chore: Pi 开发依赖跟随最新版并去掉测试里的固定版本号

- pi-hermes-memory / pi-notify / pi-permission-system / pi-subagents 的
  @earendil-works/* devDependencies 改为 latest，pi-rtk-optimizer 放开
  peer 范围；各子包锁文件刷新到 Pi 1.3.0，测试改为验证当前宿主 API。
- 新增 scripts/update-pi-devdeps.sh，统一刷新根与子包锁文件中的 Pi 版本。
- pi-subagents 适配 Pi 1.3：mention clone 改走 systemPromptOverride、
  BOM frontmatter、e2e 读取规范化后的 system 消息与工具列表。
- 根测试不再写死 ssh2 / jiti / better-sqlite3 / context-mode 版本号，
  改为精确版本格式与锁文件一致性断言；新增 latest 约束测试。

Generated with [Devin](https://devin.ai)

Co-Authored-By: Devin <158243242+devin-ai-integration[bot]@users.noreply.github.com>

pi-subagents/package.json

@@ -57,9 +57,9 @@
   },
   "devDependencies": {
     "@biomejs/biome": "^2.4.14",
-    "@earendil-works/pi-ai": "0.84.2",
-    "@earendil-works/pi-coding-agent": "0.84.2",
-    "@earendil-works/pi-tui": "0.84.2",
+    "@earendil-works/pi-ai": "latest",
+    "@earendil-works/pi-coding-agent": "latest",
+    "@earendil-works/pi-tui": "latest",
     "@types/node": "^25.5.0",
     "@vitest/coverage-istanbul": "^4.1.10",
     "typescript": "^6.0.0",

pi-subagents/src/agent-file-toggle.ts

@@ -104,15 +104,17 @@ const FENCE = /^---[ \t]*$/;
  * what `parseFrontmatter` (the load side) considers a frontmatter block.
  *
  * Lines keep their terminators, so an edit preserves the file's existing line
- * endings instead of rewriting CRLF to LF. Returns undefined when there is no
- * usable block — notably for a BOM-prefixed file, which the parser also reads
- * as having none, so writing a key into it would change nothing on load.
+ * endings instead of rewriting CRLF to LF. A leading BOM stays attached to the
+ * fence line — the parser skips it, so the block behind it is real frontmatter.
+ * Returns undefined when there is no usable block.
  */
 function splitFrontmatter(content: string):
   | { lines: string[]; openIdx: number; closeIdx: number; eol: string }
   | undefined {
   const lines = content.split(/(?<=\n)/);
-  if (lines.length === 0 || !FENCE.test(lines[0].replace(/\r?\n$/, ""))) return undefined;
+  if (lines.length === 0 || !FENCE.test(lines[0].replace(/\r?\n$/, "").replace(/^/, ""))) {
+    return undefined;
+  }
   const closeIdx = lines.findIndex((l, i) => i > 0 && FENCE.test(l.replace(/\r?\n$/, "")));
   if (closeIdx === -1) return undefined;
   return { lines, openIdx: 0, closeIdx, eol: lines[0].endsWith("\r\n") ? "\r\n" : "\n" };

