opencli-llm-advisors

name

opencli-llm-advisors

description

Use opencli to drive LLM chat apps (Gemini web, ChatGPT Desktop) as external advisors — bypasses API 429/quota by using the user's logged-in browser/desktop-app session. Triggers when: Gemini CLI returns 429 or capacity-exhausted, user wants a CCG advisor and API is unavailable, user wants to use their ChatGPT Plus subscription instead of API, or user says 'ask Gemini/ChatGPT via opencli'. Covers `/u/N` account switching, clicking the in-page mode picker to select highest-tier model (Gemini Pro 3.1), ChatGPT Desktop native control, and common failure modes. Does NOT apply for: normal API calls that work, image generation (use `opencli chatgpt image` directly), or headless/CI workflows.

OpenCLI as LLM Advisor Bridge

⚠️ Experimental / unofficial. This skill automates vendor web UIs (Gemini, ChatGPT) via opencli's browser bridge. It is inherently fragile:

• Upstream UI changes may break the DOM selectors overnight. When that happens, the adapter inside opencli needs updating — not this skill.
• Terms of Service: automating chat UIs may violate provider ToS. Use for personal productivity only; avoid high-volume scraping.
• Account safety: excessive automation can get accounts flagged. Keep per-minute rates low.

Use only as a fallback when the official API path is unavailable. If your API quota works, use that instead.

When the official API path fails (429, quota, auth), or when the user wants a higher-tier model than their API key allows, opencli can drive the web UI or desktop app of the LLM instead — reusing your logged-in browser / installed app session. No tokens consumed on the API side, no rate limits against free tiers.

Complements the official OpenCLI skills (opencli-usage, opencli-browser, opencli-adapter-author). Those cover generic opencli capabilities. This skill covers how we specifically use it as an LLM advisor in this user's environment.

When to Use

• gemini-cli (API path) returned 429 / MODEL_CAPACITY_EXHAUSTED / auth issue
• CCG flow (/oh-my-claudecode:ccg) needs a Gemini response but API is down
• User explicitly wants Gemini 3.1 Pro or GPT-5.2 Thinking via their Plus subscription
• Want to cross-validate a Claude answer with a different frontier model, no budget
• ask-gemini / ask-codex helpers failed; need fallback

When NOT to Use

• Normal programmatic calls in working CI (opencli needs a GUI Chrome session)
• Image generation → use opencli chatgpt image / opencli gemini image directly (these are already simple)
• When headless / no GUI available
• When the user's own browser Chrome window is not logged into the target service

User's Environment (specific to this setup)

Service	Preferred account	Notes
Gemini web	Your preferred Google account — use /u/N where N is the account index (/u/0 is the default-logged-in account, /u/1 is the second, etc.)	Always navigate to the correct /u/N/app URL before calling opencli gemini
ChatGPT Desktop	Native macOS app	Use opencli chatgpt-app — native, not web
Chrome profile	Default profile	opencli uses whichever profile's browser is connected via the Browser Bridge extension

Recipe 1 — Gemini 3.1 Pro (free-tier highest)

Goal: Send a prompt to Gemini with the highest-tier free model (Pro / 3.1 Pro), not the default Fast mode.

# 1. Navigate to the right account (/u/1) AND open new chat
opencli browser open "https://gemini.google.com/u/1/app"
sleep 2

# 2. Find the mode picker button (labeled "Fast" or current mode)
opencli browser find --css "bard-mode-switcher button, [aria-label='Open mode picker']"
# Note the ref number from output — likely around 29 after page loads

# 3. Click to open mode picker, then list options
opencli browser click <ref>
sleep 1
opencli browser find --css "[role=menuitem], button[role=menuitemradio]" --text-max 200

# 4. Click the "Pro" menu item (text: "Pro  Advanced math and code with 3.1 Pro")
opencli browser click <pro-ref>

# 5. Now send the prompt — adapter will use the Pro model
opencli gemini ask "your prompt here" --timeout 120

Why click instead of CLI flag: opencli gemini ask has no --model flag. Mode is set via UI state, persists for the session.

Timeout: Pro mode "thinks" — set --timeout 120 or higher. Default 60s will cut Pro mid-reasoning.

Recipe 2 — ChatGPT Desktop as advisor

Goal: Ask ChatGPT Desktop (native macOS app) for a second opinion. Uses Plus subscription, zero API spend.

