v2 — Hierarchical Dispatch
Live session
Project joke-book, session 20260502-055334-470510, May 2, 2026.
The same four-sentence prompt that drove the v1 run — a book on the universal nature of humor — entered the orchestrator again two months later, on a project that had accumulated a memory bank, a skill candidate from a prior attempt, and an installed agent roster. Where v1 satisfied its INTENT end-to-end with the project lead carrying most of the substantive work, v2 is the dispatch-and-memory demonstration: 78 dispatches across a three-level tree, six proxy escalations exercising both modes from §7, and the May 1 attempt's plan consulted as material for May 2's plan.
I would like a book on the universal nature of humor. There are certain types of humor that transcend time, culture and language (e.g. comedy wildlife photos, physical humor, affiliative humor, etc). This is a 5-7 chapter book, targeting armchair enthusiasts, and should explore this phenomena across cultural, temporal, language, belief, and technological boundaries. Thesis: Humor is what unites us.
The session ran ~10.5 hours wall-clock (05:53 → 16:34 UTC, May 2). It traversed the full CfA lifecycle (INTENT → PLAN → EXECUTE → DONE).
What this run demonstrates
Hierarchical dispatch. The dispatch tree is real and three deep: the project lead delegated to 8 writing-leads and 10 editorial-leads; those leads in turn dispatched 13 markdown-writers, 23 voice-editors, 8 fact-checkers, 8 copy-editors, and 2 style-reviewers. 78 dispatches in total. Each child ran in its own git worktree with its own claude session, and the per-parent cap of 3 concurrent children from §13 is observable end-to-end in the timing data — wide fan-outs ran as serialized batches of 3.
Proxy with ACT-R memory. The project's proxy memory — proxy.md, 17 ACT-R chunks accumulated since April 23 — was actively consulted during this session. The consolidation log records 38 ADD and 7 SKIP decisions during the run, meaning 38 new memory entries were written and 7 candidate entries were merged into existing ones. This is the memory layer from the human-proxy system operating in production.
Both escalation modes from §7. Six escalations to the proxy, all under when_unsure policy. One was answered autonomously: the proxy read the worktree, consulted proxy.md, said verbatim "I'll respond from memory.", and issued a clean RESPONSE without ever opening a human dialog. The other five involved 1–3 human turns before the proxy issued its consolidated response. Both modes from the escalation model demonstrated in one session.
Skill-candidate reuse across runs. The May 2 planning agent read skill-candidates/job-20260501-053328-911918--i-would-like-a-book-on-the-uni.md — the prior humor-book attempt's plan, captured as a skill candidate at the end of that run. May 2's PLAN.md is structurally adapted from May 1's: same phase structure, same deliverable shape, refined details. This is the planning skill's SELECT/APPLY path realized in practice — though the formal .active-skill.json artifact wasn't written this run, the consultation itself is on the bus and on disk.
Multiple checkpoints and deliverables. Phase 0 produced manuscript/architecture.md as an explicit pre-drafting checkpoint. Then five chapter drafts (manuscript/ch{1..5}.md), per-chapter research directories under research/, and a structural editorial pass (editorial/architecture-coherence-review.md). The editorial-leads ran six independent passes (style, fact, copy, voice × 3) and the proxy gated three CfA assert points.
By the numbers
| Metric | Value |
|---|---|
| Wall-clock duration | ~10h 41m (05:53 → 16:34 UTC, May 2) |
| Dispatches (total) | 78 |
| Dispatch tree depth | 3 levels |
| Distinct agent types | 9 (joke-book-lead, writing-lead, editorial-lead, markdown-writer, voice-editor, fact-checker, copy-editor, style-reviewer, proxy) |
| Proxy escalations | 6 (1 autonomous, 5 collaborative) |
| ACT-R memory chunks (pre-existing) | 17 |
| Consolidation decisions during run | 45 (38 ADD, 7 SKIP) |
| Phase artifacts | architecture.md + 5 chapter drafts + editorial review + per-chapter research |
| Final state | DONE |
CfA state machine trace
Read directly from the session log:
| Time (UTC) | Event |
|---|---|
| 05:53:34 | SESSION started — prompt accepted |
| 05:53:34 | INTENT phase started |
| 06:45:22 | INTENT phase completed |
| 06:45:22 | INTENT → PLAN |
| 06:45:35 | PLAN phase started |
| 06:58:50 | PLAN phase completed |
| 06:58:50 | PLAN → EXECUTE |
| 06:58:50 | EXECUTE phase started |
| 16:31:18 | EXECUTE phase completed |
| 16:31:18 | EXECUTE → DONE |
| 16:34:45 | SESSION completed — DONE |
Six proxy escalations fired during the lifecycle: at INTENT review, at PLAN review (multiple revision rounds), and at WORK ratification. Details in Proxy & ACT-R.
Where to read next
- Dispatch Tree — the complete dispatch interaction diagram with per-agent counts, plus a spotlight on parallel fan-out under one editorial-lead.
- Proxy & ACT-R Memory — the six escalations broken down by mode, what the proxy read at each one, and the consolidation log from this run.
- Artifacts — every file the session produced, organized by phase: planning, architecture, research, manuscript drafts, editorial review, and the proxy's memory bank.
- Takeaways — what v2 closes that v1 left open, and what's still open.