mcp-toolcall-lab stub

Static Pages evidence for MCP tool-call experiments.

mcp7 tools

Commit 4848e45b · main · 2026-09-24T00:36:25+09:00 · Merge pull request #57 from myon-bioinformatics/fix/pages-match-ironmate-web-ui-palette

This host's Wiki panel fetches Wikipedia from the browser (MediaWiki Action API, CORS origin=*, no MCP). The MCP tools and local GET /wiki form still need compose or stub_front serve.

Last Actions summary

Allowlisted snapshot from the last stub-pages GitHub Actions run (stub + MCP mock + CPU-class model on one compose network). Raw MCP logs are not on Pages. This host cannot keep Docker running.

{
  "source": "stub-pages",
  "verdict": "PASS",
  "antipattern_id": null,
  "case": "MCP_SUCCESS",
  "cpu_llm_ok": true,
  "cpu_llm_backend": "lite-stub",
  "cpu_llm_completion_ok": true,
  "stub_ok": true,
  "turn_http": 200,
  "showed_expected_fragment": true,
  "observation_n": 1,
  "antipattern_ids": []
}

Local: docker compose -f docker/stub-pages/docker-compose.yml up --build
Actions: workflow stub-pages (workflow_dispatch)
MCP: http://mcp-mock:8000/mcp · CPU model: http://cpu-llm:8080/v1

Wikipedia extract (browser → MediaWiki API)

This #wiki panel calls Wikipedia's Action API from the browser (origin=*). It is not an MCP tool call. github.io /wiki stays 404; that path is the local stdlib stub. MCP tools / local form:

python -m mcp_toolcall_lab.stub_front serve --port 8765
# then open /wiki  (http://127.0.0.1:8765/wiki)

CI screenshots of the local stub still use fixtures/wikipedia/yokohama_extract.json. A heading switch on this panel reuses the in-memory extract (no second fetch).

Browser → MediaWiki API (not MCP). For the MCP Wikipedia tools / local stub form, run compose or stub_front serve.

MediaWiki plaintext extract, HTML-escaped. This page does not scrape Wikipedia HTML or run it through markdown.py.

Back to report

Pixiv Encyclopedia search

Pixiv Encyclopedia fetching stays MCP/local-server owned. GitHub Pages provides the same search/result workspace without pretending that a static host can run MCP.

Run the local stub to execute the MCP-backed search.

fetch_pixiv_dictionary_section — Fetch one public pixiv Encyclopedia article HTML page, normalize it through vendored markdown.py, and return its hierarchical heading list or one selected section. The cached article is reused when only the heading changes.
fetch_pixiv_dictionary_article — Fetch one public pixiv Encyclopedia article HTML page and normalize it through vendored markdown.py. Returns canonical_title, Markdown, hierarchical headings and source_url.

python -m mcp_toolcall_lab.stub_front serve --port 8765

Back to report

Tool catalog

find_municipalities — Find mock municipalities by name or prefecture.
find_transaction_prices — Return mock property transaction prices. This never calls a live API.
find_stations — Find mock stations in a municipality.
fetch_wikipedia_section — Fetch a real Wikipedia article's sections over HTTP (the one family here that is not a mock). Omit `heading` to list section titles (a pulldown); pass one to get that section's body. Unmatched heading is an empty result, not an error. Same article is cached in-process so heading switches do not refetch.
fetch_wikipedia_article — Fetch a real Wikipedia article as a MediaWiki plaintext extract (not HTML). Returns canonical_title, extract, and headings from that same fetch. In-process TTL/LRU cache reuses the extract when only the heading changes.
fetch_pixiv_dictionary_section — Fetch one public pixiv Encyclopedia article HTML page, normalize it through vendored markdown.py, and return its hierarchical heading list or one selected section. The cached article is reused when only the heading changes.
fetch_pixiv_dictionary_article — Fetch one public pixiv Encyclopedia article HTML page and normalize it through vendored markdown.py. Returns canonical_title, Markdown, hierarchical headings and source_url.

Workspace provenance

web-ui contract pinned to e7d16a2ce0cee76b7744a4b6a8f8ce374491a4db.