Overnight Worktree Uv Warmup And Log Path Guardrails

Overnight worktree uv warmup and log-path guardrails

Use when launching unattended Codex runs in fresh worktrees, especially for Python repos that rely on uv and for nested repos like digitalmodel.

Trigger

• Overnight/background Codex batch in isolated worktrees
• Fresh worktree or fresh workspace
• Commands will run uv run ...
• You want reliable log capture with tee

Problem patterns observed

1. uv first-run warmup can look like a hang

• In a fresh worktree, the first meaningful uv run ... can spend 55-60 minutes compiling/importing before useful output appears.
• This can make a healthy worker look blocked or dead.
• We observed this in the digitalmodel overnight lane: the real work eventually completed and all tests passed, but the first-use warmup consumed most of the apparent runtime.

2. Buffered Codex logs can stay empty for a long time

• Codex -p ... | tee <log> may produce no visible log growth for a long period even when the process is healthy.
• Empty logs are not sufficient evidence of failure.

3. tee can fail if you create the wrong log directory

• mkdir -p logs && ... | tee /abs/path/to/logs/run.log is unsafe if the actual tee target differs from the cwd-relative logs/ you created.
• In the digitalmodel overnight lane, the command completed the task but still exited non-zero because tee could not open the intended log path.

Recommended launch pattern

1. Pre-warm uv in the exact worktree

Before the real workload, run a disposable warmup in the same worktree:

uv run python -c "pass"

Use this especially when:

• the worktree is fresh
• the repo is large
• the lane is test-heavy
• the first useful command is a long pytest run

2. Create the exact log directory for the exact log file

Do not rely on mkdir -p logs unless tee writes to ./logs/... in that same cwd.

Safe pattern:

LOG=/abs/path/to/logs/run.log
mkdir -p "$(dirname "$LOG")"
PROMPT=$(< /abs/path/to/prompt.md)
Codex -p \
  --permission-mode acceptEdits \
  --no-session-persistence \
  --output-format text \
  --max-turns 80 \
  "$PROMPT" </dev/null | tee "$LOG"

3. Monitor health by process state and artifacts, not log growth alone

Preferred signals:

• background PID still alive
• expected output/result artifacts appear
• worktree git status --short changes as files are written

Only treat the run as failed after checking those, not merely because the log is empty.

Recovery pattern when a run exits non-zero after doing useful work

If the background command ends with a logging-related error:

1. inspect the worktree, not just the exit code
2. check git status --short
3. inspect target files and test outputs
4. verify whether the actual task completed despite the shell/logging failure
5. if yes, continue from the resulting repo state and post the correct GitHub update

Practical rule

For overnight worktree batches:

• pre-warm uv
• use absolute log paths
• create the exact parent directory of the log file
• do not interpret empty Codex logs as immediate failure

These guardrails reduce false hang diagnoses and prevent task-success/logging-failure confusion in unattended runs.