Background Claude Worktree Absolute Path Launch

Background Claude Worktree Absolute-Path Launch

Use when:

• launching claude -p in the background via Hermes terminal(background=true)
• running overnight worker waves from a git worktree
• the prompt file lives under docs/plans/overnight-prompts/...

Symptom

The background process exits immediately with an error like:

bash: docs/plans/.../worker-N.md: No such file or directory

This can happen even when terminal(..., workdir=...) points at the correct worktree and the prompt file really exists.

Cause

For unattended/background launches, relative prompt-file or log-file paths are less reliable than they look. In a worktree launch, shell expansion / file redirection may resolve the relative path unexpectedly before the process does useful work.

Fix

Always use absolute paths for:

1. the prompt file read into PROMPT=$(< ...)
2. the tee log destination
3. any other launch-critical file arguments

Preferred pattern:

PROMPT=$(< /abs/path/to/worker.md)
claude -p \
  --permission-mode acceptEdits \
  --no-session-persistence \
  --output-format text \
  --max-budget-usd 20 \
  "$PROMPT" </dev/null | tee /abs/path/to/worker.log

Hermes background-process launch rule

When launching a Claude worker from Hermes, do not wrap the shell command with nohup, trailing &, disown, or setsid in a foreground terminal() call. Hermes rejects shell-level background wrappers and cannot track them reliably.

Use the tool's native background mode instead:

terminal(
  background=true,
  notify_on_complete=true,
  workdir="/abs/path/to/worktree",
  command="claude --print --dangerously-skip-permissions < /abs/path/to/prompt.md"
)

For approved issue implementation/landing work where a user asks to "use a subagent", prefer this real background Claude process over delegate_task; it writes to the intended checkout and can commit/push, while delegate_task may lose repo writes. First verify the issue number exists before dispatching; if the requested number is missing but the current active issue is obvious, state the correction in the worker prompt and proceed only on the verified active issue.

Verification step

Immediately poll the background process once after launch.

Healthy early signal:

• process status is running
• no immediate No such file or directory output

Failure signal:

• process status is exited
• output preview shows missing prompt path or missing log path

Also persist a small run manifest immediately after launch (for example logs/overnight/<run>/RUNNING.md or a shell transcript) with:

• issue / terminal mapping
• Hermes process session ID
• OS PID, if available
• prompt file path
• log file path
• expected result artifact path

Hermes process(action="list") can return an empty process list even while direct polling by a known session_id still reports running. Treat the launch-returned session IDs as authoritative and poll them directly. Do not relaunch solely because process list is empty or because Claude logs are still zero bytes; first check direct poll, OS PID liveness, and expected result artifact creation.

Recovery

If the first launch used relative paths and failed:

1. confirm the prompt file exists with a file search or read_file
2. relaunch using absolute prompt and log paths
3. poll again right away before assuming the overnight wave is healthy

Scope

This is a launch-hygiene fix for planning or implementation waves. It does not replace the other unattended-run requirements:

• --permission-mode acceptEdits for write-capable runs
• prompt passed as a positional argument when stdin is closed
• worktree isolation for zero git contention