Hare/yoi
1
1
Fork 0
Code Actions Packages Releases Activity
255a212af9
yoi/.yoi/tickets/00001KTJXS31R/thread.md

28 KiB
Raw Blame History

Created

Created by LocalTicketBackend create.

Decision

Scope refinement: avoid idle starvation, not constant iteration

The phrase "continuously iterate" should not mean that the Orchestrator constantly re-kicks itself regardless of current work. The desired behavior is narrower:

  • Do not leave newly queued work unnoticed while the Orchestrator is otherwise idle.
  • Do not leave planned queued work unstarted when there is no in-progress work and capacity/policy allows starting it.
  • Do not re-kick just because queued work exists if the Orchestrator is already waiting for an active coder/reviewer/preflight/merge step.

Refined planning model:

  • new_queued: Tickets with workflow_state = queued that have not yet been incorporated into the OrchestrationPlan.
  • planned_queued: queued Tickets that the Orchestrator has considered and placed into an explicit plan/order/waiting set, but has not yet accepted as inprogress.
  • inprogress: Tickets accepted by the Orchestrator and currently awaiting worktree/coder/reviewer/preflight/merge/cleanup progress.

Desired scheduling/re-kick semantics:

  1. If there is new_queued work and the Orchestrator is idle/not occupied by an active in-progress operation, it should be kicked or shown a bounded work list so it can incorporate the new Tickets into the plan.
  2. If there is no inprogress work and there is planned_queued work that is not blocked/capacity-limited, the Orchestrator should be kicked to start/accept the next planned Ticket rather than waiting indefinitely for user instruction.
  3. If there is active inprogress work whose next expected event is coder/reviewer/preflight/merge completion, do not re-kick merely because queued/planned work exists. The Orchestrator should wait for the active work or an explicit user action/capacity decision.
  4. If planned queued work is blocked or waiting for capacity, the plan should record that reason so the panel/user can see why nothing starts.

This should be framed as starvation prevention and explicit work-set planning, not a background scheduler loop.

Decision

Dependency ordering: planning lane before re-kick planning layer

replace-intake-state-with-planning owns the behavior for cases where work cannot be planned/implemented yet because requirements or preflight are missing: return the Ticket to the planning lane with a visible reason.

This Ticket assumes that behavior exists and builds the new_queued / planned_queued / inprogress re-kick model on top of the clarified workflow states. Implementation order should be: settle replace-intake-state-with-planning first, then implement the OrchestrationPlan/re-kick layer.

Therefore, the OrchestrationPlan/re-kick layer should not introduce its own separate preflight_needed bucket or cleanup logic.

Decision

Intake refinement: ready scope for Orchestrator routing

既存 Ticket の本文と thread を確認し、関連 work と現在の Ticket lifecycle vocabulary に合わせて、残りの実装対象を以下に整理する。

Scope after related work

  • この Ticket は、Orchestrator が queued work を見落とさず、かつ active work を待っている最中に無駄な再 kick を繰り返さないための active work set discovery / re-kick policy を実装対象にする。
  • OrchestrationPlan record/tool surface は既に別 Ticket で実装済みの前提として扱い、この Ticket で plan store 自体を再実装しない。
  • 既存本文・thread の workflow_state という語は現在 schema の authoritative field ではない。実装では Ticket frontmatter の state を authority とし、既存記録中の workflow_state = queued などは historical wording / conceptual alias として読む。

Binding decisions / invariants

  • new_queued / planned_queued / inprogress は新しい core Ticket state として追加しない。必要な work-set classification は、現在の Ticket state、OrchestrationPlan records、role/session claim、visible Pod/worktree state から導出する。
  • queued は Orchestrator が routing/start-if-unblocked を検討できる状態であり、implementation side effects は必ず queued -> inprogress が記録された後に限る。
  • Panel や local lifecycle hook が Orchestrator に attention / kick を与えることはできるが、unattended scheduler loop や常時 polling にはしない。
  • Orchestrator が active coder/reviewer/preflight/merge/cleanup 等の完了待ちである間は、queued/planned work が存在するだけでは再 kick しない。
  • planned_queued work を開始しない理由が capacity / dependency / conflict / dirty workspace 等で説明できる場合は、bounded reason を OrchestrationPlan record または既存の適切な durable artifact に残し、Panel/user が理解できるようにする。
  • duplicate start を避けるため、Ticket state、role/session claims、Plan records、visible Pod/worktree state を確認してから attention/kick/acceptance を行う。

