ellmos-ai English — Official English version of
model-strategy.
Model-Switching Strategy (English)
Multi-model orchestration: score-based model selection, cross-agent delegation, advisor pairing, escalation triggers, and cost-efficiency optimization
1. Model catalog
Claude (subagent-capable via the Agent tool)
Level 4 (Reviewer): Opus 4.8 — advisor, math review [user only: /model, /advisor]
Level 3 (Strategist): Opus 4.6 — architecture, concepts [subagent: model:"opus"]
Level 3 (Creative): Fable 5 — creative texts, stories [subagent: model:"fable"]
Level 2 (Workhorse): Sonnet 4.6— implementation, debug [subagent: model:"sonnet"]
Level 1 (Fast): Haiku 4.5 — boilerplate, formatting [subagent: model:"haiku"]
External agents (companion scripts / SSH)
Level 2-3: Gemini 3.5 pro — research, scientific databases [agy-companion CLI]
Level 2: Gemini 3.5 flash— fast research [agy-companion CLI]
Level 2-3: Codex 5.5 (GPT) — code review, code generation [codex-companion CLI]
Level 2: Codex 4.5 (GPT) — simpler code tasks [codex-companion CLI]
Local models (token-free, 24/7)
Level 1-2: Ollama (Qwen 3.5:35b-a3b) — Haiku-to-Sonnet level [<ollama-host>:11434]
Invocation: SSH + curl http://<ollama-host>:11434/v1/chat/completions
Or: delegation via an agent-system control API (if available)
Reachability matrix
| Model | LLM-startable | Invocation path | Constraints |
|---|---|---|---|
| Sonnet 4.6 | Yes | Agent(model:"sonnet") |
— |
| Opus 4.6 | Yes | Agent(model:"opus") |
— |
| Haiku 4.5 | Yes | Agent(model:"haiku") |
— |
| Fable 5 | Yes | Agent(model:"fable") |
— |
| Opus 4.8 | Advisor only | advisor() in session |
user must set /advisor |
| Gemini 3.5 | Yes (Bash) | companion-for-agy "prompt" |
Windows-only, stdout workaround |
| Codex 5.5/4.5 | Yes (Bash) | node codex-companion.mjs task "prompt" |
auth required |
| Ollama | Yes (SSH/curl) | SSH + curl to the Ollama host API | VPN/Tailscale must be active |
| Opus 4.8 as main model | No | user: /model opus 4.8 |
user action only |
| Fable 5 as main model | No | user: /model fable |
user action only |
2. Score computation
Dimensions (0-10):
CLARITY : How unambiguous is the task?
COMPLEXITY : How many components?
CREATIVITY : New solutions needed?
CONTEXT : How much prior knowledge?
CRITICALITY : How important is perfection?
SCORE = (10 - CLARITY) + COMPLEXITY + CREATIVITY + CONTEXT + CRITICALITY
Score thresholds
| Score | Model | Examples |
|---|---|---|
| 0-8 | Ollama (local host) | prompt generation, summaries, simple texts |
| 9-12 | Haiku | __init__.py, formatting, boilerplate |
| 13-22 | Sonnet | implementation, bug fixes, standard code |
| 13-22 | Gemini 3.5 | research, literature search, scientific databases |
| 13-22 | Codex 5.5 | code generation (Luau, Node.js), compute scripts |
| 23-28 | Sonnet + advisor review | complex code with quality check |
| 23-35 | Fable 5 | creative texts, marketing, storytelling |
| 29-40 | Opus 4.6 | architecture, strategy, paper writing |
| 35-50 | Opus 4.6 + advisor | proofs, architecture decisions, statistics |
| 40-50 | Opus 4.8 (user recommendation) | mathematical proof work, highest rigor |
3. Cross-agent delegation
Which external agent for what?
| Task | Best agent | Reason |
|---|---|---|
| Scientific literature search | Gemini 3.5 pro | native OpenAlex/arXiv/PubMed skills |
| Code review (second opinion) | Codex 5.5 | independent perspective |
| Simple text generation | Ollama (local host) | token-free, 24/7 |
| Creative texts, marketing | Fable 5 | strongest creative output |
| Mathematical proofs | Opus 4.8 (advisor) | highest analytical depth |
Exclusions (documented weaknesses)
- Gemini: NOT for mathematical reviews/proof work (documented direction error in a proof review, 2026-06-07)
- Codex 4.5: only when 5.5 is unavailable; otherwise always 5.5
Invocation paths
Replace the placeholders
<host>,<ollama-host>,<tailscale-ip>,<user>, and~/.ssh/<key>with your own infrastructure.
