The execution heartbeat of an eAI mission. The Lead Orchestrator drives it; workers execute; testing workers gate. Prime is not in this loop — it set the architecture and the phase plan, and it only re-enters on escalation.
┌──────────────────────────────────────────────┐
│ Lead Orchestrator (Opus 4.8) │
│ decompose phase → assign by tier │
└───────────────┬────────────────────────────────┘
│ assign (with acceptance criteria)
▼
┌───────────────────────┐
│ Worker (tier) │ senior / bulk / fast
│ implement │
└───────────┬───────────┘
│ deliverable + evidence
▼
┌───────────────────────┐
│ Testing worker │ Evidence Collector / Reality Checker /
│ validate (evidence) │ API Tester / Performance Benchmarker
└───────────┬───────────┘
PASS │ FAIL
┌───────────┴───────────┐
▼ ▼
merge gate attempt < 3 → feedback → retry (same worker)
(Lead decides) attempt = 3 → escalate to Lead:
reassign up-tier / decompose / defer
- Evidence over claims. A task passes only with proof (screenshot, test output, benchmark) — never a worker's assertion that it works.
- Max 3 retries per task. Lead tracks the count. At 3, escalate; do not loop forever.
- Tier escalation on repeated failure. A bulk-tier worker failing a correctness gate twice is a signal the task needs senior tier — Lead reassigns up rather than retrying at the same tier.
- Security diffs get a senior security pass. Any change touching auth, secrets, crypto, or trust boundaries is reviewed by a senior-tier security worker before the merge gate, regardless of who wrote it.
- No phase advances without its gate green. Lead enforces the gate at every phase boundary.
- Architectural surprises go to Prime. If a task reveals the architecture is wrong (not just the implementation), Lead escalates to Prime — that is not a retry.
| Task shape | Tier | Model |
|---|---|---|
| Security analysis, threat modeling, audit | senior | Sonnet 4.6 |
| Correctness-critical impl, complex debugging | senior | Sonnet 4.6 |
| Schema/SRE/infra judgment calls | senior | Sonnet 4.6 |
| QA verification that gates a merge | senior | Sonnet 4.6 |
| High-volume codegen, CMS/UI from spec | bulk | Kimi K2 |
| Long mechanical refactor (large context) | bulk | Kimi K2 |
| Lookup, grep/glob sweep, read-only explore | fast | Haiku 4.5 |
| Formatting, linting, doc cleanup | fast | Haiku 4.5 |
| Classification, report collation | fast | Haiku 4.5 |
| Simple, well-specified single-file fix | fast | Haiku 4.5 |
See ../routing/models.yaml for the authoritative mapping and escalation triggers.