Implementation latitude

  • どの runtime/panel boundary で idle Orchestrator を検出して kick するか、どの view-model/helper に work-set derivation を置くか、bounded payload の具体形は実装側で選んでよい。
  • 既存 OrchestrationPlan types で waiting/capacity reason が足りない場合、最小の typed extension は検討してよい。ただし core Ticket lifecycle state の増殖や旧 workflow_state 復活は避ける。
  • Panel 表示は state-first / heuristic-free の原則を維持する。title text、labels、thread event の有無、Pod 名 substrings を lifecycle authority にしない。

Acceptance criteria

  • Orchestrator が idle/not occupied のとき、new queued work を検出して plan incorporation または bounded work-list attention に進める。
  • Inprogress work が存在せず、planned queued work が unblocked かつ capacity/policy 上開始可能なとき、Orchestrator が次の acceptance/routing を行える。
  • Active inprogress operation の完了待ち中は、queued/planned work の存在だけで再 kick しない。
  • 開始しない planned queued work には、ユーザーが確認できる bounded waiting/blocking reason が残る。
  • 既存の human gate、queued -> inprogress acceptance step、dirty-workspace/dependency/conflict/capacity checks を迂回しない。
  • duplicate Orchestrator/coder/reviewer/worktree start を起こさない。

Validation

  • nix build .#yoi を通す。
  • Ticket / panel / orchestrator routing 周辺の既存テストまたは追加テストで、少なくとも idle queued detection、active-work wait suppression、waiting reason recording、duplicate-start prevention の主要分岐を確認する。
  • 実装報告では、どの authority を work-set classification に使ったか、implementation side effects が queued -> inprogress 後に限定されていることを明示する。

Readiness

  • readiness: implementation_ready
  • risk_flags: [orchestration-policy, panel-lifecycle, persistence, role-session, authority-boundary, duplicate-start]
  • blocking open questions: none

この refinement により、Orchestrator はこの Ticket を implementation candidate として routing できる。

Intake summary

Ticket 00001KTJXS31R は implementation_ready。残り範囲は既に実装済みの OrchestrationPlan record/tool surface を前提にした active work set discovery / Orchestrator re-kick policy。new_queued / planned_queued / inprogress は新しい core Ticket state ではなく、現在の Ticket state、OrchestrationPlan records、role/session claims、visible Pod/worktree state から導出する分類として扱う。Panel/lifecycle hook は idle Orchestrator に bounded attention を与えるが、unattended scheduler loop や常時 polling にはしない。queued -> inprogress acceptance 前の implementation side effects、blind spawn、duplicate start は禁止。active coder/reviewer/preflight/merge/cleanup 待ち中は queued/planned work の存在だけで re-kick しない。risk_flags: [orchestration-policy, panel-lifecycle, persistence, role-session, authority-boundary, duplicate-start]。blocking open questions はない。

State changed

Intake refinement により、既存の plan store 実装との差分、current state vocabulary、binding invariants、implementation latitude、validation focus が整理され、Orchestrator が routing できる状態になった。

State changed

Ticket を workspace-panel が queued にしました。

Decision

Routing decision: leave queued for now (waiting_capacity_note)

Reason:

  • This queue review already accepted and delegated three in-progress branches: 00001KTVJFT6F (Panel focus model), 00001KTTW04W2 (Companion progress notification), and 00001KTVJGC0Y (Ticket language guidance).
  • This Ticket concerns Orchestrator idle queued re-kick / active work-set behavior and overlaps conceptually with the active Companion progress notification and Panel lifecycle work. Starting it now would raise duplicate-start / scheduler-boundary / integration-conflict risk rather than safe independent parallelism.
  • Relation check: dependency relation target 00001KTG3MDFG is closed, so this is not blocked by an unresolved relation; the current stop reason is capacity/conflict with active work.

