diff --git a/bin/omarchy-agent-usage-claude b/bin/omarchy-agent-usage-claude
index 5b3634aef1c..012cf9c4dac 100755
--- a/bin/omarchy-agent-usage-claude
+++ b/bin/omarchy-agent-usage-claude
@@ -123,7 +123,15 @@ def scan_projects(projects_path: Path) -> dict[str, Any]:
   recent_dates = recent_date_strings()
   recent = {day: {"date": day, "messageCount": 0} for day in recent_dates}
 
-  seen: set[str] = set()
+  # A streamed response is several lines sharing one message id, and the
+  # first line's output_tokens is a placeholder while the last carries the
+  # real count. Keep each message's line with the highest output count
+  # (the last one scanned wins a tie) and tally once all lines are in. The
+  # whole line is kept: a response that fell back to another model
+  # mid-stream has lines that are separate snapshots, and mixing fields
+  # across them over-counts cache tokens and credits the wrong model.
+  # Tuple: (output_tokens, model, day, session_key, input, cache_read, cache_write).
+  seen: dict[str, tuple[int, str, str, str, int, int, int]] = {}
   sessions: set[str] = set()
   active_days: set[str] = set()
   today_sessions: set[str] = set()
@@ -158,44 +166,48 @@ def scan_projects(projects_path: Path) -> dict[str, Any]:
 
           message_id = message.get("id") or entry.get("messageId") or ""
           unique_key = str(message_id) if message_id else f"{path}:{entry.get('uuid') or entry.get('requestId') or line_number}"
-          if unique_key in seen:
-            continue
-          seen.add(unique_key)
 
           input_tokens = usage_token(usage, "input_tokens", "inputTokens")
           output_tokens = usage_token(usage, "output_tokens", "outputTokens")
           cache_read = usage_token(usage, "cache_read_input_tokens", "cacheReadInputTokens")
           cache_write = usage_token(usage, "cache_creation_input_tokens", "cacheCreationInputTokens")
-          total = input_tokens + output_tokens + cache_read + cache_write
-          if total <= 0:
+          if input_tokens + output_tokens + cache_read + cache_write <= 0:
+            continue
+
+          held = seen.get(unique_key)
+          if held is not None and output_tokens < held[0]:
             continue
 
           model = str(message.get("model") or entry.get("model") or "claude")
           day = local_date_from_timestamp(entry.get("timestamp") or message.get("timestamp"))
           session_key = str(entry.get("sessionId") or path)
-          sessions.add(session_key)
-          active_days.add(day)
-          prompts += 1
-
-          bucket = usage_by_model.setdefault(model, empty_bucket())
-          bucket["inputTokens"] += input_tokens
-          bucket["outputTokens"] += output_tokens
-          bucket["cacheReadInputTokens"] += cache_read
-          bucket["cacheCreationInputTokens"] += cache_write
-
-          if day in recent:
-            # recentDays.messageCount is actually a token total, despite the
-            # legacy name shared with synced snapshots.
-            recent[day]["messageCount"] += total
-
-          if day == today:
-            today_prompt_count += 1
-            today_sessions.add(session_key)
-            today_token_total += total
-            today_tokens[model] = today_tokens.get(model, 0) + total
+          seen[unique_key] = (output_tokens, model, day, session_key, input_tokens, cache_read, cache_write)
     except Exception as exc:
       print(f"Ignoring unreadable Claude project file {path}: {exc}", file=sys.stderr)
 
+  for output_tokens, model, day, session_key, input_tokens, cache_read, cache_write in seen.values():
+    total = input_tokens + output_tokens + cache_read + cache_write
+    sessions.add(session_key)
+    active_days.add(day)
+    prompts += 1
+
+    bucket = usage_by_model.setdefault(model, empty_bucket())
+    bucket["inputTokens"] += input_tokens
+    bucket["outputTokens"] += output_tokens
+    bucket["cacheReadInputTokens"] += cache_read
+    bucket["cacheCreationInputTokens"] += cache_write
+
+    if day in recent:
+      # recentDays.messageCount is actually a token total, despite the
+      # legacy name shared with synced snapshots.
+      recent[day]["messageCount"] += total
+
+    if day == today:
+      today_prompt_count += 1
+      today_sessions.add(session_key)
+      today_token_total += total
+      today_tokens[model] = today_tokens.get(model, 0) + total
+
   return {
     "todayPrompts": today_prompt_count,
     "todaySessions": len(today_sessions),
diff --git a/test/shell.d/agent-usage-claude-scanner-test.sh b/test/shell.d/agent-usage-claude-scanner-test.sh
index e1407f3cf2e..2f6c9c83ab2 100644
--- a/test/shell.d/agent-usage-claude-scanner-test.sh
+++ b/test/shell.d/agent-usage-claude-scanner-test.sh
@@ -29,14 +29,46 @@ pass "Claude collector counts each API message once"
   fail "Claude collector keeps mutually exclusive token categories" "$result"
 pass "Claude collector keeps mutually exclusive token categories"
 
+
 [[ $(jq -r '.id + "/" + .usageStatusText' <<<"$result") == "claude/Waiting for auth" ]] ||
   fail "Claude collector identifies itself and reports missing auth" "$result"
 pass "Claude collector identifies itself and reports missing auth"
 
