multi-model-orchestration

Files (10)

• README.md
• README.ru.md
• SKILL.md
• hero-dark.svg · assets/
• hero-light.svg · assets/
• failure-handling.md · references/
• model-card.md · references/
• prohibitions.md · references/
• routing.md · references/
• worker-contract.md · references/

SKILL.md (first lines)

# Multi-Model Orchestration

Coordinator dispatches tasks to 2+ model workers via Orca, waits for results, synthesizes.
Coordinator dispatches, waits, synthesizes, gates. Writing code in a coordinator session requires an explicit owner-pin switch to writer role — coordinator does not combine orchestration with code writing in the same task.
Coordinator = any model with this skill loaded. This skill does not prescribe coordinator family.

## TL;DR (minimum path)

**multi-model-orchestration = dispatch/review движок для ≥2 моделей.** Координатор маршрутизирует задачи воркерам (Orca), ждёт, синтезирует. Координатор не совмещает оркестрацию с написанием кода в одной задаче.

**Когда:** «обсудите с 2 моделями», cross-validate, parallel review, fidelity port, security/RLS (никогда одной моделью). **Когда нет:** §1 говорит solo, тривиал, одна модель.

**Ядро:** solo-vs-multi дерево (§1) → роутинг моделей (§2) → **cross-family rule: writer.family ≠ reviewer.family** → brief (§4) → dispatch через Orca (`dispatch --inject`, НЕ `terminal send`) → wait (`check --wait`) → синтез (§7, stricter wins). Worker contract: SUMMARY/EVIDENCE/CHANGES/RISKS/BLOCKERS + written≠persisted.

**Роутер:** новый проект → `project-bootstrap` · план спринта → `wave-spec` · 2+ модели → `multi-model-orchestration`. Lifecycle gates (Commit/PR/Merge/Deploy/On prod) — канон в `wave-spec`.

---

## 0. Preconditions

```bash
orca status --json          # runtime up?
```

If Orca unavailable → manual fallback (§0b). Do not guess flags.

Skills required: `orca-cli` (terminal ops), `orchestration` (task/dispatch/wait).

### 0b. Manual Fallback (Orca unavailable)

If `orca status` fails or Orca is not installed:

1. Coordinator opens N terminal tabs/windows manually (one per model).
2. In each terminal: launch the model's CLI (`opencode`, `codex`, etc.).
3. Copy-paste the model-specific brief (from `references/routing.md`) into each terminal.
4. Collect outputs manually when workers finish.
5. Synthesize per §7.

No `worker_done` lifecycle in manual mode — coordinator polls visually.
Skill still applies: routing (§2), brief structure (§4), synthesis (§7), write policy (§6).

---

## 1. Decision: Solo vs Multi

```
Task clear, 1 deliverable, linear, < 2-3h?
  └─ YES → SOLO. Stop. Do not open multi.

2+ independent pieces needed IN PARALLEL?
  └─ YES → multi ok.

Expensive error + need second family? (prod, security, big refactor)
  └─ YES → multi: writer ≠ reviewer.

Bulk > 1 context window?
  └─ YES → multi by layers.

Scope unclear?
  └─ solo discovery first. Then re-evaluate.

Otherwise → solo.
```