Evidence checked:

  • Ticket body/thread: active in-progress suppression, clean workspace gate, no queue-drain loop, active-work-set snapshot requirements, and escalation conditions.
  • TicketRelationQuery: two relation records checked; depends_on target 00001KTG3MDFG is closed.
  • TicketOrchestrationPlanQuery: no prior records; added waiting_capacity_note orch-plan-20260612-145604-1.
  • Workspace/Pod state: Orchestrator worktree clean; active sibling coder Pods are running for Panel focus, Companion progress notification, and Ticket language guidance.

Next action:

  • Keep queued and reconsider after at least Panel focus / Companion progress notification are merged or blocked, so this Ticket can validate against the current Panel/notification behavior without duplicate scheduler semantics.

Decision

Routing decision: implementation_ready

Reason:

  • 前回の waiting reason は、Panel focus / Companion progress notification / Ticket language guidance の active in-progress work と重なる capacity/conflict だった。現在それらは merge/close/cleanup 済みで、visible child Pod は 0、Orchestrator worktree は clean。
  • Relation blocker はない。depends_on target 00001KTG3MDFG は closed 済みで、incoming related は non-blocking context である。
  • Ticket body/thread は、session-lifetime work set discovery / re-kick policy、active_inprogress suppression、duplicate-start prevention、no scheduler/no polling/no blind spawn、queued -> inprogress acceptance gate を binding decisions として明確にしている。
  • risk flags は orchestration-policy / panel-lifecycle / persistence / role-session / authority-boundary / duplicate-start だが、bounded context check の結果、具体的な未決定 design/API/authority 判断は残っていない。実装方式は Panel/lifecycle/runtime boundary と local session work-set representation の local tactic に閉じている。
  • OrchestrationPlan に accepted plan orch-plan-20260612-154541-2 を記録済み。

Evidence checked:

  • Ticket body / thread / artifacts: requirements, acceptance criteria, binding decisions, previous waiting_capacity_note, and accepted plan を確認。
  • TicketRelationQuery: depends_on target 00001KTG3MDFG closed、incoming related non-blocking。
  • Related Ticket: 00001KTG3MDFG closed。OrchestrationPlan record/tool surface は実装済みで、この Ticket では再実装しない前提を確認。
  • Recent merged context: Panel focus model、Companion weak progress notify、Ticket language guidance は closed/merged/cleaned up 済み。
  • Code map: crates/tui/src/workspace_panel.rs, crates/tui/src/multi_pod.rs, crates/tui/src/role_session_registry.rs, Ticket state/action surfaces, Pod visibility/state surfaces, recent companion_progress weak notify path を確認。
  • Workspace/Pod state: Orchestrator branch orchestration/yoi-orchestrator is clean; visible child Pods 0。

IntentPacket:

Intent:

  • Orchestrator が idle で actionable queued work を見落とすことを防ぐ bounded starvation-prevention layer を実装する。
  • active_inprogress が導出されている間は queued/planned work があるだけでは re-kick しない。
  • re-kick は inspection / work-set incorporation / next planned acceptance を促す attention であり、scheduler や blind implementation start ではない。

Binding decisions / invariants:

  • new_queued / planned_queued / active_inprogress / actionable_inprogress は core Ticket state に追加しない。Ticket state、session-lifetime work set、role/session claims、visible Pod/worktree state、OrchestrationPlan records から導出する分類に留める。
  • Work set は session-lifetime runtime state であり、Ticket ごとの durable artifact log として積まない。durable に残すべき判断だけ Ticket comment / state transition / OrchestrationPlan record に残す。
  • 常時 polling / unattended scheduler loop / queue drain loop を作らない。
  • queued -> inprogress acceptance なしに coder/reviewer Pod spawn、worktree 作成、implementation side effect を行わない。
  • active coder/reviewer/planning-sync/merge/cleanup 等の完了待ち中は queued/planned work の存在だけで re-kick しない。
  • duplicate Orchestrator/coder/reviewer/worktree start を防ぐ。
  • Panel は authority/backend/scheduler にならず、bounded attention surface に留まる。

Requirements / acceptance criteria:

