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

作成

LocalTicketBackend によって作成されました。

---

Intake summary

Ticket 20260609-101732-001 は implementation_ready。目的は Orchestrator / multi-agent workflow guidance に parallel capacity policy を追加し、独立した queued work が複数あり capacity が空いている場合に、安全確認後は 1 件ずつ待たず並列受理・起動を優先するよう明文化すること。非目標は automatic background scheduler、unqueued Ticket の開始、OrchestrationPlan store/runtime scheduling の変更。既存安全不変条件（queued -> inprogress 前の副作用禁止、blind spawn 禁止、worktree/write-scope 分離、relation blocker 尊重、reviewer read-only 原則）は維持する。関連 Ticket 20260608-061235-001 は broader starvation-prevention/planning work であり、本 Ticket は workflow guidance に限定した non-blocking follow-up として扱う。risk_flags: [orchestration-policy, workflow-guidance, parallelism, authority-boundary, scope-isolation]。未解決の blocking question はない。

---

State changed

Intake により要件・受け入れ条件・非目標・安全不変条件・関連 work を確認した。実装は prompt/workflow resource の focused update と必要な snapshot/test 更新に限定でき、Orchestrator が routing 可能な状態。

---

State changed

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

---

State changed

Accepted queued implementation under the updated parallel-capacity policy. This is workflow guidance for starting independent queued work in parallel, matching the user's current routing preference. It is docs/workflow-oriented and expected conflicts are small/manageable.

---

Decision

Routing decision: implementation_ready_parallel

Reason:

• The user explicitly requested more parallelism when conflicts are absent or small.
• This Ticket updates Orchestrator/multi-agent workflow guidance to encode that same policy.
• It is workflow/docs-oriented and can run alongside active code changes with expected small conflict risk.

IntentPacket:

Intent:

• Update Orchestrator and multi-agent workflow guidance so independent queued work with available capacity is started in parallel after explicit safety checks, rather than waiting one Ticket at a time by default.

Binding decisions / invariants:

• Orchestrator is not an automatic background scheduler.
• Do not start unqueued Tickets.
• No implementation side effects before queued -> inprogress acceptance.
• No blind spawn from queue notifications alone.
• Respect unresolved depends_on / incoming blocker relations, do_not_parallelize, conflict records, workspace dirty state, and shared write-scope constraints.
• Use separate worktrees/branches/write scopes for parallel Coder Pods.
• Reviewer remains read-only unless explicitly scoped otherwise.
• If queued work is left idle while capacity appears available, record a bounded reason: dependency, conflict, capacity, missing planning decision, workspace dirty state, reviewer/coder bottleneck, or human gate.
• Distinguish active work waiting on coder/reviewer output from idle Orchestrator queue-review moments.

Validation:

• Focused workflow/prompt text validation or tests showing parallel start is preferred when safety checks pass.
• Validation that safety invariants remain explicit.
• git diff --check, cargo run -q -p yoi -- ticket doctor, nix build .#yoi if packaged resources/docs are touched.

---

Implementation report

Implementation routing started in parallel with the current active work set.

Worktree/branch:

• Worktree: .worktree/orchestrator-parallel-queued-work
• Branch: orchestrator-parallel-queued-work
• Base/routing commit: 5d9e227 ticket: route analytics batching and parallel guidance

Spawned sibling Coder Pod:

• coder-orchestrator-parallel-queued-work
• Scope: non-recursive read on parent workspace root plus recursive write limited to the child worktree.

Parallel-routing rationale:

• This workflow/docs Ticket encodes the user's updated parallel-capacity preference.
• It is expected to have only small/managed conflicts with active work and no runtime scheduler changes.

---

Implementation report

Coder implementation completed and is ready for sibling review.

Coder Pod:

• coder-orchestrator-parallel-queued-work
• Commit: 492fe06832c30a08a982f469080041a290b1c228 workflow: prefer parallel queued starts
• Worktree status before review: clean branch orchestrator-parallel-queued-work
• Stopped after collecting output to reclaim delegated worktree scope.

Implementation summary:

• Updated Orchestrator routing guidance to prefer starting additional independent queued Tickets during explicit queue review/routing when capacity exists, instead of waiting for the first active Ticket by default.
• Added required checks before parallel start: Ticket body/thread/artifacts, Ticket relations, OrchestrationPlan, dirty state, visible Pods, worktrees/branches, conflict/dependency notes.
• Preserved and strengthened the queued -> inprogress gate before implementation side effects.
• Explicitly disallowed starting unqueued Tickets just to fill capacity.
• Added bounded idle/defer reasons such as dependency, conflict, capacity, dirty workspace, migration boundary, and human gate.
• Updated multi-agent workflow to require separate worktrees/branches/narrow write scopes for parallel Coder Pods and read-only Reviewer default.
• Restated standing merge/validate/close/cleanup policy after reviewer approval, with migration/runtime/human gates as stop conditions.

