| name | consult-local-clis |
| description | Consult local AI CLIs (Claude Code, Gemini/Antigravity, OpenAI Codex) through the KJ CLIController gateway and bring their answers back into the conversation. Use this skill WHENEVER the user wants a second opinion from another model or asks to "consult", "ask", or "check with" the other CLIs — including Vietnamese phrasings like "tham vấn mấy ông CLI", "hỏi mấy ông kia", "hỏi claude/gemini/codex xem", "cho mấy ông CLI coi thử", "tham khảo ý kiến", or English ones like "ask the other models", "get a second opinion from gemini/codex/claude", "what do the local CLIs think", "cross-check with the other AIs". Trigger even when the user names just one model (e.g. "hỏi codex thử") or all of them. This is the right tool any time the intent is to delegate a question to the local claude/gemini/openai CLIs and synthesize what comes back. |
Consult Local AI CLIs (via CLIController gateway)
This skill lets you forward a question or task to the local AI CLIs — Claude Code,
Gemini/Antigravity (agy), and OpenAI Codex — running behind the local
KJ CLIController gateway (OpenAI-compatible API), then bring their replies back and
make sense of them for the user.
The user thinks of this as "tham vấn mấy ông CLI" — consulting the panel of CLIs.
The one tool you need: scripts/consult.py
Always go through this bundled script. It is dependency-free (Python stdlib only) and it
already handles the two things that bite people who call the gateway by hand:
- On Windows, an inline
curl -d '{...}' gets its quotes mangled by the shell. The script
builds the JSON itself, so this never happens.
- The
openai (Codex) provider on this host does not accept a system role / --system
flag — sending one returns HTTP 500. The script folds any --system text into the user
message for that provider automatically.
python scripts/consult.py --providers
python scripts/consult.py --model claude --prompt "<the question>"
python scripts/consult.py --all --prompt "<the question>"
python scripts/consult.py --model openai --prompt "<task>" --system "Be concise" --timeout 240
python scripts/consult.py --model codex --workspace . --prompt "Implement the parser"
The script prints each reply under a ===== MODEL ===== header. Pass --json if you'd
rather parse structured output. Exit code is non-zero if any provider errored.
Orchestration mode (--workspace): when you delegate a real task (not just a
question), pass --workspace <dir>. The worker runs inside that directory (via the gateway's
cwd) with file access, and the prompt is auto-wrapped so the worker reads AGENT_LOG.md
for prior context first and appends a dated summary (plus a STATUS / FILES_CHANGED / NEXT
footer) when done. This is the shared cross-step memory — see
orchestration/ORCHESTRATION.md in the gateway repo for the full protocol.
Run it from the skill directory, or give the absolute path to consult.py. The gateway
defaults to http://localhost:8080 with token my-secret-lan-token; override with
--base-url/--token or the env vars CLI_CONTROLLER_URL / CLI_CONTROLLER_TOKEN if the
user has changed them.
How to choose who to consult
Read the user's phrasing and pick the narrowest interpretation that fits:
-
User named specific model(s) → consult exactly those. "hỏi codex" → --model openai.
Note the name mapping: claude = Claude Code, gemini = Gemini/Antigravity, and
openai = OpenAI Codex (so "codex"/"openai"/"GPT" all map to --model openai).
-
User said "mấy ông CLI" / "cả ba" / "all the models" / "the panel" → use --all
and then synthesize (see below). gemini/agy is the slowest backend, so when speed
matters you can delegate to just --models claude,openai; when it doesn't, gemini is
fully capable — give it a longer --timeout (300+) and let it run.
-
Ambiguous "tham vấn" with no target → this is the default: pick the single CLI best
suited to the task rather than spamming all three. Rough guide:
- Codex (
openai, GPT-5) — pure code generation, algorithmic problems, tight
reasoning, "write/refactor this function".
- Claude Code (
claude) — agentic/codebase reasoning, multi-file design, careful
prose, nuanced judgement calls.
- Gemini (
gemini) — vision/multimodal, very large context, broad world knowledge.
Slowest backend, so give it a longer --timeout (300+) when you use it.
If the question is genuinely a "who's right / I want multiple perspectives" question
(architecture decisions, tradeoffs, reviews), prefer --all even when unprompted — that
is usually the spirit of "tham vấn".
When in real doubt about scope, ask the user once: one model or the whole panel?
Synthesizing multiple answers
When you consulted more than one CLI, don't just dump the three blocks back. Add a short
synthesis on top so the consultation is actually useful:
- Consensus — what they agree on (this is the high-confidence part).
- Divergence — where they differ, and your read on which is more credible and why.
- Recommendation — your own bottom line, informed by all of them.
Keep the raw per-model answers available below the synthesis (or offer to show them) so the
user can audit. Attribute claims to the model that made them.
Practical notes
- Default
--timeout is 120s. Bump to 240–300 for large code generation or long analyses;
the CLIs run real subprocesses and can be slow.
- If a provider returns an error, report it plainly and continue with the others — a partial
consultation still has value. A quick
--providers check explains most failures
(a CLI not installed/authenticated shows "available": false).
- This is local-first: prompts go only to the user's own machine via the gateway, never to a
cloud API directly. Reassure the user of that if relevant.
- Don't consult the CLIs for trivial questions you can answer yourself — the point is a
genuine second opinion or specialized horsepower, not a reflex.
