ls-devstral

name	ls-devstral
description	Spawn a local Devstral coding agent running on LM Studio. Use when asked to '/devstral', 'use devstral', 'ask devstral', 'local agent', 'run this on devstral', 'devstral this'. Routes tasks through the local model_router.py proxy at port 5111 which transparently forwards devstral requests to LM Studio while keeping Anthropic calls working normally.

/devstral — Local Devstral Agent via LM Studio

Requirements — before this skill can run, the following must be available on the user's machine:

• LM Studio installed and running with mistralai/devstral-small-2-2512 loaded (download from https://lmstudio.ai/)
• uv on PATH — Python package runner (install: pip install uv or https://docs.astral.sh/uv/)
• LiteCLI (optional) — universal CLI for MCP backends; set LITECLI_DIR env var to its directory if installed. If missing, Step 3 is skipped gracefully.
• model_router.py proxy — expected at .claude/plugins/kuro/proxy/model_router.py relative to your project root. This is bundled with the liteharness plugin; if absent the skill falls back to direct HTTP (Step 5b).

You are executing the /devstral skill. This routes the user's coding task to a local Devstral Small 2 (24B) agent running in LM Studio.

Safety rule: This skill ONLY activates when explicitly invoked. Never auto-invoke it. Devstral is a Mistral coding agent — optimized for code generation, refactoring, and analysis tasks.

The user's task is: $ARGUMENTS

---

Step 1 — Verify LM Studio is Reachable

Before doing anything else, confirm Devstral is available in LM Studio:

curl -s --max-time 5 http://localhost:1234/api/v0/models | python -c "
import json, sys
data = json.load(sys.stdin)
models = data.get('data', [])
devstral = [m for m in models if 'devstral' in m.get('id','').lower()]
if devstral:
    for m in devstral:
        print(f\"LOADED: {m['id']} state={m.get('state','?')}\")
else:
    print('NOT LOADED')
    print('Available:', [m['id'] for m in models[:5]])
" 2>/dev/null || echo "LM_STUDIO_UNREACHABLE"

If LM Studio is unreachable or Devstral is not loaded:

• Report: "LM Studio is not running or Devstral is not loaded. Please start LM Studio and load mistralai/devstral-small-2-2512 before running /devstral."
• Stop here.

If Devstral is listed but not in a loaded state, offer to load it:

"$USERPROFILE/.lmstudio/bin/lms.exe" load "mistralai/devstral-small-2-2512" -y -c 32768 --gpu max --identifier "devstral"

---

Step 2 — Check and Start the Model Router Proxy

The proxy (model_router.py) routes Devstral requests to LM Studio and everything else to Anthropic. It must be running before spawning the agent.

Check if proxy is running

Look for the PID file relative to the project root:

PROXY_PID_FILE="${CLAUDE_PROJECT_DIR:-.}/.claude/plugins/kuro/proxy/proxy.pid"
if [ -f "$PROXY_PID_FILE" ]; then
    PID=$(cat "$PROXY_PID_FILE" 2>/dev/null)
    if [ -n "$PID" ] && kill -0 "$PID" 2>/dev/null; then
        echo "PROXY_RUNNING pid=$PID"
    else
        echo "PROXY_STALE_PID"
        rm -f "$PROXY_PID_FILE"
    fi
else
    echo "PROXY_NOT_RUNNING"
fi

Start proxy if not running

If PROXY_NOT_RUNNING or PROXY_STALE_PID:

PROXY_SCRIPT="${CLAUDE_PROJECT_DIR:-.}/.claude/plugins/kuro/proxy/model_router.py"
uv run "$PROXY_SCRIPT" &
echo "Proxy started, waiting for startup..."
sleep 2

Verify proxy health

curl -s --max-time 5 http://127.0.0.1:5111/health

Expected response: {"status": "ok", "port": 5111}

If the health check fails after starting, report:

• "Proxy failed to start. Check that uv is installed and model_router.py exists at .claude/plugins/kuro/proxy/model_router.py."
• Stop here.

