fix(intercept): match block rules on the declared targets before any lookup

The target gate resolved each name through the cache before it matched the
asset block rules, so a call to a name that a name rule blocks still cost a
DNS lookup, and that query reached the name servers of the blocked domain.
URL rules (exact_url, fuzzy_url) were not evaluated at dispatch at all: the
gate passed no URL to the matcher, so they only applied to asset and intent
writes.

targetCheck.run now decides in this order, for a gated call:

1. URL rules, once per decision, on the URLs the call declares, as written.
2. Every block rule on the declared form of every target: the name as
   written, the IP literal, the range.
3. Only then resolution (names a derived exemption covers are still never
   resolved), and the block rules again on the resolved addresses.
4. Exemptions, pins and requests, unchanged.

A hit in 1 or 2 blocks with TARGET_RULE_BLOCKED and the same decided row as
any block, costs no lookup for any target of the call, and opens no request.

- intercept.extractDeclared returns the targets (ExtractTargets is now a thin
  wrapper, unchanged for its callers) and the URLs that name a target: the
  WebFetch url, and every URL token the command-string rules read in Bash,
  shell_open, Monitor, shell_send and the string leaves of other tools,
  including ExecuteExtraTool's inner call.
- The custom http tool's rendered URL and the URLs of a rendered command go to
  the same URL rules (gateRendered takes the URLs).
- Known limit, stated in the comments next to the matcher (db.matchOne) and
  at the dispatch match: URL rules match the declared string. exact_url misses
  any other spelling (case, percent-encoding, trailing slash, added query,
  /./), fuzzy_url misses an encoded substring, and a command can build its URL
  at run time. For a host, a domain, IP or CIDR rule is the one to write.

Unchanged: a tool-rule or judge deny still blocks before the gate is reached
(no extraction, no lookup, no request); asset block rules still win over an
allow and an ask; the gate off path does nothing; scripts stay advisory.

Tests:
- intercept: TestBlockRuleOnDeclaredNameNeedsNoLookup (exact_domain and
  fuzzy_domain; zero lookups, also for an unblocked name earlier in the same
  call), TestExtractDeclaredURLs, TestURLBlockRuleAtDispatch (WebFetch, Bash,
  MCP, ExecuteExtraTool, a pinned and an exempt host, a disabled rule, the
  rendered URL and command; zero lookups), TestExactURLRuleMatchesOnlyAsWritten
  (documents the limit).
- guard: TestBlockRuleWinsOverAllowAndAsk (name rule and URL rule, each over a
  rule allow, judge allow, rule ask and judge ask, real interceptor: blocked,
  no allow row, nobody asked, zero lookups). TestDenyRuleWinsBeforeTargetRequest
  now also covers a judge deny and asserts zero lookups with the real gate.
- server: TestBypass_CustomScriptToolTargetNotGated documents that a script
  tool's target, written in its stored code, is not visible to the gate: in a
  gated task the call is allowed and reaches the target with no lookup, no
  request and no row.

db/asset_intercept_match.go

@@ -45,6 +45,12 @@ func (r AssetInterceptRule) Reason() string {
 
 // matchOne checks whether a single enabled rule matches the given domain/IP/URL
 // candidates, and returns the concrete matched value.
+//
+// URL rules match the URL string as given, with no normalisation. Known limit:
+// exact_url misses any other spelling of the same URL (case, percent-encoding
+// such as %2F, a trailing slash, an added query, /./), and fuzzy_url misses an
+// encoded form of its substring. At tool dispatch the string is the URL the
+// call declares (intercept, targetCheck.run).
 func matchOne(r AssetInterceptRule, domains, ips, urls []string) (string, bool) {
 	p := strings.TrimSpace(r.Pattern)
 	if p == "" {

guard/target_gate_test.go

@@ -282,18 +282,27 @@ type gateDB struct {
 	task string // task id
 }
 
-// stubResolver answers every name with the addresses it holds.
+// stubResolver answers every name with the addresses it holds, and counts the
+// lookups.
 type stubResolver struct {
 	mu  sync.Mutex
 	ips []string
 	err error
+	n   int
 }
 
 func (r *stubResolver) Name() string { return "stub" }
 
+func (r *stubResolver) lookups() int {
+	r.mu.Lock()
+	defer r.mu.Unlock()
+	return r.n
+}
+
 func (r *stubResolver) LookupIPs(context.Context, string) ([]netip.Addr, error) {
 	r.mu.Lock()
 	defer r.mu.Unlock()
+	r.n++
 	if r.err != nil {
 		return nil, r.err
 	}
@@ -576,8 +585,8 @@ func TestGateOffLeavesEveryBranchUnchanged(t *testing.T) {
 	}
 }
 
