evolve-install

claude-evolve Install

This is the setup command for claude-evolve. Run it once after installing the plugin to verify Python dependencies, create a virtual environment if needed, and confirm the MCP server is registered.

When this skill is invoked, immediately execute the workflow below. Do not only restate or summarize these instructions back to the user.

Flag Parsing

Check for flags in the user's invocation:

• --help → Show Help Text (below) and stop
• --force → Skip Pre-Setup Check, run full setup from scratch
• --verify → Only run the verification phase (Phase 4), skip install/env setup
• --local → Install into a project-local .venv
• --global → Use the system Python (no venv)
• No flags → Auto-detect and run full setup if needed

Help Text

When user runs with --help, display this and stop:

claude-evolve Install - Set up evolutionary code discovery plugin

USAGE:
  /evolve-install              Install or verify (auto-detect)
  /evolve-install --local      Install into a project-local .venv
  /evolve-install --global     Use system Python (no venv)
  /evolve-install --verify     Only verify existing install
  /evolve-install --force      Force full reinstall
  /evolve-install --help       Show this help

PHASES:
  1. Python check       Verify Python >= 3.10 is available
  2. Environment setup  Create venv (if --local or auto) and install claude-evolve
  3. Claude CLI check   Verify `claude` is on PATH and authenticated
  4. MCP verification   Confirm claude-evolve MCP server registers and exposes tools
  5. Smoke test         Run a trivial claude call to confirm the bridge works

NEXT STEPS:
  After install completes, run `/evolve-interview` to define your first
  evolution task, or `/evolve` if you already have initial.py + evaluate.py.

Pre-Setup Check: Already Installed?

Before doing anything else, check if the plugin is already installed and working. Read the state file:

CLAUDE_EVOLVE_STATE="${CLAUDE_PLUGIN_ROOT:-.}/state/install-state.json"

If the state file exists AND reports "status": "ok" AND no --force flag is set, use AskUserQuestion to prompt:

Question: "claude-evolve is already installed. What would you like to do?"

Options:

1. Verify install - Run the verification phase only (quick health check)
2. Reinstall - Run the full setup wizard again
3. Cancel - Exit without changes

On "Verify install" → jump to Phase 4. On "Reinstall" → continue with Phase 1 below. On "Cancel" → exit.

Phase 1: Python Check

Use Bash to verify Python:

python3 --version 2>&1 || python --version 2>&1

Requirements:

• Python >= 3.10
• If not available, tell the user to install Python 3.10+ from python.org or their package manager and stop

Report the detected Python path and version.

Phase 2: Environment Setup

Determine installation mode:

• If --global: use system Python, skip venv creation
• If --local or default: create core/.venv/ in the plugin directory

Steps:

1. Check if core/.venv/ exists. If not (and mode is local), create it:

   python3 -m venv "${CLAUDE_PLUGIN_ROOT:-.}/core/.venv"
   

2. Install claude-evolve into that Python:

   PYTHON="${CLAUDE_PLUGIN_ROOT:-.}/core/.venv/bin/python"   # or system python3 if --global
   "$PYTHON" -m pip install -e "${CLAUDE_PLUGIN_ROOT:-.}/core"
   

3. Check for install errors. Common issues:

   • Missing build tools (setuptools, wheel)
   • Missing mcp package on PyPI
   • scipy/numpy compilation failures on older systems

If install fails, report the stderr tail and stop.

Phase 3: Claude CLI Check

Verify the Claude CLI is available and authenticated:

claude --version 2>&1

If not installed, tell the user:

The claude CLI is required. Install it from https://claude.com/code/install

Then run a tiny non-authenticated smoke:

claude --help 2>&1 | head -5

Phase 4: MCP Verification

Verify the MCP server starts and exposes all expected tools.

Important: Do NOT use the timeout or gtimeout shell commands here. They are Linux-only and GNU-coreutils-on-macOS respectively, and neither is guaranteed to exist on a fresh install. Use Python's built-in subprocess timeout instead, which works everywhere Python works (and Python is guaranteed present by Phase 2).

Run the probe by inlining a short Python script that drives the MCP handshake. Use Popen + a deadline reader — do not use subprocess.run(input=...), because closing stdin triggers FastMCP shutdown before the tools/list response is flushed.