Gemini (via companion-for-agy):
companion-for-agy --researcher --json --timeout 120000 "research prompt"
Codex (via codex-companion):
node "~/.claude/plugins/cache/openai-codex/codex/1.0.4/scripts/codex-companion.mjs" task --effort high "code prompt"
Ollama on a remote host (via SSH):
ssh -i ~/.ssh/<key> <user>@<tailscale-ip> "curl -s http://localhost:11434/v1/chat/completions -d '{\"model\":\"qwen3.5:35b-a3b\",\"messages\":[{\"role\":\"user\",\"content\":\"Prompt\"}]}'"
Delegation to an agent system with tools (example):
curl -s -X POST http://<host>:8081/api/chat -H "Content-Type: application/json" -d '{"prompt": "...", "chat_id": "claude-delegate"}'
4. Advisor pairing
Mechanics
advisor() is a session-level tool — the advisor model is set by the user via /advisor, not programmatically. This yields these pairing patterns:
| Pattern | How it works | When to use |
|---|---|---|
| Session advisor | user sets /advisor opus 4.8, agent calls advisor() |
standard for proofs/architecture |
| Orchestrator-as-reviewer | Opus main model reviews Sonnet subagent output | orchestrator is stronger than the worker |
| Counter-agent | agent A works, agent B checks adversarially | independent verification, 2 perspectives |
| User recommendation | agent recommends: "do this task with opus 4.8 + advisor" | when the current session is too weak |
When to recommend an advisor?
- Mathematical proof work (score ≥ 35)
- Architecture decisions with long-term consequences
- Statistical methodology / study design
- Complex bugs after 2+ unsuccessful debug cycles
When NOT to use an advisor?
- Routine code, content, formatting (score < 23)
- Simple feature implementation
- Well-defined, non-critical tasks
5. Escalation triggers
Ollama -> Haiku
- File access required
- Code analysis needed
Haiku -> Sonnet
- More than 2 files affected
- Decision between alternatives needed
- Unexpected error occurred
- Delete operation requested
Sonnet -> Opus
- Architecture decision required
- 3+ systems must be integrated
- Requirements contradictory/unclear
- Strategic planning needed
Sonnet -> Gemini (lateral)
- Scientific research needed
- Bibliography verification
Sonnet -> Codex (lateral)
- Code review as a second opinion
- Advisor overloaded (fallback reviewer)
Opus -> Opus + advisor
- Proof review needed
- Critical architecture decision
- Statistical methodology
De-escalation
- Concept defined -> Sonnet takes over implementation
- Task trivial/repetitive -> Haiku takes over
- Text only, no tool access -> Ollama takes over
6. Permission matrix
| Operation | Ollama | Haiku | Sonnet | Opus | Gemini | Codex |
|---|---|---|---|---|---|---|
| Read files | - | Yes | Yes | Yes | Yes* | Yes* |
| Write files | - | Yes | Yes | Yes | Yes* | Yes* |
| Delete files | - | - | Yes** | Yes | - | - |
| System commands | - | - | Yes** | Yes | Yes* | Yes* |
| Architecture decisions | - | - | - | Yes | - | - |
| Web research | - | - | Yes | Yes | Yes | - |
| Call advisor() | - | - | Yes | Yes | - | - |
*via companion script in its own sandbox mode **with user confirmation
7. Cost efficiency
Token savings through routing
| Task type | Without routing | With routing | Savings |
|---|---|---|---|
| Trivial | Opus tokens | Ollama (free) | 100% |
| Boilerplate | Opus tokens | Haiku tokens | ~80% |
| Standard code | Opus tokens | Sonnet tokens | ~50% |
| Research | Claude tokens | Gemini tokens | ~70% (different budget) |
| Code review | advisor() tokens | Codex tokens | ~60% (different budget) |
8. Golden rule
"Opus thinks, Sonnet builds, Haiku executes, Ollama saves. Gemini researches, Codex reviews, Fable narrates."
Changelog
2.0.0 (2026-06-12)
- Cross-agent delegation: Gemini, Codex, Ollama (local host) as routing targets
- Advisor pairing: 4 patterns (session advisor, orchestrator-as-reviewer, counter-agent, user recommendation)
- Reachability matrix: LLM-startable vs. user-only documented
- Ollama (Qwen 3.5:35b-a3b, Haiku-to-Sonnet level) added as level 1-2
- Lateral escalation: Sonnet -> Gemini (research), Sonnet -> Codex (review)
- Exclusions documented (Gemini not for math)
- Score thresholds extended to all models
1.0.0 (2026-03-15)
- Ported from BACH v3.8.0 (ing-strategie v2.0.0)
Ported from BACH v3.8.0 | Extended with cross-agent + advisor v2.0.0