Hare/yoi
1
1
Fork 0
Code Actions Packages Releases Activity
a40627ab9f
yoi/.yoi/tickets/closed/20260527-194421-pod-orchestration-system-guidance/thread.md

9.1 KiB
Raw Blame History

Created

Created by tickets.sh create.

Plan

Background

Pod notification / notice によって child Pod の完了や状態変化が見えても、現状の assistant はユーザーから明示的に「レビューして」「確認して」と言われるまで自発的に消化しないことがある。

AGENTS.md や workflow に multi-agent の運用は書かれているが、これは知識として読めるだけで、Pod 管理ツールが利用可能な turn における runtime 行動規範としては弱い。特に、自分が spawn した child Pod の完了通知は background signal として扱い、自然な区切りで ReadPodOutput / worktree status / diff / test を確認して次の action に進むべきである。

一方で、notification は non-blocking であり、進行中の user request を不必要に中断してまで消化すべきではない。system instruction には「自発的に follow-up するが、現在の user task を壊さない」というバランスを明示する必要がある。

Requirements

  • Pod management tools が有効な Worker にだけ、Pod orchestration 用の system guidance を注入する。
    • 例: SpawnPod / ReadPodOutput / SendToPod / StopPod / AttachOrRestorePod などが利用可能な場合。
    • Pod 管理 tool がない通常 Worker / child Pod には不要な guidance を出さない。
  • guidance 本文は resources/prompts 配下に置く。
    • prompt 文字列を Rust code に直書きしない。
  • guidance には以下を含める。
    • Pod notification / notice は、自分が処理すべき background signal として扱う。
    • 自分が spawn した child Pod の完了通知を受けたら、自然な区切りで ReadPodOutput を確認する。
    • 委譲 task が完了していれば、報告・worktree status・diff・test 結果を確認し、修正依頼 / merge / ticket 完了処理 / Pod 停止のいずれかに進む。
    • user が明示的に follow-up を要求するまで routine follow-up を放置しない。
    • ただし進行中の user request を不用意に中断しない。
    • output / diff / test を確認せずに完了扱いしない。
  • この guidance は scheduler / auto-maintainer ではない。
    • workflow を勝手に開始しない。
    • project decision / merge / cleanup は既存 workflow と user authorization に従う。
  • notification / PodEvent を context に載せる場合は、既存の history 永続化原則を破らない。
    • turn を跨げない情報を history に残さず system context にだけ差し込まない。

Acceptance criteria

  • Pod management tools が有効な Worker の system prompt に orchestration guidance が含まれる。
  • Pod management tools が無効な Worker には含まれない。
  • prompt 本文が resources/prompts にある。
  • prompt assembly の test で conditional inclusion が確認されている。
  • guidance が user request の中断を促さず、natural stopping point での follow-up を促す文言になっている。
  • cargo fmt --check と関連 crate の test が通る。

Out of scope

  • 自動 scheduler / auto-maintain loop の実装。
  • PodEvent / notification の protocol 変更。
  • spawned Pod registry restore の修正。
  • TUI notification UI の変更。

Comment

Supplemental guidance from dogfooding

Add two explicit rules to the Pod orchestration/system guidance:

  • A spawned Pod completion notification is delivered as a normal background signal. The parent does not need to keep the turn open or call tools solely to wait for it; it is acceptable to finish the current turn and handle the notification at the next natural point.
  • Do not use sleep/polling loops just to wait for a Pod's output. If there is no other useful immediate work, return control to the user instead of blocking the turn; when the notification arrives, read the Pod output then.

Rationale: during multi-agent work, waiting with sleep wastes the active turn and fights the notification model. The desired behavior is notification-driven follow-up, not artificial polling.

Plan

Preflight classification

implementation-ready.

