fix: harden worktree creation and improve subagent mode support

[mohammedahmed18]

Summary

• Worktree creation is now mandatory: setup_worktree exits with an error instead of silently skipping when the git repo check fails or worktree creation returns None. This prevents optimizations from running in an invalid state.
• Removed redundant git repo check from create_detached_worktree() — the caller in optimizer.py already performs this check, so the parameter and duplicate logic were unnecessary.
• Subagent mode parity with LSP: Skip was_function_previously_optimized checks in subagent mode (same behavior as LSP), and render optimization result trees as markdown logs instead of Rich console output.
• Disabled custom benchmark loop overrides (pytest_min_loops/pytest_max_loops) for subagent runs — commented out pending further tuning.
• Downgraded Jest benchmarking failure logs from info to debug to reduce noise.

Test plan

• Verify worktree creation fails fast when not in a git repo
• Run optimization in subagent mode and confirm markdown output is logged
• Run JS benchmarking with failures and confirm debug-level logging

🤖 Generated with Claude Code

[claude]

Claude finished @mohammedahmed18's task in 2m 28s —— View job

---

PR Review Summary

• Triage PR size — SMALL (53 lines, 5 production files)
• Run lint/typecheck — prek clean; mypy errors are pre-existing
• Resolve stale threads — none found
• Review code
• Check for duplicates — none found
• Post summary

---

Prek Checks

Clean — no ruff or formatting issues introduced by this PR. Mypy reports 174 pre-existing errors across 4 files (not caused by this PR).

---

Code Review

Bug: Failure returned when worktree already exists (optimizer.py:784-785)

if self.current_worktree:
    return Failure("There is an existing worktree, Can't create more than one per function")

The original code silently returned here (idempotency — "already initialized, skip"). This PR changes it to a Failure, causing the caller to log an error and abort run() entirely. If worktree_mode() is ever called when a worktree is already set (e.g. retry logic, multi-function runs), this will now silently abort the entire optimization run rather than continuing with the existing worktree. The silent return behavior should be preserved for this branch. Fix this →

Minor: Result[bool, str] with # noqa: FBT003 (optimizer.py:783, 798)

Since the success value is meaningless (True is just a placeholder), Result[None, str] / Success(None) is cleaner and avoids the suppressed ruff warning entirely. Fix this →

Minor: Removed helpful comment (functions_to_optimize.py:761)

The comment explaining why LSP skips was_function_previously_optimized was removed. Since subagent mode now gets the same treatment, a brief comment explaining the reason for both would help future maintainers.

---

Duplicate Detection

No duplicates detected.

---

Last updated: 2026-03-25

[aseembits93]

LGTM. i have noticed the windows tests are a bit flaky. will investigate a bit.

[aseembits93]

see if you could just return instead of sys.exit(). like here - https://github.com/codeflash-ai/codeflash/pull/1889/changes#diff-d242b27fa098f222741a48a7daa8e421433b4c5e19dd38512404cd000df144a0L489