Orchestrator Mode

⚠️ LEGACY — 2026-07-26 · NOT RECOMMENDED FOR NEW WORK

This skill is retired upstream and is shipped for reference and reversibility, not for use.

Do not invoke it against a live API key. Its cloud path issues a real, billable client.beta.messages.create request using the advisor_20260301 tool with betas=["advisor-tool-2026-03-01"], and hardcodes "model": "claude-opus-4-7". That beta has not been exercised since April 2026, so whether it still exists server-side is unknown, and the model pins throughout this skill (claude-opus-4-7, claude-sonnet-4-6, claude-haiku-4-5) are superseded. An invocation may bill your account for a request that cannot succeed.

Why it is here at all. The tier heuristic below is the durable part and is model-agnostic:

1. Can a smart intern do this in one step? → C
2. Does it need synthesis, structure, or multiple steps? → B
3. Does it involve strategy, irreversibility, or cross-cutting impact? → A
4. Would getting it wrong cost $10K+ or set a wrong direction for months? → A+

Plus two disciplines worth keeping regardless of which models you run: plan before act — the stronger model advises on the plan, the executor carries it out — and treat an advisor budget as a hard ceiling, not a suggestion.

What replaced it upstream. Modern agent harnesses route internally, so an explicit classify-then-dispatch layer has no caller left. The advisor role became a dedicated reviewer subagent; gate-and-escalate became a standing review norm. If you want the capability, wire the heuristic into whatever harness you already run rather than reviving this dispatcher.

Retired 2026-07-26. Scripts are preserved unmodified so nothing you may have built on them breaks.

Orchestrator Mode

Sequential Chain Routing | Quality-Gated Escalation | Cost-Minimizing

Orchestrator mode takes a compound task, decomposes it into subtasks, and routes each subtask to its optimal tier via advisor-mode. Between steps, a quality gate checks whether the output satisfies the subtask's acceptance criteria. If the gate fails, the subtask re-dispatches one tier up. If the gate passes, the chain continues.

Core principle:

Route around failure, don't pick the "best" model. Orchestrator keeps every subtask on the cheapest tier that passes, and only escalates when a specific gate trips.

This is the complement to council-mode. Council compares models on the same task; orchestrator runs a chain where each step picks its own tier and earns its escalation.

BRAIN CHECK — Run Before Every Orchestrator Call

Brain root: resolved from $CEREBRO_ROOT env var or ~/.cerebro/profile.yaml.

1. Verify advisor-mode is installed and provider-aware
2. Local fleet (optional): if fleet_primary is set, verify Ollama is reachable. Cloud-only deployments skip this check — all tiers route to cloud.
3. Verify the task decomposes into ≥2 discrete subtasks (otherwise use advisor-mode directly)

WHEN TO USE ORCHESTRATOR MODE

Do NOT use orchestrator for: single-shot tasks, tasks with no clear gate condition, tasks where all steps need frontier reasoning uniformly.

CHAIN MODEL

TASK → DECOMPOSE → [Step 1: cheapest tier] → GATE → PASS → [Step 2: cheapest tier] → GATE ...
                                                      → FAIL → ESCALATE (next tier up) → GATE

Tier ladder (default — cloud-only):

With local fleet: Fast tier = local Ollama model (cheapest); Haiku is mid; Sonnet is frontier. Configure via orchestrator.local_tier: true in ~/.cerebro/profile.yaml.

INVOCATION

# Standard chain (cloud tiers)
python3 {brain_root}/skills/orchestrator-mode/Scripts/orchestrator_run.py "<compound task>"

# With local fleet as fast tier
python3 {brain_root}/skills/orchestrator-mode/Scripts/orchestrator_run.py --local "<compound task>"

# Custom gate threshold
python3 {brain_root}/skills/orchestrator-mode/Scripts/orchestrator_run.py \
  --gate-threshold 0.85 "<compound task>"

OUTPUT FORMAT

Each step emits:

[STEP N] <subtask description>
  Tier: <fast|mid|frontier>
  Gate: PASS | FAIL
  Escalated: yes | no
  Output: <subtask output>

Final chain summary:

=== ORCHESTRATOR CHAIN SUMMARY ===
Steps: N   |   Escalations: M   |   Cost tiers used: fast=X mid=Y frontier=Z
Deliverable: <final stitched output>

PROFILE CONFIGURATION

orchestrator:
  local_tier: false             # true = use Ollama as fast tier
  local_host: ""                # fleet_primary value
  gate_threshold: 0.80          # quality gate pass threshold (0–1)
  max_escalations: 3            # HALT after N escalations in single chain

SCOPE CONTRACT