  • Orchestrator Pod が Idle で new_queued work があるとき、bounded work-list attention または session work set incorporation に進める。
  • active_inprogress がなく、planned queued work が unblocked/capacity-allowed のとき、next acceptance/routing に進める。
  • active_inprogress が導出される間は queued/planned work の存在だけで re-kick しない。
  • 開始しない planned queued work には session work set 上でユーザーに提示できる bounded waiting/blocking reason を保持する。
  • Ticket state、session work set、role/session claims、visible Pod/worktree state を使って duplicate start を避ける。
  • session work set が失われても queued Ticket から安全に再検出・再 inspection できる。

Implementation latitude:

  • Panel open / Orchestrator restore/spawn / explicit user action のどの boundary で idle detection と attention payload を組むかは実装側で選んでよい。
  • Session work-set の具体的な runtime representation は既存 local role/session registry や Panel app state に合わせて選んでよい。
  • Bounded attention payload の具体形は実装側で選んでよいが、full Ticket thread / unbounded Pod output / hidden context-only injection は避ける。
  • OrchestrationPlan types の最小拡張は必要なら検討してよいが、core Ticket lifecycle state や旧 workflow_state は復活させない。

Escalate if:

  • Durable scheduler state / persistent queue runner / polling loop が必要になる。
  • queued -> inprogress acceptance gate を迂回しないと実装できない。
  • Duplicate start prevention に新しい global lock/lease authority が必要になる。
  • Panel が lifecycle authority / scheduler になる必要が出る。
  • Role/session claims から active_inprogress を安全に導出できない。

Validation:

  • idle Orchestrator + queued detection。
  • active_inprogress 導出時の re-kick suppression。
  • planned queued waiting reason retention。
  • duplicate-start prevention。
  • session work set loss から queued Ticket 再検出。
  • relevant cargo test for tui / panel / role session / ticket routing paths。
  • cargo fmt --check。
  • git diff --check。
  • cargo run -p yoi -- ticket doctor または同等。
  • nix build .#yoi。

Current code map:

  • crates/tui/src/workspace_panel.rs: Ticket rows/actions and state-first Panel view model。
  • crates/tui/src/multi_pod.rs: Panel lifecycle, Companion/Orchestrator interactions, recent progress notice / header behavior。
  • crates/tui/src/role_session_registry.rs: local role/session claims。
  • crates/pod-store, crates/pod-registry, and protocol Pod status surfaces only if needed for visible Pod/worktree state derivation。
  • Ticket relation / OrchestrationPlan tool surfaces are already implemented and should be reused/read, not reimplemented as plan store。

Critical risks / reviewer focus:

  • No scheduler / queue drain loop / background polling。
  • No blind worktree/Pod spawn before queued -> inprogress acceptance。
  • active_inprogress suppression works and avoids re-kick while coder/reviewer/merge/cleanup is active。
  • duplicate start prevention uses actual Ticket state + claims + visible Pods/worktrees, not heuristic title/labels。
  • session work set is runtime/session scoped, not a new durable per-Ticket artifact log。
  • Recent Companion weak progress notify and Panel focus behavior remain coherent.

Next action:

  • queued -> inprogress を記録してから、branch ticket/orchestrator-idle-queued-rekick / worktree /home/hare/Projects/yoi/.worktree/orchestrator-idle-queued-rekick を作成し、sibling coder に narrow write scope で実装を委譲する。Reviewer は coder evidence 後に read-only で起動する。

State changed

Routing accepted for implementation. Ticket body/thread, dependency relation target, incoming non-blocking relation, previous waiting-capacity note, current Orchestrator workspace state, visible Pods, recent merged Panel/Companion/Ticket-language context, and relevant Panel/lifecycle/role-session code map were rechecked. No unresolved blocker or missing planning decision remains. Implementation side effects will start only after this accepted queued -> inprogress transition is recorded, using accepted plan orch-plan-20260612-154541-2.

Plan

Implementation worktree created for multi-agent handoff.

  • Ticket: 00001KTJXS31R
  • Branch: ticket/orchestrator-idle-queued-rekick
  • Worktree: /home/hare/Projects/yoi/.worktree/orchestrator-idle-queued-rekick
  • Base: Orchestrator workspace HEAD e72a4536 (orchestration/yoi-orchestrator), including the accepted routing record and queued -> inprogress transition.
  • Scope rule: implementation work happens only in this child worktree. Root/original workspace must not be read, written, validated, merged, cleaned, or used for git operations.
  • Next: spawn sibling coder with narrow write scope to the implementation worktree. Reviewer will be started after coder evidence is available.

