E
Egora
OverviewHow It WorksArchitectureDesignDashboard
Teacher sign inOpen dashboard
E
Egora

A teacher workspace for planning courses, reviewing drafts, and publishing approved work to students.

© 2026 Egora

How It WorksArchitectureDashboardTeacher sign in
Back to ADR index
0008
ADR · Proposed (2026-05-27 grilling session)

Pipelined-sequential dispatch with APPROVED threshold

Date
2026-05-27
Deciders
pdg (HITL during the 2026-05-27 grilling)
Companion design
docs/superpowers/specs/2026-05-27-v2-claude-code-style-redesign.md § 3.3
Related:0004 · V2 baseline0005 · Prompts-as-agents0007 · Evaluator brought back

Context

V2 dispatches sub-agents across four milestones (M1–M4) with multiple artifacts per milestone. Within each parent's children, two dispatch patterns are possible: parallel (all N fan out, fast wall-clock, minimal cross-sibling coherence) or sequential (one at a time, slower, tighter coherence as each reads prior's committed output).

V1 #158's 'per-unit dispatch' chose sequential for M3 specifically. The earlier draft of this redesign recommended parallel everywhere on uniformity grounds. The operator pushed back during the 2026-05-27 grilling:

Operator quote

'M1 (Strategy) is the foundation of all other curriculum (equivalent to the PRD or spec-kit for apps), so i would lean towards methods that produce higher quality content... For M2, while the user is reviewing Unit 1, unit 2 can be generating, decreasing token burn if the direction of units is wrong or if specific units need clarification.'

Decision

Within-parent siblings dispatch sequentially. Next sibling's sub-agent fires when prior sibling reaches APPROVED state (teacher endorsed but not yet finalized). Cross-parent siblings parallelize.

  • D11 — Pipelined-sequential dispatch within-parent siblings.
  • D12 — Trigger threshold: APPROVED (β). Not DRAFT (too aggressive), not FINALIZED (defeats overlap).
  • D13 — Cross-parent siblings parallelize. M3 modules across units fan out; M4 lessons across modules fan out.

Threshold rationale

ThresholdOverlapToken-burn riskVerdict
(α) DRAFTMaximumHigh — teacher may rejectToo aggressive for foundation
(β) APPROVEDSignificantLow — teacher endorsedBalanced — chosen
(γ) FINALIZEDNone — defeats overlapZeroDefeats purpose

The dispatch flow

Parent finalized → sibling-1 dispatched → commit (draft) → teacher reviews → approves → sibling-2 dispatched (while teacher still on sibling-1) → ... → all siblings finalized → parent gate close → next milestone or next parent.

Consequences

Positive
  • • Zero teacher idle time — system always has something for teacher to do AND is always working on next thing.
  • • Lower token burn if teacher changes course mid-milestone.
  • • Pattern uniformity — same dispatch primitive across M1–M4 within-parent.
  • • Cross-sibling coherence by construction — sub-agent reads prior sibling's committed file.
Negative
  • • Wall-clock latency for short courses — 3-unit M2 takes 3× per-unit time, not max.
  • • Implementation complexity — Orchestrator must track sibling state + dispatch on APPROVED.
  • • Historical sibling-cascade handling is retired; current follow-up is surfaced through readiness reports and teacher review.
Neutral
  • • Worker queue still required (async sub-agent dispatch).
  • • Teacher can manually trigger out-of-order dispatch via UI (future affordance).

Migration / execution

  1. 1.Wave R2: implement parent-gated cascade dispatcher in Orchestrator (on artifact → approved, check parent's child list, dispatch next sibling).
  2. 2.Implement worker-queue dispatch hooks.
  3. 3.UI reflects in-flight state — Progress tab + agent stream show 'sibling-N generating' while teacher reviews sibling-(N-1).

Open questions

  • Configurable threshold — advanced teachers opt-in to DRAFT-threshold (max overlap) per course. Defer to V2-P3.
  • Rapid-fire mode — bulk parallel within-parent for teachers who want to skip individual reviews. Defer to V2-P3.
  • Teacher-driven out-of-order dispatch — 'skip to sibling-N' before completing N-1. Defer to R4.
← Previous
0007 · Revert the V2 LLM-Evaluator deletion (commit 0e0703d)
Next →
The full spec
ADR index