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
- Sanitize tool_use_id values in OpenAI request conversion to match
Bedrock's required pattern ^[a-zA-Z0-9_-]+$. Fixes 400 errors when
LiteLLM proxies to Bedrock and tool IDs contain invalid characters.
- Parse cache usage stats from LiteLLM/OpenAI responses
(prompt_tokens_details.cached_tokens, cache_read_input_tokens,
cache_creation_input_tokens) and propagate to token usage tracking.
- Hide cache-o-meter in session status bar when provider doesn't report
cache data (LiteLLM/OpenAI), instead of showing misleading 0% stats.
- Update cost estimation to account for cache read/write pricing tiers.
typescript-language-server requires a valid TypeScript installation to
function. Previously, no initializationOptions were sent during the LSP
initialize request, causing the server to fail with:
'Could not find a valid TypeScript installation. Please ensure that
the typescript dependency is installed in the workspace or that a
valid tsserver.path is specified.'
This fix:
- Adds initializationOptions.tsserver.path resolution that searches for
TypeScript in: workspace node_modules, global npm install, and npx cache
- Wires initialization_options into the LSP startup flow via LSPServerType
- Updates the install step to proactively install TypeScript globally if
not found locally
When a skill was invoked on a clean slate (no conversation history),
the request sent to the AI provider contained only a system prompt
with zero user messages, causing Bedrock/LiteLLM to reject it with:
'Invalid Message bedrock requires at least one non-system message'
Root cause: extract_new_input_messages() had no handler for
Type::InvokeSkill, so it fell through to the _ => {} wildcard and
produced zero ConversationMessage results.
Fix adds InvokeSkill handling to three functions:
- extract_new_input_messages(): generates a User message with the
skill name and content wrapped in <skill-instructions> tags
- extract_input_messages(): persists the InvokeSkill as a proper
api::Message in task history for conversation continuity
- extract_user_query_text(): returns '/<skill-name>' for the
UserQuery proto event used in conversation title generation
- Load global rules (AIFact/AIMemory) from local CloudModel and inject
them into the Bedrock/OpenAI system prompt as a '## Global Rules' section
when memory is enabled.
- Fix rule seeding: always re-seed predefined rules when the CloudModel has
none, regardless of the has_seeded_predefined_rules flag (handles case
where flag was set but rules never persisted due to prior missing owner).
- Rename /context slash command to /copy-context: dumps the full context
window (global rules, progressive summary, message history) to the
clipboard for debugging.
Previously, predefined rules (11 system-defined behavioral rules) were
only seeded when the user navigated to the Rules settings page, and
required a cloud-authenticated owner to create cloud objects.
Since Galaxy operates without Warp cloud auth:
1. personal_drive() now returns a synthetic local owner
('local-galaxy-user') when not authenticated, allowing cloud objects
(rules, etc.) to be created and stored locally in SQLite.
2. seed_predefined_rules_if_needed() runs at app startup (after
CloudModel and UpdateManager are initialized) to seed rules on
first launch without requiring the user to visit settings.
Rules are now visible in Galaxy Drive > Rules on first launch.
The previous implementation only used the standard OpenAI /models
endpoint which often lacks context window and capability metadata,
causing a blind 200K fallback for all models.
Now fetch_openai_models_from_endpoint tries LiteLLM's /model/info
endpoint first, which returns rich metadata:
- max_input_tokens (e.g. 1,000,000 for Sonnet 4.6)
- max_output_tokens (e.g. 128,000 for max models)
- supports_vision
- supports_function_calling
- underlying model path (for provider detection)
Falls back to /models if /model/info is unavailable (e.g. non-LiteLLM
OpenAI-compatible endpoints).
This ensures the model picker and context window configuration reflect
the actual capabilities of the configured models.
Galaxy should never call Warp's cloud API for AI. This is a policy
requirement. All model availability is now determined exclusively by
locally configured providers (Bedrock and/or OpenAI/LiteLLM).
Changes:
- Disable refresh_authed_models, refresh_public_models, refresh_available_models
(now no-ops with debug log)
- Disable on_server_update and update_feature_model_choices
- Disable get_cached_models (no stale server models restored from cache)
- Replace ModelsByFeature::default() with minimal placeholder that gets
stripped by inject_bedrock_models/inject_openai_models
- Make inject_bedrock_models strip Unknown placeholders unconditionally
- Make default_llm_info() return a static fallback instead of panicking
when no models are configured (prevents null reference crashes)
- Add has_any_provider_models() for UI to check provider availability
- Make ProviderConfig::None return a user-friendly error instead of
calling Warp's cloud API (the previous fallback behavior)
Safety: if no providers are enabled, the system gracefully returns an
error message rather than crashing or silently calling Warp's servers.
