Skip to content

mcp.md

Latest commit

 

History

History
153 lines (103 loc) · 5.69 KB

mcp.md

File metadata and controls

153 lines (103 loc) · 5.69 KB

MCP capabilities

The server exposes 1 tool, 2 resources, and 3 prompts over JSON-RPC 2.0 (stdio transport), plus MCP Sampling for server-initiated LLM disambiguation.

Tool: zhtw

Unified lint / fix / gate for zh-TW text.

Parameter Type Description
text string (required) Text to check
fix_mode "none" / "orthographic" / "lexical_safe" / "lexical_contextual" Fix mode (default: "none")
max_errors integer Reject if residual errors exceed threshold
max_warnings integer Reject if residual warnings exceed threshold
profile "base" / "strict" Rule profile
relaxed boolean Relax colon and other UI-string-level rules
content_type "plain" / "markdown" / "markdown-scan-code" / "yaml" Content type (markdown-scan-code also lints inside code blocks)
political_stance "roc_centric" / "neutral" / "international" Political stance filter
ignore_terms array of strings Terms to downgrade to Info for this call
explain boolean Attach cultural/linguistic annotations
output "full" / "compact" / "tabular" / "summary" Output verbosity
include_telemetry boolean Include estimated token, cache, and Tier 2 resolution metrics in JSON responses (full, compact, summary)

Lint only (default):

{"text": "這個軟件使用了遞歸算法來遍歷鏈表"}

Returns issues with line/column position, matched term, suggestions, rule type, severity, and English anchor. Structured JSON responses also include document-level scan metadata when available:

  • coverage: active spelling rules checked and distinct rules matched
  • oral_density: spoken-style filler ratio across CJK text
  • quality_flags: coarse document signals such as spaced_acronyms, stutter_detected, asr_artifacts, high_oral_density

The above flags: 軟件 (software), 遞歸 (recursion), 算法 (algorithm), 遍歷 (traverse), 鏈表 (linked list).

Lint + fix + gate:

{"text": "請使用內存中的緩存數據", "max_errors": 0, "fix_mode": "lexical_safe"}

If residual errors exceed max_errors (or warnings exceed max_warnings), the response has "accepted": false. Otherwise "accepted": true with corrected text.

Per-call suppression:

{"text": "這個軟件很好用", "ignore_terms": ["軟件"]}

Matching issues are downgraded to Info severity for this call only.

Telemetry-enabled call:

{"text": "這個軟件很好用", "include_telemetry": true}

When enabled, the response includes a telemetry object with estimated prompt/completion tokens, cache hit/miss counts, Tier 2 local resolutions, and raw counters for the call. tabular output does not support telemetry because it is plain text rather than structured JSON.

Summary output:

{"text": "這個那個這個那個這個那個這個那個這個那個", "output": "summary"}

Returns aggregate counts only, plus any available document-level metadata such as coverage, oral_density, quality_flags, and ai_signature.

Resources

URI Description
zh-tw://style-guide/moe MoE punctuation, variant, and vocabulary standards (Markdown)
zh-tw://dictionary/ambiguous Terms requiring LLM disambiguation (JSON array)

Prompts

Name Arguments Description
normalize_tone (none) Grounds the host LLM in MoE-standard zh-TW conventions
lint_natural instruction, text Translates free-form instruction into a zhtw tool call
editorial_review text, max_iterations (default 3) Iterative review: calls zhtw, explains issues, applies fixes until accepted

Sampling

When the scanner encounters an ambiguous term (with english field indicating multiple translations) and the client supports sampling, the server sends a sampling/createMessage request for LLM disambiguation. Budget: 5 calls per invocation, 5-second timeout. On timeout, the issue is kept at original severity.

Prompt examples

Once installed, type these directly into your AI assistant's chat (Claude Code, OpenCode, etc.). The assistant will call the zhtw tool automatically.

Linting and reviewing

Check README-zh.md for Taiwan MoE zh-TW standard violations.

Review docs/api.md for zh-CN terminology and explain each issue.

Run a strict MoE lint on this markdown and list every violation with line numbers.

Auto-fixing

Auto-correct zh-CN vocabulary in src/locales/zh-TW.json and show the diff.

Fix all non-standard terms in CHANGELOG.md using safe mode.
Reject the result if any errors remain.

Output gate (strict enforcement)

Lint this article with max_errors=0 and abort if any violations are found:
[paste text]

Act as a zh-TW copy editor. For every response you write in Chinese, run zhtw
with fix_mode "lexical_safe" and max_errors 0 before sending it to me.

Git / CI workflows

Check all staged markdown files for MoE compliance before I commit.

Review every file changed in the last commit for zh-TW regressions.

Translate this English error message to Traditional Chinese, then verify with
zhtw before giving it to me.

MCP prompts and resources

Use the normalize_tone prompt so all Chinese text you produce follows MoE standards.

Load zh-tw://style-guide/moe and follow those conventions for this session.

Use the editorial_review prompt on this draft with max_iterations=2, and stop
early if zhtw returns accepted=true:
[paste text]

Profile and suppression

Check this UI copy with the relaxed flag:
[paste text]

Lint this document but ignore "軟件" for this run, explain all other issues:
[paste text]