enable-smart-prompts

Enable Smart-Recall Prompts

You help the user install a customized UserPromptSubmit hook that intercepts their prompt, uses a small model to formulate a good NeoHive query, calls memory_recall, and injects relevant results back into Claude's context.

This is a dynamic setup — every user has a different hive layout, shell, API key location, and tolerance for latency. You walk them through each choice with a strong recommended default, then write the script.

Phase 0 — Check prerequisites

Before asking anything, verify:

command -v claude >/dev/null && echo "claude-cli: OK" || echo "claude-cli: MISSING"
command -v curl   >/dev/null && echo "curl: OK"       || echo "curl: MISSING"
command -v python3>/dev/null && echo "python3: OK"    || echo "python3: MISSING"
[ -n "${ANTHROPIC_API_KEY:-}" ] && echo "ANTHROPIC_API_KEY: set" || echo "ANTHROPIC_API_KEY: not set (required for headless claude)"

If claude-cli or python3 is missing, stop and tell the user to install them. If the API key is missing, tell them to export it first and point at https://console.anthropic.com/.

Phase 1 — Gather configuration

Ask these in sequence, one per AskUserQuestion call (never combine):

1. Which hive to target

Call list_hives. Ask:

• Header: "Target hive"
• Question: "Which hive should the hook search on every prompt?"
• Options: populate from list_hives, first option (Recommended) All hives (cross-hive RRF) — this calls memory_recall without a hive param.

2. Which model drives the query rewriter

• Header: "Rewriter model"
• Question: "Which model should rewrite your prompt into a NeoHive query and filter results?"
• Options:
  • claude-haiku-4-5 (Recommended) — fast + cheap
  • claude-sonnet-4-6 — more accurate, slower, ~10x cost
  • claude-opus-4-7 — overkill, only for very noisy hives

3. Trigger policy

• Header: "When to run"
• Question: "When should the hook fire?"
• Options:
  • (Recommended) Every prompt longer than 10 chars — skips short clarifications
  • Only when prompt contains a keyword I pick
  • Every prompt — no filtering
  • Manual only — I'll trigger it via an env flag

If "keyword": ask for the keyword(s) via a follow-up AskUserQuestion with Other.

4. Install location

• Header: "Install location"
• Question: "Where should the hook live?"
• Options:
  • (Recommended) ~/.claude/hooks/neohive-smart-recall.sh — personal, all projects
  • ./.claude/hooks/neohive-smart-recall.sh — this project only
  • Just show me the script — I'll place it myself

5. Disable-flag name

• Header: "Disable flag"
• Question: "What env var should disable the hook when set to 1?"
• Options:
  • (Recommended) NEOHIVE_SMART_DISABLED
  • NEOHIVE_HOOK_DISABLED (same flag as the default dumb hook — one switch turns both off)
  • Custom — I'll type it

Phase 2 — Preview the generated script

Build the script from the template at ${CLAUDE_PLUGIN_ROOT}/skills/enable-smart-prompts/template.sh, substituting the chosen values. Show the final script to the user in a fenced code block. Summarize changes at the top:

Generated hook with:
  • Hive:          <hive-or-all>
  • Model:         <model>
  • Trigger:       <policy>
  • Install path:  <path>
  • Disable flag:  <env-var>

Ask one last AskUserQuestion:

• Header: "Install"
• Question: "Install this hook now?"
• Options:
  • (Recommended) Yes, write it and register it in settings.json
  • Yes, write it — I'll register it myself
  • No — I want to tweak the script first

If "tweak": ask what they want to change, regenerate, re-preview. If "no" at any point: stop with "Nothing written."

Phase 3 — Write the script

Create parent directories if needed. Write the script to the chosen location with mode 0755. Show:

Wrote <path> (N bytes, mode 755).

Phase 4 — Register in settings.json (if user opted in)

Edit ~/.claude/settings.json (personal) or ./.claude/settings.json (project) to add the hook under hooks.UserPromptSubmit. Preserve existing entries. If the file has an existing NeoHive UserPromptSubmit entry from the plugin itself, do not remove it — the two can coexist; warn the user that both will run.

Use python3 -c with json to do the edit — never hand-edit JSON via sed. Pattern:

python3 <<PY
import json, pathlib, os
p = pathlib.Path(os.path.expanduser("<settings-path>"))
data = json.loads(p.read_text()) if p.exists() else {}
data.setdefault("hooks", {}).setdefault("UserPromptSubmit", [])
entry = {
    "matcher": "",
    "hooks": [{"type": "command", "command": "<hook-path>", "timeout": 15000}]
}
if entry not in data["hooks"]["UserPromptSubmit"]:
    data["hooks"]["UserPromptSubmit"].append(entry)
p.parent.mkdir(parents=True, exist_ok=True)
p.write_text(json.dumps(data, indent=2))
print(f"Registered: {p}")
PY

Phase 5 — Verification

Tell the user:

Restart Claude (or run /reload-plugins) for the hook to take effect.

Test it: start a new session and ask about something you know is in your hive. You should see a block starting with "NeoHive smart context:" before Claude's reply.

Disable temporarily: export <DISABLE_FLAG>=1 in your shell. Disable permanently: remove the entry from your settings.json, or delete the script.

Important rules

• Never overwrite an existing hook at the target path without confirmation. If the file exists, show its contents and ask whether to replace.
• Never put the API key in the generated script. The script reads $ANTHROPIC_API_KEY at runtime.
• Never hardcode the hive UUID in the script. It discovers the MCP URL the same way the default hook does (via .mcp.json / ~/.claude.json).
• Always set a --max-time on every curl and claude -p call. A slow hook blocks every prompt.
• Gracefully exit 0 on any failure. A broken hook must never block the user's prompt from reaching Claude.