v2.4.5-agent.1

New

• Automatic context compaction: when a conversation crosses a usage threshold you set, older turns are summarized so long agent runs keep going instead of dying at the context limit. Recent tool calls and their results are kept, and summaries are stored alongside the original messages rather than replacing them, so nothing you sent disappears from the transcript. Can also be triggered by hand.
• Sub-agents can run on a model you pick: name a model when dispatching and that sub-agent uses it, so bulk or cheap work does not have to run on your main model. An unknown or ambiguous name now fails the dispatch and lists the valid choices instead of quietly falling back to the parent's model.
• Telegram bot proxy: reach Telegram through a SOCKS5 or HTTP proxy where it is blocked, with optional credentials on HTTP proxies. Configure it in the bot's settings while the bot is stopped. Only the bot's own traffic is routed through it.
• Interrupted model streams are retried automatically with a backoff, on every provider. The attempt limit lives in model settings, and 0 turns it off.

Fixes

• Tool-heavy conversations on the Google Cloud Code Assist and Antigravity backend no longer fail with schema errors before reaching the model. Tool definitions from MCP servers are reduced to what the backend accepts, which also clears the same failure on Claude models served through it.
• Replies that never arrived and never failed: a failed stream could leave the request hanging open, showing an endless loading indicator with no error. Failures now surface as errors.
• Minutes of silence after sending a message: a stream carrying content the app did not recognize was mistaken for a dropped connection and retried again and again. Unrecognized content is now skipped, and retries are reserved for streams that genuinely delivered nothing.
• Thinking steps render on Cloud Code Assist instead of being dropped.
• Reasoning models served through Cloud Code Assist no longer reject requests whose thinking budget sat below their minimum.
• An unfinished reply and its in-progress tool calls survive switching conversations or restarting the app, instead of reverting to your last message.
• Broad reliability hardening across streaming persistence, sub-agent dispatch, and Telegram delivery.

Known limitations

• Automatic context compaction has not been verified on a device for this build. Manual compaction is available if automatic triggering misbehaves. A verification pass is planned.
• The Telegram proxy does not support SOCKS5 with a username and password, because those credentials cannot be scoped to the bot alone without affecting the rest of the app. Use an HTTP proxy with credentials, or an unauthenticated SOCKS5 proxy.

Thanks

• @WangWei-CM for automatic context compaction, the stream retry mechanism, and the long-running chat reliability work (#21).