v0.5.1

Features

• Coalesce queued messages into one turn (#67). When multiple DMs (or messages in a passive group) arrive while a turn is in flight, they merge into a single follow-up SDK turn instead of firing one turn per message. Lets the agent see "do X" → "wait" → "actually nevermind" together and skip wasted work on retracted requests. Mention-required groups bypass coalescing because per-message mention filtering would be lost. Passive group batches use sender-prefixed lines (Alice: ... / Bob: ...) plus a header noting the messages came from a group. Each user message still appends individually to the on-disk transcript; only the SDK prompt is merged.

Bug fixes

• Telegram NO_REPLY no longer leaks via the streaming first frame (#66). The streaming placeholder was sent before the model finished, so a NO_REPLY response left an empty/partial message in the chat. StreamingMessage now exposes cancel(), called in the NO_REPLY path (and the new batched path) after the SDK turn completes — cancels the placeholder before any visible content lands.
• Channel ingress was serializing against the SDK turn (#67). grammy's bot.on(...) awaits its handler body before reading the next update, and the iMessage dispatch did the same — both then awaited enqueueMessage's task-completion promise, so the channel layer waited a full SDK turn before delivering the next message. Result: messages never piled up in the queue, and the new coalescing was effectively dead in production. Three layers of fix (defense-in-depth): telegram.ts and imessage.ts dispatch fire handlers without awaiting; agent.enqueueMessage returns once queued, not when the turn completes. Regression test models the serial channel-loop pattern.