Orchestration Map

By Yad Konrad — @0bserver07

Hierarchy of sessions and the wave-by-wave timeline.

Topology

graph TD
    Yad["Yad<br/>(terminal)"]
    Orch["orchestrator<br/>63285119<br/>40 Yad-typed prompts<br/>1026 assistant turns"]
    Team[("team: schmidhuber-impl<br/>TeamCreate × 1")]
    Workers["worker sessions × 58<br/>one per stub<br/>spawned via Agent(team_name=...)"]
    Audits["Explore audits × 15<br/>1 initial survey<br/>12 per-wave audits<br/>2 BUILD_NOTES extracts"]
    PRs["13 wave + meta PRs"]

    Yad -->|prompts| Orch
    Orch -->|1 TeamCreate| Team
    Orch -->|58 Agent dispatches| Workers
    Orch -->|15 Agent dispatches| Audits
    Orch -->|69 SendMessage| Workers
    Workers -->|one branch per stub| PRs
    Audits -->|verdict| PRs
    PRs -->|merged in 90s burst| Yad

    classDef root fill:#fff3e0,stroke:#e65100,color:#000
    classDef orch fill:#e3f2fd,stroke:#1565c0,color:#000
    classDef team fill:#f3e5f5,stroke:#6a1b9a,color:#000
    classDef worker fill:#e8f5e9,stroke:#2e7d32,color:#000
    classDef audit fill:#fce4ec,stroke:#ad1457,color:#000
    classDef output fill:#fff9c4,stroke:#f57f17,color:#000
    class Yad root
    class Orch orch
    class Team team
    class Workers worker
    class Audits audit
    class PRs output

The orchestrator created one persistent team (schmidhuber-impl) via TeamCreate and routed work to teammates via Agent(team_name=schmidhuber-impl, name=<stub>-builder, ...) calls. The orchestrator’s SendMessage calls were used to nudge specific teammates mid-build.

• TeamCreate calls in orchestrator: 1
• Agent dispatches: 73
• SendMessage calls: 69
• TaskCreate / TaskUpdate (orchestrator’s own todo tracking): 15 / 34

Wave-by-wave timeline (UTC)

One wave, sequenced

sequenceDiagram
    autonumber
    participant Y as Yad
    participant O as Orchestrator
    participant W as Workers (N parallel)
    participant A as Audit (Explore)
    participant G as GitHub

    Y->>O: "trigger wave N" (sometimes implicit)
    O->>W: Agent × N (one per stub, into team schmidhuber-impl)
    par each worker, isolated worktree
        W-->>W: build stub, commit LOCAL ONLY
    end
    W->>O: SendMessage(summary)
    O->>A: Agent (Explore, audit all wave-N stubs)
    A->>O: verdict
    O->>G: assemble wave/N branch + open wave PR
    O->>W: SendMessage(shutdown_request) × N
    Note over O,Y: orchestrator continues to wave N+1

Every wave followed the same protocol:

1. Orchestrator emits N parallel Agent calls (one per stub) into the schmidhuber-impl team.
2. Workers each build their stub in a sandboxed worktree.
3. After workers finish, orchestrator launches one Explore audit agent that reads all stubs and flags inconsistencies.
4. Orchestrator opens the wave PR.
5. Move to the next wave.

The interesting bit: PRs were all merged in a tight burst on 2026-05-08 15:49–15:50 UTC, after every wave was built and audited. The build itself ran across 2026-05-06 23:09 → 2026-05-08 14:51 UTC, ~40 hours wall-clock.

Sidechain note

The orchestrator session has 0 sidechain turns: in-process subagent traffic (mostly the Explore agent reads). Worker sessions appear as separate JSONLs in the project directory because agent-teams spawns them as distinct Claude Code instances, even though their cwd is still SutroYaro/.