Core Call Chain

Every Commander execution follows a structured pipeline:

1. Deliberation

CLI / HTTP / API
  │
  ├─ deliberation.ts     ← "What kind of task is this?"
  │   └─ TaskComplexityAnalyzer

The deliberation engine analyzes the task's complexity, dependency graph, and domain requirements to determine the optimal execution strategy.

2. Effort Scaling

  ├─ effortScaler.ts     ← "How many agents?"

Based on complexity, Commander scales across 1–20 agents. Simple tasks get a single agent; complex research tasks get a team.

3. Topology Routing

  ├─ topologyRouter.ts   ← "Which topology fits?"

Selects the optimal execution topology from 5 canonical options:

• SINGLE — Simple tasks, one agent
• CHAIN — Dependent steps, chain-of-thought (legacy: SEQUENTIAL)
• DISPATCH — Independent subtasks, max throughput (legacy: PARALLEL)
• ORCHESTRATOR — Lead agent delegates to specialists (legacy: HIERARCHICAL / HYBRID)
• REVIEW — Generate → critique → refine loop (legacy: DEBATE / ENSEMBLE / EVALUATOR-OPT)

4. Atomization

  ├─ atomizer.ts         ← "Break into subtasks"

ROMA-style decomposition splits the task into atomic, dependency-aware subtasks.

5. Execution

  ├─ agentRuntime.ts.execute(ctx)
      │
      ├─ acquireSlot()              ← Concurrency semaphore
      ├─ [Tenant check]             ← Rate limit + concurrency quota
      ├─ resolve tenant storage     ← Per-tenant isolation
      │
      ├─ [Retry loop: 0..maxRetries]
      │   ├─ callWithTimeout()      ← LLM provider call
      │   ├─ [Tool execution loop]
      │   │   ├─ toolCache.get()    ← SHA-256 hash lookup
      │   │   ├─ planner.plan()     ← Dependency-aware plan
      │   │   ├─ executeTool()      ← StepErrorBoundary
      │   │   └─ toolCache.set()    ← Cache result
      │   ├─ verification.check()   ← 5 quality gates
      │   └─ checkpoint()           ← Atomic state save
      │
      └─ → AgentExecutionResult

6. Quality Gates

After execution, results pass through 5 verification gates:

• Hallucination detection
• Consistency check
• Completeness verification
• Accuracy validation
• Safety check

7. Completion

Results are flushed to trace store, metrics are recorded, and the execution summary is returned.