refactor(sandbox): tidy up service, API, and test code; truncate shell output

[Simpleyyt]

Summary

Cleans up the sandbox service without changing its API contract (verified against the routes the backend's docker_sandbox.py calls), and adds output truncation to the shell API.

Shell output truncation (new)

• shell/view and console records now truncate output to max_length (default 10000 chars), keeping the most recent tail with a (truncated) prefix; shell/exec results get the same default
• New optional max_length field on the shell/view request, mirroring file/read
• Stored per-session output is capped at 1MB (with 64KB slack to avoid copying on every 128-byte append), discarding the oldest output — previously a chatty long-running process grew session memory without bound

app/services/shell.py

• Replace the untyped Dict[str, Dict[str, Any]] session store with a typed ShellSession dataclass
• Extract _get_session (deduplicates 5 copies of the "session not found" check) and _terminate_process (deduplicates graceful-terminate/force-kill logic used by both exec_command and kill_process)
• Remove dead code: unused ShellTask model, shell_tasks store, unused subprocess/Tuple imports, and an unused get_console_records call in exec_command
• Fix the console parameter being shadowed by a local variable in view_shell
• Straighten out exec_command control flow (single create-process path, no nested try/except swallowing) and guard the output reader so a replaced process can't write into the new command's output
• Compile the ANSI-escape regex once at module level

app/services/supervisor.py

• Connect to supervisord lazily (ServerProxy connects on first RPC call), so importing the app no longer requires a running supervisord
• Deduplicate activate_timeout / extend_timeout into a shared _schedule_shutdown
• Fix the import-time shutdown timer: it previously created an asyncio task on a never-running loop (silently doing nothing); now it uses get_running_loop() and falls back to threading.Timer when no loop is running
• Centralize timer cancellation in _cancel_timer, remove bare except: blocks, unify auto_expand/auto_extend naming

app/services/file.py

• Remove unused imports (subprocess, mimetypes, BinaryIO), rename misleading *_async inner functions (they are sync functions run via asyncio.to_thread), simplify ensure_file and exception re-raising

API layer

• Move TimeoutRequest out of the route file into app/schemas/supervisor.py
• Serialize all supervisor responses with model_dump() for consistency; add response_model=Response to the upload route
• Drop redundant empty-session-ID checks now handled by the service layer; use os.path.basename for the download filename

Misc

• main.py: remove duplicated log-level setup
• config.py: migrate deprecated class Config to SettingsConfigDict
• Tests: upload to /tmp instead of the mounted repo directory — this removes sandbox/resource/test_upload_unique.txt, which a previous test run had accidentally committed into the repo; clean up unused imports and reuse the client fixture

Testing

• ✅ cd sandbox && uv run pytest — 3 passed against the running dev sandbox container
• ✅ Manual curl verification of every endpoint against the dev container: shell/exec (fast + long-running + session reuse), shell/view (with console records), shell/wait (timeout 400), shell/write (interactive cat), shell/kill, file/read|write|replace|search|find|upload|download (incl. sudo read, invalid regex 400, missing file 404), supervisor/status, supervisor/timeout/activate|extend|cancel|status
• ✅ Truncation verified against the dev container: small output untouched; 200KB seq output returns 10011 chars starting with (truncated) and keeping the last line; max_length=50 honored on shell/view; console records truncated too; 5MB yes output stored at ~1MB in memory with the tail (DONE_MARKER) preserved
• ✅ Full-stack end-to-end GUI test (frontend + backend + mockserver + refactored sandbox): sent a task in the web UI at localhost:5173; the mock-LLM-driven agent flow completed — plan, notify message, web search, three file/writes, shell/exec (ls -a), browser navigation, completion with attachments. Sandbox access logs show all file/read, file/write, shell/exec, shell/view calls returning 200, and the files exist in the sandbox with the expected content.

e2e_agent_flow_shell_file_tools.mp4

Shell tool view showing ls -a output with the created files
Completed task with attachment cards

_{To show artifacts inline, enable in settings.}