# Pre-flight: ensure ChatGPT Desktop is running
opencli chatgpt-app status
# If not running: launch ChatGPT.app via Spotlight or `open -a "ChatGPT"`

# 1. Open a new chat
opencli chatgpt-app new

# 2. (Optional) Switch model to thinking mode for harder questions
opencli chatgpt-app model thinking   # or: 5.2-thinking / 5.2-instant

# 3. Send the prompt
opencli chatgpt-app send "your prompt here"

# 4. Wait for the response (no built-in wait — poll read)
sleep 15
opencli chatgpt-app read
# If still generating, sleep longer and read again

Gotcha: chatgpt-app read returns the last visible message. If ChatGPT is still streaming, you may get partial text. Poll until content stops growing.

Recipe 3 — Account switching (general pattern)

OpenCLI has no --account flag. It uses whatever session is active in the current browser tab.

To use a non-default Google account (e.g., /u/1 instead of /u/0):

# Always pre-navigate to the /u/N variant of the URL
opencli browser open "https://gemini.google.com/u/1/app"
# Then run the normal adapter command — it operates on whatever tab is active
opencli gemini ask "..."

Same pattern for any Google-authenticated site.

Recipe 4 — CCG (Claude-Codex-Gemini) via opencli fallback

When omc ask gemini fails with 429, the original /oh-my-claudecode:ccg flow can still complete by substituting opencli for the Gemini leg:

# Codex leg (API path, usually fine)
omc ask codex "<codex prompt>"

# Gemini leg via opencli (no API, no 429)
opencli browser open "https://gemini.google.com/u/1/app"
# ... select Pro per Recipe 1 ...
opencli gemini ask "<gemini prompt>" --timeout 120 > /tmp/gemini-response.md

# Claude synthesizes both as usual

Failure Modes & Workarounds

Symptom	Likely cause	Fix
opencli gemini ask hangs or times out at 60s	Pro mode thinking longer	Add --timeout 120 or --timeout 180
Wrong Google account replies	Active tab is on /u/0, not /u/1	Always browser open the /u/N URL first
Mode picker button ref not 29	Page layout changed or lazy load	Re-run browser find --css "bard-mode-switcher button" to get fresh ref
"Upgrade" shown instead of Pro option	Free-tier quota for Pro used up that day	Switch to Thinking mode (still high-tier) or use chatgpt-app instead
ChatGPT Desktop returns partial message	Still streaming	Sleep longer, re-read until stable
opencli chatgpt-app send but nothing happens	App not focused / not running	Check chatgpt-app status; focus window manually if needed
Adapter targets wrong tab	opencli binds to the "site workspace" tab, not the one you have open	Use opencli browser open to seed the correct URL into the site workspace first

Model Tier Reference (as of 2026-04)

Gemini web (via opencli gemini ask):

• Fast (default) — low-effort answers
• Thinking — medium, good for most advisor queries
• Pro (3.1 Pro) — highest free-tier, use for architecture/review tasks
• Upgrade — paid, skip unless user has Gemini Advanced

ChatGPT Desktop (opencli chatgpt-app model <mode>):

• auto — ChatGPT picks
• instant / 5.2-instant — fast
• thinking / 5.2-thinking — higher reasoning, recommended for advisor role

Integration with Existing Skills

Task	Skill to use
First-time setup / CLI install questions	opencli-usage
Writing a new adapter for a new site	opencli-adapter-author
Fixing a broken built-in adapter	opencli-autofix
Low-level browser primitives	opencli-browser
Using an LLM as advisor/fallback (this skill)	opencli-llm-advisors

Anti-patterns

• Don't use opencli when the API path works. API is faster, scriptable, idempotent. opencli is a fallback, not a default.
• Don't leave chat history accumulating in the browser. Gemini/ChatGPT web context can bleed between calls. Always browser open .../app fresh or chatgpt-app new before a new advisor query.
• Don't assume the model persists. If the user restarts the browser, mode resets to Fast / default. Re-run the mode picker flow.
• Don't send sensitive prompts through the chat UI without thinking about it — these go through the web service with regular chat-history retention, unlike opaque API calls.
• Don't wait forever. If Pro mode takes > 3 min, something's wrong — abort, retry with smaller prompt.

Quick one-liner wrappers

Add to ~/.zshrc for fast access:

# Gemini Pro on account /u/1
ask-gemini-pro() {
  opencli browser open "https://gemini.google.com/u/1/app" > /dev/null
  opencli gemini ask "$@" --timeout 180
}

# ChatGPT Desktop with thinking mode
ask-gpt-think() {
  opencli chatgpt-app new > /dev/null
  opencli chatgpt-app model thinking > /dev/null
  opencli chatgpt-app send "$@"
  sleep 20
  opencli chatgpt-app read
}