map-reduce

name	map-reduce
description	Process a list of items by dispatching one sub-agent per item in parallel (up to max_parallel concurrent), then aggregate results into a final summary. Same per-item sub-task and final summary contract as for-each — only the dispatch order differs. Only invoke when the user explicitly asks for the map-reduce skill or pattern by name. Do not auto-select for general tasks.
metadata	{"author":"Yiheng Tao","version":"0.0.1"}

Map-Reduce (Parallel)

Use when items are fully independent and throughput / wall-clock time matters. If items must be processed in order, or later items depend on earlier results, use for-each instead. The per-item sub-task and final summary contracts below are identical between the two skills.

Workflow

1. Identify the item list. If the user did not provide it, ask once via ask_user. Confirm before proceeding if there are more than 5 items.
2. Dispatch all per-item sub-agents concurrently, up to max_parallel (default 5) at a time. As each one settles, dispatch the next pending item until every item has been mapped.
3. After every dispatched mapper has settled (success or recorded error), produce the final summary.

Per-item sub-task (identical to for-each)

Dispatch one sub-agent per item with:

• A stable item id plus enough metadata to act on it (path, URL, identifier).
• The user-provided task description, scoped strictly to that item.
• An expected result shape: short structured output with key findings, evidence pointers, and a confidence note.
• Instruction: do not summarize across items, do not process items outside the assigned one, and on unrecoverable failure return { id, error } instead of raising.

Use claude-sonnet-4.6 or gpt-5.4 for mappers. Prefer the explore agent for read-only work and general-purpose for mutations.

Final summary (identical to for-each)

Dispatch a single summary sub-agent (or produce inline) that reads every per-item result and writes a markdown report containing:

• A header line "N of M items succeeded".
• A table or sectioned list keyed by item id with key findings.
• A dedicated "Errors" section listing every { id, error } result.
• Actionable next steps if any.

Use claude-opus-4.6 for the summary. Present the final report to the user.