repo-research

Defaults by intent:

• Concept question in one GitHub repo ("how does lib X do Y") → mcp__plugin_ora_morph__github_codebase_search. Citations are real; the synthesis can misread the code — verify by reading cited files before relaying load-bearing conclusions.
• Exact-identifier search across repos, GitLab/Bitbucket targets, symbol navigation, git history → Sourcegraph MCP (keyword_search, nls_search, read_file, go_to_definition, find_references, commit_search, diff_search).
• Deep dive (3+ files, branching follow-ups, ast-grep) → bash scripts/git-clone.sh <repo> — shallow-clones into a ~/.cache/clio-repos/ cache and echoes the path. Accepts owner/repo, HTTPS, or SSH; --refresh for the latest commit.
• Issues / PRs / releases / single known file → gh CLI: gh issue view, gh pr view, gh release view|list, gh api repos/.../contents/<path>.

Don't WebFetch github.com or raw.githubusercontent.com — gh returns the same content structured. The releases page is the sharpest trap: its relative timestamps ("2 weeks ago") get hallucinated into training-era dates.

Sourcegraph specifics:

• keyword_search is AND + literal — one wrong term zeroes results; use for concrete identifiers. nls_search is OR + stemming, not semantic — pass 2–5 extracted keywords, not sentences.
• Anchor repo filters: repo:^github\.com/foo/bar$ — unanchored repo:foo/bar matches forks too.
• Search results annotated repo@revision: pass repo and revision separately to read_file, or you silently read default-branch HEAD.
• Code newer than ~24 h isn't indexed; unindexed repo or 401/403 → fall back to git-clone.