Plan

Coder delegated.

  • Coder Pod: yoi-coder-idle-queued-rekick
  • Worktree: /home/hare/Projects/yoi/.worktree/orchestrator-idle-queued-rekick
  • Branch: ticket/orchestrator-idle-queued-rekick
  • Scope: write /home/hare/Projects/yoi/.worktree/orchestrator-idle-queued-rekick; read child worktree plus minimal non-recursive runtime-root read required by spawn validation.
  • Task: implement bounded idle-Orchestrator queued-work attention / session work-set rediscovery without scheduler/polling/queue-drain behavior, preserve queued -> inprogress acceptance gate, suppress re-kick while active_inprogress work exists, prevent duplicate starts, and add focused tests plus required validation.
  • Constraint: coder was instructed not to use root/original workspace for decision evidence or git/validation/cleanup, and not to create generated memory/local/runtime/log/lock/secret-like .yoi paths.

Implementation report

Coder completed implementation.

Coder Pod:

  • yoi-coder-idle-queued-rekick

Implementation branch/worktree:

  • Branch: ticket/orchestrator-idle-queued-rekick
  • Worktree: /home/hare/Projects/yoi/.worktree/orchestrator-idle-queued-rekick
  • Commit: d2fae81a tui: add idle queued orchestrator attention
  • Worktree status: clean

Changed surfaces:

  • crates/tui/src/multi_pod.rs
  • resources/prompts/panel/orchestrator_idle_queue_notice.md

Implementation summary:

  • Adds session-scoped OrchestratorWorkSet derived at Panel reload boundary.
  • Derives active_inprogress, new/actionable queued, planned queued, waiting reasons, and duplicate guards from Panel Ticket rows, local role-session claims, related Pods, blocked reasons, and live Orchestrator status.
  • Sends bounded Orchestrator attention only when Orchestrator is live/reachable/idle and queued work is actionable.
  • Suppresses attention when active_inprogress exists or duplicate-start guard indicates existing claim/Pod/worktree activity.
  • Retains waiting reason / planned queued information in Panel header detail for session-visible explanation.
  • Adds prompt resource resources/prompts/panel/orchestrator_idle_queue_notice.md for the attention payload, preserving the prompt-resource boundary.
  • Does not add scheduler loop, polling loop, queue drain loop, core Ticket states, or persistent work-set artifact store.

Implementation note for reviewer attention:

  • Coder report used the phrase workflow_state for TicketPanelEntry.workflow_state; this is existing Panel model naming over the typed Ticket state enum, not reintroduction of obsolete Ticket frontmatter workflow_state.

Orchestrator-verified evidence from child worktree:

  • git status --short --branch: clean on ticket/orchestrator-idle-queued-rekick
  • cargo test -p tui queued_attention: pass, 3 tests
  • cargo test -p tui planned_queued_prompts: pass, 1 test
  • cargo test -p tui rediscovered_queued_work: pass, 1 test
  • cargo test -p tui active_inprogress_suppresses: pass, 1 test
  • cargo test -p tui idle_orchestrator_gets_bounded_attention: pass, 1 test
  • cargo test -p tui workspace_panel: pass, 12 tests
  • cargo check -p tui: pass
  • cargo fmt --check: pass
  • git diff --check orchestration/yoi-orchestrator...HEAD: pass
  • cargo run -p yoi -- ticket doctor: doctor: ok
  • nix build .#yoi: pass

Known broad-suite note from coder:

  • cargo test -p tui still fails unrelated existing/neighboring expectation tests:
    • multi_pod::tests::orchestrator_launch_context_uses_orchestration_root_for_runtime_workspace
    • spawn::tests::profile_choices_use_project_registry_default
    • spawn::tests::profile_choices_include_builtin_and_project_default_marker
  • Focused tests for this Ticket pass; reviewer should confirm broad failures are unrelated and not introduced by this Ticket.

Next:

  • Spawn read-only sibling reviewer against recorded intent, binding decisions/invariants, acceptance criteria, commit d2fae81a, diff, validation evidence, and known broad-suite failures.

Plan

Reviewer delegated.

