| name | batch |
| description | Batch dispatch — read queued tasks from ~/.claude/batch/inbox.md, fix each task's spec contract, fan them out to background subagents per case, and leave the user free until review time. Also surfaces unreported client sessions (~/.claude/batch/unreported.md). Use when the user says "batch", "morning batch", "朝バッチ", "タスク投入", "/batch", "/morning-batch", or at the start of a work day with queued tasks. |
| user-invocable | true |
| allowed-tools | Read, Write, Edit, Glob, Grep, Bash, Agent, Skill |
| argument-hint | [dispatch (default) | status | add <case>: <task>] |
batch
Turns the working day into "invest 15 min in the morning, review at noon and evening". The user's role is dispatcher/reviewer, not implementer.
Files
~/.claude/batch/inbox.md — task queue. One task per line under ## Queue:
- [ ] <case>: <task> | 成果物: <format> | 完了条件: <criteria>
(成果物/完了条件 may be omitted; see Phase 1.)~/.claude/batch/log-YYYY-MM-DD.md — dispatch log for the day~/.claude/batch/unreported.md — sessions in tracked orgs (~/.claude/tracked-orgs.txt) that ended without leaving a report (appended by the SessionEnd hook; client orgs use the sreaas marker, personal orgs can use reflect/learn)~/.claude/batch/kpi.md — weekly KPI table (appended by the weekly sreaas draft routine — see ~/.claude/scheduled-tasks/weekly-sreaas-report/)
Verb: dispatch (default)
- Read
~/.claude/batch/inbox.md. If empty, say so and stop (do not invent work).
- Contract check: for tasks missing 成果物 or 完了条件, infer them when obvious from the task text; otherwise collect ALL open questions into ONE AskUserQuestion round (batched — never one question per task).
- Dispatch: for each task, spawn a background subagent (
run_in_background: true; general-purpose, or Explore for pure investigation):
- Prompt = task + 成果物 + 完了条件 + "write the deliverable to
~/.claude/batch/out/YYYY-MM-DD-<case>-<slug>.md" (for code tasks: work in the case repo, prepare a branch + diff, do NOT push or open PRs)
- Machine verification is mandatory: every dispatch prompt must include "run the repo's mechanical checks (test / lint /
terraform|tofu plan as applicable) and attach the results to the deliverable; a deliverable without verification output is incomplete"
- Model routing (cost/limit stretching):
Explore+haiku for lookups and inventory scans, sonnet for implementation and report drafting; reserve the default (inherited) model for tasks needing judgment
- Client-repo policy applies: no GitHub writes from background agents, ever. Artifacts land locally for human review.
- Independent tasks are dispatched in a single message (parallel). Tasks in the SAME repo working tree must use
isolation: "worktree" or run sequentially.
- Mark dispatched tasks
- [~] in inbox with agent IDs; write ~/.claude/batch/log-YYYY-MM-DD.md (task → agent ID → expected artifact path).
- If
~/.claude/batch/unreported.md is non-empty, list its entries and propose per-session /sreaas:task report (or /sreaas:task report all for the whole iteration).
- Final message: dispatch table (case / task / agent ID / artifact path) + "レビューは昼と夕方に" reminder. Then stop — do not babysit the agents.
Verb: status
Check background task outputs (TaskOutput / artifact files under ~/.claude/batch/out/), summarize per task: done (artifact path) / running / failed (reason). Mark finished tasks - [x] in inbox and move their line to the day's log.
Verb: add <case>: <task>
Append to ## Queue in inbox (create the contract fields inline if given). No dispatch.
Notes
- Friday variant: after dispatch, remind the user to queue next week's tasks (金曜仕込み — prevents the Monday rush).
- Never dispatch more than 6 concurrent background agents; queue the rest and say so.