Changed files:

• .yoi/workflow/ticket-orchestrator-routing.md
• .yoi/workflow/multi-agent-workflow.md

Coder validation reported passed:

• Focused text validation for parallel preference.
• Focused text validation for safety checks / bounded defer reasons.
• Focused text validation for no scheduler / no unqueued start.
• git diff --check
• cargo run -q -p yoi -- ticket doctor
• nix build .#yoi

cargo check --workspace was not run because only workflow resource/docs changed.

---

Review: approve

Review result: approve

Evidence checked:

• Reviewed commit 492fe06832c30a08a982f469080041a290b1c228 on branch orchestrator-parallel-queued-work; HEAD matches the reviewed commit and worktree status is clean.
• Reviewed .yoi/workflow/ticket-orchestrator-routing.md and .yoi/workflow/multi-agent-workflow.md against the Ticket invariants.
• ticket-orchestrator-routing.md now states that queued notification is not an unattended scheduler trigger, that queued -> inprogress must be recorded before worktree/Pod/implementation side effects, and that unqueued Tickets are not capacity-fill candidates.
• The new Parallel acceptance pass requires checking Ticket body/thread/artifacts, TicketRelationQuery, TicketOrchestrationPlanQuery, workspace/worktree dirty state, visible Pods, branches, and conflict/dependency notes before accepting additional queued work.
• The same pass permits parallel start only for independently queued work with no blocking relation/dependency, no applicable do_not_parallelize/conflict, disjoint or low-risk mechanical surfaces, coder/reviewer capacity, committed acceptance records, and separate worktree/branch/scope.
• multi-agent-workflow.md reinforces separate worktree/branch/narrow write scope, reviewer read-only default, and bounded defer reasons when queued work remains idle despite apparent capacity.
• Both workflows explicitly reject scheduler/background-runner/resource-graph-solver/automatic-queue-drain semantics.

Validation run:

• Focused fixed-string validation for scheduler/no-background/no-queue-drain language, queued -> inprogress pre-side-effect acceptance, no unqueued start, required inspection surfaces, bounded idle reasons, separate worktrees/branches/scopes, and reviewer read-only default: passed.
• git diff --check 492fe06832c30a08a982f469080041a290b1c228^ 492fe06832c30a08a982f469080041a290b1c228: passed.
• cargo run -q -p yoi -- ticket doctor: doctor: ok.
• nix build .#yoi: passed.

Residual risks:

• This is workflow guidance only; enforcement still depends on Orchestrator following the documented routing/acceptance checks rather than runtime scheduler constraints.
• cargo check --workspace was not rerun because the reviewed changes are workflow/resource text only and nix build .#yoi passed.

---

Review: approve

Review result: approve

Evidence:

• Review target 492fe06832c30a08a982f469080041a290b1c228 changes only workflow guidance files:
  • .yoi/workflow/ticket-orchestrator-routing.md
  • .yoi/workflow/multi-agent-workflow.md
• Orchestrator routing guidance states queued notifications are not unattended scheduler triggers.
• It requires durable queued -> inprogress acceptance before worktree creation, Pod spawn, or implementation side effects.
• It forbids starting unqueued Tickets merely to fill capacity.
• Parallel acceptance guidance checks Ticket body/thread/artifacts, Ticket relations, OrchestrationPlan, workspace/worktree dirty state, visible Pods, branches, and conflict/dependency notes.
• It explicitly says this is not a scheduler/background runner/resource graph solver/automatic queue drain loop.
• Multi-agent workflow guidance prefers parallel starts for independent queued Tickets when blockers/conflicts are absent, surfaces are disjoint or low-risk, capacity exists, and record commits can be made.
• Parallel Coder Pods require separate worktrees/branches/narrow write scopes; Reviewer remains read-only by default.
• Idle queued work with visible capacity requires a bounded reason.

Reviewer validation:

• Focused fixed-string validation for no scheduler/background runner/queue drain loop, pre-side-effect acceptance, no unqueued start, relation/orchestration-plan/dirty-state/visible-Pod checks, bounded idle reasons, separate worktree/branch/narrow scope, reviewer read-only default.
• git diff --check 492fe06832c30a08a982f469080041a290b1c228^ 492fe06832c30a08a982f469080041a290b1c228
• cargo run -q -p yoi -- ticket doctor
• nix build .#yoi

All passed. Residual note: this is workflow guidance, not runtime scheduler enforcement, which is consistent with the Ticket scope.

---