fix: use correct subagent flag in checkpoint guard

[aseembits93]

Summary

• The checkpoint guard in optimizer.py referenced args.agent, which is never defined on the CLI args namespace — the actual flag is --subagent (args.subagent).
• Since getattr(self.args, "agent", False) always returned False, the condition was effectively just self.args.all, meaning checkpointing was never skipped in subagent mode as intended.

Test plan

• Verify --all --subagent skips checkpoint creation
• Verify --all without --subagent still creates checkpoint

🤖 Generated with Claude Code

[claude]

PR Review Summary

Prek Checks

✅ All checks pass — ruff check and ruff format both passed with no issues.

Mypy

⚠️ 22 pre-existing mypy errors in optimizer.py — none introduced by this PR. All errors exist on main as well (type variance issues, unreachable code, Path | None mismatches, etc.).

Code Review

✅ No issues found. This is a clean, correct 1-line fix:

• Before: getattr(self.args, "agent", False) — agent attribute never exists on args, so getattr always returns False, making the checkpoint guard effectively just self.args.all
• After: self.args.subagent — correctly references the --subagent CLI flag defined at cli_cmds/cli.py:134
• Consistent with existing usage in checkpoint.py:144 which already uses getattr(args, "subagent", False)
• The fix ensures checkpointing is properly skipped when running in subagent mode with --all --subagent

Test Coverage

File	Stmts	Miss	Cover
codeflash/optimization/optimizer.py	451	365	19%

• The changed line (607) is not covered by unit tests — it falls within the run() method (lines 528-715) which is integration-level orchestration code
• This is pre-existing — the same line was uncovered on main (just with the old getattr call)
• No coverage regression — this is a 1-line attribute name fix that doesn't change the coverage profile
• 8 test failures observed in tests/test_tracer.py — these are pre-existing and unrelated to this PR (tracer API changes)

Overall Assessment

✅ LGTM — Straightforward bug fix. The old code silently failed (always returned False), this correctly references the actual CLI flag.

---

Last updated: 2026-02-25

[KRRT7]

@aseembits93 you're bumping the benchmark package, not codeflash here

[aseembits93]

@KRRT7 fixed

[KRRT7]

@aseembits93 big jump?

[aseembits93]

@aseembits93 big jump?

i only see a single line of change, i think it was out of sync with main before?

[KRRT7]

oops, I misread the update