-// A deny by a rule blocks first, as before: the gate is not consulted and no
-// target request exists.
+// A deny by a rule or by the judge blocks first, as before: the gate is not
+// consulted, nothing is resolved and no target request exists.
 func TestDenyRuleWinsBeforeTargetRequest(t *testing.T) {
 	g := newGateDB(t)
 	m := marker(t)
@@ -601,8 +610,77 @@ func TestDenyRuleWinsBeforeTargetRequest(t *testing.T) {
 	if block, _ := pre(guard, g.ctx(true, time.Minute), "Bash", `{"command":"curl https://app.example.com/ # `+m+`"}`); !block {
 		t.Fatal("deny did not block")
 	}
-	if len(g.pending()) != 0 || len(g.pins()) != 0 {
-		t.Fatalf("pending=%+v pins=%+v, want none", g.pending(), g.pins())
+	if len(g.pending()) != 0 || len(g.pins()) != 0 || g.res.lookups() != 0 {
+		t.Fatalf("pending=%+v pins=%+v lookups=%d, want none", g.pending(), g.pins(), g.res.lookups())
+	}
+
+	// A deny by the judge takes the same branch.
+	g.judge("deny")
+	input := `{"command":"curl https://judged.example.com/"}`
+	gate = &fakeGate{t: t, check: unpinnedOnce}
+	guard = NewWithInterceptor(g.ic)
+	guard.gate = gate
+	before := len(g.rows())
+	if block, msg := pre(guard, g.ctx(true, time.Minute), "Bash", input); !block || !strings.Contains(msg, "judge says deny") || strings.Contains(msg, "TARGET_") {
+		t.Fatalf("judge deny: block=%v msg=%q", block, msg)
+	}
+	if gate.checks != 0 || len(gate.asked) != 0 {
+		t.Fatalf("judge deny: the gate was reached: %d checks, %d requests", gate.checks, len(gate.asked))
+	}
+	if rows := g.rows()[before:]; len(rows) != 1 || rows[0].status != "denied" || rows[0].decidedBy != "model" || rows[0].targetKey != "" {
+		t.Fatalf("judge deny: rows = %+v, want the denied judge row only", rows)
+	}
+	guard = NewWithInterceptor(g.ic)
+	if block, _ := pre(guard, g.ctx(true, time.Minute), "Bash", input); !block {
+		t.Fatal("judge deny did not block")
+	}
+	if len(g.pending()) != 0 || len(g.pins()) != 0 || g.res.lookups() != 0 {
+		t.Fatalf("judge deny: pending=%+v pins=%+v lookups=%d, want none", g.pending(), g.pins(), g.res.lookups())
+	}
+}
+
+// An asset block rule wins over an allow and an ask, of a rule or of the
+// judge: the call is blocked, the allow is not logged, no human is asked, and
+// the rule hits on what the call declares, so nothing is resolved. A name rule
+// and a URL rule behave the same.
+func TestBlockRuleWinsOverAllowAndAsk(t *testing.T) {
+	const input = `{"command":"curl https://app.example.com/blocked-path # %s"}`
+	for _, block := range []struct{ name, kind, pattern string }{
+		{"name rule", "exact_domain", "app.example.com"},
+		{"URL rule", "fuzzy_url", "/blocked-path"},
+	} {
+		for _, v := range []struct{ name, by, action string }{
+			{"rule allow", "rule", "allow"},
+			{"judge allow", "judge", "allow"},
+			{"rule ask", "rule", "ask"},
+			{"judge ask", "judge", "ask"},
+		} {
+			t.Run(block.name+" over "+v.name, func(t *testing.T) {
+				g := newGateDB(t)
+				id, _ := strconv.ParseInt(g.task, 10, 64)
+				if _, err := g.d.Assets().CreateTaskInterceptRule(id, "block", block.kind, block.pattern, "guard test", true); err != nil {
+					t.Fatal(err)
+				}
+				m := marker(t)
+				if v.by == "rule" {
+					g.rule(v.action, m)
+				} else {
+					g.judge(v.action)
+				}
+				guard := NewWithInterceptor(g.ic)
+				r := g.result(call(guard, g.ctx(true, time.Minute), "Bash", fmt.Sprintf(input, m)))
+				if !r.block || !strings.Contains(r.msg, intercept.TargetRuleBlocked+": ") || !strings.Contains(r.msg, block.pattern) {
+					t.Fatalf("block=%v msg=%q, want %s", r.block, r.msg, intercept.TargetRuleBlocked)
+				}
+				rows := g.rows()
+				if len(rows) != 1 || rows[0].status != "denied" || rows[0].decidedBy != "system" || rows[0].ruleID != 0 {
+					t.Fatalf("rows = %+v, want the one block row of the gate: no allow logged, nobody asked", rows)
+				}
+				if n := g.res.lookups(); n != 0 || len(g.pins()) != 0 {
+					t.Fatalf("%d lookups, pins %+v; want none", n, g.pins())
+				}
+			})
+		}
 	}
 }
 

