From Dev10x
Dispatches parallel worktree-isolated background agents to process GitHub issues/PRs experimentally, testing Write/Edit constraints. Invoke explicitly for parallel testing.
npx claudepluginhub dev10x-guru/dev10x-claude --plugin Dev10xThis skill is limited to using the following tools:
> **EXPERIMENTAL**: This skill tests whether worktree-isolated
Creates isolated Git worktrees for feature branches with prioritized directory selection, gitignore safety checks, auto project setup for Node/Python/Rust/Go, and baseline verification.
Executes implementation plans in current session by dispatching fresh subagents per independent task, with two-stage reviews: spec compliance then code quality.
Dispatches parallel agents to independently tackle 2+ tasks like separate test failures or subsystems without shared state or dependencies.
EXPERIMENTAL: This skill tests whether worktree-isolated background agents can perform Write/Edit operations. The assumptions from PR #584 (GH-549, GH-555, GH-562) may no longer hold. Use this skill to validate and collect evidence.
Announce: "Using Dev10x:fanout-parallel to process [N] work items with parallel worktree agents (EXPERIMENTAL)."
This is a parallel variant of Dev10x:fanout that dispatches
implementation tasks to background agents with worktree
isolation, testing whether the Write/Edit permission
constraints documented in PR #584 still apply.
Key difference from Dev10x:fanout: Where fanout routes
all write-requiring tasks to the main session sequentially,
this skill dispatches them as parallel worktree-isolated
background agents with mode: "dontAsk".
What we're testing:
isolation: "worktree" agents use Write/Edit tools?mode: "dontAsk" enable file operations in subagents?Skill() calls?REQUIRED: Create tasks before ANY work. Execute at startup:
TaskCreate(subject="Scan: discover work items", activeForm="Scanning")TaskCreate(subject="Classify: dependency and conflict analysis", activeForm="Classifying")TaskCreate(subject="Dispatch: launch parallel agents", activeForm="Dispatching")TaskCreate(subject="Collect: gather results and evidence", activeForm="Collecting")TaskCreate(subject="Report: summarize capability findings", activeForm="Reporting")Skip when: Session config exists at .claude/Dev10x/session.yaml.
Otherwise, prompt via AskUserQuestion (ALWAYS_ASK):
Persist to .claude/Dev10x/session.yaml.
Identical to Dev10x:fanout Phase 1. Accept space-separated
list of URLs, issue numbers, or PR numbers. Classify each:
| Pattern | Type |
|---|---|
https://github.com/.../issues/N | github-issue |
https://github.com/.../pull/N | github-pr |
GH-N | github-issue |
#N | github-pr |
Fetch details via mcp__plugin_Dev10x_cli__issue_get or
gh pr view. Create one subtask per discovered item.
Build conflict graph (same as Dev10x:fanout Phase 2):
Present execution plan showing parallel groups.
Supervisor gate:
REQUIRED: Call AskUserQuestion (do NOT use plain text).
Options:
This is where the experiment happens. For each parallel group, dispatch worktree-isolated background agents simultaneously.
For each non-conflicting issue in a parallel group:
Agent(
subagent_type="general-purpose",
description="Implement GH-{N}: {title}",
isolation="worktree",
mode="dontAsk",
run_in_background=true,
prompt="""
EXPERIMENTAL PARALLEL AGENT — Evidence Collection Required
You are testing whether worktree-isolated agents can perform
full implementation work. Your primary goal is implementing
the issue, but you MUST also collect capability evidence.
Issue: {issue_url}
Title: {title}
Description: {body_summary}
## Implementation Instructions
1. Create a feature branch: git checkout -b {branch_name}
2. Read the relevant code to understand the change needed
3. Implement the fix/feature
4. Run tests if applicable
5. Commit changes with proper gitmoji format
## Capability Evidence (REQUIRED)
After each operation, record whether it succeeded or failed.
Return your results in this exact format at the end:
CAPABILITY_REPORT:
- Write tool: {SUCCESS|FAILED|NOT_ATTEMPTED} — {detail}
- Edit tool: {SUCCESS|FAILED|NOT_ATTEMPTED} — {detail}
- Skill() tool: {SUCCESS|FAILED|NOT_ATTEMPTED} — {detail}
- MCP tools: {SUCCESS|FAILED|NOT_ATTEMPTED} — {detail}
- Bash: {SUCCESS|FAILED|NOT_ATTEMPTED} — {detail}
- Read: {SUCCESS|FAILED|NOT_ATTEMPTED} — {detail}
- Glob/Grep: {SUCCESS|FAILED|NOT_ATTEMPTED} — {detail}
- git operations: {SUCCESS|FAILED|NOT_ATTEMPTED} — {detail}
- Branch created: {branch_name or NONE}
- Commits made: {count}
- Files changed: {count}
- Worktree path: {path}
If Write/Edit fails, attempt the Bash workaround:
Bash: printf '%s' 'content' > file.py
Record whether the workaround succeeded.
## Fallback Protocol
If Write/Edit tools are blocked:
1. Record the failure in the capability report
2. Attempt Bash-based file creation as workaround
3. If Bash also fails, record that and return findings only
4. Do NOT hang or retry indefinitely
"""
)
For sequential chains (conflicting items), process in the
main session via Skill(Dev10x:work-on) — same as standard
fanout. The experiment only applies to parallel groups.
For PRs ready to merge (no code changes needed), dispatch as background agents WITHOUT worktree isolation — same as standard fanout.
Before dispatching, create one subtask per agent:
TaskCreate(
subject="Agent: GH-{N} — {title}",
metadata={"type": "parallel-agent", "item": "GH-{N}",
"dispatch_mode": "worktree-isolated"})
REQUIRED: Launch all agents in a single tool-call block so they run concurrently. Do NOT dispatch sequentially.
As each background agent completes, parse its result for:
For each agent result, update the subtask:
TaskUpdate(taskId, status="completed",
metadata={"capability_report": {parsed_report},
"implementation_success": true|false})
If an agent's worktree has uncommitted changes (implementation failed mid-way), note the worktree path for manual recovery.
For agents that succeeded:
Skill(Dev10x:gh-pr-monitor)Synthesize all capability reports into a summary:
| Capability | Agent 1 | Agent 2 | Agent 3 | Conclusion |
|------------|---------|---------|---------|------------|
| Write | ? | ? | ? | ? |
| Edit | ? | ? | ? | ? |
| Skill() | ? | ? | ? | ? |
| MCP tools | ? | ? | ? | ? |
| Bash | ? | ? | ? | ? |
| Read | ? | ? | ? | ? |
| Glob/Grep | ? | ? | ? | ? |
| git ops | ? | ? | ? | ? |
Based on evidence, determine:
Can we re-enable parallel implementation? If Write/Edit
works in worktree agents with mode: "dontAsk", the
Permission-Aware Dispatch table in task-orchestration.md
and Dev10x:fanout can be updated.
Which workarounds are needed? If Write/Edit fails but Bash file creation works, document the reliable workaround pattern.
What hasn't changed? If the constraints still hold, document that PR #584's assumptions remain valid with the current Claude Code version.
Present findings via AskUserQuestion:
git worktree remove <path>.mode: "dontAsk",
background agents likely cannot call Skill(). The experiment
tests this explicitly.