opencli-reader

name: opencli-reader description: > Read-only access to the long tail of sites Alice's own tools do NOT cover, via the community opencli CLI (~160 site adapters): social sentiment (Reddit, HackerNews, Twitter/X, Bluesky, Xueqiu 雪球, Weibo), options flow (Barchart), crypto long-tail (CoinGecko, DeFiLlama, Binance), global news frontpages (Bloomberg, Reuters, BBC), CN money-flow (Eastmoney 东方财富 northbound / longhu / money-flow / hot-rank), research (arXiv, PubMed, Google Scholar), and a generic web read fallback. Triggers: "what's reddit / 雪球 saying about X", "unusual options flow on Y", "TVL of Z", "Bloomberg headlines", "northbound flow today", "search arXiv for…", any read from a site Alice has no tool for. READ-ONLY — never invoke write commands. opencli is NOT bundled with OpenAlice: if it's missing and a task would benefit, you MUST say so and ask the user — never install silently, never silently work with thinner data.

opencli — read the sites Alice doesn't ship (optional, read-only)

opencli is an independent community CLI that turns ~160 websites and desktop apps into a uniform opencli <site> <command> surface. OpenAlice does not bundle or install it — this skill teaches you to use it when the user has it, and to ask for it when a task needs data only it can reach.

Where it sits next to Alice's own tools

The boundary is capability-based, not site-based: if Alice ships the number, use Alice's tool — that keeps data 口径 consistent with what the trading engine sees.

When unsure whether Alice covers something, check traderhub --help and alice --help first — that boundary moves as Alice grows; the CLIs are the source of truth, not this table.

Hard rule: opencli data never directly feeds a trading decision. Don't size a position, route an order, or set a limit price off an opencli number. Use it for narrative, sentiment, screening leads, and research context; when a lead matters, re-anchor the actual numbers through Alice's tools before acting.

Behavior contract (read this even if you skip the rest)

1. Never install silently. Installation touches the user's machine globally (npm install -g). Propose it; the user decides.
2. Never silently lack data. If the task would materially benefit from a source opencli covers and it isn't installed (or the needed login isn't set up), say exactly what's missing, what it would unlock, and how to set it up — then ask. Proceeding quietly with thinner data is a failure mode.
3. Read-only, always. Never invoke commands that mutate state: post, reply, comment, like, unlike, upvote, downvote, save, subscribe, unsubscribe, follow, unfollow, block, delete, bookmark, send, create-draft, reply-dm, accept, and anything whose description suggests a mutation. Unsure → don't run it.

Step 1 — Is it installed?

command -v opencli

Not installed? Don't stop at "I can't." Tell the user:

This question would benefit from (e.g. Reddit sentiment, Eastmoney northbound flow), which Alice's own tools don't cover. The community opencli CLI can read it. Install with: npm install -g @jackwener/opencli (needs Node ≥ 20) Want me to proceed without it, or will you install it?

Then respect the answer. If they decline, note in your output which angle is missing — don't let the gap disappear.

Step 2 — Discover the command. Never guess.

The registry has 500+ commands across ~160 sites and changes weekly. Flags and names must come from live discovery, not memory:

opencli list -f json            # full registry, machine-readable
opencli list | grep -i <site>   # filter to a site
opencli <site> --help           # a site's commands
opencli <site> <command> --help # args, flags, defaults

Each opencli list -f json entry carries site, name, description, strategy, args, columns — see references/discovery.md for the schema and for telling read from write commands.

Step 3 — Check strategy before running

If a browser-backed adapter is needed and the user isn't set up, that's a Behavior-contract-#2 moment: explain (install extension from the Web Store, log into the site in Chrome, opencli doctor to verify the bridge) and ask. opencli doctor diagnoses the browser bridge only — PUBLIC/LOCAL adapters don't need it green.

Step 4 — Execute

opencli <site> <command> [args] [flags] -f json

• Always -f json for processing; -v for debug logs.
• Start with a small --limit (10–20) to validate the shape before pulling more. Command-specific flags are not universal — check --help.

opencli reddit subreddit wallstreetbets --limit 20 -f json
opencli hackernews top --limit 20 -f json
opencli eastmoney hot-rank -f json
opencli xueqiu hot-stock -f json
opencli barchart flow AAPL -f json
opencli defillama --help          # discover before first use
opencli arxiv search "volatility surface" --limit 10 -f json
opencli web read "https://example.com/article" -f json

Step 5 — When it fails

Sites change; adapters break. There's a built-in repair loop:

OPENCLI_DIAGNOSTIC=1 opencli <site> <command> <args>

This emits a structured RepairContext. Suggest the user file it at https://github.com/jackwener/opencli/issues. Never fall back to hand-rolled curl/fetch scraping — that hides the breakage from the registry and gives you a parser that rots in a week. Empty results can also just be rate limits; wait and retry once before declaring breakage.

Step 6 — Present, don't dump

• Summarize for the user's actual question; don't paste raw JSON.
• Attribute each item (site + URL where available).
• Posts/news: headline, timestamp, key quotes. Papers: title, authors, abstract, link.
• Label the provenance honestly: "Reddit sentiment via opencli", not "the market thinks". And per the hard rule above — if a finding should drive a trade, re-anchor through traderhub/alice first.
• Browser sessions are private: never echo cookies, CDP endpoints, or auth tokens into output.

Maintained by OpenAlice. Doctrine follows the official jackwener/opencli skills; structure informed by opencli-reader in himself65/finance-skills (MIT). opencli itself is an independent project — adapter bugs go upstream.