Conductor Orchestrator

Conductor Orchestrator

The single entry for closed-loop, spec-driven build work. It runs conducty's conductor cycle, but with the one organ conducty lacks: each task is dispatched to the right skill-cluster via the resolver, and deferred clusters are activated on demand. Read conductor-core for the full integration contract; this routes the loop.

The loop (conducty cycle) → spokes

Execute — the integration (THIS is the new organ)

conducty's Execute dispatches to general-purpose. Instead, resolve each task to a cluster and load that cluster's orchestrator into the subagent:

1. Resolve the dispatch plan: node ~/.agents/skill-clusters/scripts/resolve-task.mjs <tasks.md> <plan.md> --json → per task: {cluster, dispatch: <cluster>-orchestrator, tier, activate?, spokes[], confidence}.
2. Classifier proposes, resolver validates (the chosen triage): for each task, propose the cluster/skill from the task intent, then trust the resolver's validation — it rejects phantom clusters (only real ones in skill-index.json), flags low-confidence/unresolved → escalate to human, and marks deferred clusters to activate.
3. Activate deferred clusters the wave touches: node ~/.agents/skill-clusters/scripts/tier.mjs --activate <cluster> --apply.
4. Dispatch a conducty-execute subagent whose loaded skill is the resolved <cluster>-orchestrator (the hub then routes to the spoke on demand). Respect [P] for parallel waves; tracer-first per group.
5. Taste — the right brain (je-ne-sais-quoi). For visual/frontend clusters (creative-frontend · astro · react-native · expo · design · electron · tauri UI), consult the taste engine before the subagent builds: node ~/.agents/skill-clusters/taste/scripts/taste-resolve.mjs "<task>" --brand <brand> --json → inject the returned taste brief (aesthetic classification · top-K on-brand Codrops exemplars + code links · conformance directive · steerable axes) into the subagent's context. The left brain picks the technique; the right brain says how it should feel, on-brand. Skip for non-visual clusters.
6. Gate via PAI hooks (SkillClusterResolver denies non-enumerated skills; fail-closed exit(2) gates; commit-on-criterion). The orchestrator never writes the verdict — it synthesizes the subagents' status. After ship, record the cycle for taste learning: node taste/scripts/taste-feedback.mjs --record ….

Delivery modality — local subagents vs GitHub multi-agent

Before Execute, pick the modality (see conductor-core §4). Propose it from the work shape; let the resolver validate it (work-shape signals + orchestrator availability):

node ~/.agents/skill-clusters/scripts/resolve-task.mjs <tasks.md> <plan.md> --modality github-delivery --json

• local (default) — Execute as above: conducty-execute dispatches subagents loaded with the resolved <cluster>-orchestrator.
• github-delivery — work is issue-tracked / multi-agent / Copilot-dispatchable / team-scale. Hand the loop to the user's two orchestrators (kept in place; full trees in their source repos):
  1. Plan → swarm-architect — decompose into phase→wave→swarm (~80 schema'd tasks), freeze contracts, map tasks to GitHub issues, mark copilot_eligible. (Source: thoughtseed/swarm-architect-skill/ — its runbooks/playbooks/schemas live there, not in the trimmed spoke copy.)
  2. Execute → github-next-wave-orchestrator — repo reality scan → ranked next wave → dispatch each task to its lane: human or autonomous copilot-swe-agent[bot] (via the agent-ready label). resolve-task still resolves each task's cluster, so every worker — human or bot — carries the right capability loadout; the resolver also emits the per-task lane hint (👤/🤖).
  3. Gate → ship-battery.mjs (fail-closed) wraps swarm verification-gates + the next-wave review checkpoint. Never auto-merge a Copilot PR; a second reviewer (not the issue author) approves.
  4. Close → loop-feedback.mjs records the cycle (now with modality + the human/copilot lane split) alongside swarm OpenViking memory capture.

The conductor proposes the modality; resolve-task validates (real orchestrators? signals agree?); a phantom or low-signal mismatch → escalate to the human, exactly like the cluster-name audit.

Standard flow

spec-kit /specify → /plan → /tasks → conductor-orchestrator: shape→plan from the spec → trace → (resolve) execute per cluster → verify/checkpoint → debug (leverage points) → improve (Failure Patterns) → code-review → ship → close: write outcomes to the vault graph so the next cycle is sharper.

Guardrails

See conductor-core. In short: tracer before multiplying; independent verification per task (no self-attested "done"); leverage-point debug (fix at plan level, not code, when 2+ tasks fail); 3-strike circuit breaker; never auto-merge (ship is advisory until PAI gates pass); resolve to real clusters only (phantom-proof); escalate low-confidence tasks to a human.

Loading spokes on demand

This cluster's spokes (the conducty-* loop skills) are not separately enumerated — only this orchestrator and conductor-core are. Load a spoke on demand from ~/.agents/skill-clusters/skills/<spoke-name>/SKILL.md.