redraft-fetch-docs

docs/ is a local, gitignored cache (organizing protocol §2) — it never travels with a clone or redraft sync. This skill restores it from the source_url recorded on each artifact node, entirely with grep/curl/mkdir/ls — no MCP server required.

Follow these steps in order. Do not skip ahead or combine steps.

1. Find the graph root. Check whether graph/nodes/ exists at the project root / current working directory — the same detection redraft-init step 1 uses. If it doesn't, this isn't a graph repo; stop and say so.

2. Scan for fetchable artifacts. Run:

   grep -nE "doc_path:|source_url:" graph/nodes/*.md
   

   If that reports "No such file or directory," graph/nodes/ has no node files yet — there's nothing to fetch; report 0 fetched / 0 already-present / 0 failed and stop.

   Otherwise each output line is <file>:<line>:<content>. Group lines by <file>: a node only qualifies when both doc_path: and source_url: appear for it (that pairing is the artifact-node convention from organizing protocol §2 — a file with only one of the two is a bare citation or a node mid-edit, not this skill's job). Read the value after the colon on each qualifying line — strip a wrapping " or ' if the YAML dumper added one — to get that node's doc_path and source_url.

3. Resolve each qualifying pair, one at a time. Quote paths and URLs in every command below (doc_path can contain spaces).

   • ls "<graph-root>/<doc_path>" — if it lists the file, skip: never re-download, never overwrite an existing local copy, even one you suspect is stale.

   • If ls reports the file missing: mkdir -p the parent directory of <doc_path> (compute it yourself from the path string, e.g. the parent of docs/papers/qin-2022.pdf is <graph-root>/docs/papers — don't shell out to dirname, it isn't one of this skill's allowed commands), then fetch:

     curl -L --fail --remove-on-error -o "<graph-root>/<doc_path>" "<source_url>"
     

     --remove-on-error matters: without it, a failed fetch (404, paywall, network error) can leave a zero-byte or truncated file sitting at doc_path, and a later run's ls check would then wrongly treat that corrupt stub as "already present" and skip it forever.

4. Report a compact summary, not a blow-by-blow: fetched N, already-present M, failed K. For each failure, name the node title, the source_url, and the error (curl's exit status or stderr text) — a paywalled, moved, or unreachable source is expected for some entries; report it and move on to the rest of the list, never stop on the first failure.

What NOT to do

• Never overwrite a file that already exists at its doc_path (step 3).
• Never fetch a URL that isn't recorded as a source_url on some artifact node — this skill restores what's recorded, it doesn't go discover new material.
• Never touch graph/nodes/, and never commit or run snapshot — this skill only populates the gitignored docs/ cache; it has no graph-write or git side effects.
• A failed fetch is reportable, not fatal — keep going through the rest of the list rather than aborting on the first bad URL.