Allow working without a worktree

[cnwangjie]

Summary

Currently each task runs in its own git worktree. I'd like an option to run a task directly in the existing working
directory instead, skipping worktree creation entirely.

Motivation

Worktrees are great for isolating parallel work, but they aren't always the right fit:

• Small / sequential changes — for a quick fix I just want to work where I already am, without the overhead of spinning
  up (and later cleaning up) a separate checkout.
• Non-git or single-checkout setups — some projects, tooling, or environments don't play well with multiple worktrees
  (path-dependent configs, native build caches, node_modules, virtualenvs, IDE indexes, etc.).
• Disk / setup cost — each worktree can mean re-installing dependencies or rebuilding, which is wasteful when I only
  have one task in flight.
• Working on an already-checked-out branch — sometimes I'm mid-change and just want the agent to continue in place
  rather than branching off into an isolated copy.

Proposed behavior

• A per-task (and/or global default) option like "Run in current directory" / "No worktree".
• When enabled, the task operates against the existing working directory and current branch — no new worktree is created and
  none is cleaned up afterward.
• Make the tradeoff clear in the UI (no isolation; changes happen in place), ideally with a warning if there are uncommitted
  changes or multiple in-place tasks would overlap.

Possible enhancements

• Guard against running two in-place tasks on the same directory at once.
• Remember the choice per repository.

Alternatives considered

• Always using worktrees and manually merging back — adds friction for small changes and doesn't help with environments that
  can't use worktrees.

[seanmcilvenna]

I struggle with the working tree... I like to have my IDE open with the code base, and when a task completes, launch the app within the IDE and review/test. I can't do this with working trees, because I have to open a separate folder for the working trees each time in the IDE, which makes it very inefficient.

[gae-rafatjah]

The issue with forcing the worktrees is the sequential tasks. Not having the artifacts of each task (As not everything get's committed of course) makes it hard to continue.

Process logs > List issues > Review and fix each

This can be so huge that you might require a ton of tasks and scales up easily with the Parallel tasks.