---

Step 3 — Discover Available LiteCLI Tools

LiteCLI is a universal CLI generator that turns MCP server tools into Click commands. Devstral's primary way of interacting with the system is through these generated CLIs.

Run discovery to find available backends and tools:

cd "${LITECLI_DIR:-<your-litecli-dir>}" && uv run -m litecli --json admin discover

This returns JSON with each backend's id, prefix, state, tool_count, and description. Example output:

[
  {
    "id": "my-mcp-backend",
    "type": "http",
    "prefix": "mcp",
    "state": "connected",
    "enabled": true,
    "tool_count": 25,
    "description": "Your MCP Core"
  }
]

For each connected backend, you can list its tools by running:

cd "${LITECLI_DIR:-<your-litecli-dir>}" && uv run -m litecli --json <prefix> --help

Format a tool catalog for the Devstral prompt. Example:

AVAILABLE LITECLI TOOLS:
Backend: mcp (Your MCP Core) — 25 tools
  Invocation: litecli --json mcp <tool_name> [args]

Backend: tools (Your Tools MCP) — 12 tools
  Invocation: litecli --json tools <tool_name> [args]

If LiteCLI is not installed or discovery fails, proceed without the tool catalog — Devstral can still do code analysis and generation tasks without it.

Important: The --json flag ensures machine-parseable output. Always include it when Devstral runs litecli commands.

---

Step 4 — ANTHROPIC_BASE_URL Warning

IMPORTANT: To route Devstral through the proxy, ANTHROPIC_BASE_URL must be set to http://127.0.0.1:5111 for the Claude Code session. This affects ALL API calls in the session. The proxy handles routing correctly:

• Model contains "devstral" or "mistral" → routed to LM Studio at http://localhost:1234
• Everything else → routed to Anthropic at https://api.anthropic.com

Proxy resilience caveat: If the proxy process dies mid-session, ALL API calls will fail (including your own calls back to Anthropic). The proxy is lightweight and stable, but worth noting.

Check if already set:

echo "ANTHROPIC_BASE_URL=${ANTHROPIC_BASE_URL:-<not set>}"

If ANTHROPIC_BASE_URL is not already pointing at the proxy, inform the user:

"Note: To spawn Devstral via the Agent tool, ANTHROPIC_BASE_URL must be set to http://127.0.0.1:5111. This requires either restarting the session with that env var set, or setting it in your shell before launching Claude Code."

"Alternatively, I can query Devstral directly via HTTP (no env var needed) — this works for single-turn tasks but not for agent-style multi-turn work."

If the user wants to proceed with direct HTTP (simpler, no env var required), jump to Step 4b. If the user wants full agent spawning, they must restart with ANTHROPIC_BASE_URL=http://127.0.0.1:5111 set.

---

Step 5a — Spawn Devstral Agent (if ANTHROPIC_BASE_URL is set)

Build the full agent prompt combining:

1. The user's task
2. The LiteCLI tool catalog from Step 3
3. Available file context (paths, code snippets if mentioned)
4. System instructions tuned for Devstral's coding-agent behavior

Then spawn using the Agent tool:

Agent(
  subagent_type="devstral-worker",
  prompt="<assembled prompt — see below>"
)

Prompt assembly template

You are Devstral Small 2, a coding agent running locally on LM Studio.

TASK:
<user's task from $ARGUMENTS>

LITECLI TOOLS:
You can interact with MCP servers via LiteCLI. Always use --json for parseable output.
Invocation: litecli --json <prefix> <tool_name> [args]

<insert the tool catalog from Step 3 here, e.g.:>
Backend: mcp (Your MCP Core) — 25 tools
  Example: litecli --json mcp rag_query --query "search term"
  Example: litecli --json mcp read_file --path "src/main.py"

INSTRUCTIONS:
- Focus solely on the task above
- Use litecli --json commands via Bash to interact with MCP tools
- Do not emit raw [TOOL_CALLS] tokens — use Bash + litecli instead
- If writing code, include the full file path at the top as a comment
- Be concise and direct. No preamble.
- If the task requires reading files, use the Read tool or litecli

CONTEXT:
<include any file content, error messages, or code snippets the user referenced>

CRITICAL for Devstral: If the tool catalog is empty (LiteCLI not available), omit the LITECLI TOOLS section and instruct Devstral to use Read/Glob/Grep/Bash directly instead.

