diff --git a/docs/changes/unreleased/1771-web-fetch-cached-snapshot.md b/docs/changes/unreleased/1771-web-fetch-cached-snapshot.md
new file mode 100644
index 0000000000..e0eae318f1
--- /dev/null
+++ b/docs/changes/unreleased/1771-web-fetch-cached-snapshot.md
@@ -0,0 +1,14 @@
+---
+kind: fixed
+title: web fetch asks the reader for a fresh copy and rejects its cached snapshot
+pr: 1771
+surface: [chat, engine]
+invalidates:
+  - "`codeaf web fetch https://example.com` returned r.jina.ai's stale cached snapshot of the page — Jina's own 'cached snapshot' page — because the reader serves its 3600s page cache by default. It now sends the reader's no-cache header and, when the reader still marks its answer a cache hit, fetches the target directly, so what comes back is the page at that address now."
+---
+
+The Jina reader r.jina.ai serves a 3600s page cache by default, and codeaf's
+fetch passed its answer through verbatim — so a model asked to read a page
+could be handed someone else's cache of it without knowing. The fetch now
+asks for a fresh copy on every request, and discards an answer the reader
+still marks as cached (X-Cache: HIT), fetching the target itself instead.
diff --git a/internal/search/jina.go b/internal/search/jina.go
index d6de6c5660..b777e70138 100644
--- a/internal/search/jina.go
+++ b/internal/search/jina.go
@@ -6,6 +6,7 @@ import (
 	"io"
 	"net/http"
 	"net/url"
+	"regexp"
 	"strings"
 )
 
@@ -162,6 +163,11 @@ func (j jina) Fetch(ctx context.Context, target string) (string, error) {
 		req.Header.Set("Authorization", "Bearer "+key)
 	}
 	req.Header.Set("Accept", "text/plain")
+	// The reader keeps a 3600s page cache and answers from it by default;
+	// X-No-Cache asks it not to. The ask is not a guarantee — a HIT can still
+	// come back (see [cacheHit] below), but without the header every answer
+	// would be the cache's.
+	req.Header.Set("X-No-Cache", "true")
 
 	resp, err := j.opts.client().Do(req)
 	if err != nil {
@@ -177,6 +183,18 @@ func (j jina) Fetch(ctx context.Context, target string) (string, error) {
 		return "", fmt.Errorf("jina: %s%s", resp.Status, detail(body))
 	}
 