pi-subagents/src/mention-clone.ts

@@ -67,6 +67,7 @@ import {
   createAgentSession,
   DefaultResourceLoader,
   type ExtensionContext,
+  type ExtensionToolContext,
   getAgentDir,
   SessionManager,
   type ToolDefinition,
@@ -127,7 +128,10 @@ export async function runMentionClone(opts: MentionCloneOptions): Promise<Mentio
         { ...(params as Record<string, unknown>), run_in_background: true } as typeof params,
         signal,
         onUpdate,
-        ctx,
+        // Pi 1.3 narrowed tool ctx to ExtensionToolContext; the two added
+        // members (tools, executeTool) serve tool-to-tool calls, which the
+        // Agent tool never makes — runAgent reads only session fields.
+        ctx as ExtensionToolContext,
       );
     },
   };
@@ -148,6 +152,11 @@ export async function runMentionClone(opts: MentionCloneOptions): Promise<Mentio
     // the settings level instead, which is what a session that never ran
     // `/think` is on anyway. Same shim shape as `modelRuntime` below.
     const thinkingLevel = (ctx as { thinkingLevel?: ThinkingLevel }).thinkingLevel;
+    // The clone rebuilds a system prompt from cwd and agentDir, which is close
+    // but not the live one — extensions contribute to it per turn. Copy the
+    // real thing via the loader override, so the copy reasons under the
+    // instructions the user's model is actually working under.
+    const systemPrompt = ctx.getSystemPrompt?.();
     // This off-screen clone must expose only its synthetic Agent tool. Loading
     // host/project extensions would let active-tool owners such as Tool Search
     // replace the clone's allowlist during session_start, and would also run
@@ -159,6 +168,7 @@ export async function runMentionClone(opts: MentionCloneOptions): Promise<Mentio
       noSkills: true,
       noPromptTemplates: true,
       noThemes: true,
+      ...(systemPrompt !== undefined && { systemPromptOverride: () => systemPrompt }),
     });
     const created = await runInChildSessionContext(() =>
       createAgentSession({
@@ -185,13 +195,6 @@ export async function runMentionClone(opts: MentionCloneOptions): Promise<Mentio
     );
     session = created.session;
 
-    // The clone rebuilds a system prompt from cwd and agentDir, which is close
-    // but not the live one — extensions contribute to it per turn. Copy the
-    // real thing, so the copy reasons under the instructions the user's model
-    // is actually working under.
-    const systemPrompt = ctx.getSystemPrompt?.();
-    if (systemPrompt) session.agent.state.systemPrompt = systemPrompt;
-
     // The conversation itself. Pushed rather than assigned so the array the
     // session was built around stays the one it goes on using.
     session.agent.state.messages.push(...conversation.messages);

pi-subagents/test/agent-file-toggle.test.ts

@@ -124,15 +124,15 @@ describe("disableInContent", () => {
     expect(loaderSeesDisabled(content)).toBe(true);
   });
 
-  it("reports no-frontmatter for a BOM-prefixed file, matching what the loader sees", () => {
-    // The parser doesn't accept a BOM either — it reports an empty frontmatter
-    // and treats the whole file as body. So refusing here is agreement with the
-    // read side, not a gap: inserting the key would have no effect on loading.
+  it("disables a BOM-prefixed file, matching what the loader sees", () => {
+    // Pi ≥1.3's parser skips a leading BOM, so the block behind it is real
+    // frontmatter — refusing here would leave a file the loader reads enabled
+    // while reporting "nothing to do". The edit keeps the BOM in place.
     const src = "---\ndescription: Scout\n---\n\nBody.\n";
-    expect(parseFrontmatter<Record<string, unknown>>(src).frontmatter).toEqual({});
     const { content, outcome } = disableInContent(src);
-    expect(outcome).toBe("no-frontmatter");
-    expect(content).toBe(src);
+    expect(outcome).toBe("disabled");
+    expect(content.startsWith("")).toBe(true);
+    expect(loaderSeesDisabled(content)).toBe(true);
   });
 });
 

pi-subagents/test/e2e/usage-reaches-session-stats.e2e.test.ts

@@ -112,20 +112,19 @@ describe("subagent usage reaches the parent session's stats (real pi)", () => {
     }
   });
 
-  it("leaves the context-window percentage alone", async () => {
-    // pi derives context usage from assistant messages only. If that ever
-    // changed, a delegating session would look like it was filling its context
-    // with work that happened somewhere else entirely — and users would compact
-    // for no reason.
+  it("counts only the result's own text toward context, not its reported usage", async () => {
+    // Since Pi 1.3 the percentage reflects the transcript's real contents, so
+    // the message's own few tokens do register. What must not register is the
+    // reported usage: 150k delegated tokens on a 200k window would read as
+    // ~75% occupied, and a delegating session would look like it was filling
+    // its context with work that happened somewhere else entirely.
     const session = await realSession();
     try {
-      const before = session.getSessionStats().contextUsage?.percent ?? null;
-
       const pool = new PendingUsagePool();
       pool.add({ input: 150_000, output: 400, cacheWrite: 100, cost: 1.5 });
       session.sessionManager.appendMessage(toolResultCarrying(pool.drain()) as any);
 