---

Step 5b — Direct HTTP Query (fallback, no env var required)

When ANTHROPIC_BASE_URL is not set or the user wants a simpler path, query Devstral directly via the LM Studio API. This works for single-turn tasks (write this function, review this code, explain this error).

python -c "
import json, urllib.request, sys

task = '''<USER_TASK_HERE>'''

payload = json.dumps({
    'model': 'devstral',
    'messages': [
        {
            'role': 'system',
            'content': 'You are a precise coding assistant. Answer directly. Do not use any tools or tool calls. Output only code or analysis as requested.'
        },
        {
            'role': 'user',
            'content': task
        }
    ],
    'max_tokens': 4096,
    'temperature': 0.1
}).encode()

req = urllib.request.Request(
    'http://localhost:1234/v1/chat/completions',
    data=payload,
    headers={'Content-Type': 'application/json'}
)
try:
    resp = json.loads(urllib.request.urlopen(req, timeout=120).read())
    print(resp['choices'][0]['message']['content'])
except Exception as e:
    print(f'ERROR: {e}', file=sys.stderr)
    sys.exit(1)
"

For large code files or content with special characters, write to a temp file first:

cat > /tmp/devstral_task.py << 'SCRIPT_EOF'
import json, urllib.request, sys

task = open('/tmp/devstral_input.txt').read()

payload = json.dumps({
    'model': 'devstral',
    'messages': [
        {
            'role': 'system',
            'content': 'You are a precise coding assistant. Answer directly. Do not use any tools or tool calls.'
        },
        {'role': 'user', 'content': task}
    ],
    'max_tokens': 4096,
    'temperature': 0.1
}).encode()

req = urllib.request.Request(
    'http://localhost:1234/v1/chat/completions',
    data=payload,
    headers={'Content-Type': 'application/json'}
)
resp = json.loads(urllib.request.urlopen(req, timeout=120).read())
print(resp['choices'][0]['message']['content'])
SCRIPT_EOF

python /tmp/devstral_task.py

---

Step 6 — Present Results

When Devstral returns:

1. Present the response clearly with attribution: "Devstral (local):"
2. If code was returned, show it in proper fenced code blocks with language tags
3. If the task involved editing files, apply the changes using the Edit tool
4. Note approximate inference time if remarkable (Devstral 24B @ RTX 5090 is typically 10-30s for most tasks)
5. If Devstral's response included [TOOL_CALLS] artifacts, strip them and re-run with a stronger "no tool calls" instruction

---

Proxy Lifecycle Notes

Scenario	Action
Proxy already running	Skip startup, proceed directly
Proxy PID file exists but process gone	Delete stale PID, restart proxy
Proxy failed to start	Report error with uv path, stop
LM Studio unreachable	Warn user, stop
Devstral not loaded	Offer to load via lms.exe, wait ~12s
ANTHROPIC_BASE_URL not set	Warn, offer direct HTTP fallback

To stop the proxy manually:

kill $(cat "${CLAUDE_PROJECT_DIR:-.}/.claude/plugins/kuro/proxy/proxy.pid")

To check proxy logs (if started in this session):

# Proxy logs go to stderr — check terminal where Claude Code was launched

---

Quick Reference

Item	Value
Proxy port	http://127.0.0.1:5111
LM Studio API	http://localhost:1234
Devstral model ID	mistralai/devstral-small-2-2512
Devstral identifier (LMS)	devstral
VRAM usage	~15.2 GB (Q4_K_M)
Context length	32768 (recommended), 65536 (max, solo)
Proxy script	.claude/plugins/kuro/proxy/model_router.py
PID file	.claude/plugins/kuro/proxy/proxy.pid
uv path	uv (must be on PATH — install from https://docs.astral.sh/uv/)