ghc-search

Web and X search through a local ghc-proxy Responses API. The script owns request building, parsing, inline-citation stripping, and URL dedup — you get an answer plus a clean source list.

Use it

The engine is a required first argument: gpt searches the web, x searches X/Twitter. They hit different corpora, so there is no default.

The script is scripts/ghc_search.py in this skill's own directory — under ~/.claude/skills/ghc-search/ for a global Claude Code install, and the equivalent skills/ dir for any other runtime npx skills installed it into.

S=~/.claude/skills/ghc-search/scripts/ghc_search.py   # adjust per runtime

python "$S" gpt "what is the current stable .NET SDK version"
python "$S" x --handle SpaceXAI "newest post"
python "$S" gpt --domain dotnet.microsoft.com --json "latest .NET SDK version"
python "$S" x --handle dotnet --from-date 2026-08-01 "what shipped this month"

A flag from the wrong engine is rejected before any network call.

Configuration

Defaults assume a ghc-proxy on 127.0.0.1:4141. Override without flags via environment variables (a different proxy host, or different default models):

--model and --endpoint flags win over env vars.

What you need to know

• Latency ~3-20s measured; date-filtered X searches are the slow end.
• Model must support the responses endpoint. Claude models return HTTP 400 here.
• --handle restricts results but does not steer the model, so the script appends the handles to the prompt for you. That is why a terse x --handle X "newest post" works; without the appending it draws a clarifying question instead of a search.
• An empty source list is possible — the model sometimes answers from its own knowledge without searching. No sources means no citation; judge the answer accordingly, and add --domain or ask it to cite if you need provenance.
• A response cut off by --max-tokens keeps its answer, flagged [truncated], rather than being discarded. Raise --max-tokens and re-run if the tail matters.
• Exit code is non-zero with a one-line stderr message on any failure; a partial parse is never printed as success.

Prerequisite

ghc-proxy must be running locally — this skill is useless without it:

curl -sS -o /dev/null -w "%{http_code}\n" http://127.0.0.1:4141/v1/models \
  -H "authorization: Bearer dummy"

200 = up, and that endpoint also lists the models you can pass to --model. The script prints ghc-proxy not reachable at ... (with the configured endpoint) when it is not.

When this is not the right tool

Without a reachable ghc-proxy this skill cannot run; fall back to whatever browser automation your runtime offers (e.g. drive Bing with &setlang=en; Google may serve a block page).

Reach for something else when you already have what you need: fetch the page directly when you have the URL, and prefer a documentation source (context7 MCP) over general search when looking up a named library or API.

Check

python "$(dirname "$S")/test_ghc_search.py"   # offline, no network required