The ticket affects prompt/system guidance and conditional prompt assembly, but the desired product behavior is already specified in the ticket thread and item: include orchestration guidance only when Pod management tools are available, keep the prose in resources/prompts, and explicitly avoid sleep/polling or turn-blocking waits for child Pod output.

Current code map

  • resources/prompts/: prompt text sources; new guidance text should live here.
  • Prompt assembly code/tests: locate the system prompt construction path that already conditionally includes memory/workflow/tool guidance and add a tool-availability gate for Pod orchestration guidance.
  • Tool registry / available tool list: use existing tool availability rather than hard-coding a Worker kind if possible.
  • Existing prompt assembly tests: add inclusion/exclusion coverage for Pod management tools available/unavailable.

Requirements / invariants

  • Guidance is conditional on Pod management tools being available.
  • Guidance is not shown to Workers without Pod management tools.
  • Guidance must not imply an auto scheduler or unauthorized workflow start.
  • Guidance must say notifications can be handled at the next natural point and the parent need not keep the turn open just to wait.
  • Guidance must say not to use sleep/polling loops merely to wait for Pod output.
  • Do not change PodEvent/notification protocol, TUI notification UI, spawned registry restore, or workflow semantics.

Escalate if

  • The only available hook requires injecting notification-derived context without durable history.
  • Conditional tool-availability detection would require broad ToolRegistry redesign.
  • The implementation would change runtime notification delivery or Pod lifecycle behavior instead of prompt guidance.

Validation

  • Focused prompt assembly tests for conditional inclusion/exclusion.
  • Any touched crate tests relevant to prompt loading/assembly.
  • cargo fmt --check.

Review: approve

External reviewer recommendation: approve.

Summary:

  • Static Pod orchestration guidance was added under resources/prompts/common/pod-orchestration.md and registered through the prompt catalog.
  • The guidance is appended to the materialized system prompt only when available tool names include Pod-management capabilities.
  • The gate uses registered tool definitions, not Worker kind, matching the ticket boundary.
  • The prompt explicitly covers background notifications, natural stopping points, not keeping a turn open solely to wait, no sleep/polling loops for Pod output, evidence-before-completion, and no scheduler/authorization bypass.

Intent / requirement mapping:

  • Included when Pod management tools are enabled: satisfied.
  • Omitted when Pod management tools are disabled: satisfied.
  • Prompt body lives under resources/prompts: satisfied.
  • Conditional prompt assembly tests exist: satisfied.

Invariant check:

  • No changes to PodEvent/notification protocol, TUI notification UI, spawned registry restore, Pod lifecycle behavior, scheduler/auto-maintain behavior, or notification-derived context injection.
  • The implementation adds static guidance based on durable tool availability, not transient notification state.

Blockers: none.

Non-blocking follow-ups:

  • A future small test could pin the intended "any Pod-management tool is enough" semantics with a single representative tool.
  • Tool-name class recognition could be centralized later if more prompt gates need it.

Reported validation from coder was considered sufficient:

  • cargo test -p pod pod_orchestration
  • cargo test -p pod prompt::catalog
  • cargo fmt --check

Closed

Merged and completed.

Implementation:

  • Added resource-backed Pod orchestration guidance at resources/prompts/common/pod-orchestration.md.
  • Registered the guidance through the prompt catalog and internal prompt resources.
  • Added conditional system prompt assembly based on registered Pod-management tool names.
  • Guidance is included for Workers with Pod management tools and omitted otherwise.
  • Guidance explicitly says Pod notifications are background signals handled at natural stopping points, that the parent does not need to keep a turn open solely to wait, and that agents should not use sleep/polling loops just to wait for Pod output.
  • Guidance also preserves evidence-before-completion and no scheduler/authorization-bypass constraints.

Review:

  • External reviewer approved with no blockers.

Validation after merge:

  • cargo test -p pod pod_orchestration passed.
  • cargo test -p pod prompt::catalog passed.
  • cargo fmt --check passed.
  • ./tickets.sh doctor passed.