intercept/target_declared_test.go

@@ -0,0 +1,193 @@
+package intercept
+
+import (
+	"encoding/json"
+	"slices"
+	"strings"
+	"testing"
+
+	"github.com/Autumn-27/artex/db"
+)
+
+// The declared forms of a call decide first: the URL block rules on the URLs
+// the call writes, then the block rules on every name, IP literal and range as
+// written. Nothing is resolved before that, so a hit costs no lookup.
+
+func TestExtractDeclaredURLs(t *testing.T) {
+	raw := func(v any) []byte {
+		b, _ := json.Marshal(v)
+		return b
+	}
+	for _, c := range []struct {
+		name  string
+		tool  string
+		input []byte
+		want  []string
+	}{
+		{"WebFetch", "WebFetch", raw(map[string]string{"url": " https://app.example.com/admin?x=1 "}), []string{"https://app.example.com/admin?x=1"}},
+		{"WebFetch without a target", "WebFetch", raw(map[string]string{"url": "http:///path"}), nil},
+		{"Bash, query kept", "Bash", raw(command("curl -s 'https://a.example.com/admin?x=1'")), []string{"https://a.example.com/admin?x=1"}},
+		{"Bash, two URLs", "Bash", raw(command("curl http://a.example.com/x && curl https://b.example.com:8443/y")),
+			[]string{"http://a.example.com/x", "https://b.example.com:8443/y"}},
+		{"shell_send", "shell_send", raw(map[string]string{"text": "curl http://a.example.com/x\n"}), []string{"http://a.example.com/x"}},
+		{"Monitor", "Monitor", raw(command("watch curl http://198.51.100.9:8080/status")), []string{"http://198.51.100.9:8080/status"}},
+		{"MCP string leaves", "mcp__scan__run", []byte(`{"target":"http://a.example.com/p","opts":{"list":["https://b.example.com/q"]}}`),
+			[]string{"http://a.example.com/p", "https://b.example.com/q"}},
+		{"ExecuteExtraTool unwrap", executeExtraTool, raw(map[string]any{"tool_name": "WebFetch", "params": map[string]string{"url": "https://a.example.com/z"}}),
+			[]string{"https://a.example.com/z"}},
+		{"tool without a target", "web_search", raw(map[string]string{"query": "https://a.example.com/admin"}), nil},
+		{"wrapped tool without a target", executeExtraTool, raw(map[string]any{"tool_name": "web_search", "params": map[string]string{"query": "https://a.example.com/admin"}}), nil},
+		{"authority that names no host", "Bash", raw(command("curl http://$HOST/x https://${H}:8443/admin")), nil},
+	} {
+		d := extractDeclared(c.tool, c.input)
+		if !slices.Equal(d.urls, c.want) {
+			t.Errorf("%s: urls = %q, want %q", c.name, d.urls, c.want)
+		}
+		// The targets are those ExtractTargets returns.
+		if !slices.Equal(d.targets, ExtractTargets(c.tool, c.input)) {
+			t.Errorf("%s: targets differ from ExtractTargets", c.name)
+		}
+		if len(d.urls) > 0 && len(d.targets) == 0 {
+			t.Errorf("%s: a URL without a target", c.name)
+		}
+	}
+}
+
+// A name that a name rule blocks is blocked on the name as written: no lookup
+// is made for it, nor for another name of the same call that no rule blocks.
+func TestBlockRuleOnDeclaredNameNeedsNoLookup(t *testing.T) {
+	for _, rule := range []db.AssetInterceptRule{
+		{Enabled: true, Kind: "exact_domain", Pattern: "blocked.example.com"},
+		{Enabled: true, Kind: "fuzzy_domain", Pattern: "blocked.example"},
+	} {
+		g := newGate(t)
+		g.res.set("ok.example.com", "203.0.113.7")
+		g.res.set("blocked.example.com", "203.0.113.8")
+		g.store.rules = []db.AssetInterceptRule{rule}
+		for name, v := range map[string]verdict{
