Problem

The claude adapter fork (packages/agent/src/adapters/claude) was last synced with upstream claude-agent-acp at v0.44.0. Upstream is now at v0.54.1 with a major ACP SDK release (1.1.0), a rearchitected prompt loop that fixes turns that never end while background tasks run, and a batch of streaming, permission and model fixes we want.

Closes #1919
Closes #2414

Changes

SDK bumps: @anthropic-ai/claude-agent-sdk 0.3.170 to 0.3.197, @agentclientprotocol/sdk 0.25.0 to 1.1.0, @anthropic-ai/sdk 0.104.1 to 0.109.0. The ACP major is source-compatible for us: the deprecated AgentSideConnection/ClientSideConnection classes still ship and still route the extMethod/extNotification surface our _posthog/* extensions use, so the agent() builder migration is recorded as an intentional divergence instead of ported.

Ported from upstream (details and commit refs in UPSTREAM.md):

• Persistent consumer + turn queue (feat: slimmed down focus ui and move default terminal position #780): prompt() now enqueues a Turn deferred and a single long-lived consumer drains the SDK stream for the session's life. Turns settle at their terminal result instead of the trailing idle, between-turn and background output streams live, and a dead query stream rejects new prompts with a clear "session has ended" error. Adapted to the fork's single session, steer mode, interruptReason, broadcast timing, unsupported-slash-command gate and refreshSession() (new queryGeneration guard). This is what closes monitors not working #1919 and handle ScheduleWakeUp tool call properly #2414: wake-up output (background tasks, Monitor, ScheduleWakeup) now streams in live instead of sitting unread until the next user message.
• Content-based streamed-text dedupe (feat: add spinner for tasks loading #785/feat(signals): Display proactive tasks (signal reports) as suggested tasks #789/Instead of "no content" start a new shell session in new tabs when using split #800, ported at the final Instead of "no content" start a new shell session in new tabs when using split #800 state): the consolidated assistant message prefix-diffs each text/thinking block against what actually streamed and forwards only the remainder, so non-streaming gateways keep their answers and cut-short streams recover their tails.
• Skip empty (signature-only) thinking chunks (If there’s an error view in the session view, don’t display scroll to bottom please. #793).
• Emit the tool_call before its permission request (Ripgrep missing in codex #820) via a shared emittedToolCalls set; whichever of the stream and the permission flow runs second refines with a tool_call_update.
• Permission requests carry a cancellationSignal (fix: o(n) requests on project switch + perisstance #801) so cancelling a turn dismisses the open dialog instead of leaving the await hanging.
• Bash tool fixes: errors keep rendering as terminal output (The session view won't keep auto scrolling to the bottom as the agent works #776) and image output surfaces as ACP content instead of being dropped (Slow workspace creation takes user to blank screen for a long time #617).
• Surface the SDK's informational system notices (fix: Fix auto-scroll race condition during agent streaming #777) so hook-blocked stops are visible.
• Sonnet 5 model-version matching (Individual file search #826): single-number family versions resolve and [1m] context hints no longer read as versions.
• Session title pushes at turn end (Ugly Codex bash command rendering #812) via session_info_update.
• Fast mode config option (Trim workspace/repo path in permissions, tool calls etc #828), adapted to gateway models (Opus 4.8/4.7): on/off select wired to applyFlagSettings({ fastMode }), intent retained across model switches and reconciled with SDK-reported state.

Skipped with reasons logged in UPSTREAM.md: elicitation fixes, ACP logout, version flag, agent persona dropdown and the availableModels/1M-inference fixes that only apply to upstream's SDK-settings model pipeline.

Also adds fork-side regression tests (claude-agent.wake-up.test.ts) locking in the wake-up behavior: messages forward live after a settled turn, wake-up streaming dedupes, a task-notification result cannot settle the next queued turn and a normal turn runs after wake-up output.

How did you test this?

• pnpm --filter agent typecheck and pnpm --filter agent build
• pnpm typecheck (whole repo)
• pnpm --filter agent test: 834 passed, including new unit tests for the content-based dedupe (tail forwarding, id-mismatch dedupe, residue clearing, empty-delta stalls), the consumer cancel paths (queued-turn settle, orphan accounting, force-cancel backstop) and Sonnet 5 model resolution
• pnpm --filter code test: 179 passed
• biome check --write on all changed files
• Wake-up regression suite (4 tests) plus live repros of both issues against the real SDK on this branch: monitors not working #1919 (background sleep 10, turn ends with end_turn, "the test succeeded" streams in live 9s later) and handle ScheduleWakeUp tool call properly #2414 (ScheduleWakeup 60s out, turn ends, the wake-up turn streams in live when the timer fires), both with no follow-up prompt

Automatic notifications

• Publish to changelog?
• Alert Sales and Marketing teams?