Model Brain

Route each incoming message to the right Bankr/OpenClaw model or to a zero-LLM path based on task type, risk, and cost. Use when you need per-message model s...

Security Scan

Model Brain is an internally consistent, instruction-only model-routing skill that contains local Python/Bash routing logic, requests no credentials, and does not perform network I/O or install external code.

Purpose & Capability

The name/description match the included files: a deterministic, rule-based router implemented in scripts/route_message.py and scripts/select_model.sh. The requested resources (none) and included artifacts are proportionate to a routing skill. The references to 'bankr' model names are consistent with the stated purpose.

Instruction Scope

SKILL.md directs the agent to run the included scripts and to consult local reference docs. The runtime instructions and code operate only on input text and return routing metadata; they do not read external files, environment secrets, or call external endpoints. Note: routing is purely rule-based (regexes and simple heuristics) and may misclassify edge cases (e.g., 'script' appears in multiple pattern sets, the <=40-word length rule forces a low-cost route, and keywords like 'send'/'swap' trigger wallet routing). These are correctness/coverage concerns rather than security issues, but they can cause accidental escalation or cost if not validated.

Install Mechanism

No install spec or external downloads; the skill is instruction-only with bundled scripts. No archives are extracted and no external package managers are invoked.

Credentials

The skill requires no environment variables, credentials, or config paths. The scripts export only a transient JSON_OUT environment variable at runtime. There are no requests for tokens, keys, or unrelated credentials.

Persistence & Privilege

always:false and user-invocable:true (defaults) — the skill does not request persistent presence or modify other skills' configs. It doesn't write persistent files or store credentials.

Assessment

This skill appears to be what it says: a local, deterministic router. Before installing, test it with representative inputs to ensure routing behavior matches your expectations (especially for short messages, 'script' overlaps, and wallet-related keywords that force wallet/escalation). Verify that the Bankr model names used here map to models your agent can call and that downstream model invocations have appropriate access controls — routing should not be treated as authorization to perform transactions. If you rely on deterministic zero-LLM bypasses for safety, review and adjust ZERO_LLM_PATTERNS and wallet/code pattern lists so sensitive requests aren't accidentally routed to lower-cost models.

Use this skill when the goal is to choose the right model for a single message or action.

Core policy

• Prefer zero-llm when a deterministic skill or script can do the job.
• Use the cheapest adequate model for low-risk work.
• Use bankr/claude-sonnet-4.5 for routine wallet and treasury operations that are important but still normal.
• Escalate to bankr/claude-opus-4.6 only for explicitly high-stakes, security-sensitive, or ambiguous wallet flows.
• Use bankr/gpt-5.2-codex for coding-heavy implementation work.
• Keep routing deterministic and rule-based before adding any LLM-on-LLM behavior.
• Treat the output as a routing recommendation, not as transaction authority.

Default routes

• zero-llm
  • deterministic skills like pet-me-master
  • shellable tasks with no reasoning need
• bankr/minimax-m2.5
  • casual chat, lightweight rewriting, simple summaries, low-risk classification
• bankr/claude-sonnet-4.5
  • general reasoning, planning, product thinking, routine wallet ops
• bankr/gpt-5.2-codex
  • coding-heavy patching, repo surgery, implementation details
• bankr/gemini-3-pro
  • long-context synthesis and broad document digestion
• bankr/gemini-3-flash
  • lightweight vision and quick multimodal triage
• bankr/claude-opus-4.6
  • explicitly high-stakes wallet actions, security reviews, tricky architecture, final escalation

Routing workflow

1. Check whether the request can be handled by a deterministic skill or script.
2. Classify the task: chat, code, wallet, vision, long-context, or deterministic.
3. Mark whether the request is high-stakes.
4. Route to the cheapest safe model.
5. Return the primary route, fallback route, and a short reason.

Quick start

python3 {baseDir}/scripts/route_message.py --text "pet all my 53 gotchis" --json
python3 {baseDir}/scripts/route_message.py --text "rewrite this x thread shorter" --json
python3 {baseDir}/scripts/route_message.py --text "swap ETH to USDC and send to treasury" --json
bash {baseDir}/scripts/select_model.sh --text "build this feature in the repo" --mode summary

Wrapper entrypoint

Use scripts/select_model.sh when you want a simple aaigotchi-friendly wrapper before model selection. It can emit just the route, full JSON, or shell-style env lines.

References

• Read references/routing-table.md for the routing rules and escalation thresholds.
• Read references/bankr-models.md for the current Bankr model inventory and aaigotchi default.