Changelog — 2026-06-12

Features

• Platform telemetry adapter (PostHog). The new @stndrds/adapter-posthog package wires PostHog as an opt-in sink for agent, platform, and query metrics plus exception capture. Call initStandardsTelemetry(...) once at boot; tenant ids are hashed with a deployment-local salt before they leave the install, and STANDARDS_TELEMETRY=off keeps the client from ever being constructed. See adapters.
• Agent output limits and large-content file references. Each agent response now has a default output-token budget (32 000 tokens, clamped to the model's limit). To write content larger than a few KB, an agent saves it to a sandbox file and passes { "$fromSandboxFile": "/workspace/<file>" } to create_record / update_record, and the server substitutes the file content — bulk tools excepted. See agents.
• Structured tool error codes. A tool part cut short by the output limit or an interruption now carries a typed errorCode (output_limit_reached / interrupted) so the chat UI renders a localized explanation instead of a raw error string.
• Full-page form fill view. The form fill experience moved from a modal with cards to a full-page overlay with a section-navigation sidebar, scroll-spy, instance rows with status and progress, and a submit guard. Workflows you render with FormFlow pick this up automatically. See workflows.
• CLI autofill backfill. The @stndrds/cli gains an autofill backfill command to dry-run estimate (default) or enqueue an AI autofill backfill for one or more objects via --objects, --max-records, and --confirm.

Fixes

• Human-in-the-loop questions no longer get stuck. The end-of-stream reconciliation that closes dangling tool calls now exempts ask_questions, so the question widget appears, the session reaches waiting_human, and the chain-of-thought step flips to "Questions answered" the moment the user replies. A "Skip all questions" action lets the agent resume even when nothing is answered.