From qe-framework
Orchestrates parallel execution of independent atomic checklist items by dispatching them to multiple Haiku teammates and merging results. Use when a task has 5+ low-complexity items with non-overlapping file ownership.
How this skill is triggered — by the user, by Claude, or both
Slash command
/qe-framework:Qatomic-runThe summary Claude sees in its skill listing — used to decide when to auto-load this skill
A coordination skill that orchestrates multiple **Haiku Teammates** to execute atomic checklist items in parallel. It acts as the "Lead" session that partitions work and merges results.
A coordination skill that orchestrates multiple Haiku Teammates to execute atomic checklist items in parallel. It acts as the "Lead" session that partitions work and merges results.
Read the TASK_REQUEST and identify items suitable for parallel execution:
Create an Agent Team using the Agent tool:
git commit immediately after completing their specific item.SUMMARY_{Item#}.md under the active plan's phase directory — resolve the slug via .qe/state/current-session.json → .qe/planning/.sessions/{session_id}.json → .qe/planning/ACTIVE_PLAN, then write to .qe/planning/plans/{slug}/phases/{X}/SUMMARY_{Item#}.md. Legacy projects with no slug resolvable write to .qe/planning/phases/{X}/SUMMARY_{Item#}.md.haiku model for maximum speed and efficiency.As Haiku teammates complete their tasks:
SUMMARY_*.md files.After all atomic items are done, determine the next step based on task type:
type: code → trigger /Qcode-run-task for test → review → fix quality loop.type: docs / type: analysis / deletion-heavy tasks → run SIVS Loop verification (VERIFY_CHECKLIST check + supervision) directly, skip /Qcode-run-task.TASK_REQUEST into execution waves. Wave N+1 starts only after Wave N is verified.haiku for teammates. If an item requires Sonnet, it's not "Atomic" and should be handled by standard /Qrt.ContextMemo to ensure teammates have current state without redundant I/O.Before spawning Haiku teammates, check SIVS engine configuration:
.qe/sivs-config.json from the project root (via scripts/lib/codex_bridge.mjs → loadSivsConfig()).implement.engine value:
"claude" (default): Proceed with the standard Haiku swarm execution. No changes."codex": Delegate implementation to Codex via codex-plugin-cc instead of Haiku swarm:
resolveEngine("implement", config) to check availability./codex:rescue with the full TASK_REQUEST checklist as a single task. Codex handles all items internally (no wave splitting needed).detectLegacyConfig(). If non-null, display migration warning.Note: When using Codex engine, wave-based parallelism is not used — Codex handles task partitioning internally. The Verify stage (validation) and quality loop (/Qcode-run-task) still run after Codex completes.
Codex Materialization Check (Mandatory after Codex Done):
Codex may return Done before files are actually written (async companion pattern). The notification hook (notification.mjs) handles initial detection and writes state to unified-state.json under the codex_materialization key.
After every Codex Done, execute this sequence:
Read unified state — check .qe/state/unified-state.json → codex_materialization field:
status: "completed" → files written. Run git diff --stat, proceed to Verify.status: "failed" → report error, offer retry or Claude fallback.status: "running" → poll watcher active, proceed to step 2.Read signal file — cat .qe/agent-results/codex-ready.signal 2>/dev/null:
"detected": true → files written. Run git diff --stat, proceed to Verify."timeout": true → no changes after 1h. Go to step 3.Fallback — use AskUserQuestion:
Results are logged to .qe/agent-results/codex-materialization.md automatically.
Fallback guarantee: Missing .qe/sivs-config.json → all stages default to Claude. Zero impact on existing workflows.
After all Wave items are complete, resolve the active plan's ROADMAP (session binding → .qe/planning/ACTIVE_PLAN → flat fallback) and display execution summary + handoff. Read .qe/planning/plans/{slug}/ROADMAP.md when a slug resolves; fall back to flat .qe/planning/ROADMAP.md for legacy projects. Use the standard handoff format from QE_CONVENTIONS.md (vertical table, [x]/[>]/[ ] markers, single code block, lines under 60 chars).
Execution Complete: {TaskName}
Type: {code / docs / analysis}
Waves: {N}
Items: {X}/{Y} completed
Teammates: {Z}
type: code{slug} · Phase {X}: {PhaseName} — Implementation complete
Roadmap
[x] Phase 1: {Name1}
[>] Phase {X}: {PhaseName}
[ ] Phase {X+1}: {NextName}
PSE: [x] Plan [x] Spec [x] Execute [>] Verify
{TaskDescription — 다음 작업 내용 한 줄 요약}
Next: /Qcode-run-task {UUID}
type: docs / type: analysis / deletion-heavyAfter performing SIVS verification inline (VERIFY_CHECKLIST check + supervision gate):
{slug} · Phase {X}: {PhaseName} — Complete
Roadmap
[x] Phase 1: {Name1}
[>] Phase {X+1}: {NextName}
[ ] Phase {X+2}: {FutureName}
PSE: [x] Plan [x] Spec [x] Execute [x] Complete
{NextPhaseDescription — 다음 Phase 작업 내용 한 줄 요약}
{Next label — 사용자 입력 언어로, 예: "다음:" / "Next:"}: /Qgs {slug}: {짧은 별칭, 최대 6단어}
(Fallback line 금지 — /Qgs는 /Qgenerate-spec의 alias이므로 중복이다. Legacy flat-file projects drop the {slug} · prefix and use /Qgs Phase {X+1}: {짧은 별칭}.)
When all Phases are complete:
All phases done. Finalize with /Qcommit
npx claudepluginhub inho-team/qe-framework --plugin qe-frameworkCreates, edits, and optimizes skills for Claude Code, including drafting, evaluating with test prompts, iterating on performance, and improving skill descriptions for better triggering accuracy.