llm-review

name	llm-review
description	Reference for local-LLM scripts. Calling skills run the bash commands inline in the main conversation — no dedicated agents. Covers the availability check, model loading, script execution, and unloading lifecycle.

LLM Review

Full lifecycle wrapper for local LLM operations. Other skills run LLM bash commands directly in the main conversation — no dedicated agents.

How Other Skills Use This

Calling skills run the bash commands directly using the Bash tool. The typical pattern:

1. Check availability: CRAFT_SCRIPTS=$(find ...) && curl -s --max-time 2 http://127.0.0.1:1234 ... && echo "LLM_AVAILABLE:$CRAFT_SCRIPTS" || echo "LLM_UNAVAILABLE"
2. Run task in background: bash "$CRAFT_SCRIPTS/llm-agent.sh" "<task>" <working-dir> (with run_in_background: true)
3. Unload when done: bash "$CRAFT_SCRIPTS/llm-unload.sh"

Each calling skill has the exact commands inline. This skill serves as the detailed reference.

Task types:

• explore "<task>" <working-directory> — Autonomous investigation (LLM reads files itself, saves the most tokens)
• review <file> "<focus>" — Single-file review (focus positional)
• review <file1> <file2> ... --focus "<focus>" — Multi-file review (use --analyze "<task>" <files...> for task-framed analysis; absorbs the old analyze)

Input

The user input is: $ARGUMENTS

• Task type + arguments: As described above
• File path(s) + focus: /llm-review src/domain/auth/feature/LoginPage.tsx "security"
• Empty: Ask the caller what to review

Process

Step 1: Locate Scripts

The scripts directory is provided at session start (bootstrap context) as craft-skills scripts directory: <path>. Use that path if available. Otherwise fall back to:

CRAFT_SCRIPTS=$(find ~/.claude/plugins -name "llm-agent.sh" -path "*/craft-skills/*" -exec dirname {} \; 2>/dev/null | head -1)

If neither is available, return: "LLM scripts not found — craft-skills plugin may not be installed."

Step 2: Check Availability

curl -s --max-time 2 ${LLM_URL:-http://127.0.0.1:1234} > /dev/null 2>&1 && echo "LLM_AVAILABLE" || echo "LLM_UNAVAILABLE"

If LLM_UNAVAILABLE, return: "LLM_UNAVAILABLE: LM Studio server not running."

Step 3: Run the Task

Note: Scripts auto-detect and fix context length (reload with 64K if loaded with less). No manual check needed.

Choose the script based on task type:

review — file content(s) passed to LLM with thinking mode:

# single file (focus positional)
bash "$CRAFT_SCRIPTS/llm-review.sh" <file-path> "<focus>"
# multiple files (focus as a flag; --analyze "<task>" <files> for task-framed analysis)
bash "$CRAFT_SCRIPTS/llm-review.sh" <file1> <file2> ... --focus "<focus>"

explore — autonomous agent with file access tools (saves the most tokens):

bash "$CRAFT_SCRIPTS/llm-agent.sh" "<task description>" <working-directory>

If the script returns LLM_ERROR or exits non-zero, report the error — do not retry. If the script returns LLM_THINKING_OVERFLOW, report: "Model used all tokens on reasoning. File may be too large for review mode — try explore mode instead."

Step 4: Triage Findings

Before returning findings, filter out known false positives:

• The local model doesn't understand Claude Code plugins, skills, or agent dispatch
• Findings about "missing files" may be false if files exist in the plugin directory
• Focus on: structural issues, logic bugs, type mismatches, architecture violations, consistency

Step 5: Unload Model

bash "$CRAFT_SCRIPTS/llm-unload.sh"

Skip unloading if the caller passes Keep loaded: true — this means another LLM step is expected soon.

Step 6: Return

Return the triaged findings to the caller. Be concise — the caller will triage further against their conversation context.

Configuration

Variable	Default	Description
LLM_URL	http://127.0.0.1:1234	LM Studio server URL
LLM_MODEL	qwen3.6-35b-a3b	Model identifier

Notes

• Thinking mode is enabled by default — fewer false positives
• Context window is set to 64K tokens automatically on load (scripts handle this)
• Typical time: 30-90 seconds for review, 1-3 minutes for explore
• Model uses ~22GB RAM when loaded — always unload when done unless told otherwise