Add the 'Crosscheck Work' experiment to the Agents settings. When enabled,
a reviewer sub-agent is spawned after the main agent finishes a turn (with
no pending tool calls). The reviewer critiques the output using a dedicated
system prompt focused on correctness, simplicity, and code quality. If the
reviewer does not respond with 'LGTM!', its feedback is injected as a
synthetic user query back to the main agent, which must address it. This
loop continues until the reviewer approves or max iterations is reached.
Components:
- Feature flag: CrosscheckWork (enabled in DOGFOOD_FLAGS)
- Settings: agents.experiments.crosscheck_enabled,
agents.experiments.crosscheck_model_id,
agents.experiments.crosscheck_max_iterations
- Settings UI: new 'Experiments' subpage under Agents
- Crosscheck module: app/src/ai/crosscheck/ with prompt, reviewer model
- Controller integration: hooks into AfterStreamFinished when no actions
are queued, triggers reviewer, handles feedback injection
- Provider support: OpenAI-compatible and Bedrock direct invocation
- Safety: max iteration guard, error handling, reset on new user query
Major features:
- Auto-compact: triggers conversation summarization when context window >= 85%,
compacts Bedrock message history to a summary pair, and tracks live context tokens
- Bedrock summarization: plumbs `is_summarization` flag through translator/client/response
pipeline, handles SummarizeConversation input type, and marks `summarized` in metadata
- Session restore: rebuilds bedrock_message_history from persisted task messages via
newly-public `convert_proto_message`, preventing empty history on reconnect
- Subagent orchestration: adds SubagentQuestion/Answer/CompletionSummary event types,
parent-child question routing with depth limits, retry counting, and drain methods
- Summarization UI: inline SummarizationView in AI blocks with progress/finished states
Refactors:
- Rename WarpTheme → GalaxyTheme across ~100 files (rebrand continuation)
- Rename warp_home_config_dir → galaxy_home_config_dir and related path functions
- Predefined rules: replace "System Defined Rule #N" with descriptive names
(e.g. "Correctness Over Speed", "Never Guess") and add lookup helpers
- Usage view: replace cumulative input/output token display with live context tokens,
cache hit rate calculation, and separate cache read/write stats
- Telemetry: remove verbose doc comments, simplify trait definitions
- Facts view: simplify delete permission check (always allow local deletion)
- Remove warp_managed_paths_watcher.rs (dead code)
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Bedrock was calling suggest_next_prompt tool which created a SuggestPrompt
action that waited forever on a oneshot channel for UI interaction that
never fires in the Bedrock path, keeping the conversation permanently
InProgress. Fixed by filtering the tool from the Bedrock tool list and
skipping it at the stream level when the LLM calls it from context history.
Also includes: Bedrock cache token tracking, cost estimation, LSP
improvements, conversation usage view updates, and external config support.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Ryan Ward
rkw6086