yoi/.yoi/tickets/20260609-101732-001/item.md

title	state	created_at	updated_at	assignee	queued_by	queued_at
Orchestratorに独立queued workの並列開始を促す	inprogress	2026-06-09T10:17:32Z	2026-06-09T10:36:44Z	null	workspace-panel	2026-06-09T10:31:16Z

背景

現行の Orchestrator workflow は安全側の制約が強く、並列稼働は許可されているが積極的に促されていない。

現状の強い指示:

• Queue notification だけで blind spawn しない。
• queued -> inprogress acceptance 前に worktree / coder Pod / reviewer Pod を起動しない。
• dependency / conflict / capacity / risk / current workspace state を確認する。
• OrchestrationPlan は automatic scheduler ではなく、Orchestrator が読んで明示判断する。

一方で、以下の指示が弱い:

• independent queued Tickets が複数あり capacity が空いているなら、1件ずつ完了待ちするのではなく並列に受理・起動する。
• idle capacity を残さないよう active work set を更新する。
• Orchestrator は safety checks を満たした範囲で parallel work を優先する。

結果として、Orchestrator が保守的に 1 Ticket ずつ処理し、coder/reviewer capacity を使い切らない傾向がある。

ゴール

Orchestrator / multi-agent workflow guidance を更新し、独立した queued work があり capacity が空いている場合は、明示的な safety checks を満たした上で並列稼働を優先するようにする。

要件

• ticket-orchestrator-routing に parallel capacity policy を追加する。
• multi-agent-workflow の並列運用 guidance を、注意事項だけでなく積極的な scheduling/acceptance stance として補強する。
• Orchestrator は automatic scheduler ではないが、明示的に起動された routing / queue review / panel kick の中では idle capacity を残さないように判断する。
• 複数 queued Tickets がある場合、Orchestrator は以下を確認した上で並列開始を優先する:
  • each Ticket is queued and human-authorized for routing;
  • each Ticket has no unresolved depends_on / incoming blocking relation;
  • no do_not_parallelize / conflict / shared write-scope constraint blocks concurrent work;
  • separate worktree / branch / write scope can be created;
  • coder/reviewer capacity is available and trackable;
  • each Ticket can be accepted with queued -> inprogress before side effects;
  • existing active inprogress work is not waiting for a bottleneck that makes new work unsafe.
• If parallel start is not chosen despite available queued work, Orchestrator should record a bounded reason:
  • dependency;
  • conflict;
  • capacity;
  • missing planning decision;
  • workspace dirty state;
  • reviewer/coder bottleneck;
  • user/human gate.
• OrchestrationPlan records may help express waiting/capacity/conflict decisions, but they are not automatic scheduler authority.
• The guidance should distinguish:
  • active work that is waiting on coder/reviewer completion, where re-kick should not thrash;
  • idle Orchestrator with queued/planned work, where starting another independent Ticket is preferred.
• Keep all existing safety invariants:
  • no implementation side effects before queued -> inprogress;
  • no blind spawn from notification alone;
  • no shared write scope between coder Pods;
  • reviewer remains read-only unless explicitly scoped;
  • dependency/conflict relations are respected.

非目標

• Implementing an automatic background scheduler.
• Starting unqueued Tickets.
• Ignoring relation blockers or workspace conflicts.
• Implementing OrchestrationPlan store changes.
• Changing Pod runtime scheduling.

受け入れ条件

• Orchestrator workflow says independent queued work with available capacity should be started in parallel after explicit checks, rather than waiting one Ticket at a time by default.
• Multi-agent workflow includes active parallel work-set management guidance, not only safety notes.
• If Orchestrator leaves queued work idle while capacity appears available, it records a reason.
• Safety invariants around queued -> inprogress, worktree isolation, scope separation, relation blockers, and review loop remain explicit.
• Focused prompt/workflow tests or snapshot tests are updated if present.
• target/debug/yoi ticket doctor and git diff --check pass.