Execution

OpenClaw uses a worktree-first execution model. preqstation-openclaw prepares an isolated workspace, writes the PREQ prompt to disk, then launches the requested coding agent with a short bootstrap instruction.

Execution Flow

1. Parse message
   → engine key: claude-code / codex / gemini-cli (default: claude-code)
   → task ID (optional)
   → project key / absolute path
   → branch_name (optional)
   → objective
   → progress mode: sparse (default) or live

2. Resolve project_cwd
   → absolute path in message
   → OpenClaw agent memory
   → task-prefix key match

3. Create isolated worktree
   → git -C <project_cwd> worktree add -b <branch_name> <cwd> HEAD

4. Write full PREQ prompt
   → <cwd>/.preqstation-prompt.txt

5. Launch coding agent with bootstrap prompt
   → Claude:  claude --dangerously-skip-permissions "Read and execute instructions from ./.preqstation-prompt.txt ..."
   → Codex:   codex exec --dangerously-bypass-approvals-and-sandbox "Read and execute instructions from ./.preqstation-prompt.txt ..."
   → Gemini:  GEMINI_SANDBOX=false gemini -p "Read and execute instructions from ./.preqstation-prompt.txt ..."

6. Monitor background process
   → list / poll / log / write / submit / kill

7. Clean up worktree
   → git -C <project_cwd> worktree remove <cwd> --force
   → git -C <project_cwd> worktree prune

Prompt Transport

OpenClaw does not pass the full PREQ prompt through argv or stdin.

Instead it:

renders the full prompt into <cwd>/.preqstation-prompt.txt
launches the agent with a short instruction telling it to read that file

This keeps long prompts out of shell history and avoids argument-length problems.

Bootstrap Prompt

All engines get the same idea:

Read and execute instructions from ./.preqstation-prompt.txt in the current workspace.
Treat that file as the source of truth.
If a Task ID is present there, call preq_get_task first, then preq_start_task before substantive work.

Engine Commands

Claude Code

claude --dangerously-skip-permissions "Read and execute instructions from ./.preqstation-prompt.txt in the current workspace. Treat that file as the source of truth. If a Task ID is present there, call preq_get_task first, then preq_start_task before substantive work."

Codex CLI

codex exec --dangerously-bypass-approvals-and-sandbox "Read and execute instructions from ./.preqstation-prompt.txt in the current workspace. Treat that file as the source of truth. If a Task ID is present there, call preq_get_task first, then preq_start_task before substantive work."

Gemini CLI

GEMINI_SANDBOX=false gemini -p "Read and execute instructions from ./.preqstation-prompt.txt in the current workspace. Treat that file as the source of truth. If a Task ID is present there, call preq_get_task first, then preq_start_task before substantive work."

Background Execution

OpenClaw should launch agents with pty:true and background:true by default.

Useful process actions:

process action:list
process action:poll sessionId:<id>
process action:log sessionId:<id>
process action:write sessionId:<id> data:"..."
process action:submit sessionId:<id> data:"..."
process action:kill sessionId:<id>

Lifecycle Handoff

OpenClaw should not redefine the PREQ lifecycle. It hands off task branching to the core preqstation skill.

The only execution-specific rule to remember here is:

if a task ID is present, call preq_get_task() first
call preq_start_task() before substantive work
then follow the branch defined in Task Lifecycle

Environment Constraints

The executing agent:

works only inside the per-task worktree cwd
must not modify the primary checkout directly
must use preq_* tools for lifecycle tracking
must clean up the worktree when finished

Output Format

Success:

completed: <task or N/A> via <engine> at <cwd>

Failure:

failed: <task or N/A> via <engine> at <cwd or N/A> - <short reason>

Example:

completed: PROJ-284 via claude-code at /tmp/openclaw-worktrees/proj/preqstation-proj-rate-limit