diff --git a/claude_agent_pipe_sandboxed.py b/claude_agent_pipe_sandboxed.py
index 8bea235..226a237 100644
--- a/claude_agent_pipe_sandboxed.py
+++ b/claude_agent_pipe_sandboxed.py
@@ -279,6 +279,12 @@ def _claude_command(prompt: str, cfg: _ClaudeRunConfig) -> str:
     # overwrites with the real credential. The placeholder value is never
     # seen by Anthropic.
     env_parts.append("ANTHROPIC_AUTH_TOKEN=proxied")
+    # Per-chat CLAUDE_CONFIG_DIR isolates session JSONL, settings.json, and
+    # any file state the CLI writes. Without this, two concurrent chats for
+    # the same OWUI user race each other on ~/.claude/. With the proxy
+    # holding credentials, the config dir stores only session history —
+    # safe to shard per chat and discard with the workdir.
+    env_parts.append(f"CLAUDE_CONFIG_DIR={shlex.quote(cfg.workdir + '/.claude')}")
     # `claude --dangerously-skip-permissions` still refuses root unless told
     # it's sandboxed. open-terminal runs processes as the per-user UID so this
     # rarely matters, but setting it is harmless.
@@ -337,6 +343,7 @@ async def _stream_claude_events(
     # that fail become synthetic _raw events so the caller can log them
     # without disrupting the agent loop.
     buf = ""
+    exited_cleanly = False
     try:
         async for stream, chunk in client.stream_output(handle, session_id=session_id):
             if stream == "exit":
@@ -347,6 +354,7 @@ async def _stream_claude_events(
                     except json.JSONDecodeError:
                         yield {"type": "_raw", "text": buf.strip()}
                     buf = ""
+                exited_cleanly = True
                 yield {"type": "_exit", "code": int(chunk)}
                 return
             buf += chunk
@@ -359,9 +367,16 @@ async def _stream_claude_events(
                     yield json.loads(line)
                 except json.JSONDecodeError:
                     yield {"type": "_raw", "text": line}
-    except asyncio.CancelledError:
-        await client.kill(handle, force=True)
-        raise
+    finally:
+        # Reap orphan claude processes on *any* abnormal exit path —
+        # cancellation, httpx timeout, pipe generator GC'd mid-stream,
+        # unhandled exception. Prior to this, only CancelledError
+        # triggered the kill, so a dropped browser with OWUI still
+        # running left claude billing tokens for up to EXECUTE_TIMEOUT.
+        # `kill` is best-effort (swallows HTTP errors) so double-kills
+        # on already-exited processes are harmless.
+        if not exited_cleanly:
+            await client.kill(handle, force=True)
 
 
 # ============================================================================
diff --git a/sandbox/Dockerfile b/sandbox/Dockerfile
index 49c8b81..bb79d35 100644
--- a/sandbox/Dockerfile
+++ b/sandbox/Dockerfile
@@ -1,9 +1,18 @@
 FROM ghcr.io/open-webui/open-terminal:latest
 
+# Pin claude-code CLI so two clean builds from the same source produce
+# identical behaviour. Override at build time:
+#   docker compose build --build-arg CLAUDE_CODE_VERSION=X.Y.Z open-terminal
+# When bumping the default here, keep the image rebuild + test cycle
+# together — the stream-json schema can change between versions.
+ARG CLAUDE_CODE_VERSION=2.1.114
+
 # Base image runs as uid=1000(user) with npm prefix /usr, which requires
 # root to write. Switch to root for the global install, then back so the
 # container's runtime user and entrypoint stay identical to upstream.
 USER root
-RUN npm install -g @anthropic-ai/claude-code@latest \
+RUN npm install -g "@anthropic-ai/claude-code@${CLAUDE_CODE_VERSION}" \
     && claude --version
+COPY cleanup.sh /opt/cleanup.sh
+RUN chmod +x /opt/cleanup.sh
 USER user
diff --git a/sandbox/README.md b/sandbox/README.md
index b66aabf..933e5c7 100644
--- a/sandbox/README.md
+++ b/sandbox/README.md
@@ -73,3 +73,33 @@ The renderer's `_tool_preview` / `_tool_input_block` helpers already work on the
 - **Image context:** when the user attaches images in the chat, `client.write_file()` them into the workdir before invoking claude, then reference by path in the prompt.
 - **Session resume:** store `claude_session_id` per `chat_id` in-process (same as the current pipe). Needs testing that `claude --resume` works cleanly across separate `POST /execute` calls — each call is a fresh process, but claude persists session state to `~/.claude/` inside the user's home.
 - **Cold start:** first request per user spawns `useradd`; measure and decide whether to pre-warm on Open WebUI login.
+
+## Operations
+
+### Pinning the Claude Code CLI version
+
+The Dockerfile pins via an ARG. Two clean builds produce identical `claude --version`:
+
+```sh
+docker compose build --build-arg CLAUDE_CODE_VERSION=2.1.120 open-terminal
+```
+
+Bump the default in the Dockerfile when you want the repo to track a new version.
+
+### Disk cleanup
+
+Nothing is auto-deleted. `/opt/cleanup.sh` is installed in the image for explicit runs:
+
+```sh
+# Dry-run (safe): see what would be deleted, nothing touched.
