128 lines
8.2 KiB
Markdown
128 lines
8.2 KiB
Markdown
---
|
|
title: "Prevent idle starvation in Ticket orchestration planning"
|
|
state: 'ready'
|
|
created_at: "2026-06-08T06:12:35Z"
|
|
updated_at: '2026-06-09T11:35:29Z'
|
|
---
|
|
|
|
## Background
|
|
|
|
The current Panel Queue automation mostly handles the transition-time event:
|
|
|
|
```text
|
|
ready -> queued
|
|
-> notify workspace Orchestrator
|
|
```
|
|
|
|
That is not enough for robust orchestration. Queued Tickets can remain after missed notifications, Orchestrator restarts, planning returns, capacity limits, or multi-ticket coordination. The Orchestrator also needs a lightweight way to remember planned queued work across turns without relying only on session memory.
|
|
|
|
There is an existing related Ticket:
|
|
|
|
- `ticket-orchestration-plan-tool`
|
|
|
|
That Ticket asks for a TaskStore-like surface for Ticket ordering/dependency/conflict/capacity/accepted-plan records. This Ticket folds that need together with queued-backlog re-kick semantics into a narrower operational requirement:
|
|
|
|
> If runnable queued work exists and the Orchestrator is otherwise idle, the system should not wait indefinitely for another user instruction. The Orchestrator should be kicked with a bounded work set so it can either incorporate new queued work into the plan or start the next planned queued Ticket.
|
|
|
|
This is starvation prevention and explicit work-set planning, not a constant background scheduler loop.
|
|
|
|
## Goal
|
|
|
|
Implement an Orchestrator attention/re-kick policy and planning store for active Ticket work: distinguish new queued work from planned queued work and accepted in-progress work, persist the plan, and kick the Orchestrator only when work can progress and no active Orchestrator-managed operation is already being waited on.
|
|
|
|
## Planning model
|
|
|
|
The OrchestrationPlan store should distinguish at least:
|
|
|
|
- `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/planning-sync/merge/cleanup progress.
|
|
|
|
The names do not need to become final public API names, but the state distinction is required.
|
|
|
|
## Requirements
|
|
|
|
### Active work set discovery / re-kick
|
|
|
|
- Provide a mechanism to identify Tickets that need Orchestrator attention, including at least:
|
|
- `workflow_state = queued` Tickets not yet present in the OrchestrationPlan (`new_queued`);
|
|
- planned queued Tickets that are not blocked/capacity-limited and can be started when there is no active in-progress work;
|
|
- `workflow_state = inprogress` Tickets accepted by Orchestrator whose next action is not merely waiting for an active coder/reviewer/planning-sync/merge step;
|
|
- queued Tickets left behind after Orchestrator restart, missed notification, or previous capacity stop.
|
|
- On Panel open/Orchestrator restore/spawn, or explicit user action, surface a bounded work list to the Orchestrator when there is actionable work.
|
|
- Avoid unbounded background polling. Prefer explicit events, Panel lifecycle kick, and explicit user/Orchestrator actions.
|
|
- Prevent duplicate starts: re-kick should prompt inspection/planning or acceptance of the next planned item, not blindly start coder Pods.
|
|
|
|
### Re-kick / starvation-prevention semantics
|
|
|
|
- If `new_queued` work exists and the Orchestrator is idle/not occupied by an active in-progress operation, kick or notify the Orchestrator so it can incorporate those Tickets into the plan.
|
|
- If no active `inprogress` work exists and runnable `planned_queued` work exists, kick or notify the Orchestrator so it can accept/start the next planned Ticket rather than waiting indefinitely for user instruction.
|
|
- If active `inprogress` work exists and the next expected event is coder/reviewer/planning-sync/merge completion, do not re-kick merely because queued/planned queued work also exists.
|
|
- If planned queued work is blocked, dependency-waiting, conflict-waiting, or capacity-limited, record the reason so the Panel/user can see why nothing starts.
|
|
- A re-kick is an attention signal plus bounded context, not authority to bypass `queued -> inprogress` acceptance or spawn implementation Pods without inspection.
|
|
|
|
### Orchestration plan record
|
|
|
|
- Provide or define a TaskStore-like but Ticket-domain planning surface for Orchestrator use.
