flower
/
All briefs
idea draft note flower

Dispatched workers that spawn + orchestrate their own sub-agents (solo-playbook-guided) to deliver larger tasked work

Dispatch

canonical · plan

Spec

markdown

hand-off · dispatch

Dispatch

Auto-dispatch

when it reaches planned

Design-loop

design pass before build

Direct dispatch — no refine required. The packet tells the agent to ask questions only if the request is blocked by ambiguity.

kind

No dispatch requests yet — dispatch above to generate a copy-paste packet.

provenance · append-only

Trace

live
or paste a screenshot uploading…
  1. note added 1d ago

    ## Idea (operator, 2026-07-04) When the orchestrator dispatches a brief, it should be able to spawn the worker with explicit instructions + a POINTER to the Solo playbook (`~/Documents/code/_solo-playbook.md`, and the flower/worktrees playbooks) telling that worker to ITSELF spawn and coordinate sub-agents to deliver its tasked work. I.e., a dispatched worker acts as a mini-orchestrator for its own decomposed task — fanning out to sub-agents (Solo MCP `spawn_agent`, and/or Claude Code's own Task/subagent tooling) for parallel implementation, adversarial self-review, breadth search, etc. — then aggregates the result back to the brief. ## Why Some briefs are large/decomposable; one linear agent is slower and burns more of a single context window than a small fan-out. A worker that delegates downward (like the orchestrator delegates) can deliver faster + higher-quality (parallel work, independent review) and keeps each context lean. This is the WORKER-level analog of the orchestrator's group/order/chain discretion (#155). ## To design / decide (this is a design brief — refine before build) 1. **When** to enable sub-agent delegation for a dispatch — size/complexity heuristic (brief priority/label/estimated scope), or a per-dispatch flag the orchestrator sets, or the worker's own judgment. 2. **How the dispatch packet/kickoff points the worker at the playbook** — reference `_solo-playbook.md` (readiness/PTY protocol, delegation loop, cleanup hygiene) + worktree conventions (never edit MAIN, don't run daemons, worktree pin bug #86) so sub-agents inherit the same guardrails. 3. **Guardrails** — a sub-agent spawn budget/cap (avoid runaway fan-out + cost blowups), worktree/cost accounting, and how sub-agent concurrency composes with the orchestrator's own `auto_max_concurrent`. 4. **Provenance roll-up** — the worker aggregates sub-agent output and reports upward via brief_append / brief_dispatch_complete; do sub-agents get their own flower/session provenance, or only the parent worker? 5. **Which harness** — Solo `spawn_agent` sub-agents (separate Solo processes, cross-harness) vs Claude Code Task subagents (in-process, cheaper, no Solo overhead) — likely both, chosen by task shape. ## Relation Complements #155 (orchestrator grouping/chaining discretion) + #199 (auto-dispatch reconciler) — those make the orchestrator delegate autonomously; this makes WORKERS delegate downward. Provenance: operator directive 2026-07-04.

    agent · flower-orchestrator
  2. participant joined 1d ago
    system · flower-orchestrator

epic · dependencies

Relationships

epic parent

depends on

No dependencies — dispatchable once planned.

agents · waves

Participants

  • flower-orchestrator participant · active

trace · graph

Links

No links yet — they accrue as agents work the brief.

scope

Projects

  • flower · primary

dogfood · read-only

Agent’s-eye view

The literal recall_brief payload an agent gets — same service path as the MCP tool.