+	// A HIT is the reader serving its OWN 3600s page cache, which is someone
+	// else's snapshot of the page, not the page. It is deliberately not
+	// passed through: an agent that was handed it would believe it had read
+	// the page. Discard it and fetch the target itself.
+	if cacheHit(resp.Header) {
+		text, err := jinaDirectFetch(ctx, j.opts, target)
+		if err != nil {
+			return "", fmt.Errorf("jina: cached snapshot rejected; direct fetch of %s: %w", target, err)
+		}
+		return text, nil
+	}
+
 	// Returned whole. Capping is the RENDERER's job ([RenderFetch]) so that a
 	// caller doing something other than prompting a model — a summariser, a
 	// grep — is not handed a page someone already cut for a different budget.
@@ -186,3 +204,75 @@ func (j jina) Fetch(ctx context.Context, target string) (string, error) {
 	}
 	return text, nil
 }
+
+// cacheHit reports whether the reader answered from its own page cache: the
+// X-Cache response header is HIT. Trimmed and case-insensitive, because a
+// header's casing and whitespace are the answerer's, not ours.
+func cacheHit(header http.Header) bool {
+	return strings.EqualFold(strings.TrimSpace(header.Get("X-Cache")), "HIT")
+}
+
+// jinaDirectFetch is the freshness fallback: GET the target itself when the
+// reader answered from its cache, and reduce its HTML to text so the result
+// has the same readable shape a fresh reader answer would have had.
+func jinaDirectFetch(ctx context.Context, opts Options, target string) (string, error) {
+	req, err := http.NewRequestWithContext(ctx, http.MethodGet, target, nil)
+	if err != nil {
+		return "", err
+	}
+	req.Header.Set("User-Agent", "Mozilla/5.0 (compatible; codeaf/1.0)")
+
+	resp, err := opts.client().Do(req)
+	if err != nil {
+		return "", err
+	}
+	defer resp.Body.Close()
+
+	body, err := io.ReadAll(io.LimitReader(resp.Body, 8<<20))
+	if err != nil {
+		return "", fmt.Errorf("read response: %w", err)
+	}
+	if resp.StatusCode >= http.StatusBadRequest {
+		return "", fmt.Errorf("http %d%s", resp.StatusCode, detail(body))
+	}
+
+	text := htmlToText(string(body))
+	if text == "" {
+		return "", fmt.Errorf("empty body")
+	}
+	return text, nil
+}
+
+// HTML stripping for the direct fetch, modeled on the resident's helper of
+// the same shape: crude on purpose, because the consumer is a language model
+// that tolerates ragged text far better than it tolerates paying for markup.
+var (
+	// Written out per element rather than with a backreference: Go's regexp is
+	// RE2, which has none, and a pattern that compiles everywhere except here
+	// would fail at init rather than at review.
+	dropElements = regexp.MustCompile(`(?is)<script[^>]*>.*?</script\s*>` +
+		`|<style[^>]*>.*?</style\s*>` +
+		`|<noscript[^>]*>.*?</noscript\s*>` +
+		`|<svg[^>]*>.*?</svg\s*>` +
+		`|<head[^>]*>.*?</head\s*>`)
+	blockBreaks = regexp.MustCompile(`(?i)</?(p|div|br|li|tr|h[1-6]|section|article)[^>]*>`)
+	anyTag      = regexp.MustCompile(`(?s)<[^>]+>`)
+	manyBlanks  = regexp.MustCompile(`\n{3,}`)
+	manySpaces  = regexp.MustCompile(`[ \t]{2,}`)
+)
+
+// htmlToText strips markup well enough to read. It is intentionally crude:
+// the consumer is a language model, which tolerates ragged text far better
+// than it tolerates paying for markup it will ignore.
+func htmlToText(page string) string {
+	page = dropElements.ReplaceAllString(page, " ")
+	page = blockBreaks.ReplaceAllString(page, "\n")
+	page = anyTag.ReplaceAllString(page, " ")
+	page = strings.NewReplacer(
+		"&nbsp;", " ", "&amp;", "&", "&lt;", "<", "&gt;", ">",
+		"&quot;", `"`, "&#39;", "'", "&mdash;", "—", "&ndash;", "–",
+	).Replace(page)
+	page = manySpaces.ReplaceAllString(page, " ")
+	page = manyBlanks.ReplaceAllString(page, "\n\n")
+	return strings.TrimSpace(page)
+}
diff --git a/internal/search/search_test.go b/internal/search/search_test.go
index e2087b1126..fd1b9583b1 100644
--- a/internal/search/search_test.go
+++ b/internal/search/search_test.go
@@ -448,6 +448,140 @@ func TestJinaFetchIsCappedByTheRenderer(t *testing.T) {
 	}
 }
 