+			"WebFetch": g.fetch("https://blocked.example.com/"),
+			// The unblocked name comes first in the call.
+			"two names": g.check("Bash", `{"command":"curl https://ok.example.com/ https://blocked.example.com/"}`),
+		} {
+			res := g.blocked(rule.Kind+", "+name, TargetRuleBlocked, v)
+			if res.Host != "blocked.example.com" || !strings.Contains(res.Reason, rule.Pattern) {
+				t.Fatalf("%s, %s: result = %+v", rule.Kind, name, res)
+			}
+		}
+		if n := g.res.total(); n != 0 {
+			t.Fatalf("%s: %d lookups, want none", rule.Kind, n)
+		}
+		for _, row := range g.store.logged {
+			if a := row.Audit.Targets[0]; a.Outcome != "rule_blocked" || len(a.ResolvedIPs) != 0 {
+				t.Fatalf("%s: logged target = %+v, want rule_blocked with no address", rule.Kind, a)
+			}
+		}
+	}
+}
+
+// URL rules apply at dispatch, on the URLs the call declares, before anything
+// is resolved. A hit blocks with the code of every block rule, opens no
+// request, and wins over a pin and an exemption.
+func TestURLBlockRuleAtDispatch(t *testing.T) {
+	g := newGate(t)
+	g.res.set("app.example.com", "203.0.113.7")
+	g.res.set("pypi.org", "151.101.0.223")
+	g.store.exempt = []db.TargetExemption{{Host: "pypi.org", Port: 443}}
+	g.store.approve(TargetRequest{Host: "app.example.com", Port: 443, ResolvedIPs: []string{"203.0.113.7"}})
+	g.passes("pinned", g.fetch("https://app.example.com/admin"))
+	g.passes("exempt", g.fetch("https://pypi.org/admin"))
+
+	rule := func(kind, pattern string, enabled bool) []db.AssetInterceptRule {
+		return []db.AssetInterceptRule{{Enabled: enabled, Kind: kind, Pattern: pattern}}
+	}
+	// A disabled URL rule blocks nothing.
+	g.store.rules = rule("exact_url", "https://app.example.com/admin", false)
+	g.passes("disabled rule", g.fetch("https://app.example.com/admin"))
+
+	lookups, rows := g.res.total(), len(g.store.logged)
+	for _, c := range []struct {
+		name       string
+		rules      []db.AssetInterceptRule
+		v          func() verdict
+		host       string
+		port       int
+		matchedURL string
+	}{
+		{"exact_url, WebFetch of a pinned host", rule("exact_url", "https://app.example.com/admin", true),
+			func() verdict { return g.fetch("https://app.example.com/admin") }, "app.example.com", 443, "https://app.example.com/admin"},
+		{"fuzzy_url, exempt host", rule("fuzzy_url", "/admin", true),
+			func() verdict { return g.fetch("https://pypi.org/admin") }, "pypi.org", 443, "https://pypi.org/admin"},
+		{"fuzzy_url, Bash with a query", rule("fuzzy_url", "/admin", true),
+			func() verdict { return g.check("Bash", `{"command":"curl -s 'https://app.example.com/admin?x=1'"}`) }, "app.example.com", 443, "https://app.example.com/admin?x=1"},
+		{"fuzzy_url, second URL of the call", rule("fuzzy_url", "/admin", true),
+			func() verdict {
+				return g.check("Bash", `{"command":"curl https://app.example.com/ && curl http://198.51.100.9:8080/admin"}`)
+			}, "198.51.100.9", 8080, "http://198.51.100.9:8080/admin"},
+		{"fuzzy_url, MCP input", rule("fuzzy_url", "/ADMIN", true),
+			func() verdict { return g.check("mcp__scan__run", `{"target":"https://app.example.com/admin/users"}`) }, "app.example.com", 443, "https://app.example.com/admin/users"},
+		{"exact_url, wrapped WebFetch", rule("exact_url", "https://app.example.com/admin", true),
+			func() verdict {
