Godfetch

godfetch

Unified external research — look up library documentation, search source code in any git repository, and check package versions from a single skill.

Routing

| Intent | Primary tool | Fallback | | ---------------------------------------------------------------------------- | ----------------------------------------------------------------- | -------------------------------------------------------- | | Library docs, API reference | llms-probe → WebFetch llms.txt | context7 if no llms.txt published | | Changelogs, breaking changes | llms-probe → WebFetch llms.txt | gh api contents for CHANGELOG.md | | Cross-repo code search (exact identifier) | Sourcegraph MCP keyword_search | gh search code, then git-clone for follow-up | | Semantic / concept query in a GitHub-hosted dep ("how does X work in lib Y") | mcp__plugin_ora_morph__github_codebase_search | git-clone if the question keeps branching across files | | Concept query across multi-host or GitLab/Bitbucket repos | Sourcegraph MCP nls_search with 2-5 extracted keywords | keyword_search after picking a literal term | | Deep dive in known repo (3+ files) | git-clone + shell tools | Sourcegraph read_file for one-off reads | | GitHub issues | gh issue view | gh search issues for discovery | | GitHub PRs | gh pr view | gh search prs for discovery | | GitHub releases (versions, dates, notes) | gh release view --repo owner/repo | gh release list --repo owner/repo for browsing | | Single file (known repo + path) | Sourcegraph MCP read_file | gh api repos/.../contents/ (GitHub only) | | Symbol navigation (def, references) | Sourcegraph go_to_definition / find_references | git-clone + ast-grep | | Git history / diff search across repos | Sourcegraph commit_search / diff_search / compare_revisions | gh api /repos/.../commits | | Package version, deprecation | deps-dev | npm view for npm-only metadata | | npm package info (non-version) | npm view (Bash) | WebSearch for community sentiment | | General web lookup | WebSearch → WebFetch | — | | Comparison / decision | llms-probe per lib + WebSearch | context7 for additional snippets |

For mixed requests, launch all relevant tools in parallel. Probe and clone are I/O-bound — start them in the background and run WebFetch/context7/WebSearch concurrently to mask latency.

GitHub access rules

Do not use WebFetch on github.com or raw.githubusercontent.com URLs — use the right tool:

| GitHub content | Use | Never | | --------------------------------- | ----------------------------------------------------------------------- | ------------------------------------------------------------------------------------------------------------- | | Source code (exploration) | git-clone + shell tools | browsing files via gh api contents | | Source file (known path) | gh api repos/.../contents/ | WebFetch raw.githubusercontent.com | | Issues | gh issue view --repo owner/repo | WebFetch github.com/.../issues/N | | Pull requests | gh pr view --repo owner/repo | WebFetch github.com/.../pull/N | | Issue/PR search | gh search issues "q" --repo ... | WebFetch github.com/issues?q=... | | Releases (versions, dates, notes) | gh release view/list --repo owner/repo or gh api repos/.../releases | WebFetch github.com/.../releases — relative timestamps on the HTML get hallucinated into training-era years | | CHANGELOG.md | gh api repos/.../contents/CHANGELOG.md | WebFetch blob/ or raw URLs |

Search discipline

• deps-dev for versions: when checking latest version, deprecation, or comparing installed vs latest — always use deps-dev first. Only fall back to npm view or WebSearch if deps-dev errors or the package is private.
• llms.txt first, context7 fallback: for library docs, run scripts/llms-probe.sh against the docs domain before reaching for context7. Author-published llms.txt has no community-curation lag and no enrichment layer that can hallucinate. Fall back to context7 only when probe returns nothing.

llms.txt — Author-Canonical Library Documentation

Many doc sites publish llms.txt (Markdown index of doc pages) and llms-full.txt (concatenated full content). These are author-published — no enrichment layer, no community-curation lag — so they reflect the deployed docs version exactly. Prefer them over context7 when available.

Step 1: Probe for availability

bash scripts/llms-probe.sh 

Outputs TSV kind \t url \t size for any found files. Probes root + common nested paths (/docs/, /en/), follows redirects, dedupes. Returns non-zero exit if nothing found.

| kind | Meaning | | ------ | ------------------------------------------------- | | index | llms.txt — Markdown list of doc page URLs | | full | llms-full.txt — entire docs corpus concatenated |

Size shows ? when the CDN strips both Content-Length and Content-Range headers (Vercel does this on react.dev) — treat ? as unknown and prefer the index path.

Step 2: Fetch based on what's there

| Found | Action | | ------------------------------------- | --------------------------------------------------------------------- | | llms-full.txt ≤ ~500 KB | WebFetch it directly — single round trip, full corpus | | llms-full.txt > ~500 KB or size ? | WebFetch llms.txt first, pick relevant section links, fetch those | | Only index (no full) | WebFetch the index, then fetch individual page links | | Probe failed | Fall back to context7 (next section) |

For multi-section pulls, dispatch the page WebFetch calls in parallel.

Known publishers

Confirmed live (April 2026): React (react.dev), Next.js (nextjs.org, content under /docs/), Vercel, Anthropic (docs.anthropic.com → platform.claude.com), Cloudflare (docs.cloudflare.com → developers.cloudflare.com), Supabase, Drizzle (orm.drizzle.team), Hono (hono.dev), Zod (zod.dev), Expo (docs.expo.dev), tRPC (trpc.io), shadcn/ui (ui.shadcn.com). Most Mintlify- and GitBook-hosted docs auto-publish.

Tailwind, most pre-1.0 libraries, and many community packages do not publish — those go straight to context7.

