--- name: web-fetch description: Fetch and extract one web source within an explicit allowlist, returning the content by digest with full provenance. runx: category: research --- # Web Fetch Fetch one URL, prove it was allowed, extract the part the caller asked for, and return that slice by digest with the provenance needed to trust it later. ## What this skill does `web-fetch` resolves a single URL against a host allowlist, retrieves it, extracts text, metadata, or links, and seals the result so a downstream step can cite the fetch without re-fetching. It checks the final host against the allowlist before and after redirects, retrieves up to `max_bytes`, and returns the final URL, the HTTP status, a `content_digest` over the retrieved body, the extracted slice, and a provenance block recording when it ran, every redirect hop, and how many bytes it read. The body is referenced by digest; only the extracted slice is inlined. This is the primitive an agent reaches for when it has already decided which page to read. The decision it makes easier is "can I read this page, and what did it actually say", with the answer backed by a digest instead of a remembered paraphrase. It differs from the research family: `research` and `deep-research` decide *which* sources matter and synthesize across them, while `web-fetch` retrieves exactly one source and refuses anything off the allowlist. ## When to use this skill - An agent has chosen a specific page and needs its content bound to a `content_digest` so a later step can cite it without re-fetching. - A research pass needs each source retrieved through a single bounded `net:allowlist` fetch with a complete redirect chain and byte count. - A review must later prove what a page said at fetch time. - A follow-on skill (prior-art, vuln-scan, brief) needs one source extracted as `text`, `metadata`, or `links`. ## When not to use this skill - To judge, rank, or synthesize across sources. That is the research family's job; `web-fetch` retrieves exactly one source and refuses to reason over many. - To reach a host the caller did not declare in `allowlist`, including a host reached only through a redirect. - To write anything. The only scope is `net:allowlist`; there is no repo, file, wallet, or send authority here. - To inline a large raw body. The extracted slice is the payload; the full body lives behind `content_digest`. - To carry secrets. Request headers may reference a credential by `${secret}` handle, but no header value, cookie, token, or auth string appears in the output or the receipt. ## Procedure 1. Require `url` and `allowlist`. Either missing returns `needs_agent`; the fetch cannot run without a target and a declared scope. 2. Match the URL host against `allowlist`. On a miss, return `policy_denied` before any network call, recording the attempted host and the allowlist it was checked against. 3. Fetch, following redirects, re-checking each redirect target's host against the same `allowlist`. A redirect that lands off-allowlist halts the fetch, returns `policy_denied` with the hop that failed, and discards partial bodies. Cap the read at `max_bytes` when set. 4. Compute `content_digest` over the retrieved body. 5. Extract per `extract`: `text` (readable body text, default), `metadata` (title, description, canonical, declared language, content type), or `links` (absolute hrefs found in the document). 6. Return `fetch_result` with the final URL, status, digest, extracted slice, and provenance. Flag truncated reads in provenance; never return a clipped read as if whole. ## Edge cases and stop conditions - **Missing `url` or `allowlist`:** return `needs_agent`; the fetch has no target or no scope to check against. - **Host off the allowlist:** stop with `policy_denied` before any network call; record the attempted host, not a response body (there is none). - **Redirect off the allowlist:** halt the fetch, return `policy_denied` naming the hop that failed, and discard the partial body. - **Read clipped by `max_bytes`:** flag `truncated: true` in provenance; the digest is over the bytes actually retrieved. - **Large raw body:** never inline beyond the extracted slice; anything bigger than the requested view is reachable only through `content_digest`. - **Credential in a header:** reference it by `${secret}` handle only; no header value, cookie, or token reaches the output or the receipt. ## Output schema ```yaml fetch_result: decision: ready | needs_agent | policy_denied final_url: string # URL after redirects, the one the digest is over status: number # HTTP status of the final response content_digest: string # digest of the retrieved body, algorithm prefix included extract_mode: text | metadata | links extracted: string | object | array # string for text, object for metadata, array of hrefs for links provenance: fetched_at: string # timestamp of the fetch redirects: array # ordered host hops, each re-checked against the allowlist bytes: number # bytes read truncated: boolean # true when max_bytes clipped the read policy: allowlist_decision: allowed | denied attempted_host: string # set on policy_denied allowlist_checked: array # the hosts the request was checked against ``` The sealed `runx.receipt.v1` carries the final URL, status, `content_digest`, byte count, the redirect chain, and the allowlist decision. It carries no header values, no cookies, and no raw body beyond the digest. ## Worked example Input: `url` of the HTTP Semantics RFC, an `allowlist` of `www.rfc-editor.org` and `rfc-editor.org`, `extract: text`, and `max_bytes: 200000`. Output: `decision: ready`; the host matched the allowlist before the request left; no redirects; status `200`; `content_digest` is taken over the retrieved body; `extracted` holds the readable text slice; provenance records `fetched_at`, an empty redirect chain, `184302` bytes, and `truncated: false`. The receipt seals with the final URL, status, digest, byte count, and the allowlist decision; no header value reaches it. ## Inputs - `url` (required): the single URL to fetch; its host must match the allowlist. - `allowlist` (required): permitted hosts or host patterns; the URL and every redirect target must match an entry. - `extract` (optional): `text`, `metadata`, or `links`. Defaults to `text`. - `max_bytes` (optional): cap on bytes read; a clipped read is flagged `truncated` in provenance.