+				return g.check(executeExtraTool, `{"tool_name":"WebFetch","params":{"url":"https://app.example.com/admin"}}`)
+			}, "app.example.com", 443, "https://app.example.com/admin"},
+	} {
+		g.store.rules = c.rules
+		g.res.set("app.example.com", "198.51.100.200") // a lookup would make a pinned host ask
+		res := g.blocked(c.name, TargetRuleBlocked, c.v())
+		if res.Host != c.host || res.Port != c.port || !strings.Contains(res.Reason, c.matchedURL+" is blocked by an asset rule") ||
+			!strings.Contains(res.Reason, c.rules[0].Pattern) {
+			t.Errorf("%s: result = %+v", c.name, res)
+		}
+	}
+	if n := g.res.total() - lookups; n != 0 {
+		t.Fatalf("%d lookups for calls a URL rule blocks, want none", n)
+	}
+	if got := g.store.outcomes()[rows:]; len(got) != 6 || slices.ContainsFunc(got, func(o string) bool { return o != "denied/rule_blocked" }) {
+		t.Fatalf("logged rows = %v, want 6 rule_blocked rows", got)
+	}
+
+	// The rendered URL of a custom http tool and the URLs of a rendered command
+	// meet the same rules, with no lookup.
+	g.store.rules = rule("exact_url", "https://app.example.com/admin", true)
+	if res := GateRenderedURL(g.ctx(), "probe_http", "https://app.example.com/admin"); res.Code != TargetRuleBlocked || res.Host != "app.example.com" {
+		t.Fatalf("rendered URL: %+v", res)
+	}
+	g.store.rules = rule("fuzzy_url", "/admin", true)
+	if res := GateRenderedCommand(g.ctx(), "probe_cmd", "curl 'https://app.example.com/admin'"); res.Code != TargetRuleBlocked {
+		t.Fatalf("rendered command: %+v", res)
+	}
+	if n := g.res.total() - lookups; n != 0 {
+		t.Fatalf("%d lookups for rendered calls a URL rule blocks, want none", n)
+	}
+}
+
+// Documents the known limit of URL rules: they match the URL as the call
+// writes it. Another spelling of the same URL is not blocked. For a host, a
+// domain, IP or CIDR rule is the one to write.
+func TestExactURLRuleMatchesOnlyAsWritten(t *testing.T) {
+	g := newGate(t)
+	g.res.set("app.example.com", "203.0.113.7")
+	g.store.approve(TargetRequest{Host: "app.example.com", Port: 443, ResolvedIPs: []string{"203.0.113.7"}})
+
+	g.store.rules = []db.AssetInterceptRule{{Enabled: true, Kind: "exact_url", Pattern: "https://app.example.com/admin"}}
+	g.blocked("as written", TargetRuleBlocked, g.fetch("https://app.example.com/admin"))
+	for _, u := range []string{
+		"https://APP.example.com/admin",     // case of the host
+		"https://app.example.com/Admin",     // case of the path
+		"https://app.example.com/admin/",    // trailing slash
+		"https://app.example.com/admin?x=1", // added query
+		"https://app.example.com/%61dmin",   // percent-encoding
+		"https://app.example.com/./admin",   // dot segment
+		"https://app.example.com:443/admin", // default port written out
+	} {
+		g.passes("exact_url, "+u, g.fetch(u))
+	}
+
+	g.store.rules = []db.AssetInterceptRule{{Enabled: true, Kind: "fuzzy_url", Pattern: "/admin"}}
+	g.blocked("fuzzy, as written", TargetRuleBlocked, g.fetch("https://app.example.com/x/admin"))
+	for _, u := range []string{"https://app.example.com/%61dmin", "https://app.example.com/adm%69n", "https://app.example.com//%2fadmin"} {
+		if strings.Contains(strings.ToLower(u), "/admin") {
+			t.Fatalf("%s contains the pattern; the case documents an encoded form", u)
+		}
+		g.passes("fuzzy_url, "+u, g.fetch(u))
+	}
+}