+# A streamed response is several lines sharing one message id. The first
+# line's output_tokens is a placeholder and the last line has the real count,
+# so the message is counted from the line with the highest output.
+STREAM_HOME=$(mktemp -d)
+trap 'rm -rf "$TEST_HOME" "$STREAM_HOME"' EXIT
+stream_projects="$STREAM_HOME/.claude/projects/example"
+mkdir -p "$stream_projects"
+cat >"$stream_projects/session.jsonl" <<EOF
+{"timestamp":"$timestamp","type":"assistant","sessionId":"session-1","uuid":"event-1","message":{"id":"message-1","role":"assistant","model":"claude-test","usage":{"input_tokens":8,"cache_creation_input_tokens":100,"cache_read_input_tokens":2000,"output_tokens":1}}}
+{"timestamp":"$timestamp","type":"assistant","sessionId":"session-1","uuid":"event-2","message":{"id":"message-1","role":"assistant","model":"claude-test","usage":{"input_tokens":8,"cache_creation_input_tokens":100,"cache_read_input_tokens":2000,"output_tokens":1}}}
+{"timestamp":"$timestamp","type":"assistant","sessionId":"session-1","uuid":"event-3","message":{"id":"message-1","role":"assistant","model":"claude-test","usage":{"input_tokens":8,"cache_creation_input_tokens":100,"cache_read_input_tokens":2000,"output_tokens":260}}}
+{"timestamp":"$timestamp","type":"assistant","sessionId":"session-1","uuid":"event-4","message":{"id":"message-2","role":"assistant","model":"claude-test","usage":{"input_tokens":10,"cache_creation_input_tokens":300,"cache_read_input_tokens":4000,"output_tokens":1}}}
+{"timestamp":"$timestamp","type":"assistant","sessionId":"session-1","uuid":"event-5","message":{"id":"message-2","role":"assistant","model":"claude-fallback","usage":{"input_tokens":10,"cache_creation_input_tokens":0,"cache_read_input_tokens":3500,"output_tokens":40}}}
+EOF
+
+result=$(HOME="$STREAM_HOME" XDG_CACHE_HOME="$STREAM_HOME/.cache" XDG_DATA_HOME="$STREAM_HOME/.local/share" \
+  "$ROOT/bin/omarchy-agent-usage-claude" --force)
+
+[[ $(jq -c '.modelUsage["claude-test"]' <<<"$result") == '{"cacheCreationInputTokens":100,"cacheReadInputTokens":2000,"inputTokens":8,"outputTokens":260}' ]] ||
+  fail "Claude collector counts a streamed message from its final usage line" "$result"
+pass "Claude collector counts a streamed message from its final usage line"
+
+# A response that fell back to another model mid-stream has lines that are
+# separate snapshots: the final line's model and cache figures are kept whole
+# rather than a maximum taken per field across both.
+[[ $(jq -c '.modelUsage["claude-fallback"]' <<<"$result") == '{"cacheCreationInputTokens":0,"cacheReadInputTokens":3500,"inputTokens":10,"outputTokens":40}' ]] ||
+  fail "Claude collector keeps the final line whole after a mid-stream fallback" "$result"
+[[ $(jq -r '.totalPrompts' <<<"$result") == "2" ]] ||
+  fail "Claude collector keeps the final line whole after a mid-stream fallback" "$result"
+pass "Claude collector keeps the final line whole after a mid-stream fallback"
+
 # A machine with no transcripts and no stats-cache still gets today's counts
 # from history.jsonl alone.
 HISTORY_HOME=$(mktemp -d)
-trap 'rm -rf "$TEST_HOME" "$HISTORY_HOME"' EXIT
+trap 'rm -rf "$TEST_HOME" "$STREAM_HOME" "$HISTORY_HOME"' EXIT
 mkdir -p "$HISTORY_HOME/.claude"
 
 now_ms=$(($(date +%s) * 1000))
@@ -56,7 +88,7 @@ pass "Claude collector falls back to history.jsonl without a stats-cache"
 # A subscription burned entirely through opencode has no ~/.claude transcripts;
 # usage must come from opencode's message database, filtered to Anthropic.
 OPENCODE_HOME=$(mktemp -d)
-trap 'rm -rf "$TEST_HOME" "$HISTORY_HOME" "$OPENCODE_HOME"' EXIT
+trap 'rm -rf "$TEST_HOME" "$STREAM_HOME" "$HISTORY_HOME" "$OPENCODE_HOME"' EXIT
 
 python3 - "$OPENCODE_HOME/.local/share/opencode/opencode.db" <<'PY'
 import json
@@ -106,7 +138,7 @@ pass "Claude collector ignores prefix-colliding providers, user messages, and ma
 # Pi and omp can both spend a Claude subscription without writing native
 # Claude Code transcripts. Their compatible JSONL sessions must be included.
 PI_HOME=$(mktemp -d)
-trap 'rm -rf "$TEST_HOME" "$HISTORY_HOME" "$OPENCODE_HOME" "$PI_HOME"' EXIT
+trap 'rm -rf "$TEST_HOME" "$STREAM_HOME" "$HISTORY_HOME" "$OPENCODE_HOME" "$PI_HOME"' EXIT
 mkdir -p "$PI_HOME/.pi/agent/sessions/project" "$PI_HOME/.omp/agent/sessions/project"
 
 cat >"$PI_HOME/.pi/agent/sessions/project/pi.jsonl" <<EOF