-      expect(session.getSessionStats().contextUsage?.percent ?? null).toBe(before);
+      expect(session.getSessionStats().contextUsage?.percent ?? 0).toBeLessThan(1);
     } finally {
       session.dispose?.();
     }

pi-subagents/test/helpers/print-mode-runner.ts

@@ -195,6 +195,46 @@ function resolveReply(
   return typeof reply === "function" ? (reply as (c: Context) => FauxReply)(ctx) : reply;
 }
 
+/**
+ * Names of the tools the session offered the model for this call. Pi ≥1.3
+ * normalizes `Context.tools` into the transcript's system messages
+ * (`toolsAdded`/`toolsRemoved`) before the provider sees the request, so read
+ * both shapes.
+ */
+export function contextToolNames(context: Context): string[] {
+  const names = new Set<string>((context.tools ?? []).map((t) => t.name));
+  for (const m of context.messages) {
+    if (m.role !== "system") continue;
+    for (const t of m.toolsAdded ?? []) names.add(t.name);
+    for (const t of m.toolsRemoved ?? []) names.delete(t.name);
+  }
+  return [...names];
+}
+
+/**
+ * The system-prompt text the session offered for this call. Pi ≥1.3 normalizes
+ * `Context.systemPrompt` into the transcript's system messages, so read both
+ * shapes: `context.systemPrompt` plus every system message's `content` and
+ * named `sections` values.
+ */
+export function contextSystemPrompt(context: Context): string {
+  const parts: string[] = [];
+  if (context.systemPrompt) parts.push(context.systemPrompt);
+  for (const m of context.messages) {
+    if (m.role !== "system") continue;
+    const content = m.content;
+    if (typeof content === "string") {
+      parts.push(content);
+    } else if (Array.isArray(content)) {
+      for (const b of content) if (b.type === "text" && b.text) parts.push(b.text);
+    }
+    for (const v of Object.values(m.sections ?? {})) {
+      if (typeof v === "string") parts.push(v);
+    }
+  }
+  return parts.join("\n");
+}
+
 /**
  * The common single-spawn flow as a responder. Routes by inspecting the calling
  * session's own context:
@@ -210,7 +250,7 @@ export function routeBySession(routes: {
   subagent: FauxReply | ((ctx: Context) => FauxReply);
 }): FauxResponder {
   return (context) => {
-    const isParent = (context.tools ?? []).some((t) => t.name === "Agent");
+    const isParent = contextToolNames(context).includes("Agent");
     if (!isParent) return resolveReply(routes.subagent, context);
     const spawned = context.messages.some(
       (m) => m.role === "toolResult" && (m as { toolName?: string }).toolName === "Agent",

pi-subagents/test/mention-clone.test.ts

@@ -17,10 +17,14 @@ import { beforeEach, describe, expect, it, vi } from "vitest";
 
 // Hoisted: vi.mock's factory is lifted above the imports, so it cannot close
 // over ordinary top-level consts.
-const { buildSessionContext, createAgentSession, inMemory } = vi.hoisted(() => ({
+const { buildSessionContext, createAgentSession, inMemory, loaderOptions } = vi.hoisted(() => ({
   buildSessionContext: vi.fn(),
   createAgentSession: vi.fn(),
   inMemory: vi.fn(),
+  // The clone's prompt reaches the session through the loader's
+  // systemPromptOverride (Pi 1.3 made AgentState.systemPrompt read-only), so
+  // capture what runMentionClone hands DefaultResourceLoader.
+  loaderOptions: { current: undefined as Record<string, unknown> | undefined },
 }));
 
 vi.mock("@earendil-works/pi-coding-agent", async () => {
@@ -29,6 +33,12 @@ vi.mock("@earendil-works/pi-coding-agent", async () => {
     ...actual,
     buildSessionContext,
     createAgentSession,
+    DefaultResourceLoader: class extends actual.DefaultResourceLoader {
+      constructor(options?: Record<string, unknown>) {
+        loaderOptions.current = options;
+        super(options);
+      }
+    },
     SessionManager: { ...actual.SessionManager, inMemory },
   };
 });
@@ -44,6 +54,7 @@ const CONVERSATION = [
 