intercept/target_extract.go

@@ -62,6 +62,13 @@ func toolSet(names ...string) map[string]bool {
 // names on its own. The gate detects a declared target; it does not prevent an
 // undeclared connection.
 func ExtractTargets(toolName string, input []byte) []Target {
+	return extractDeclared(toolName, input).targets
+}
+
+// extractDeclared reads what a tool call declares: its targets, as
+// ExtractTargets returns them, and the URLs it writes, as written, for the URL
+// block rules. A URL is kept only when it names a target.
+func extractDeclared(toolName string, input []byte) targetList {
 	if toolName == executeExtraTool {
 		var in struct {
 			ToolName string          `json:"tool_name"`
@@ -72,13 +79,15 @@ func ExtractTargets(toolName string, input []byte) []Target {
 		}
 	}
 	if noTargetTools[toolName] {
-		return nil
+		return targetList{}
 	}
 	var out targetList
 	switch toolName {
 	case "WebFetch":
-		if t, ok := NormalizeTarget(declaredFields(input).URL); ok {
+		u := declaredFields(input).URL
+		if t, ok := NormalizeTarget(u); ok {
 			out.add(t)
+			out.urls = append(out.urls, strings.TrimSpace(u))
 		}
 	case "Bash", "shell_open", "Monitor":
 		out.scan(declaredFields(input).Command)
@@ -87,7 +96,7 @@ func ExtractTargets(toolName string, input []byte) []Target {
 	default:
 		out.scanStrings(input)
 	}
-	return out.targets
+	return out
 }
 
 // declaredInput holds the fields of the built-in tools that carry a target.
@@ -109,6 +118,7 @@ func declaredFields(input []byte) declaredInput {
 type targetList struct {
 	targets []Target
 	seen    map[Target]bool
+	urls    []string // the URLs that name a target, as written
 }
 
 func (l *targetList) add(t Target) {
@@ -207,6 +217,10 @@ func (l *targetList) scan(s string) {
 				found = true
 			}
 		}
+		if found {
+			// The whole token, path and query included, from the original string.
+			l.urls = append(l.urls, s[m[0]:m[1]])
+		}
 		// The scheme and what follows the authority are never read again. An
 		// authority that named no host is: an address written in it still counts.
 		blank(m[0], m[4])

intercept/target_gate.go

@@ -93,12 +93,12 @@ func (i *Interceptor) CheckTargets(ctx context.Context, scope TargetScope, toolN
 	if scope.Enabled == nil || !scope.Enabled() {
 		return nil, TargetResult{}
 	}
-	targets := ExtractTargets(toolName, input)
-	if len(targets) == 0 {
+	declared := extractDeclared(toolName, input)
+	if len(declared.targets) == 0 {
 		return nil, TargetResult{}
 	}
-	c := &targetCheck{i: i, ctx: ctx, scope: scope, tool: toolName, input: input}
-	return c.run(targets)
+	c := &targetCheck{i: i, ctx: ctx, scope: scope, tool: toolName, input: input, urls: declared.urls}
+	return c.run(declared.targets)
 }
 
 // targetCheck is one run of the gate decision.
@@ -109,6 +109,7 @@ type targetCheck struct {
 	ref   db.TargetScopeRef
 	tool  string
 	input []byte
+	urls  []string // the URLs the call declares, as written, for the URL block rules
 }
 
 // checkedTarget is one target of the call and what the gate learned about it.
@@ -146,23 +147,47 @@ func (c *targetCheck) run(targets []Target) ([]TargetRequest, TargetResult) {
 		return c.block(TargetGateUnavailable, "gate_unavailable", first, "the exemption list could not be read: "+err.Error())
 	}
 
