Commit Graph
7 Commits
Author SHA1 Message Date
Jon Wingender 0d26f630e6 fix(preview): hung-frame watchdog + require approval for guessed action calls
A runaway preview script (sync infinite loop, or a re-render loop
outpacing the bootstrap's own coalescing) had nothing detecting it -
the frame just spun. The bootstrap now heartbeats every second, and the
parent tears the iframe down if it goes _WATCHDOG_MS silent, whatever
the cause.

_coerce_tool_calls recovers a tool call guessed from `content` for
models with no native tool_calls field. That guess is weaker evidence
than the API's own structured field - a model can land on JSON shaped
like a call while only meaning to describe one - so an action tool
recovered this way now always requires approval, even under the
"allow" policy that lets a native tool_calls field run unattended.
2026-08-26 13:08:58 -05:00
Athena 952ef8a0c4 refactor(preview): simplify compile and validation paths
Load Sucrase only when a JSX/TSX preview is opened, remove the hand-written transform, and leave subjective render evaluation to the reader while retaining structural fence validation.
2026-08-26 03:37:29 -05:00
AthenaandCursor 656c14caf3 feat(preview): add sandboxed live code previews
Render validated HTML, SVG, JSX, and TSX fences locally while preserving tool context and preventing explanatory JSON from triggering actions.

Co-authored-by: Cursor <cursoragent@cursor.com>
2026-08-26 03:34:19 -05:00
AthenaandCursor 1f6beed0d2 feat(security): SSRF guard on fetch_url + single-use tool-approval tokens
Two tool/agent-layer hardening changes:

* fetch_url now resolves the target host and refuses to connect if any
  resolved address is loopback, private (RFC1918/ULA), link-local (incl. the
  169.254.169.254 cloud-metadata endpoint), multicast, reserved, or
  unspecified. IPv4-mapped IPv6 is unwrapped first, and the guard re-runs on
  every redirect hop so a public URL cannot 302 its way to an internal target.

* /chat/approve now requires a single-use token minted when the stream pauses
  for approval and delivered only in that stream's tool_request event, compared
  in constant time. Previously the pending approval was keyed solely on a
  client-supplied conversation_id, so anyone who could enumerate a
  conversation_id could approve another client's pending action.

The frontend threads the token from the tool_request event into the approve
call.

Co-authored-by: Cursor <cursoragent@cursor.com>
2026-08-07 09:41:09 -05:00
jonandClaude Opus 4.8 52b3c5c3f0 feat(tools): per-call approval for action tools
3-way action_tool_policy (off/ask/allow). In "ask", the chat stream stays
open and the tool loop awaits approval: emits event:tool_request, the UI
shows Approve/Deny, POST /chat/approve resumes the same stream. Declined
actions return a denied result; a timeout denies.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
2026-07-23 14:39:00 -05:00
jonandClaude Opus 4.8 f509034fdd feat: playbook tools, RAG, vision, voice, chat controls, faster boot
- Tool-using playbooks: read-only tool registry (search_memory,
  search_history, search_documents, list_models, get_time), per-playbook
  allowlist, agentic loop, and a "running tool" status indicator.
- Document ingest / RAG: documents table + chunker + embed/cosine retrieval
  reusing the existing stack; Documents page (upload/paste, viewer, delete);
  top-k chunks injected into the system prompt.
- Vision chat: attach images, base64 into /api/chat.
- Voice I/O: Web Speech dictation + read-aloud (browser-native, no backend).
- Chat controls: stop, regenerate, edit-and-resend; num_ctx knob in Settings.
- Per-playbook model override.
- History polish: per-conversation + bulk ShareGPT export.
- Faster ncp start: UI up first, Ollama warms in the background, with a
  per-phase timing readout.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
2026-07-23 13:31:48 -05:00
Jon Wingender 714b9fc890 Initial commit: NexusOS - local AI assistant platform 2026-07-21 22:48:38 -05:00