"$PYTHON" - <<'PYEOF'
import json, subprocess, sys, time

init = {"jsonrpc":"2.0","id":1,"method":"initialize",
        "params":{"protocolVersion":"2024-11-05","capabilities":{},
                  "clientInfo":{"name":"evolve-install","version":"1.0"}}}
notif = {"jsonrpc":"2.0","method":"notifications/initialized"}
list_req = {"jsonrpc":"2.0","id":3,"method":"tools/list","params":{}}

payload = "\n".join(json.dumps(m) for m in [init, notif, list_req]) + "\n"

proc = subprocess.Popen(
    [sys.executable, "-m", "claude_evolve.server"],
    stdin=subprocess.PIPE,
    stdout=subprocess.PIPE,
    stderr=subprocess.PIPE,
    text=True,
    bufsize=1,
)

proc.stdin.write(payload)
proc.stdin.flush()

deadline = time.time() + 15
tools_line = None
try:
    while time.time() < deadline:
        line = proc.stdout.readline()
        if not line:
            break
        if '"id":3' in line and '"tools"' in line:
            tools_line = line
            break
finally:
    try:
        proc.stdin.close()
    except Exception:
        pass
    proc.terminate()
    try:
        proc.wait(timeout=3)
    except subprocess.TimeoutExpired:
        proc.kill()

if tools_line is None:
    err = proc.stderr.read()[:500] if proc.stderr else ""
    print(f"MCP_PROBE_FAIL: no tools/list response within deadline. stderr={err}")
    sys.exit(1)

data = json.loads(tools_line)
tools = [t["name"] for t in data["result"]["tools"]]
expected = {"evolve_start", "evolve_status", "evolve_stop", "evolve_visualize", "evaluator_create"}
missing = expected - set(tools)

if missing:
    print(f"MCP_PROBE_FAIL: missing tools {sorted(missing)}. Got: {tools}")
    sys.exit(1)

print(f"MCP_PROBE_OK: {len(tools)} tools exposed: {sorted(tools)}")
PYEOF

Parse the output — look for MCP_PROBE_OK on success or MCP_PROBE_FAIL: on failure. If it fails, report the stderr tail and stop.

Also run claude mcp list 2>&1 and confirm an entry matching claude-evolve (or the short name e) appears with ✓ Connected. If it doesn't, the user may need to restart Claude Code or run claude mcp add manually — provide the exact command.

Phase 5: Smoke Test (Optional)

Run one real Claude CLI call through the bridge to confirm end-to-end:

"$PYTHON" -c "
import asyncio
from claude_evolve.ensemble.bridge import query_claude_async
result = asyncio.run(query_claude_async(arm='haiku/low', prompt='Respond with just OK', timeout=60))
print('Bridge smoke test:', result.content[:50])
"

If this succeeds, the full stack works.

If it fails with an auth error, tell the user to run claude once interactively to authenticate.

Phase 6: HUD Setup (Status Line)

Set up the claude-evolve HUD that displays evolution progress in the Claude Code status bar. This appends a second line below any existing HUD (e.g. OMC's).

Steps:

1. Read the current statusLine.command from ~/.claude/settings.json (or $CLAUDE_CONFIG_DIR/settings.json).

2. If a statusLine already exists (e.g. OMC's HUD), save it to ~/.claude-evolve/original-hud.json so the combiner can chain it:

   {"command": "<the existing statusLine.command value>"}
   

3. If no statusLine exists yet, write ~/.claude-evolve/original-hud.json with {"command": null}.

4. Set the statusLine command to the claude-evolve combiner:

   {
     "statusLine": {
       "type": "command",
       "command": "node ${CLAUDE_PLUGIN_ROOT}/bridge/hud-combiner.mjs"
     }
   }
   

   IMPORTANT: Use the statusline-setup agent or directly patch ~/.claude/settings.json to update the statusLine field. Do NOT overwrite other settings.json fields.

5. If the current statusLine already points to hud-combiner.mjs, skip (already configured).

What the HUD shows during an active evolution run:

evolve │ ⟳ gen 15 │ best 2.4103 │ 30 calls │ 12 progs

Status icons: ⟳ (yellow, running) | ✓ (green, complete) | ✗ (red, crashed/dead)

When no evolution run is active, the evolve line is hidden — only the original HUD shows.

Save State

After all phases succeed, write state to state/install-state.json:

{
  "status": "ok",
  "installed_at": "<ISO timestamp>",
  "python_path": "<path>",
  "python_version": "<version>",
  "venv_path": "<path or null>",
  "mcp_tools": 5,
  "claude_cli_version": "<version>",
  "plugin_version": "0.1.0"
}

Report Success

Tell the user:

✅ claude-evolve installed and verified.

• Python: at
• Venv: <path or "system">
• MCP tools: 5 (evolve_start, evolve_status, evolve_stop, evolve_visualize, evaluator_create)
• Claude CLI:
• HUD: <configured / already configured / skipped>

Next step: Run /evolve-interview to define your first optimization task interactively, or /evolve if you already have initial.py and evaluate.py.

Graceful Interrupt Handling

If any phase fails, save partial state to state/install-state.json with "status": "partial" and the phase that failed. On next invocation, offer to resume from the failed phase.

Troubleshooting

• "Module mcp not found" → pip install mcp into the target Python
• "scipy build error" → install a pre-built wheel or upgrade pip
• MCP tools list empty → the Python import failed; check python -m claude_evolve.server directly for traceback
• Claude CLI "Not logged in" → remove --bare flag if present; run claude interactively once