When LiteLLM routes OpenAI-format requests to AWS Bedrock, the provider
rejects conversations ending with an assistant message with:
'This model does not support assistant message prefill.'
Add ensure_ends_with_user_message() as the final sanitization step in
sanitize_messages_for_openai(). If the conversation ends with an assistant
message, a minimal 'Continue.' user message is appended.
Updated tests to reflect the new behavior and added dedicated tests for
the new function.
Rewrote the skill from a single-feature porting tool into a full upstream
merge workflow that:
- Fetches and merges latest warp/master
- Resolves conflicts preserving Galaxy's AI providers (Bedrock/OpenAI-LiteLLM)
- Strips any Warp API/auth/telemetry additions
- Iteratively repairs the build until cargo build succeeds
- Runs clippy/format/leak checks before committing
- Update version from 1.6.3 to 2.0.0 in app/Cargo.toml and Cargo.lock
- Add install-galaxy.sh upload step to build-and-deploy-hermes script
- Include pending AI provider and agent changes
Keychain profile approach doesn't work from Galaxy terminal (no interactive
auth entitlement). Switched to inline Apple ID/team/password in notarize.sh
and package.sh. Added migration-docs/ for session continuity on new machines.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
- Construct ServerTask directly for Bedrock CLI subagents so messages route
correctly without needing a server CreateTask upgrade
- Handle CliAgentUserQuery input type in both request translators, including
running command context and terminal output
- Add force_cancel_all_streaming_exchanges fallback for when Ctrl+C finds no
in-flight streams (stuck subagent / unexpected stream end)
- Cancel active conversation on Ctrl+C in agent view compose state
- Skip agent view entry when agent is tagged-in for a running command
- Set root_task_id on Bedrock requests for proper optimistic task upgrade
- Persist app state on will_terminate to avoid losing sessions
- Trust persisted CWD without is_dir() recheck (fixes network mount restore)
- Log warning instead of silently dropping tabs with unreadable root nodes
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Remove duplicate popup rendering from render_vertical_tabs_panel.
The popup was rendered both inside the panel's stack AND at the
workspace level in a Dismiss overlay, causing event dispatch conflicts
due to shared MouseStateHandle instances between the two identical
popup trees.
- Fix start_agent tool mapping: route to Tool::StartAgent instead of
dead-end Tool::Subagent so tool calls become executable actions
- Block start_agent until child finishes: parent waits for child
conversation to complete and receives full output as tool result
- Fix tool result delivery: add StartAgent/StartAgentV2 cases to
extract_tool_result_content so the model actually sees agent output
- Support parallel agent spawning: change StartAgent action phase from
Serial to Parallel, and track multiple pending agents via Vec
- Mark child conversations as Success on EndTurn: emit
ConversationStatus::Success when a child stream ends with no actions
- Skip orchestration SSE in local mode: prevent app freeze from trying
to connect to non-existent server
- Remove send_message_to_agent and suggest_next_prompt from tool list:
these require server infrastructure that doesn't exist in OSS mode
- Add "Waiting for sub-agents..." status message while agents process
- Fix child agent pane close: actually dismiss instead of re-hiding,
track dismissed IDs to prevent re-creation on restart
- Fix cache_miss_tokens calculation and show per-block cache stats
- Add [tool-debug] logging throughout tool invocation pipeline
Bump version to 1.6.0.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Two changes addressing agent context visibility and error recovery:
1. Enable DOGFOOD_FLAGS in oss.rs — this activates AgentViewBlockContext,
which auto-attaches user-executed command output (stdout/stderr/exit code)
to the AI conversation context. Previously the LLM was blind to any
commands the user ran via `!` prefix in agent view.
2. Add instant "Fix this error" prompt suggestion when a user-run command
fails (non-zero exit) in agent view. This lets users hand off errors to
the agent with one click instead of manually copy-pasting output.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Move token/cache/cost metrics from the expandable per-block usage button
into a persistent status bar rendered in the agent input area. This gives
always-visible feedback without requiring user interaction.
- Add render_session_status_bar to agent input showing context %, cache
hit rate, and cost
- Persist cache/cost totals in ConversationUsageMetadata for session restore
- Remove render_usage_button, ToggleIsUsageFooterExpanded action, and
UsageFooterToggled event
- Remove "Request tokens: --" debug block footer overlay
- Fix has_footer() to return false so blocks no longer reserve phantom
footer space (root cause of padding/margin issue)
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Module is scaffolded but not yet wired into callers — allow dead code
at the module level until integration is complete.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Resolves merge conflicts integrating compact_summary parameter into
build_converse_request alongside the existing CachingConfig approach.
Adds the new prompt_builder module for structured prompt construction.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Ryan Ward
rkw6086