+docker compose exec \
+  -e CHAT_TTL_DAYS=30 -e SESSION_TTL_DAYS=90 -e CLEANUP_DRY_RUN=true \
+  open-terminal /opt/cleanup.sh
+
+# Execute:
+docker compose exec \
+  -e CHAT_TTL_DAYS=30 -e SESSION_TTL_DAYS=90 -e CLEANUP_DRY_RUN=false \
+  open-terminal /opt/cleanup.sh
+```
+
+Schedule nightly via host cron if desired. Defaults to dry-run to prevent surprise deletions.
diff --git a/sandbox/cleanup.sh b/sandbox/cleanup.sh
new file mode 100755
index 0000000..de069a7
--- /dev/null
+++ b/sandbox/cleanup.sh
@@ -0,0 +1,96 @@
+#!/bin/sh
+# Reclaim disk inside the open-terminal /home volume.
+#
+# Not installed into the image's startup — shipped as an explicit-action
+# script so nothing gets surprise-deleted. Run from the host via cron/timer,
+# or `docker compose exec open-terminal /opt/cleanup.sh`.
+#
+# Env vars (all default to "disabled" — no destructive defaults):
+#   CHAT_TTL_DAYS       delete ~/chat-<id>/ dirs whose newest-file mtime
+#                       exceeds this many days. 0 = skip.
+#   SESSION_TTL_DAYS    delete session JSONLs under
+#                       ~/*/.claude/projects/*/*.jsonl older than this
+#                       many days. 0 = skip. Note: we now put .claude
+#                       *inside* each chat-<id>/, so this is redundant
+#                       once CHAT_TTL_DAYS is in use — kept for pre-
+#                       migration homes that still have a flat ~/.claude.
+#   CLEANUP_DRY_RUN     if "true", print what would be deleted. Default true
+#                       the first time you run this — flip explicitly after
+#                       reviewing the output.
+#
+# Exit 0 on normal completion (including dry-run). Non-zero only on hard
+# errors (bad env, cannot read /home).
+
+set -eu
+
+CHAT_TTL_DAYS="${CHAT_TTL_DAYS:-0}"
+SESSION_TTL_DAYS="${SESSION_TTL_DAYS:-0}"
+DRY_RUN="${CLEANUP_DRY_RUN:-true}"
+
+log() { printf "[%s] %s\n" "$(date -u +%Y-%m-%dT%H:%M:%SZ)" "$*"; }
+
+if [ "$CHAT_TTL_DAYS" = 0 ] && [ "$SESSION_TTL_DAYS" = 0 ]; then
+    log "both TTLs are 0; nothing to do. Set CHAT_TTL_DAYS and/or SESSION_TTL_DAYS."
+    exit 0
+fi
+
+log "cleanup start — DRY_RUN=$DRY_RUN CHAT_TTL_DAYS=$CHAT_TTL_DAYS SESSION_TTL_DAYS=$SESSION_TTL_DAYS"
+
+chat_deleted=0
+chat_kept=0
+session_deleted=0
+
+# Iterate OWUI users. Skip root, distro accounts, etc.
+for home in /home/owui_*; do
+    [ -d "$home" ] || continue
+    user=$(basename "$home")
+
+    # --- chat-<id>/ dirs ---
+    if [ "$CHAT_TTL_DAYS" -gt 0 ]; then
+        # `find ... -type d` can't filter on newest-file-mtime cheaply;
+        # emulate with a per-dir check. One process per user, bounded by
+        # #chats (rarely >100).
+        for chatdir in "$home"/chat-*; do
+            [ -d "$chatdir" ] || continue
+            # Most recent mtime of any file *inside* (not the dir itself,
+            # which gets touched by subdir creation).
+            newest=$(find "$chatdir" -type f -printf '%T@\n' 2>/dev/null | sort -nr | head -1)
+            if [ -z "$newest" ]; then
+                # Empty chat dir. Treat as old.
+                newest=0
+            fi
+            now=$(date +%s)
+            age_days=$(awk "BEGIN { printf \"%d\", ($now - $newest) / 86400 }")
+            if [ "$age_days" -gt "$CHAT_TTL_DAYS" ]; then
+                if [ "$DRY_RUN" = "true" ]; then
+                    log "  would delete $chatdir (idle ${age_days}d)"
+                else
+                    rm -rf -- "$chatdir"
+                    log "  deleted $chatdir (idle ${age_days}d)"
+                fi
+                chat_deleted=$((chat_deleted + 1))
+            else
+                chat_kept=$((chat_kept + 1))
+            fi
+        done
+    fi
+
+    # --- legacy flat ~/.claude/projects/ ---
+    if [ "$SESSION_TTL_DAYS" -gt 0 ] && [ -d "$home/.claude/projects" ]; then
+        # Find JSONLs older than TTL; delete whole parent slug dirs if
+        # they're fully aged out.
+        while IFS= read -r jsonl; do
+            if [ "$DRY_RUN" = "true" ]; then
+                log "  would delete $jsonl"
+            else
+                rm -f -- "$jsonl"
+                log "  deleted $jsonl"
+            fi
+            session_deleted=$((session_deleted + 1))
+        done <<EOF
+$(find "$home/.claude/projects" -type f -name '*.jsonl' -mtime "+$SESSION_TTL_DAYS" 2>/dev/null)
+EOF
+    fi
+done
+
+log "cleanup done — chat_deleted=$chat_deleted chat_kept=$chat_kept session_deleted=$session_deleted"