v2026.06.19
v2026.06.19
Highlights
See Exactly What the Agent Looked At — Data Provenance & the Sources Panel
Every external data source the agent touches is now captured and laid out for you to inspect — automatically, and guaranteed on every turn. Because capture lives in the platform itself, the trace stays complete no matter how the model works: web searches, fetched pages, SEC filings, market-data pulls, your files, memos, memory reads, and the MCP calls the agent makes inside its sandbox all land in the record. Each completed answer gets a Sources (N) pill, and opening it brings up the new Sources panel with everything that turn pulled in — flip between this turn and all sources across the whole conversation.
Open any source for the full detail: a content fingerprint (checksum, size, and a snippet of what came back) so you can confirm the agent actually read what it claims, an Open link / Open file jump to the original, and clear attribution when a subagent fetched it. When the agent hit the same ticker several ways in one turn — overview, daily prices, options chain — those collapse into a tidy stack you can fan open.
It all updates live as the turn streams, survives a reload, and any secrets in tool arguments are redacted before they leave the server. Sources stay owner-only — they never appear in public shared chats.
Stop Is Now an Immediate Hard Cancel
We've retired the soft-interrupt mechanism — steering already covers everything it offered. The chat Stop button is now an immediate hard cancel: it stops the main agent and any running subagents, marks the turn with a ⏹ Stopped chip, and flushes a checkpoint so your next message resumes from the last completed step.
Changelog
New Features
- provenance: trace + surface the external data the agent accesses (#274) (9adc4ed)
- retire soft interrupt — make stop an immediate hard cancel (#273) (3156e83)
Bug Fixes
Improvements
2 features, 1 fix, 1 improvement — 4 commits total