fix(workflows): make agent startup abortable

[CuSO41108]

Summary

Closes #116.

Make workflow child-agent startup responsive to cancellation during both session creation and extension binding.

When cancellation wins the startup race, runAgent() now returns an explicit aborted outcome without waiting indefinitely. Any session created after cancellation is still owned and cleaned up, and late startup rejections are observed.

Changes

• Install the abort listener before creating the child session.
• Race session creation and extension binding against cancellation.
• Skip session creation when the signal is already aborted.
• Abort and dispose sessions that are created after cancellation.
• Reuse the existing bounded, idempotent child-session cleanup path.
• Prevent prompting after startup cancellation.
• Classify startup cancellation as aborted: true.
• Observe late factory and binding rejections to avoid unhandled rejections.

bindExtensions() does not currently accept an AbortSignal, so its internal initialization may still finish in the background. This change stops the workflow from waiting on it indefinitely and safely handles its eventual result.

Tests

Added deterministic regression coverage for:

• cancellation before startup;
• cancellation while session creation is pending;
• cleanup of a session created after cancellation;
• a late session-factory rejection;
• cancellation while extension binding is pending;
• preventing prompts and duplicate disposal after cancellation.

Validation:

• extensions/workflows/runner.test.ts: 21/21 passed
• bun run lint: passed
• bun run typecheck: passed

Local full-suite validation is affected by existing Windows-specific failures involving process handling, path separators, file modes, and symlink permissions. The runner test suite and all new regression cases pass.

[Copilot]

🟢 Approval recommended

The cancellation behavior changes appear correct and well-covered by targeted regression tests, with only a minor error-message clarity nit remaining.

Pull request overview

This PR hardens workflow child-agent lifecycle handling by making runAgent() startup (session creation + extension binding) responsive to AbortSignal cancellation, ensuring cancellation produces an explicit aborted outcome and any late-created sessions are still properly cleaned up.

Changes:

• Installs the abort listener before child-session creation and races startup steps against cancellation.
• Ensures late session creation after cancellation is still owned and disposed via the bounded cleanup path.
• Adds deterministic regression tests covering startup cancellation races and late failures.

File summaries

File	Description
extensions/workflows/runner.ts	Makes startup abortable by wiring abort handling earlier and racing session creation/binding against cancellation, with safe cleanup of late sessions.
extensions/workflows/runner.test.ts	Adds regression coverage for pre-abort, abort during creation/binding, late factory/binding failures, and prompt suppression after startup cancellation.

Review details

Suppressed comments (1)

extensions/workflows/runner.ts:636

• The startup try/catch covers more than session creation (it also races extension binding and installs tool guards), but this error message still says "Failed to create agent session". If binding/tool-guard setup fails, the message is misleading and makes diagnosis harder.

    return {
      ok: false,
      output: "",
      error: `Failed to create agent session: ${errorText(error)}${cleanupError ? `; cleanup failed: ${cleanupError}` : ""}`,
      aborted: false,

• Files reviewed: 2/2 changed files
• Comments generated: 0
• Review effort level: Lite

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.

[tt-a1i]

Thanks for the focused lifecycle hardening. I reviewed the startup cancellation races across pre-aborted startup, pending session creation, late factory settlement, extension binding, prompting, and bounded idempotent cleanup. The branch also integrates cleanly with current main; bun run check, the 21 runner tests, and the full suite (874 node tests plus 30 file-search tests) pass. No blocking issues from my side.