[agegr]

Problem

pi's AgentSessionRuntime.fork() emits session_before_fork before it copies anything. A handler can return { cancel: true }, or ask the user something first.

• pi's git-checkpoint example offers to restore the code to that point.
• pi's dirty-repo-guard example refuses a fork while the repository has uncommitted changes.

Pi Web's New session on a user message (send({ type: "fork" })) copied the branch on disk without asking. Those extensions never ran.

Repro on main: an extension with pi.on("session_before_fork", async (e, ctx) => (await ctx.ui.confirm("Fork this session?", "…")) ? undefined : { cancel: true }). Click New session on a message: it forks at once, with no dialog.

Fix

• The wrapper's fork emits session_before_fork ({ entryId, position: "before" }), but only when a handler listens, exactly as emitBeforeFork() does.
• { cancel: true } returns { cancelled: true }, and nothing is written.
• A handler's dialog shows in the page while the fork waits.
• The hook waits for the extensions to be bound first, as pi only forks a bound session. Before binding, a handler sees ctx.hasUI as false and takes its non-interactive branch; dirty-repo-guard, for example, cancels silently there.
• The hook runs before the copy's session-replacement lock, so while its dialog waits, a running source still takes Stop and steers. The entry and a running ! command are checked again afterwards.
• Unlike pi, the source is not torn down after the hook, because a running source keeps its run in Pi Web. A handler that restores code therefore does so beside it.
• The sidebar's Fork stays a runtime-free copy on disk, as documented.
• session_fork has no counterpart here: the copy gets its runtime later, when it is opened.

Verification

• New test in lib/rpc-manager.test.mjs: a cancelled fork writes no file, and the next one goes through.
• npm test, tsc --noEmit, npm run lint.
• Headless Chrome on an isolated dev server with the extension above:
  • New session showed the "Fork this session?" dialog;
  • Cancel kept the page on the source session;
  • Confirm opened the copy;
  • on a freshly started server (fork right after the session's wrapper started), the dialog still appeared.

🤖 Generated with Claude Code

[agegr]

Review

The finding below was checked twice: once by tracing the code and once in a wrapper test.

Must fix: the guard fails open after a Stop

forkCancelledByExtension() emits while extensionUiAbortController may still be aborted: Stop aborts it, and only prompt and reload renew it. Two cases follow:

• After a Stop with no later prompt (stop a bad answer, then "New session" on an earlier message):
  1. The handler's ctx.ui.confirm / select rejects at once with "Extension UI cancelled by Stop".
  2. ExtensionRunner.emit swallows the error, so the result is undefined.
  3. The fork is written with no dialog, and an error toast shows.
• Stop pressed while the fork's dialog waits approves the fork the same way.

pi treats a closed dialog as undefined, so guards such as confirm-destructive and dirty-repo-guard cancel the fork there.

Fix: renew the controller if it is aborted before emitting. Capture its signal, and treat an abort during the hook as cancelled.

The same root cause already affects session_before_tree (navigate_tree) on main, so one helper could cover both.

Follow-up, not a defect of this PR

• /clone still skips the hook. pi's /clone, in both the TUI and RPC mode, emits session_before_fork with position: "at". pi-web's clone still doesn't, so a guard can now block a message fork but not /clone. A follow-up should:
  • ask the hook before the clone lock;
  • let the client tell an extension's cancel apart from "Cannot clone an empty or unsaved session".

Checked, fine

• The event payload, the hasHandlers check and { cancel: true } match pi's emitBeforeFork.
• The sidebar's on-disk Fork and side chats skip the hook, which is documented.
• No conflict with feat(chat): a message's New session starts with that message in the composer #1539, and the two compose: the draft is written only when !cancelled.

[agegr]

Pushed 0f419a2 and bd7ac55 for the review above:

1. The hook fails closed on Stop.
   • renewExtensionUiStop() renews a controller an earlier Stop left aborted.
   • A Stop while the hook runs counts as cancelled.
   • The reviewers' repro passes in all three cases: baseline, after Stop, and during the dialog.
2. The same renew now runs before session_before_tree (navigate_tree) and session_before_compact (manual /compact). Both had the same root cause on main.
3. /clone asks session_before_fork with position: "at", as pi's /clone does. The client tells an extension's cancel apart from "Cannot clone an empty or unsaved session".

tsc, eslint and npm test pass (3277 pass, 1 skipped).

Documented, not changed:

• A Stop during the tree hook still lets the move through: pi moves inside navigateTree() right after the hook. Stop shows then only if a prompt was sent while the dialog waited.
• A Stop that lands after the fork dialog was answered, while the handler finishes, also drops the copy. The window is narrow.