-	// Block rules come first, on the name and on what it resolves to: an exempt
-	// or pinned host that matches one is blocked. A name the configuration
-	// exempts is never resolved.
+	// Block rules come first, and an exempt or pinned host that matches one is
+	// blocked. They are matched on what the call declares before anything is
+	// resolved: a hit there costs no lookup, for any target of the call.
+	//
+	// URL rules (exact_url, fuzzy_url) are matched once per decision, on the
+	// URLs as the call writes them. That is a known limit: exact_url compares
+	// the string as written, so another case, encoding (%2F), a trailing slash,
+	// an added query or /./ does not match, and fuzzy_url is evaded by encoding
+	// the substring; a command can also build its URL at run time. To keep an
+	// agent away from a host, a domain, IP or CIDR rule is the one to write.
+	if rule, u, hit := db.MatchAssetInterceptRules(rules, nil, nil, c.urls); hit {
+		ct := &checkedTarget{Target: targets[0]}
+		if t, ok := NormalizeTarget(u); ok {
+			ct.Target = t
+		}
+		return c.block(TargetRuleBlocked, "rule_blocked", ct, fmt.Sprintf("%s is blocked by an asset rule. %s", u, rule.Reason()))
+	}
+	// Then the declared form of every target: the name as written, the IP
+	// literal, the range.
 	checked := make([]*checkedTarget, len(targets))
 	for n, t := range targets {
 		ct := &checkedTarget{Target: t}
 		checked[n] = ct
-		switch {
-		case t.Kind != TargetDomain:
+		if t.Kind != TargetDomain {
 			ct.ips, ct.resolver, ct.resolvedAt = []string{t.Host}, literalResolver, time.Now().UTC()
-		case !ex.derived[db.TargetKey(t.Host, t.Port)]:
-			c.resolve(ct, false)
 		}
 		if rule, hit := blockRule(rules, ct); hit {
 			return c.block(TargetRuleBlocked, "rule_blocked", ct, fmt.Sprintf("%s is blocked by an asset rule. %s", t, rule.Reason()))
 		}
 	}
+	// Only then are the names resolved, and the rules matched on the addresses.
+	// A name the configuration exempts is never resolved.
+	for _, ct := range checked {
+		if ct.Kind != TargetDomain || ex.derived[db.TargetKey(ct.Host, ct.Port)] {
+			continue
+		}
+		c.resolve(ct, false)
+		if rule, hit := blockRule(rules, ct); hit {
+			return c.block(TargetRuleBlocked, "rule_blocked", ct, fmt.Sprintf("%s is blocked by an asset rule. %s", ct.Target, rule.Reason()))
+		}
+	}
 
 	i.mu.RLock()
 	proxied := i.upstreamProxy
@@ -306,7 +331,8 @@ func (c *targetCheck) pinUsed(ct *checkedTarget, pin *db.TargetPin) {
 }
 
 // blockRule matches the asset block rules against a target: its name, and its
-// addresses. A range is matched as written and by its network address.
+// addresses. A range is matched as written and by its network address. URL
+// rules are not matched here: run matches them once, on the declared URLs.
 func blockRule(rules []db.AssetInterceptRule, ct *checkedTarget) (db.AssetInterceptRule, bool) {
 	var domains []string
 	ips := ct.ips

intercept/target_rendered.go

@@ -40,7 +40,7 @@ func GateRenderedURL(ctx context.Context, toolName, rawURL string) TargetResult
 	if err != nil {
 		return undeterminedTarget(ctx, scope, toolName, input, "the target of the rendered URL could not be read ("+err.Error()+")")
 	}
-	return gateRendered(ctx, scope, toolName, input, []Target{t})
+	return gateRendered(ctx, scope, toolName, input, []Target{t}, []string{rawURL})
 }
 
 // GateRenderedCommand is the target gate for the command a custom command tool
@@ -62,7 +62,7 @@ func GateRenderedCommand(ctx context.Context, toolName, command string, readings
 	for _, r := range readings {
 		l.scan(r)
 	}
-	return gateRendered(ctx, scope, toolName, renderedInput("command", command), l.targets)
+	return gateRendered(ctx, scope, toolName, renderedInput("command", command), l.targets, l.urls)
 }
 
 func renderedScope(ctx context.Context) (TargetScope, bool) {
@@ -79,9 +79,10 @@ func renderedInput(field, value string) []byte {
 
