v0.0.464

Version 0.0.464 (July 17, 2026)

• Rolling context compaction for Agent v2 (#689) — long conversations no longer hit the context ceiling: the agent compacts older turns into a rolling summary automatically when the window fills (and on demand), keeps answering past the watermark, and the report chat shows a subtle "compacted" divider where the fold happened. Compaction is visible live via a context.compacted SSE and the context-usage estimate refreshes immediately.
• Queue and steer prompts during a run (#690) — typing while the agent is working no longer means waiting: Queue holds the prompt (shown as removable chips under the thinking indicator) and runs it when the current completion finishes, while Steer injects it into the running completion at the next observation point — hard-interrupting in-flight planning — with a visible acknowledgment on the message once the agent picks it up. Queued prompts stay out of the model's conversation window until they actually run.
• Multi-pick clarifying questions (#693) — the clarify tool supports select-all-that-apply: the agent can ask one question with multiple selectable options, the answer card renders checkboxes with a single confirm, and the selections rehydrate correctly on page refresh.
• Conversation cost and tokens in the trace view (#694) — TraceModal's header now shows the conversation's total LLM tokens and cost, and each turn's summary strip shows per-turn planner tokens — with full token and cost breakdowns sourced from usage events when licensed.
• Fix empty review for NEW instructions pending approval (#695) — reviewing a newly suggested instruction no longer opens an empty diff: the review payload for instructions that never had a published version now carries the proposed content instead of nothing.
• Anti-overfit guard for learned instructions (#696) — create_instruction/edit_instruction pass through a generality gate that rejects one-off, prompt-parroting rules before they pollute the instruction base, backed by an objective bait/control overfitting benchmark with a deterministic DB-based scorer (provider errors count as failed trials, not behavioral outcomes).
• Smarter instruction loading with an on-demand catalog (#697) — intelligent instructions are matched by coverage-based scoring (light stemming, title/label/table-name boosts) instead of brittle exact-word overlap, zero-score instructions fill remaining capacity instead of silently vanishing, and over-capacity ones become a compact catalog the planner can pull from via a new read_instruction tool (scoped to the report's data sources and the user's table access); search_instructions gains a compact chat mode.
• Collapsible table-search results in chat — the "Searched …" table results in the conversation now collapse to a single line by default, with each found table shown in the header with its data source icon (deduped, "+N more" past six); clicking the line expands the detailed per-table list as before.