Orphan Issue Detection¶

Problem¶

Conductor's dispatch model is pull-based: agents poll GET /api/assignments/next and receive a capability-matched issue. This works well at scale but has a blind spot: if no polling agent matches an issue (e.g. the target repo isn't in any agent's scope), the issue sits in the queue indefinitely with no alert.

Example: dashecorp/rig-memory-mcp#3 was labeled agent-ready on 2026-04-17 and sat four days unclaimed without anyone being paged.

Solution¶

OrphanScanService runs as a background service inside conductor, scanning the issue queue every hour and applying a dead-letter pattern.

Thresholds¶

Age	OrphanStatus	Action
> 24h	`null`	Alert Discord `#admin`, apply `orphan` label, emit `ISSUE_ORPHANED`
> 48h	`"orphan"`	Escalate: alert `#admin` at high urgency, apply `needs-human`, remove `agent-ready`, emit `ISSUE_NEEDS_HUMAN`
Any	`"needs_human"`	No further action — human must decide

Labels¶

Label	Color	Meaning
`orphan`	`#cc317c` (magenta)	No agent has claimed this in >24h
`needs-human`	`#e11d48` (red)	No agent has claimed in >48h; auto-dispatch stopped

Both labels are created automatically in the target repo on first use.

Events¶

Two new domain events are emitted into the issue's stream:

ISSUE_ORPHANED   — first orphan detection (>24h)
ISSUE_NEEDS_HUMAN — escalation (>48h, still unclaimed)

These events update the IssueStatus.OrphanStatus read model field: - "orphan" after ISSUE_ORPHANED - "needs_human" after ISSUE_NEEDS_HUMAN - null when the issue is assigned or re-approved (automatically reset)

Read Model Fields Added to IssueStatus¶

Field	Type	Description
`QueuedAt`	`DateTimeOffset`	When the issue last entered "queued" state. Reset on `IssueApproved` and `IssueUnassigned`.
`OrphanStatus`	`string?`	`null`, `"orphan"`, or `"needs_human"`. Reset to `null` on `IssueAssigned`.

Discord Alerts¶

24h alert (🔶)¶

🔶 Orphan issue detected — `rig-memory-mcp#3`
**"feat: add memory store"**
Queue age: 4.1d — no agent has claimed this issue.
https://github.com/dashecorp/rig-memory-mcp/issues/3
Label `orphan` applied. If no agent picks this up within 24h it will escalate to `needs-human`.

48h escalation (🚨)¶

🚨 ESCALATION — needs human — `rig-memory-mcp#3`
**"feat: add memory store"**
Queue age: 5.0d — no agent has claimed this issue after 48h.
https://github.com/dashecorp/rig-memory-mcp/issues/3
Labels: `agent-ready` removed, `needs-human` applied. Auto-dispatch stopped — human must decide.

Configuration¶

Env Var	Required	Description
`DISCORD_ADMIN_WEBHOOK_URL`	Optional	Webhook for `#admin` alerts. If unset, alerts are logged only.
`GITHUB_TOKEN`	Optional	Token for label operations. If unset, labels are skipped.

Both degrade gracefully — missing env vars disable that action without crashing the service.

Implementation¶

File	Role
`src/ConductorE.Api/Services/OrphanScanService.cs`	Background service — scans queue, applies labels, posts Discord alerts
`src/ConductorE.Core/Domain/Events.cs`	`IssueOrphaned`, `IssueNeedsHuman` event records
`src/ConductorE.Core/Domain/ReadModels.cs`	`QueuedAt`, `OrphanStatus` fields on `IssueStatus`
`src/ConductorE.Api/Adapters/MartenProjections.cs`	Projection handlers for new events and fields
`src/ConductorE.Core/UseCases/SubmitEvent.cs`	`ISSUE_ORPHANED` / `ISSUE_NEEDS_HUMAN` event mapping

Multi-tenancy (PR-4)¶

OrphanScanService runs its scan once per active tenant via ITenantWorkRunner.ForEachTenantAsync: each tenant gets its own DI scope with ITenantContext.TenantId set, so the IIssueQuery read and the ISSUE_ORPHANED / ISSUE_NEEDS_HUMAN SubmitEvent writes bind to that tenant's database. In single-tenant operation today this runs exactly once (for invotek) — behaviour-neutral — but it means a tenant's orphan detection only ever sees and writes that tenant's queue. The service no longer opens its own DI scope; the tenant-scoped provider is supplied by the runner.

Scan Cycle¶

OrphanScanService (hourly)
  └─ for each active tenant (ITenantWorkRunner):
       └─ GetQueueAsync()          ← all state=="queued" issues
       └─ for each issue:
            QueuedAt == default → skip  (pre-feature issues)
            age < 24h           → skip
            OrphanStatus==null  → HandleOrphan()
            OrphanStatus==orphan AND age≥48h → HandleEscalation()
            OrphanStatus==needs_human → skip (human handles it)

Non-Goals¶

Does not switch to push-based assignment (pull model is correct).
Does not implement a capability registry — this is an observability layer, not routing.
Does not auto-reassign orphaned issues to random agents; relaxed-match dispatch is a separate concern.