On failure (non-zero exit, or "status":"error" in the JSON): if the message is empty plan content on stdin, retry the save exactly once with a proper quoted-heredoc call. On a repeated error — or any other error — print ⚠️ <error.message> and stop. Never retry endlessly.

If file editing is unavailable in the current mode, print ⚠️ Agent cannot write files in this mode. Switch to a mode with write access. and stop.

Steps

Context. Run .agents/skills/pflow-task-plan/scripts/plan-context.sh → JSON {status,current_task,previous_tasks,next_tasks}. Plan only current_task. Use previous_tasks (up to 2 just completed) and next_tasks (up to 2 upcoming) only to understand what came before and what follows — never plan or pull in those tasks.
Clarify. Resolve anything ambiguous before planning. Answer from the codebase and project docs when you can; only ask the user about genuine decision points. Ask one question at a time, walk each decision branch to its end, and offer a recommended answer for every question (via AskUserQuestion). If the user passed extra detail when invoking the skill, treat it as task input.
Plan. Decompose the task into steps small enough for another agent to execute correctly. Write the plan as Markdown with the sections in Format below.
Save. Save the plan with one non-interactive Bash call, feeding the full Markdown to plan-save.sh through a quoted heredoc → JSON {status,plan_filename,bytes}:
```
.agents/skills/pflow-task-plan/scripts/plan-save.sh <<'__PFLOW_PLAN_EOF__'
<full plan Markdown>
__PFLOW_PLAN_EOF__
```

pflow-task-plan

Steps