+// TestJinaFetchAsksForAFreshCopy: the reader keeps a 3600s page cache and
+// serves it by default, so every fetch must ask it not to, key or no key.
+func TestJinaFetchAsksForAFreshCopy(t *testing.T) {
+	const page = "# Go 1.25\n\nThe release notes, as markdown."
+	var (
+		gotNoCache string
+		gotAuth    string
+	)
+	server := serve(t, &jinaBaseURL, func(w http.ResponseWriter, r *http.Request) {
+		gotNoCache = r.Header.Get("X-No-Cache")
+		gotAuth = r.Header.Get("Authorization")
+		fmt.Fprint(w, page)
+	})
+	jinaBaseURL = server.URL + "/"
+
+	fetcher := jina{}.Bind(Options{})
+	if _, err := fetcher.Fetch(context.Background(), "https://go.dev/doc/go1.25"); err != nil {
+		t.Fatalf("Fetch error = %v", err)
+	}
+	if gotNoCache != "true" {
+		t.Errorf("X-No-Cache = %q, want the no-cache ask", gotNoCache)
+	}
+	if gotAuth != "" {
+		t.Errorf("Authorization = %q, want none without a key", gotAuth)
+	}
+
+	// A key only raises the rate ceiling, but it must ride the same ask.
+	if _, err := (jina{}).Bind(Options{JinaKey: "jina-key"}).Fetch(context.Background(), "https://go.dev"); err != nil {
+		t.Fatalf("keyed Fetch error = %v", err)
+	}
+	if gotNoCache != "true" {
+		t.Errorf("keyed X-No-Cache = %q, want the no-cache ask", gotNoCache)
+	}
+	if gotAuth != "Bearer jina-key" {
+		t.Errorf("Authorization = %q, want Bearer jina-key", gotAuth)
+	}
+}
+
+// TestJinaFetchFallsBackWhenTheReaderServedItsCache: when the reader marks
+// its answer a HIT it is serving someone ELSE's snapshot, so the fetch must
+// discard it and GET the target directly, reducing the page to text.
+// The handler tells the two routes apart by path: the target is the server's
+// own /page, while the reader route prefixes it with the reader base URL.
+func TestJinaFetchFallsBackWhenTheReaderServedItsCache(t *testing.T) {
+	const page = `<html><head><title>Example Domain</title></head><body><h1>Example Domain</h1><p>This domain is for use in illustrative examples in documents.</p><footer>Ignore &amp; forget me</footer></body></html>`
+	server := serve(t, &jinaBaseURL, func(w http.ResponseWriter, r *http.Request) {
+		if strings.HasPrefix(r.URL.Path, "/http") {
+			// The reader route: a cached snapshot of the page.
+			w.Header().Set("X-Cache", "HIT")
+			fmt.Fprint(w, "# stale cached snapshot, not the page")
+			return
+		}
+		if r.URL.Path != "/page" {
+			t.Errorf("unexpected route %q", r.URL.Path)
+			return
+		}
+		w.Header().Set("Content-Type", "text/html")
+		fmt.Fprint(w, page)
+	})
+	jinaBaseURL = server.URL + "/"
+
+	target := server.URL + "/page"
+	text, err := jina{}.Bind(Options{}).Fetch(context.Background(), target)
+	if err != nil {
+		t.Fatalf("Fetch error = %v", err)
+	}
+	if strings.Contains(text, "stale cached snapshot") {
+		t.Errorf("Fetch = %q, returned the reader's snapshot instead of the page", text)
+	}
+	if !strings.Contains(text, "Example Domain") || !strings.Contains(text, "illustrative examples") {
+		t.Errorf("Fetch = %q, want the direct page's own text", text)
+	}
+	if strings.Contains(text, "<h1>") || strings.Contains(text, "&amp;") {
+		t.Errorf("Fetch = %q, want the markup stripped and entities decoded", text)
+	}
+}
+
+// TestJinaFetchNamesTheFallbackWhenTheDirectFetchFails: a HIT must never be
+// returned, so when the direct fetch itself fails the error has to name that
+// fallback rather than hand the snapshot over silently.
+func TestJinaFetchNamesTheFallbackWhenTheDirectFetchFails(t *testing.T) {
+	server := serve(t, &jinaBaseURL, func(w http.ResponseWriter, r *http.Request) {
+		if strings.HasPrefix(r.URL.Path, "/http") {
+			w.Header().Set("X-Cache", "HIT")
+			fmt.Fprint(w, "# stale cached snapshot")
+			return
+		}
+		w.WriteHeader(http.StatusInternalServerError)
+		fmt.Fprint(w, "boom")
+	})
+	jinaBaseURL = server.URL + "/"
+
+	target := server.URL + "/page"
+	_, err := jina{}.Bind(Options{}).Fetch(context.Background(), target)
+	if err == nil {
+		t.Fatal("Fetch of a HIT whose direct fetch 500s returned no error")
+	}
+	msg := err.Error()
+	if !strings.Contains(msg, "cached snapshot rejected") || !strings.Contains(msg, "direct fetch of "+target) {
+		t.Errorf("Fetch error = %q, want the rejected snapshot and its direct fallback named", msg)
+	}
+	if strings.Contains(msg, "stale cached snapshot") {
+		t.Errorf("Fetch error = %q, leaked the snapshot body", msg)
+	}
+}
+
+// TestJinaFetchEmptyDirectPageIsAnError: an empty direct response means the
+// fallback found nothing to read, and the error must say so — never fall back
+// to the snapshot it just rejected.
+func TestJinaFetchEmptyDirectPageIsAnError(t *testing.T) {
+	server := serve(t, &jinaBaseURL, func(w http.ResponseWriter, r *http.Request) {
+		if strings.HasPrefix(r.URL.Path, "/http") {
+			w.Header().Set("X-Cache", "HIT")
+			fmt.Fprint(w, "# stale cached snapshot")
+			return
+		}
+		w.Header().Set("Content-Type", "text/html")
+	})
+	jinaBaseURL = server.URL + "/"
+
+	target := server.URL + "/page"
+	_, err := jina{}.Bind(Options{}).Fetch(context.Background(), target)
+	if err == nil {
+		t.Fatal("Fetch of a HIT whose direct body is empty returned no error")
+	}
+	msg := err.Error()
+	if !strings.Contains(msg, "cached snapshot rejected") || !strings.Contains(msg, "direct fetch of "+target) {
+		t.Errorf("Fetch error = %q, want the rejected snapshot and its direct fallback named", msg)
+	}
+	if strings.Contains(msg, "stale cached snapshot") {
+		t.Errorf("Fetch error = %q, leaked the snapshot body", msg)
+	}
+}
+
 // --- resolution -----------------------------------------------------------
 
 func TestResolveAuto(t *testing.T) {