diff --git a/internal/chat/command_output.go b/internal/chat/command_output.go
index 5669af5..dd22e2c 100644
--- a/internal/chat/command_output.go
+++ b/internal/chat/command_output.go
@@ -12,7 +12,9 @@ import (
 
 // Each stream holds at most one bounded lexical token. Tokens spanning writes
 // are scrubbed together; PEM bodies and oversized secret tokens are discarded before
-// reaching storage. No raw output is spooled to disk.
+// reaching storage or the agent. No raw output is spooled to disk. The agent
+// always gets the scrubbed head and tail of the command it asked to run; only
+// snippets capture stores the full scrubbed output.
 type outputStream struct {
 	parent             *commandOutput
 	name, token        string
@@ -50,10 +52,12 @@ func (o *commandOutput) emit(ctx context.Context, stream, data string) {
 	if data == "" || o.err != nil {
 		return
 	}
-	if o.err = o.st.AppendCommandOutput(ctx, o.action, stream, o.seq, []byte(data)); o.err != nil {
-		return
+	if o.snippets {
+		if o.err = o.st.AppendCommandOutput(ctx, o.action, stream, o.seq, []byte(data)); o.err != nil {
+			return
+		}
+		o.seq++
 	}
-	o.seq++
 	o.total += len(data)
 	if len(o.head) < 2048 {
 		n := min(2048-len(o.head), len(data))
@@ -97,9 +101,6 @@ func (s *outputStream) Write(p []byte) (int, error) {
 	if o.finished {
 		return 0, errors.New("command output closed")
 	}
-	if !o.snippets {
-		return len(p), nil
-	}
 	var safe strings.Builder
 	for _, b := range p {
 		if b == ' ' || b == '\n' || b == '\r' || b == '\t' {
@@ -140,9 +141,7 @@ func (o *commandOutput) Finish(ctx context.Context) (string, error) {
 	defer o.mu.Unlock()
 	if !o.finished {
 		for _, s := range o.streams {
-			if o.snippets {
-				o.emit(ctx, s.name, s.flush())
-			}
+			o.emit(ctx, s.name, s.flush())
 		}
 		o.finished = true
 	}
diff --git a/internal/chat/command_output_test.go b/internal/chat/command_output_test.go
index dc202df..a0ce66b 100644
--- a/internal/chat/command_output_test.go
+++ b/internal/chat/command_output_test.go
@@ -49,8 +49,12 @@ func TestCommandOutputCapture(t *testing.T) {
 			if err = st.ReadCommandOutput(context.Background(), "a", &full); err != nil {
 				t.Fatal(err)
 			}
+			// The agent gets the scrubbed head and tail under every policy.
+			if strings.Contains(summary, "ghp_") || !strings.Contains(summary, "HEAD") || !strings.Contains(summary, "[redacted]") || !strings.Contains(summary, "TAIL") || len(summary) > 5000 {
+				t.Fatal("agent summary", summary)
+			}
 			if !snippets {
-				if full.Len() != 0 || summary != "" {
+				if full.Len() != 0 {
 					t.Fatal("titles retained output")
 				}
 				return
diff --git a/site/src/content/docs/guides/chat.mdx b/site/src/content/docs/guides/chat.mdx
index c1cc0b7..33c9534 100644
--- a/site/src/content/docs/guides/chat.mdx
+++ b/site/src/content/docs/guides/chat.mdx
@@ -180,15 +180,17 @@ interrupted work is marked interrupted after a restart and never replayed to
 the agent.
 
 Your question is scrubbed and sent to Claude. Context and tool results pass the
-capture policy first. Under the default `capture_text = "titles"` Driga keeps
+capture policy first, except the output of commands you approve, described
+below. Under the default `capture_text = "titles"` Driga keeps
 no chat message bodies; `snippets` keeps scrubbed 200-character snippets.
 Claude's own session history is managed by Claude and is not controlled by
 Driga's capture policy.
 
 Command results join that action log under the same capture policy and
-30-day retention. Long command output keeps its head and tail in the
-conversation; retained full output is scrubbed and follows the capture
-policy (`titles` does not retain command output). Process result counts and
+30-day retention. Claude gets the scrubbed head and tail of each command you
+approve, up to 2 KB of each, under every capture policy, since reading that
+output is why it asked to run the command. Retained full output is scrubbed
+and follows the capture policy (`titles` does not retain command output). Process result counts and
 per-target outcomes appear on the card. External `driga mcp` clients and
 `driga ask` have no command actions.