parallel

name	parallel
description	Execute multiple independent backlog tasks simultaneously using subagents with worktree isolation. Use when the user says "run tasks in parallel", "do these at the same time", or when multiple non-overlapping tasks can be worked on concurrently.
argument-hint	["task IDs separated by comma"]
allowed-tools	["Read","Write","Bash","Glob","Grep","Agent"]
model	sonnet
effort	medium

Parallel Tasks

Execute multiple independent backlog tasks simultaneously using subagents with worktree isolation.

Arguments

$ARGUMENTS — comma-separated task IDs (e.g., 1,2,3 or TSK-001,TSK-002). Optional.

Prerequisites

Verify .backlog/ exists.

Workflow

1. Identify Candidates

If $ARGUMENTS provided, parse comma-separated IDs and validate each exists in tasks/pending/ or tasks/active/ and is not blocked.

If no argument:

• List all pending non-blocked tasks
• Parse file references from each task
• Identify independent groups (no overlapping files)

2. Detect File Overlaps (Informational)

Parse each task for file references. Note overlaps — with worktree isolation these won't conflict at runtime but may need manual merge afterward.

3. Present Selection

## Independent Tasks

| ID | Title | Priority | Files | Notes |
|----|-------|----------|-------|-------|

Recommended groups:
- Group A: TSK-001 + TSK-002 (no overlaps)

Which tasks? (comma-separated)

4. Start All Selected Tasks

For each selected task, activate it via CLI:

node "${CLAUDE_PLUGIN_ROOT}/dist/bin/flowstate.js" task-move {{ID}} --to active

The CLI handles frontmatter updates, file moves, and index updates.

5. Load Context for Subagents

Run a single combined search using all unique tags and titles from the selected tasks:

node "${CLAUDE_PLUGIN_ROOT}/dist/bin/flowstate.js" learning-search --tags "{{ALL_UNIQUE_TAGS}}" --query "{{ALL_TITLES_CONCATENATED}}" --limit 5 --json true

Distribute results to subagents by tag overlap: include a learning in a subagent's prompt if its tags or reasons reference terms from that task's title or tags. A single learning may appear in multiple prompts if relevant.

Also scan .backlog/reports/pending/ once for any reports related to the selected tasks' scope.

6. Launch Subagents

Use the Agent tool to launch ALL subagents in a single message for true parallel execution.

Each subagent gets isolation: "worktree".

Subagent prompt:

Complete Task TSK-{{ID}}: {{TITLE}}

## Task Description
{{DESCRIPTION}}

## Acceptance Criteria
{{CRITERIA}}

## Relevant Learnings              ← only if matches found
- LRN-XXX: {{TITLE}}
  {{INSIGHT_SUMMARY}}

## Known Issues                    ← only if related reports found
- RPT-XXX: {{TITLE}} ({{SEVERITY}})

## Instructions
1. Read project documentation (README, CLAUDE.md, etc.) first
2. Apply the learnings above — they capture past mistakes and proven patterns
3. Implement each acceptance criterion
4. Verify changes work (build, lint, test as applicable)
5. **Capture learnings as you go** — whenever you hit a non-obvious root cause, undocumented behavior, gotcha, or a pattern worth reusing, create a learning immediately using the CLI (do NOT wait for task completion):
   ```bash
   cat <<'BODY' | node "${CLAUDE_PLUGIN_ROOT}/dist/bin/flowstate.js" learning-create --title "{{TITLE}}" --tags "{{TAGS}}" --task TSK-{{ID}} --body -
   ## Context
   {{what you were doing}}
   ## Insight
   {{the non-obvious why}}
   ## Application
   {{what to do or avoid next time}}
   BODY

Skip routine work and anything obvious from the code. 6. Create a commit referencing TSK-{{ID}}

### 7. Collect Results

Parallel Execution Complete

Task	Result	Branch/Worktree

Next Steps

• Review changes from each worktree
• /flowstate:complete-task for successful tasks
• /flowstate:block-task for failed tasks

## Error Handling

- Agent fails: block the task via `node "${CLAUDE_PLUGIN_ROOT}/dist/bin/flowstate.js" task-block {{ID}} --reason "{{failure summary}}"`, continue others
- All fail: summarize errors, suggest reviewing task definitions