|
|
- The plan should be scoped to Ticket orchestration and support records such as:
|
|
- current active target set;
|
|
- state bucket: `new_queued` / `planned_queued` / `inprogress` or equivalent;
|
|
- ordering: Ticket A before Ticket B;
|
|
- dependency/blocker: A blocks B / B blocked by A;
|
|
- conflict: do not run A and B in parallel;
|
|
- capacity/waiting notes;
|
|
- accepted work plan: worktree/branch/coder/reviewer plan;
|
|
- current next action for each target.
|
|
- Distinguish durable project-relevant routing decisions from local runtime/session claims.
|
|
- Project-relevant decisions should live in Ticket records/thread/artifacts or a typed Ticket orchestration record under project authority.
|
|
- Local Pod/session claims remain in the local role session registry.
|
|
- Records should survive compaction and be queryable by Ticket id/slug and relation kind.
|
|
- Keep the first version lightweight; do not implement a full scheduler/graph solver.
|
|
|
|
### Plan update semantics
|
|
|
|
- The Orchestrator should update the plan at meaningful routing boundaries:
|
|
- new queued work incorporated into the plan;
|
|
- queued -> inprogress acceptance;
|
|
- inprogress -> blocked/waiting/planning/done;
|
|
- capacity stop -> leave planned queued/waiting with reason;
|
|
- merge-ready/done -> mark complete and consider the next planned queued Ticket if no active work remains.
|
|
- Each update should produce a bounded, inspectable record of:
|
|
- what was considered;
|
|
- what was incorporated into the plan;
|
|
- what was accepted/started;
|
|
- what was blocked/deferred/returned to planning;
|
|
- what remains planned queued/waiting.
|
|
- Re-kick should use the current plan/work set so the Orchestrator does not forget leftover queued Tickets between turns.
|
|
|
|
### Relationship to existing work
|
|
|
|
- This Ticket should either subsume or update `ticket-orchestration-plan-tool` so there is one coherent plan/re-kick design.
|
|
- It should coordinate with:
|
|
- `replace-intake-state-with-planning` as a prerequisite that defines the planning lane before this plan/re-kick layer builds on it;
|
|
- `panel-close-done-tickets` for done -> closed handling;
|
|
- local role session registry for active Pod/session ownership;
|
|
- direct/delegation authority work for actual child Pod spawning.
|
|
|
|
## Non-requirements
|
|
|
|
- Do not turn the Panel itself into the scheduler.
|
|
- Do not auto-start unqueued Tickets.
|
|
- Do not re-kick continuously while active coder/reviewer/planning-sync/merge work is already in progress.
|
|
- Do not blindly spawn coder Pods from re-kick without Orchestrator inspection and `queued -> inprogress` acceptance.
|
|
- Do not implement a full dependency graph solver in the first version.
|
|
|
|
## Acceptance criteria
|
|
|
|
- The system can distinguish new queued work, planned queued work, and accepted in-progress work.
|
|
- New queued Tickets are not left unnoticed while the Orchestrator is otherwise idle.
|
|
- Runnable planned queued Tickets are not left unstarted when there is no active in-progress work and capacity/policy allows progress.
|
|
- The system does not re-kick merely because queued/planned work exists while Orchestrator-managed in-progress work is waiting on coder/reviewer/planning-sync/merge completion.
|
|
- Missed/stale queued Tickets can be surfaced to the Orchestrator without requiring the user to manually requeue each one.
|
|
- The Orchestrator can record and query a lightweight Ticket orchestration plan covering active targets, order/dependency/conflict/capacity, state bucket, and next actions.
|
|
- Plan records survive compaction and do not rely solely on session-lifetime TaskStore state.
|
|
- Re-kick/plan updates leave an auditable record of what was incorporated, started, blocked, returned to planning, or left waiting.
|
|
- Duplicate implementation starts are prevented by consulting current Ticket state, local role/session claims, and plan records.
|
|
- Relevant workflows/prompts/docs are updated.
|
|
- Focused tests, `target/debug/yoi ticket doctor`, `cargo fmt --check`, and `git diff --check` pass.
|