 // gateRendered runs the gate decision on the targets of a rendered call, asks
 // for the approval of those no pin covers, and checks again after the
-// decisions. It never merges an ask: the rules and the judge decided at the
-// hook, before the tool ran.
-func gateRendered(ctx context.Context, scope TargetScope, toolName string, input []byte, targets []Target) TargetResult {
+// decisions. urls are the URLs the call renders, for the URL block rules. It
+// never merges an ask: the rules and the judge decided at the hook, before the
+// tool ran.
+func gateRendered(ctx context.Context, scope TargetScope, toolName string, input []byte, targets []Target, urls []string) TargetResult {
 	if len(targets) == 0 {
 		return TargetResult{}
 	}
@@ -95,7 +96,7 @@ func gateRendered(ctx context.Context, scope TargetScope, toolName string, input
 		if !scope.Enabled() {
 			return nil, TargetResult{}
 		}
-		c := &targetCheck{i: i, ctx: ctx, scope: scope, tool: toolName, input: input}
+		c := &targetCheck{i: i, ctx: ctx, scope: scope, tool: toolName, input: input, urls: urls}
 		return c.run(targets)
 	}
 	unpinned, res := check()

server/customtool_target_gate_test.go

@@ -571,6 +571,49 @@ func TestDenyRuleWinsBeforeRenderedTarget(t *testing.T) {
 	}
 }
 
+// Documents a known limit: script tools stay advisory (D9). A custom script
+// tool's target is in its stored code, which neither the hook nor the tool
+// reads, so in a gated task the call is allowed and the script reaches a target
+// nobody approved: no lookup, no request, no row. This test changes only if
+// that decision changes.
+func TestBypass_CustomScriptToolTargetNotGated(t *testing.T) {
+	if detectPython() == "" {
+		t.Skip("python3 unavailable")
+	}
+	g := newGateServer(t)
+	on := true
+	task := g.newTask(t, db.TaskCreateOptions{TargetGate: &on})
+	res := &countingResolver{answers: map[string][]string{"localhost": {"127.0.0.1"}}}
+	g.m.interceptor.SetResolver(res)
+	_, hits, _, port := countingServer(t)
+
+	code := "import urllib.request\n" +
+		"opener = urllib.request.build_opener(urllib.request.ProxyHandler({}))\n" +
+		"print(opener.open('http://localhost:" + port + "/script', timeout=20).read().decode())\n"
+	tool := g.s.buildCustomTool(&db.Tool{Key: "probe_script", Kind: "script", Exec: execJSON(t, map[string]any{"code": code})})
+	// The gate sees no target in the call: only the parameters reach the hook.
+	if got := intercept.ExtractTargets("probe_script", []byte(`{}`)); len(got) != 0 {
+		t.Fatalf("the hook reads %+v; the test needs a target only the stored code names", got)
+	}
+
+	p := &toolScript{calls: []scriptedCall{{"probe_script", `{}`}}, results: make(chan string, 16)}
+	done := runScriptedWorker(t, g, task, p, tool)
+	if r := p.next(t); !strings.Contains(r, "probe-reached") || strings.Contains(r, "TARGET_") {
+		t.Fatalf("script tool in a gated task: %q, want it allowed and the target reached", r)
+	}
+	select {
+	case <-done:
+	case <-time.After(60 * time.Second):
+		t.Fatal("the worker run did not end")
+	}
+	if hits.Load() != 1 {
+		t.Fatalf("%d requests reached the unapproved target, want 1 (the gap)", hits.Load())
+	}
+	if n, rows := res.lookups(), targetRowCount(t, g.m.pg, task.ID); n != 0 || rows != 0 {
+		t.Fatalf("%d lookups, %d target rows; the gate did see the script's target", n, rows)
+	}
+}
+
 // A task with the gate off (the default) runs as before: a worker's WebFetch
 // and the custom tools reach their targets, and the gate writes no row and
 // resolves no name.