Model Routing Skill

Automatically routes requests to the optimal model based on task type, prioritizing code, reasoning, simplicity, and normal tasks with fallback and concurren...

Security Scan

Instruction-only routing logic that is internally consistent with its stated purpose; no installs or credentials requested, but a few implementation details are underspecified (telemetry storage, admin tuning, model-name assumptions).

Purpose & Capability

The name/description match the SKILL.md routing rules. All declared requirements (none) align with a pure instruction-only routing policy; nothing requested is extraneous to model routing.

Instruction Scope

Instructions stay within routing behavior (model selection, fallbacks, concurrency, caps). Two items are underspecified and warrant attention: the telemetry line ('record {taskType, chosenModel, capsApplied} to internal trace') does not describe storage location/retention/access controls, and 'may be tuned by admin' implies external configuration but does not say how the skill obtains or enforces those admin settings. Also the skill assumes specific model IDs (gpt-5-codex, gpt-5, gpt-5-mini, gpt-4.1-nano); if those IDs don't exist in your environment the routing could fail or route incorrectly.

Install Mechanism

No install spec and no code files: the skill is instruction-only so nothing is written to disk or downloaded.

Credentials

Skill requests no environment variables, credentials, or config paths. This is proportional to its stated purpose.

Persistence & Privilege

always:false and no self-install behavior. The skill mentions admin-tunable defaults and internal traces but does not request persistent privileges or system-wide changes in the current text.

Assessment

This skill is coherent and low-risk in that it only provides routing instructions and requires no installs or credentials. Before enabling it: 1) confirm the model IDs used by the skill match the models available in your environment (otherwise update the SKILL.md or mapping); 2) ask where 'internal trace' telemetry is stored, who can access it, and how long it is retained—avoid recording PII in taskType if possible; 3) verify how 'admin tuning' and org/channel/session model locks are surfaced to the agent so the skill won't conflict with enforced policies; 4) test fallback, retry, concurrency, and budget behaviors in a safe environment to ensure they behave as expected and respect billing limits. If those operational details are acceptable, the skill appears benign and appropriate to use.

Model Routing Skill (Compact)

Automatically route each request to the best model.

Routing rules:

• Use gpt-5-codex for coding, debugging, build errors, stack traces, logs, repo edits, and implementation tasks.
• Use gpt-5 for deep reasoning, architecture, large-context analysis, complex planning, and hard diagnosis.
• Use gpt-4.1-nano for short trivial requests with no code and no deep reasoning.
• Use gpt-5-mini for everything else.

Priority:

1. code_heavy
2. heavy_reasoning
3. simple
4. normal

Fallbacks:

• gpt-5-codex -> gpt-5 -> gpt-5-mini -> gpt-4.1-nano
• gpt-5 -> gpt-5-mini -> gpt-4.1-nano
• gpt-5-mini -> gpt-4.1-nano

Retry on:

• rate limit
• TPM exceeded
• timeout
• temporary unavailable
• capacity errors

Concurrency:

• max 2 concurrent requests
• reduce to 1 for large prompts or repeated TPM failures
• queue excess work

Respect explicit user overrides and any org/channel/session model locks first. Do not override a forced model. Only auto-route when routing is enabled and no explicit model is set.

Safety caps (defaults, may be tuned by admin):

• Max input tokens per class: simple 4k, general 8k, code 12k, heavy 20k
• Max output tokens per class: simple 400, general 800, code 1200, heavy 1500
• Per-turn budget ceiling: $0.02 (downgrade on breach)

Telemetry (lightweight): record {taskType, chosenModel, capsApplied} to internal trace; do not expose to users unless they ask.

Fallbacks: never upgrade to a more expensive model on retry; only downgrade tiers.