  • Reviewer Pod: yoi-reviewer-idle-queued-rekick
  • Target branch/worktree: ticket/orchestrator-idle-queued-rekick at /home/hare/Projects/yoi/.worktree/orchestrator-idle-queued-rekick
  • Commit under review: d2fae81a tui: add idle queued orchestrator attention
  • Scope: read-only child worktree plus minimal non-recursive runtime-root read required by spawn validation.
  • Review basis: recorded IntentPacket, binding decisions/invariants, acceptance criteria, implementation summary, diff, focused validation evidence, nix build evidence, and known broad-suite failures.
  • Reviewer focus: no scheduler/polling/queue-drain behavior, no acceptance-gate bypass, active_inprogress suppression, duplicate-start prevention, session-scoped work set, prompt-resource boundary, recent Panel/Companion behavior compatibility, and robust focused tests.

Review: approve

Approve.

Review target:

  • Worktree: /home/hare/Projects/yoi/.worktree/orchestrator-idle-queued-rekick
  • Commit: d2fae81a tui: add idle queued orchestrator attention

Evidence:

  • 差分は crates/tui/src/multi_pod.rs と resources/prompts/panel/orchestrator_idle_queue_notice.md のみ。
  • queued work への通知は idle かつ reachable な Orchestrator Pod に限定されている。
  • Method::Notify { auto_run: true } は bounded attention / re-kick として使われるだけで、Ticket 状態変更・Pod spawn・worktree 作成・queue drain は追加されていない。
  • Scheduler / queue drain / background worker は追加されていない。既存 Panel reload 後に fingerprint で重複抑制された通知を送る構成。
  • queued -> inprogress acceptance は bypass されていない。通知 payload は Ticket state verification と通常 acceptance 記録を明示し、implementation side effect は行わない。
  • active_inprogress suppression は実装されている。TicketWorkflowState::InProgress が存在する場合、queued attention は送られず、waiting reason として session work set に残る。
  • Duplicate-start guard は title/label ではなく、Ticket id に紐づく local claim / related visible Pods / worktree 表示情報から導出している。
  • Session work set は MultiPodApp 内の runtime/session-scoped state のみで、永続 per-Ticket artifact は追加されていない。
  • Prompt 本文は resources/prompts/panel/orchestrator_idle_queue_notice.md に置かれており、Rust 側は include/render の境界に留まっている。
  • workflow_state 命名は既存の typed TicketWorkflowState / Panel internal naming の利用であり、この差分で obsolete frontmatter authority を再導入しているものではない。

Reviewer-run validation:

  • cargo test -p tui queued_attention -- --nocapture: pass, 3 tests
  • cargo test -p tui planned_queued_prompts -- --nocapture: pass
  • cargo test -p tui rediscovered_queued_work -- --nocapture: pass
  • cargo test -p tui idle_orchestrator_gets_bounded_attention -- --nocapture: pass
  • cargo test -p tui active_inprogress_suppresses -- --nocapture: pass
  • cargo test -p tui workspace_panel -- --nocapture: pass, 12 tests
  • cargo check -p tui: pass
  • git diff --check orchestration/yoi-orchestrator...HEAD && cargo fmt --check: pass

Known broad-suite failures:

  • cargo test -p tui: expected broad-suite failure, 312 passed / 3 failed.
  • Failing tests:
    • multi_pod::tests::orchestrator_launch_context_uses_orchestration_root_for_runtime_workspace
    • spawn::tests::profile_choices_use_project_registry_default
    • spawn::tests::profile_choices_include_builtin_and_project_default_marker
  • The two spawn failures are in untouched crates/tui/src/spawn.rs.
  • The orchestrator_launch_context... failure is in multi_pod.rs, but this diff does not touch build_orchestrator_launch_context or that test area; it remains the existing/neighboring expectation mismatch reported by coder. Not a blocker for this Ticket.

Non-blocking risks:

  • Failed delivery is not fingerprint-marked, so a repeatedly reachable-but-rejecting socket can retry on subsequent Panel reloads; this is consistent with existing notice behavior but may be noisy.
  • Duplicate guard depends on the Panel’s existing visible claim/pod/worktree derivation, so it prevents duplicate starts for locally visible state but is not a full global scheduler/lock. This matches the Ticket’s bounded Panel-attention scope.