Code Search And Navigation With The Atlassian MCP

Adapted from the TWG CLI code-search.md reference, for agents that reach code search through the Atlassian MCP server instead of twg search-code.

When To Use

Use this when an answer needs source unavailable locally, including follow-up questions without an explicit search request. Start from the known service, behavior, error, API, or symbol; discover the repository if unknown. Verify implementation claims against source. Skip when local code, supplied context, or general knowledge answers the question.

The Atlassian MCP covers code as well as Jira and Confluence. In the default v2 tool list, code search is reached through discover and execute. This skill teaches that workflow; it does not register additional MCP tools.

Operations

• searchCode: query plus optional mode (hybrid default, semantic, lexical), namespace (GitHub organization or Bitbucket workspace), repo (slug only, not owner/repo), language, extension, path, maxResults. Alternates are separated with |. Hits carry workspace, repo, path, version (a commit) and the matched spans.
• getCodeFile: workspace, repo, path, commitHash. Pass the hit's workspace, repo, path and version; do not infer the workspace from the search filter.
• listCodeDirectory: workspace, repo, commitHash, optional path, maxDepth, page.
• scanCodeRepo: workspace, repo, commitHash (a branch name works), optional filter (glob), root (directory prefix), limit, cursor. Paths only.

Use exact parameter names from the live schema. Treat the descriptions above as a navigation guide when the live contract differs. Use responseFields only when the advertised schema supports it.

Search And Navigate

Choose the route from anchor precision:

• Known repository, exact path and commit: read it with getCodeFile.
• Known repository and exact path without a commit: search within that repository for the file or a known symbol. Use a matching hit's version as commitHash. listCodeDirectory also requires a commit, so it cannot supply the initial one.
• Known repository and symbol or class: run one repository-scoped lexical or hybrid searchCode, then read only the selected files.
• Unknown repository or conceptual behavior: begin with one unscoped hybrid searchCode using the concrete topic, symbol, or behavior.
• Cross-repository or reverse-dependency question: search without a namespace or repo filter so every indexed repository participates, unless the user limits the scope.

Honor an explicitly named repository or code host as the requested scope. Widen beyond it only when the user asks for cross-repository coverage, the repository identity is ambiguous, or scoped evidence is demonstrably incomplete.

When a scoped search returns no matches

An empty result does not establish whether a repository is indexed or the code is absent. Check the repository slug, filters and query before widening the search. When the requested scope permits it, remove filters incrementally and stop when the next search cannot resolve a specific gap. Look for candidate mirrors or forks among the results. Verify their relationship to the requested repository before using them as evidence. Identify the source repository and commit, and note that its contents may differ from the original. Report an indexing gap only when confirmed; otherwise report that the search found no matches.

Reverse Dependencies

Search the exact API or symbol first. If that does not establish direct consumption, search the package or import anchor, group candidates by repository, and inspect the manifest or import plus the call site needed to verify selected consumers. Prefer stable literal anchors; do not invent delimiter-heavy code fragments merely to vary the query. Keep each search to a compact result set and use the returned repository and path metadata for selection; a result cap is not a reason to claim completeness.

Efficiency

Treat search results as a batch. Do not rerun the same symbol separately for each repository or usage pattern. Select a small, diverse set of direct consumers, read the call-site files needed to establish each integration role, and fetch a manifest only when the call-site evidence does not establish the package. Stop when another result would repeat an already verified integration role, not while a material role is unverified.

After discovery establishes repository, path and commit, read independent files concurrently in one tool round. Do not wait for one file before requesting another unless its result determines the next path.

Reporting

Rank a bounded set of source-backed implementation locations. If the anchored search is incomplete, returns only generated documentation, or exposes only one part of the requested capability, widen while the next result can close that gap. Fetch only the source files needed for symbol and behavior context. Cite workspace/repo and path, and the commit when you read a file. Deduplicate mirror results only when the retrieved code establishes equivalence. Say which source appears canonical only when the evidence supports it. Never infer behavior from a filename, or completeness from one repository's empty result. Use the fewest calls that establish the answer; for a single known repository and symbol, one search and one read round are the normal starting point.