Rules

• Probe before assuming. Adoption is uneven and paths vary (root vs /docs/ vs redirects). Always run llms-probe.sh and act on the TSV — never hardcode URLs.
• Watch file size before fetching full. Cloudflare's llms-full.txt is ~46 MB and Supabase reports ? (chunked). A blind fetch of either blows the context window. The 500 KB threshold is a heuristic — adjust to remaining context budget.
• Index → page chain for big corpora. Treat llms.txt as a routing table: parse section headings, fetch only the page URLs that match the question.

context7 — Library Documentation (Fallback)

Reach for context7 when llms-probe.sh returns nothing — the library doesn't publish llms.txt, or its docs domain isn't reachable. Coverage spans ~33K libraries via community-curated indexes; tradeoff is an enrichment layer that can introduce inaccuracies the author-published llms.txt avoids.

Two-step workflow via the official ctx7 CLI. Requires bunx ctx7@latest login once (no API key env var).

Step 1: Resolve library ID

bunx ctx7@latest library  [query]

Lists library candidates with their Context7 IDs (e.g. /websites/react_dev), trust scores, and snippet counts. The optional [query] re-ranks results by relevance — pass it whenever you already know the topic. Add --json for machine-readable output.

Step 2: Fetch documentation

bunx ctx7@latest docs  ""

Returns markdown snippets ranked by relevance. Add --json for structured output. If the first answer is shallow or off-topic, retry with --research — it spins up sandboxed agents that read the source repo and run a live web search, at higher cost.

Rules:

• One-time setup: bunx ctx7@latest login (interactive). Verify with bunx ctx7@latest whoami.
• Always resolve the library ID first — IDs are not guessable.
• Write specific queries — "useState hook with objects" beats "hooks". The query drives relevance ranking on both commands.
• Use --research only as a retry when the default answer was insufficient, not by default — it's slower and more expensive.

Reference: references/context7.md

Sourcegraph — Cross-Repo Search & Code Navigation (MCP)

Sourcegraph public instance exposes an HTTP MCP server at https://sourcegraph.com/.api/mcp providing 13+ tools for cross-repo search, code navigation, and git history. Configured in ora plugin's .mcp.json — tools auto-available to Clio.

Indexes 2M+ OSS repos across GitHub + GitLab + Bitbucket. Sub-second cross-repo queries, no local clone overhead.

Tools by intent:

| Intent | Tool | | --------------------------------- | ------------------------------------------------------------------------------------------------ | | Search by exact identifier (AND) | keyword_search — literal match, ALL terms required. Filters: repo:, file:, lang:, rev: | | Search by concept / vocab unknown | nls_search — OR logic + word stemming (handler matches handle). NOT semantic embedding | | Read single file | read_file — 128KB cap; repo, path, optional revision / startLine / endLine | | List directory | list_files | | Find repos | list_repos | | Symbol navigation | go_to_definition, find_references | | Git history / diffs | commit_search, diff_search, compare_revisions | | Contributor lookup | get_contributor_repos | | AI synthesis over codebase | deepsearch — heavy; spawns subagents |

Rules:

• Pick keyword_search vs nls_search by what you know. Concrete identifier (function/class/constant name, error string, literal phrase) → keyword_search. Vocab unknown / conceptual ("how does X work", cross-library pattern) → nls_search. Why: keyword_search is AND + literal — missing one term zeroes the result; nls_search is OR + stemming — broader recall but noisier ranking.
• Extract keywords before nls_search. Strip question words (how, what, where, does, is) and articles. Pass 2-5 meaningful nouns/verbs, not full sentences. Example: "how does the router match incoming requests" → router match request handler. Why: tool stems and OR-matches each term — natural-language fillers dilute the signal.
• Scope with repo:^...$ anchors to avoid matching forks. repo:foo/bar matches foo/bar-fork too; repo:^github\.com/foo/bar$ does not.
• Pattern: search → verify path → read_file. When a search result is annotated repo@revision (e.g. github.com/foo/bar@abc123), split it and pass repo + revision separately to read_file. Omitting revision silently reads HEAD of the default branch and may return different content.
• Parallelize independent searches. Three repos to compare → fire three keyword_search/nls_search calls in one message, not sequentially.
• Discovery vs forensics — keyword_search / nls_search for "find repos that…"; git-clone for tracing flow through 5+ files in one repo.
• Index lag — Code published within the last ~24 hours may not be indexed. Fall back to git-clone --refresh for just-released versions.
• Repo not indexed = fallback — If read_file returns null repository, drop to git-clone or gh api contents.
• deepsearch is heavy — Spawns AI subagents. Use only when keyword_search + read_file can't synthesize the answer and the target spans multiple repos or non-GitHub hosts. For GitHub-only synthesis use mcp__plugin_ora_morph__github_codebase_search instead — lighter and shaped for the same job.
• Plan tier caveat — Sourcegraph docs note MCP access is part of Enterprise plans; public sourcegraph.com MCP endpoint's free-tier behavior is unverified. If first tool call returns 401/403, drop to git-clone.

Reference: Sourcegraph MCP docs

Morph github_codebase_search — default for GitHub-hosted semantic queries

mcp__plugin_ora_morph__github_codebase_search runs Morph's WarpGrep subagent against a target public GitHub repo and returns a synthesized answer with file:line citations. Reach for it first when the question is shaped like "how does library X do Y", "where does package Z handle case W", "what's the data flow

…

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: trancong12102
• Source: trancong12102/agentskills
• License: MIT

Install and usage instructions live in the source repository linked above.