 beforeEach(() => {
   createAgentSession.mockReset();
+  loaderOptions.current = undefined;
   inMemory.mockReset();
   inMemory.mockReturnValue({ kind: "in-memory-session-manager" } as any);
   buildSessionContext.mockReset();
@@ -104,6 +115,13 @@ function cloneSession(turn?: (tool: any) => Promise<void> | void) {
     dispose: vi.fn(),
   } as any;
   createAgentSession.mockImplementation(async (opts: any) => {
+    // What the real wiring does: the loader's systemPromptOverride rewrites the
+    // rebuilt base before the session's state gets it.
+    const override = loaderOptions.current?.systemPromptOverride as
+      | ((base: string | undefined) => string | undefined)
+      | undefined;
+    const overridden = override?.(session.agent.state.systemPrompt);
+    if (overridden !== undefined) session.agent.state.systemPrompt = overridden;
     const tools = visibleTools(opts);
     session.prompt.mockImplementation(async () => {
       // No tool, no tool call: the model can only answer in prose.

pi-subagents/test/nested-delegation-e2e.test.ts

@@ -26,6 +26,7 @@ import { loadCustomAgents } from "../src/custom-agents.js";
 import { encodeCwd } from "../src/output-file.js";
 import {
   agentCall,
+  contextToolNames,
   type FauxReply,
   type PrintModeRun,
   runPrintMode,
@@ -101,7 +102,7 @@ describe("nested delegation e2e (real pi-mono, faux model)", () => {
 
     const respond = (context: Context): FauxReply => {
       const text = firstUserText(context);
-      const names = (context.tools ?? []).map((t) => t.name);
+      const names = contextToolNames(context);
 
       // Leaf: no nested tools (it never opted in) — just answer.
       if (text.includes("Do the leaf work")) {

pi-subagents/test/subagents-nested-print-mode-e2e.test.ts

@@ -22,6 +22,7 @@ import { registerAgents } from "../src/agent-types.js";
 import { loadCustomAgents } from "../src/custom-agents.js";
 import {
   agentCall,
+  contextToolNames,
   type FauxResponder,
   type PrintModeRun,
   runPrintMode,
@@ -47,7 +48,7 @@ function userPrompt(ctx: Context): string {
 }
 
 function tools(ctx: Context): string[] {
-  return (ctx.tools ?? []).map((tool) => tool.name);
+  return contextToolNames(ctx);
 }
 
 function toolResults(ctx: Context, name: string): string[] {

pi-subagents/test/subagents-print-mode-e2e.test.ts

@@ -21,6 +21,8 @@ import {
   agentCall,
   agentToolCalls,
   agentToolResults,
+  contextSystemPrompt,
+  contextToolNames,
   conversationText,
   invokedToolNames,
   type PrintModeRun,
@@ -92,7 +94,7 @@ describe.skipIf(LIVE)("subagents print-mode e2e (scripted faux, real pi-mono)",
     //     finishes → the child's own model turn actually runs (≥3 calls).
     const sleep = (ms: number) => new Promise((r) => setTimeout(r, ms));
     const respond = async (ctx: Context) => {
-      const isParent = (ctx.tools ?? []).some((t) => t.name === "Agent");
+      const isParent = contextToolNames(ctx).includes("Agent");
       if (!isParent) {
         await sleep(80); // child takes long enough that a non-held parent exits first
         return "CHILD_BG_RAN";
@@ -151,7 +153,7 @@ describe.skipIf(LIVE)("subagents print-mode e2e (scripted faux, real pi-mono)",
         parentFinal: "Reported.",
         // The child reflects whether the frontmatter body reached its own prompt.
         subagent: (ctx: Context) =>
-          `child saw: ${ctx.systemPrompt?.includes(MARKER) ? MARKER : "MISSING"}`,
+          `child saw: ${contextSystemPrompt(ctx).includes(MARKER) ? MARKER : "MISSING"}`,
       }),
     });
 
@@ -185,7 +187,7 @@ describe.skipIf(LIVE)("subagents print-mode e2e (scripted faux, real pi-mono)",
         }),
         parentFinal: "Reported.",
         subagent: (ctx: Context) =>
-          `child saw: ${ctx.systemPrompt?.includes(MARKER) ? MARKER : "MISSING"}`,
+          `child saw: ${contextSystemPrompt(ctx).includes(MARKER) ? MARKER : "MISSING"}`,
       }),
     });