ sessionId: "child-3", cwd: "/repos/a", model: "openai/gpt-5" });
+    expect(firstText(result.content)).toBe("Started tracked subsession child-3 in /repos/a using model openai/gpt-5. Continue other work, then join with yield_to_subsessions; do not poll.");
+  });
+
+  it("spawn_subsession teaches the model parameter format and the #provider/model-id reference convention", () => {
+    const { spawn: spawnTool } = tools({});
+
+    expect(spawnTool.parameters).toMatchObject({
+      properties: {
+        // eslint-disable-next-line @typescript-eslint/no-unsafe-assignment -- stringMatching yields `any` against the loosely typed tool schema.
+        model: { description: expect.stringMatching(/provider\/model-id.*#provider\/model-id.*Omit to inherit/s) },
+      },
+    });
+  });
+
   it("list_subsessions reports the caller's subsessions and their status", async () => {
     const list = vi.fn(() => Promise.resolve([
       { sessionId: "child-1", cwd: "/repos/a", status: "working" as const },

src/server/sessions/spawnSubsessionTool.ts

@@ -8,6 +8,8 @@ export type SubsessionStatus = "working" | "idle" | "error" | "unknown";
 export interface SpawnSubsessionResult {
   sessionId: string;
   cwd: string;
+  /** Model the child session runs with, as `provider/id`; absent when unknown. */
+  model?: string;
 }
 
 export type SpawnSubsessionModel = NonNullable<ExtensionContext["model"]>;
@@ -23,6 +25,8 @@ export interface SpawnSubsessionInvocation {
   cwd: string | undefined;
   /** Current model from the dispatching session, used as the spawned session's default. */
   model?: SpawnSubsessionModel;
+  /** Strict `provider/model-id` requested by the parent; overrides {@link model} when set. */
+  modelSpec?: string;
 }
 
 export interface SubsessionSummary {
@@ -72,6 +76,9 @@ const SpawnSubsessionParams = Type.Object({
   cwd: Type.Optional(Type.String({
     description: "Child workspace in the same project (worktree or root); defaults to the parent's directory.",
   })),
+  model: Type.Optional(Type.String({
+    description: 'Model for the child session, as an exact "provider/model-id" such as "anthropic/claude-sonnet-4-5". When the user references a model as #provider/model-id in their request, forward it here. An unknown value is rejected. Omit to inherit this session\'s model.',
+  })),
 });
 
 const ListSubsessionsParams = Type.Object({});
@@ -197,9 +204,11 @@ export function createSubsessionToolDefinitions(spawningCwd: string, deps: Subse
         prompt: params.prompt,
         cwd: params.cwd,
         ...(ctx.model === undefined ? {} : { model: ctx.model }),
+        ...(params.model === undefined ? {} : { modelSpec: params.model }),
       });
+      const modelNote = result.model === undefined ? "" : ` using model ${result.model}`;
       return {
-        content: [{ type: "text", text: `Started tracked subsession ${result.sessionId} in ${result.cwd}. Continue other work, then join with yield_to_subsessions; do not poll.` }],
+        content: [{ type: "text", text: `Started tracked subsession ${result.sessionId} in ${result.cwd}${modelNote}. Continue other work, then join with yield_to_subsessions; do not poll.` }],
         details: result,
       };
     },