Query Router

Unified query router combining content-type detection, complexity scoring, and prefix-based routing. Routes queries to the optimal model with safety features...

Security Scan

The skill's code and instructions are consistent with a local query-router: it classifies queries and calls a local 'openclaw' status endpoint, writes local audit logs, and does not request external credentials or unexpected network access.

Purpose & Capability

Name/description (route queries to optimal models) match the code and SKILL.md: prefix/content-type/complexity routing, mapping to cloud/local models. The use of the 'openclaw' CLI and a localhost model tags endpoint is coherent for a router that needs current model state and available models.

Instruction Scope

Runtime instructions and the code operate within the claimed domain: classify input, decide model, optionally call 'openclaw' to check/verify status, and write local audit logs. The skill references only local files (.locks, .logs) and localhost HTTP; it does not exfiltrate to external servers or request unrelated system secrets.

Install Mechanism

No install spec is provided (instruction-only). The package includes Python scripts but does not download or extract remote artifacts. This is the lowest-risk install pattern for a script-based skill.

Credentials

The skill declares no required env vars or credentials and the code does not read secrets. It invokes a local CLI ('openclaw') and contacts http://localhost:11434 which are proportional to discovering models; no unrelated service tokens are requested.

Persistence & Privilege

always is false and the skill does not attempt to modify global skill config. It creates local lock and log files under scripts/.locks and scripts/.logs, which is expected for routing/audit behavior and does not indicate elevated privileges.

Assessment

This skill appears coherent for routing queries: it classifies queries, consults a local model endpoint (http://localhost:11434/api/tags) and the 'openclaw' CLI, and writes audit logs under scripts/.logs and lock files under scripts/.locks. Before installing, verify you trust the skill source (homepage unknown) and that running 'openclaw' and a local model registry on localhost is expected in your environment. Review the included Python files yourself or run the scripts in a safe/test workspace; note they log up to the first 100 characters of queries to a JSONL file, so avoid sending sensitive secrets through queries if you enable this skill.

Routes every query to its best-fit model using three routing strategies combined:

1. Prefix routing — explicit user intent (@codex, @mini, etc.)
2. Content-type detection — auto-detect vision/voice/tool/code/reasoning/qa
3. Complexity scoring — simple/moderate/complex tier routing

Usage

Classify only

python3 skills/query-router/scripts/classify.py [--json] [--attachment .ext] <query>

Route (classify + switch model)

python3 skills/query-router/scripts/router.py [--json] [--dry-run] [--attachment .ext] <query>

Safety & Audit

python3 skills/query-router/scripts/router.py --check           # list available models
python3 skills/query-router/scripts/router.py --audit            # show recent routing log
python3 skills/query-router/scripts/router.py --no-lock          # disable lock protection
python3 skills/query-router/scripts/router.py --dry-run --no-verify  # preview without verify

Library usage

from router import route_query

result = route_query(
    query="@codex write a complex python script",
    has_attachment=False,
    attachment_ext="",
    dry_run=True,       # True = preview, False = actually switch
    enable_lock=True,   # prevent concurrent routing
    enable_verify=True, # readback verify after switch
    enable_rollback=True,  # revert on failure
)
print(result["action"])        # route | recommend | skip | blocked
print(result["to_model"])      # recommended model name
print(result["classification"]) # full classification details

Routing Priority

Priority	Signal	Example	Behavior
1	Prefix	@codex write a script	Force switch to @codex model
2	Content-type (≥60%% conf)	describe this image	Auto-switch to vision model
3	Complexity (<60%% conf)	analyze Q3 data	Route by complexity tier
4	Skip (<30%% conf)	hi how are you	Keep current model

Prefix Routing

Prefix	Model	Aliases
@codex	openai-codex/gpt-5.3-codex	@c
@mini	minimax-m2.7:cloud	Cloud primary
@cl	qwen3-coder-next:cloud	Cloud code
@q	minimax-m2.7:cloud	Cloud fast Q&A
@llava	qwen3.5:cloud	Cloud vision
@whisper	whisper	—

Content Types

Type	Detected By	Best Model
vision	Image attachment or "describe/identify"	qwen3.5:cloud
voice	Audio file or "transcribe/speech-to-text"	whisper
tool	run, fetch, send, check, sync	minimax-m2.7:cloud
code	script, function, import, debug	qwen3-coder-next:cloud
reasoning	analyze, compare, plan, think	minimax-m2.7:cloud
qa	what is, how does, define	minimax-m2.7:cloud

Complexity Tiers

Tier	Pattern	Model
simple	greetings, one-liners, short questions	minimax-m2.7:cloud
moderate	emails, analysis, search, translate	minimax-m2.7:cloud
complex	multi-file code, deep research, architect	minimax-m2.7:cloud

Safety Features

• Lock protection — prevents concurrent routing ops (fcntl)
• Verify-after-switch — confirms model actually changed
• Rollback — reverts to previous model on failure
• Audit log — JSONL log at scripts/.logs/router.log.jsonl

Merged Approaches

Source	Technique
openclaw-model-router-skill	Prefix syntax @codex/@mini, lock + verify + rollback
task-complexity-router	Complexity tier scoring (simple/moderate/complex)
query-router (original)	Content-type detection (vision/voice/tool/code/qa)

All three combined into one unified router with confidence-based priority.