Agent Orchestra

Agent Orchestra

Agent Orchestra coordinates exactly five commander-led groups for high-stakes repo work. Each commander owns one deterministic manifest, may launch bounded sub-agents, collaborates through append-only ledgers, writes live telemetry for Agent Pulse, and emits a terminal bundle for Shadow Score and Fleet Scorecard synthesis.

Use user-facing terminology sub-agents. The compatibility ledger filename child-agents.jsonl may appear in paths, but responses should say sub-agents.

Command Grammar

Default repo is the current working directory. Mission text after : is required for a launch.

Defaults

• model_tier: premium
• scale: max
• commander_count: 5
• dashboard: Agent Pulse + Stampede monitor

Do not ask tier, scale, or dashboard questions when absent. Ask only one question at a time for missing launch inputs:

1. Ask for mission when no text after : was provided.
2. Ask for repo only when no on REPO was provided and the current working directory is not the intended target or is not a usable git repository.

Every Agent Orchestra metaswarm run launches exactly five commanders: commander-001 through commander-005. Never silently downscale. If five commanders cannot launch, mark the run partial or failed and explain why.

Model Tiers

Premium rotation:

claude-opus-4.7,gpt-5.5,claude-opus-4.6,gpt-5.4,claude-opus-4.5,claude-sonnet-4.6,gpt-5.3-codex,claude-sonnet-4.5

Standard rotation:

claude-sonnet-4.6,gpt-5.4,claude-sonnet-4.5,gpt-5.3-codex

Never silently use cheap or mini models for Agent Orchestra sub-agents:

claude-haiku-4.5,gpt-5.4-mini,gpt-5-mini,gpt-4.1

SQL Tracking

Initialize these tables on every invocation:

CREATE TABLE IF NOT EXISTS agent_orchestra_runs (
    run_id TEXT PRIMARY KEY,
    repo_path TEXT NOT NULL,
    mission TEXT NOT NULL,
    model_tier TEXT NOT NULL,
    scale TEXT NOT NULL,
    commander_count INTEGER NOT NULL,
    status TEXT DEFAULT 'running',
    created_at TEXT DEFAULT (datetime('now')),
    updated_at TEXT DEFAULT (datetime('now'))
);

CREATE TABLE IF NOT EXISTS agent_orchestra_events (
    run_id TEXT NOT NULL,
    ts TEXT DEFAULT (datetime('now')),
    event TEXT NOT NULL,
    detail TEXT
);

Preflight

Before launch, verify the local hardened runtime exists:

command -v copilot
command -v tmux
command -v python3
command -v git
test -x ~/bin/stampede.sh
test -x ~/bin/stampede-monitor.sh
grep -q "stampede_synthesize_commander_bundle" ~/bin/stampede.sh
grep -q "Your exact commander task is pre-claimed" ~/bin/stampede.sh
grep -q "requires exactly 5 commanders" ~/bin/stampede.sh

If the launcher or monitor is missing, stop and tell the user to run the Agent Orchestra installer from https://github.com/DUBSOpenHub/agent-orchestra. If a hardened-launcher marker is missing, stop before launch and reinstall; stale launchers can start panes without deterministic pre-claim or terminal bundle synthesis.

Run Layout

Agent Orchestra stores coordination inside the target repo:

REPO/.stampede/RUN_ID/
  state.json
  fleet.json
  queue/
  claimed/
  results/
  commanders/commander-###/
    manifest.json
    context-capsule.json
    assignments.json
    swarm-state.json
    child-agents.jsonl
    bundle.json
    atoms/
    logs/
  collab/
    protocol.json
    proposals.jsonl
    reviews.jsonl
    improvements.jsonl
    consensus.jsonl
    broadcasts.jsonl
  shadow-score/
    sealed/criteria.json
    seal.sha256
  orchestrator-commentary.json
  orchestrator-commentary.jsonl

Launch Workflow

1. Parse repo, mission, model tier, and scale.
2. Create run-YYYYMMDD-HHMMSS under REPO/.stampede/.
3. Write state.json with product: "agent-orchestra", profile: "metaswarm", commander_count: 5, model tier, repo path, and mission.
4. Create a sealed Shadow Score envelope before commanders start. Hash shadow-score/sealed/criteria.json into shadow-score/seal.sha256.
5. Create the collaboration bus ledgers: proposals.jsonl, reviews.jsonl, improvements.jsonl, consensus.jsonl, and broadcasts.jsonl.
6. Ensure exactly five commander manifests exist and each has matching task_id and commander_id.
7. Launch Terminal Stampede with metaswarm mode:

STAMPEDE_OBJECTIVE="$MISSION" \
STAMPEDE_METASWARM_NO_GATES=1 \
STAMPEDE_MODEL_POLICY="$MODEL_TIER" \
STAMPEDE_PREMIUM_MODEL_POOL="$MODEL_POOL" \
STAMPEDE_BANNED_CHILD_MODELS="claude-haiku-4.5,gpt-5.4-mini,gpt-5-mini,gpt-4.1" \
STAMPEDE_SQUAD_LEADS_PER_COMMANDER=50 \
STAMPEDE_WORKERS_PER_SQUAD_LEAD=5 \
STAMPEDE_WORKERS_PER_COMMANDER=250 \
~/bin/stampede.sh \
  --metaswarm \
  --run-id "$RUN_ID" \
  --count 5 \
  --repo "$REPO_PATH" \
  --models "$MODEL_POOL" \
  --no-attach

8. Open Agent Pulse scoped to the repo:

osascript -e 'tell application "Terminal" to do script "cd ~/copilot-cli-agent-pulse && AGENT_PULSE_SCAN_ROOTS='"$REPO_PATH"' python3 agent_pulse.py --no-splash"' \
  -e 'tell application "Terminal" to activate'

Status Output

For status, read:

• orchestrator-commentary.json
• fleet.json
• commanders/*/swarm-state.json
• commanders/*/child-agents.jsonl
• collab/*.jsonl
• results/commander-*.json

Use this compact shape:

AGENT ORCHESTRA — EXECUTION
RUN run-...  STATUS running|partial|failed|stopped

Commander       Model              Phase                Sub-agents        State
commander-001   claude-opus-4.7    launching_workers    120/250          active
commander-002   gpt-5.5            complete             250/250          success

Totals: cmd 2/5 active · sub-agents 42 running / 370 done / 620 seen · results 3/5
Decision: wait | stop | synthesize partial | relaunch

Expected terminal bundle count is always exactly five. If fewer than five result bundles exist, report running, partial, or failed; do not begin final synthesis.

Teardown

For teardown:

~/bin/stampede.sh --teardown --run-id "$RUN_ID" --repo "$REPO_PATH"

Report process cleanup, tmux session state, queue/claim counts, result bundle count, and final run state.

Final Synthesis and Shadow Score

After exactly five results/commander-*.json bundles exist:

1. Verify each result filename, task_id, and commander_id match.
2. Verify each commander has launch proof in swarm-state.json and child-agents.jsonl; missing proof prevents success.
3. Load collaboration ledgers.
4. Recompute and verify the sealed Shadow Score hash.
5. Score requirement coverage, collaboration quality, evidence quality, test/validation impact, and synthesis usefulness.
6. Write shadow-score/scorecard.json.
7. Produce a concise final answer with commander status, collaboration stats, Shadow Score summary, best findings, caveats, and next action.

Do not expose sealed criteria while a run is active.