**Not multi:** trivial fix, single short report, no patience for wait loops.

---

## 2. Routing — Which Model for What

| Work | Route to | Family | Why |
|------|----------|--------|-----|
| **Orchestration / dispatch / handoff** | **DeepSeek V4 Flash** | DeepSeek | Role lock: не совмещает оркестрацию с написанием кода в одной задаче. Dispatch → wait → synthesize → gate. Full role |
| **Implement (primary writer/coder)** | **DeepSeek V4 Flash** (`A/agent1st_v37.3-flash`) | DeepSeek | Default writer — fast ($0.14/$0.28 per 1M), strong on 0731 benchmarks (SWE-bench 79.0 Flash-Max, LiveBench Coding 69.2). Single-file, bulk code, tests |
| **Multi-file implement (3+ files)** | **GLM 5.2** (`A/agent1st_glm`) | Zhipu | 1M state continuity, long-horizon multi-file specialty (~15 min implement, ~10 min fix-round) |
| **Review / architecture / business analysis (default reviewer)** | **Qwen 3.8 Max** (`A/agent1st_qwen-3.8`) | Alibaba | 2.4T/95B active, native vision, CoT. Vendor benches 2026-08-03 (qwen.ai/blog): GPQA-D 92.6, Terminal-Bench 2.1 86.6 — indep verification pending. Owner empirical: level ~ Kimi K3, **сильнее GLM 5.2 в architecture и depth**. NE основной кодер (slow) |
| Complex analysis, multimodal, cross-examination | **Qwen 3.8 Max** | Alibaba | 2.4T/95B active, native vision, CoT |
| Security / RLS / auth / Storage review | **gpt-5.6-luna** (launch: `codex --model gpt-5.6-luna -c model_reasoning_effort="max"`; «Luna Max» — shorthand, НЕ имя модели) + DeepSeek V4 Flash (parallel) | OpenAI + DeepSeek | Gate + implementer; cross-family required (GPT-5.5 = optional opt-in strict gate для критичных) |
| **Behavioral regression gate** (optional) | **GPT-5.5** (opt-in security gate; NOT default Strong pair) | OpenAI | Historical unique role — caught abort/cost MAJOR GLM missed. NOT "just another reviewer"; retained as opt-in for highly sensitive cases |
| **Fidelity port** (generator, vectorizer, reference→platform) | **DeepSeek V4 Flash** write (default) / **GLM 5.2** (multi-file) · **gpt-5.6-luna** ∥ **Qwen 3.8 Max** review | DeepSeek/Zhipu · OpenAI ∥ Alibaba | Evidence (production waves): Flash bulk implement; GLM exhaustive static parity; gpt-5.6-luna behavioral regression lens; Qwen architect/analyst. Dual review mandatory — complementary, not redundant |
| Cross-QA (review someone else's work) | **Different family** | — | Blind-spot detection |

**Cross-family rule:** writer.family ≠ reviewer.family. Qwen Code (Alibaba) writer → reviewer must be DeepSeek/Zhipu/OpenAI, NOT OpenCode Qwen (also Alibaba). Full pairs: `references/routing.md` → Cross-Family Pairs.

**Writer replaceable:** owner phrase «сейчас writer=X» (Flash or GLM) pins writer instantly — no skill rewrite. Default writer: DeepSeek V4 Flash (fast, cheap). GLM = swap для multi-file. Qwen 3.8 Max = default reviewer/architect (NE writer — slow). See `references/routing.md` for the full routing table with evidence anchors.

Synthesis: MAJOR from any reviewer → fix before merge; prefer stricter severity on contradictions.

Full routing details + brief templates: `references/routing.md`.

---

## 2b. Deploy / smoke gate

**Canonical definition:** `skills/wave-spec/SKILL.md` §"Deploy Probe (Universal Pattern)".
**Canonical examples:** `skills/wave-spec/SKILL.md` §Deploy Probe — curl dual-pattern + skills-repo check.

**SUMMARY:** After merge/deploy (or skill install), prove the shipped surface exists before handoff. The probe checks existence, not correctness — correctness is verified at In Review. Adapt the probe to the target stack (web curl, CLI binary, package install, skill load path). Principle: one observable command that proves the artifact reached production. No probe = RESIDUAL-RISK-OWNER-SMOKE.

Evidence: production waves had untracked route files — if deployed without tracking, 404.

---

## 2c. Fidelity Port Rules

**Canonical definition:** `skills/wave-spec/SKILL.md` §"Fidelity Dual Review".

**SUMMARY (5 rules):**
1. **Dual review mandatory:** static-parity reviewer ∥ behavioral-semantics reviewer as two complementary lenses from **different model families**. Single-reviewer fidelity ports are NOT accepted.
2. **Writer ≠ reviewer.** Cross-family mandatory (writer.family ≠ reviewer.family).
3. **MAJOR from any reviewer → fix round before In Review.** Stricter severity wins.
4. **No live smoke = RESIDUAL-RISK-OWNER-SMOKE** in handoff. Document explicitly.
5. **Deploy gate:** prove shipped surface before handoff (§2b).

Evidence (production waves): GLM-only would have missed abort MAJOR; Codex-only would have missed a numeric edge case + refine path.

For non-fidelity waves: dual review is recommended but not mandatory. At minimum, writer ≠ reviewer.

---

From dimkurilo/opencode-skills, MIT.

More from dimkurilo/opencode-skills