菜单
Use when Miller search misses expected results, ranks surprising hits, returns noisy output, or needs mode/content/region troubleshooting.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
Use before changing existing indexed files with Miller edit, especially symbol rewrites, text replacements, renames, or refactors.
Use when an agent needs extractor-recognized code-shape facts such as routes, htmx attributes, Alpine directives, SQL DDL/DML, or data-document structure.
Use when starting a Miller task, choosing a Miller tool, or choosing a Miller search mode.
Use when tracing Miller bridge paths in supported providers: dotnet-web URL literals and fetch/axios client requests to ASP.NET signals, client requests to Next.js route handlers or Nuxt server routes, Next.js/Nuxt route references to file routes, Vue/React references to route definitions, or backend-http client requests to Express/FastAPI/Flask/Django/Spring/Go/gin/echo/Rails/NestJS/Laravel/Phoenix/axum/actix/Symfony/Ktor route facts.
Use when orienting on an unfamiliar code area with Miller, explaining a module, finding entry points, or gathering context before a change.
Use when assessing blast radius with Miller, choosing tests for a change, answering who uses a symbol, or planning a refactor.
| name | miller-search-debug |
| description | Use when Miller search misses expected results, ranks surprising hits, returns noisy output, or needs mode/content/region troubleshooting. |
| user-invocable | true |
| arguments | <query and expected result> |
| allowed-tools | mcp__miller__search, mcp__miller__inspect, mcp__miller__context, mcp__miller__trace, mcp__miller__workspace |
Diagnose search behavior by checking the query mode, index freshness, projection limits, and whether the expected result exists.
search(query="<original>", limit=20)
search(query="<symbol-ish>", mode="symbol")
search(query="<path-ish>", mode="file")
search(query="<docs or prose>", mode="content")
search(query="<source body text>", mode="source")
search(query="<imported log text>", mode="external")
search(query="<imported web text>", mode="web")
search(query="<broad text>", mode="all-text")
search(query="<comment or literal>", regions="comment|string_literal|doc_comment")
search(query="<known area>", file_pattern="src/ui/**", language="typescript")
exclude_tests=false when tests are expected.name + signature; docs/prose belong in mode=content.mode=source; imported logs/reports and web markdown belong in mode=external or mode=web.content search "<term>" --workspace-id all --kind source|docs|config|external_file|web.mode=file when auto mode looks noisy.file_pattern and language before raising limit.workspace(operation="refresh").search(query="<exact symbol or file>", mode="auto", exclude_tests=false)
inspect(target="<expected-symbol-or-file>")
trace(target="<nearby-symbol>")
context(query="<workflow around expected result>")
Use this shape:
Search Debug: "<query>"
Expected: <symbol/file/text>
Observed:
- <top results or failure>
Diagnosis:
- <wrong mode, stale index, hidden tests, projection limit, region disabled, or true miss>
Next:
- <specific query, refresh, or implementation follow-up>
Do not widen Miller's symbol ranking as a first response. First prove that mode=content, regions=..., file mode, or refresh does not solve the workflow.