ticket: refine orchestrator planning queue
This commit is contained in:
parent
6e77b1cccd
commit
c835c16f1d
|
|
@ -6,9 +6,9 @@ status: open
|
|||
kind: task
|
||||
priority: P1
|
||||
labels: [ticket, orchestrator, planning, workflow, tools]
|
||||
workflow_state: intake
|
||||
workflow_state: 'ready'
|
||||
created_at: 2026-06-07T03:57:10Z
|
||||
updated_at: 2026-06-07T03:57:10Z
|
||||
updated_at: '2026-06-08T06:27:17Z'
|
||||
assignee: null
|
||||
legacy_ticket: null
|
||||
---
|
||||
|
|
|
|||
|
|
@ -5,3 +5,64 @@
|
|||
Created by LocalTicketBackend create.
|
||||
|
||||
---
|
||||
|
||||
<!-- event: decision author: intake at: 2026-06-08T06:09:11Z -->
|
||||
|
||||
## Decision
|
||||
|
||||
## Intake refinement: readiness and boundaries
|
||||
|
||||
Readiness classification: `implementation_ready`, with design preflight recommended before implementation because this introduces a new Ticket/orchestration-domain tool/API surface and durable record format.
|
||||
|
||||
Binding decisions / invariants:
|
||||
- This is a Ticket/orchestration-domain capability, not a generic `TaskStore` replacement and not a scheduler.
|
||||
- Durable project-relevant ordering/dependency/conflict/capacity decisions must survive compaction and be queryable later by Ticket id/slug and relation kind.
|
||||
- Local runtime claims, Pod/session assignment details, sockets, and other ephemeral execution state must remain in the local role-session/runtime registry rather than git-tracked Ticket metadata.
|
||||
- The first version must remain a lightweight note/plan surface; it must not silently introduce automatic topological scheduling or unattended queue reordering.
|
||||
- Orchestrator prompts/workflows should consult these records before accepting queued Tickets, but human Queue and `queued -> inprogress` acceptance semantics remain unchanged.
|
||||
|
||||
Implementation latitude:
|
||||
- The implementer may choose the concrete storage shape within the Ticket backend, such as typed thread events, artifacts, or a small typed record file, if it satisfies queryability, validation, and durability requirements.
|
||||
- The tool naming and exact relation schema can be adjusted during preflight/implementation, provided the supported relation classes remain ordering, dependency/blocker, conflict, capacity/waiting, and accepted-plan summary.
|
||||
- The first implementation can favor explicit CRUD/query operations over graph solving or automatic route selection.
|
||||
|
||||
Escalation conditions:
|
||||
- Escalate before implementing if the design would change Ticket workflow-state semantics, mutate local role-session claims, require a migration of existing Ticket records, grant mutating capabilities to Companion, or make Orchestrator start work without the explicit Queue gate.
|
||||
- Escalate if relation records need to include secret/private workspace details or raw session/model output; Ticket records should store bounded project-relevant decisions only.
|
||||
|
||||
Validation expectations:
|
||||
- Add focused tests for creating/querying relation records by Ticket and relation kind.
|
||||
- Add tests or prompt/workflow coverage showing Orchestrator can consult the records before queued-ticket acceptance.
|
||||
- Run the normal package validation for code/runtime-resource changes, including `nix build .#yoi` before completion.
|
||||
|
||||
---
|
||||
|
||||
<!-- event: intake_summary author: intake at: 2026-06-08T06:09:18Z -->
|
||||
|
||||
## Intake summary
|
||||
|
||||
Existing Ticket `ticket-orchestration-plan-tool` was refined for readiness. It is implementation-ready as a bounded Ticket/orchestration-domain tool surface for recording/querying ordering, dependency/blocker, conflict, capacity/waiting, and accepted-plan decisions. Needs preflight: true, because it introduces a new durable Ticket-domain API/storage surface and touches Orchestrator routing workflow semantics. Risk flags: ticket-domain-tool, persistence, workflow-routing, feature-boundary, compaction. Key invariants: do not replace TaskStore, do not implement a scheduler/graph solver, keep local Pod/session claims out of git-tracked Ticket metadata, preserve the human Queue gate and queued->inprogress acceptance contract.
|
||||
|
||||
---
|
||||
|
||||
<!-- event: state_changed author: intake at: 2026-06-08T06:09:18Z from: intake to: ready reason: intake_ready field: workflow_state -->
|
||||
|
||||
## State changed
|
||||
|
||||
Intake has clarified readiness, boundaries, implementation latitude, escalation conditions, and validation expectations. The Ticket is ready for human queueing and Orchestrator routing/preflight; implementation should not start until the ready Ticket is queued and accepted through the Orchestrator workflow.
|
||||
|
||||
---
|
||||
|
||||
<!-- event: decision author: hare at: 2026-06-08T06:27:17Z -->
|
||||
|
||||
## Decision
|
||||
|
||||
## Dependency ordering: planning lane before OrchestrationPlan tool
|
||||
|
||||
`replace-intake-state-with-planning` owns the workflow cleanup for cases where implementation cannot be planned yet: such Tickets should return to the planning lane with a visible reason instead of remaining in a separate `preflight_needed` state/bucket.
|
||||
|
||||
This Ticket should assume that planning-lane behavior exists before the OrchestrationPlan tool is implemented. The dependency/order should be: settle `replace-intake-state-with-planning` first, then implement the OrchestrationPlan tool on top of the clarified planning/queued/inprogress workflow.
|
||||
|
||||
Do not add a Plan-tool-specific `preflight_needed` state or requirement here. At most, the tool may observe the resulting planning/queued/inprogress Ticket states produced by the workflow model.
|
||||
|
||||
---
|
||||
|
|
|
|||
|
|
@ -0,0 +1,135 @@
|
|||
---
|
||||
id: '20260608-061235-orchestrator-active-ticket-plan-iteration'
|
||||
slug: 'orchestrator-active-ticket-plan-iteration'
|
||||
title: 'Prevent idle starvation in Ticket orchestration planning'
|
||||
status: 'open'
|
||||
kind: 'task'
|
||||
priority: 'P1'
|
||||
labels: ['ticket', 'orchestrator', 'planning', 'panel', 'taskstore', 'queue']
|
||||
workflow_state: 'intake'
|
||||
created_at: '2026-06-08T06:12:35Z'
|
||||
updated_at: '2026-06-08T06:27:33Z'
|
||||
assignee: null
|
||||
legacy_ticket: null
|
||||
---
|
||||
|
||||
## 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, preflight 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/preflight/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/preflight/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/preflight/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/preflight/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/preflight/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.
|
||||
|
|
@ -0,0 +1,50 @@
|
|||
<!-- event: create author: LocalTicketBackend at: 2026-06-08T06:12:35Z -->
|
||||
|
||||
## Created
|
||||
|
||||
Created by LocalTicketBackend create.
|
||||
|
||||
---
|
||||
|
||||
<!-- event: decision author: hare at: 2026-06-08T06:15:37Z -->
|
||||
|
||||
## 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.
|
||||
|
||||
---
|
||||
|
||||
<!-- event: decision author: hare at: 2026-06-08T06:27:33Z -->
|
||||
|
||||
## 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.
|
||||
|
||||
---
|
||||
Loading…
Reference in New Issue
Block a user