Unreleased changes

• Fixed MCP servers staying unavailable in project sessions after Test connection reported success. Testing now uses the selected project's working directory and clears that server's idle failed workspace connections without interrupting active tool calls or handshakes.

• Fixed the composer context ring staying at the pre-compaction figure after /compact or an automatic compaction until another turn ran (#2015). The ring now shows the compacted context's estimated size as soon as the checkpoint is installed, and after reopening the session.

• Fixed prompt cache misses after a conversation's agent runtime is rebuilt (#1972). Rebuilding happens after a restart or when the thinking level, mode, project rules, or model settings change. Assistant replies now keep the provider's thinking, reasoning, and Gemini function-call signatures in the transcript, so Claude, OpenAI Responses, Gemini, and other reasoning models receive the same history as before the rebuild instead of a rewritten one.

• Fixed duplicate /、 when opening slash commands with a Chinese IME in an empty composer. Keep pre-edit text intact and apply the slash alias only after composition commits, preserving recovery from a missing end event.

• Add OpenAI Fast mode (#1657). Allow it per model in Settings → Models for a ChatGPT subscription or an OpenAI Responses connection, then type /fast to turn it on for the current conversation. Requests ask for the priority service tier, which replies faster and spends plan limits faster or bills more per token; the Composer model chip shows a Fast mark while it is on.

• Fixed an older regenerated prompt appearing after the latest reply when its branch archive completed outside the loaded history window (#1870). Pager acknowledgements now update existing rows without appending historical roots.

• Generate inline visualizations faster. The visualization shell now preloads the bundled D3 and a viz chart helper layer (container sizing, resize and theme redraw, axes, tick overlap, legends, and a declarative viz.render for bar, line, area, point, heatmap, and donut charts), so model-written fragments are shorter. The visualize skill asks for a single write followed by the reply instead of reading the file back. Fragments written under the earlier skill keep rendering. When a fragment's script fails, the card now says so and shows the error instead of staying blank, and the model's write result flags syntax errors and missing chart containers so it can fix them with one edit.

• Fixed cramped inline visualizations. The visualization shell now places a fragment in the kit's #widget container, which spaces its top-level blocks 12px apart, and the visualize skill asks for the title, controls, chart, legend, and captions as top-level blocks instead of one wrapping element. Inline controls, block styling, and roots that lay themselves out keep their geometry; cards written earlier inside one wrapping element render as before.