public/galaxy
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Bump version to 1.6.3

Browse Source 
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
This commit is contained in:
Ryan Ward
2026-06-12 14:17:06 -05:00
co-authored by Claude Opus 4.6
parent 4ba9706e35
commit 59cfd0e2f5
152 changed files with 8276 additions and 1664 deletions
Download Patch File Download Diff File Expand all files Collapse all files
GALAXY.md
+34 -2
View File
@@ -68,6 +68,18 @@ Global rules are loaded from `~/.galaxy-ai/rules/*.md` (filename = rule name, co
Project rules are loaded from `GALAXY.md` or `AGENTS.md` files found in the project directory tree.
**Model Discovery** (`app/src/ai/bedrock/discovery.rs`):
At startup (and on manual refresh from the Bedrock settings page), Galaxy discovers available models using the user-selected auth settings (profile/SSO/static keys — no external config fallback) via:
1. **STS GetCallerIdentity** — validates AWS credentials before proceeding
2. **ListInferenceProfiles** (system-defined) + **ListFoundationModels** (TEXT output, ON_DEMAND) — fetched in parallel; results are deduplicated by underlying foundation model ID with inference profiles taking priority (they include cross-region routing)
3. Filters out legacy/deprecated models and non-LLM models (embedding, image gen, etc.)
4. **Invoke probe** — each candidate gets a minimal `Converse` call (max_tokens=1); only a successful probe (or a transient throttling/capacity error) counts as accessible. Validation errors (no Converse/on-demand support), access denials, and missing resources are rejected to avoid listing unusable models
5. Tags Claude 4.6+ models with `[1m]` suffix to signal 1M context window support
The settings page refresh (`RefreshAwsBedrock` in `app/src/settings_view/ai_page.rs`) clears the cached model list immediately, shows a shimmer progress indicator while discovery runs, and repopulates the list when it completes.
The `[1m]` suffix is an internal marker stripped by `strip_context_marker()` in `client.rs` before API calls. It's used by `context_window_for_model()` in `response_translator.rs` to report the correct context window size.
**External Config Fallback** (`app/src/ai/bedrock/external_config.rs`):
When Galaxy's own Bedrock settings are at defaults, it falls back to configurations from:
1. **Claude Code** (`~/.claude/settings.json`) — reads `env.AWS_PROFILE`, `env.AWS_REGION`, and `env.DCP_MODEL_MAP` (ARN-based model mappings)
@@ -75,11 +87,31 @@ When Galaxy's own Bedrock settings are at defaults, it falls back to configurati
Priority: Galaxy explicit settings > Claude Code > OpenCode > hardcoded defaults. Fallback only applies when profile is `"default"` (for profile) or empty (for region/models). External model ARNs are merged with Galaxy's built-in default model list.
**Token Usage & Cost Tracking** (`app/src/ai/bedrock/stream.rs`):
The Bedrock stream extracts full token metadata from responses: `input_tokens`, `output_tokens`, `cache_read_input_tokens`, `cache_write_input_tokens`. These flow through `build_stream_finished` → `TokenUsage` struct → `conversation.update_cost_and_usage_for_request()`. Cost is estimated using Sonnet-tier Bedrock pricing as a conservative default. Displayed in:
**Token Usage & Cost Tracking** (`app/src/ai/bedrock/response_translator.rs`):
The Bedrock stream extracts full token metadata from responses: `input_tokens`, `output_tokens`, `cache_read_input_tokens`, `cache_write_input_tokens`. These flow through `build_stream_finished` → `TokenUsage` struct → `conversation.update_cost_and_usage_for_request()`. Cost is estimated per-model using Bedrock pricing. Displayed in:
- **Agent management cards** — total token count in metadata row
- **Conversation usage footer** — full breakdown (input/output/cache read/cache write) + estimated cost
**Progressive Summarization** (`app/src/ai/blocklist/controller.rs`):
When context window usage reaches 85%, Galaxy automatically summarizes older messages while keeping the last 100 `ConversationMessage` entries verbatim. This replaces the old `/compact` hard-compact that destroyed all context. The summary is prepended to the messages array (not the system prompt) as a `<conversation-history-summary>` block. Key details:
- Trigger: automatic at 85% context usage, >100 messages in history
- Uses Sonnet via `BedrockClient::converse_collect` in a background `ctx.spawn`
- Cost attributed to conversation totals transparently
- No UI shown — user only sees context usage drop
- `recall_tool_history` tool lets the agent retrieve past tool outputs that were summarized away
**Failed Tool Call Visibility** (`app/src/ai/bedrock/response_translator.rs`):
When the model calls an unknown/hallucinated tool name, the response translator now emits a visible `AgentOutput` text message to the UI (via `build_add_agent_output_message`) showing what tool was attempted and the error. Previously, synthetic error results were only stored in history (for Bedrock message ordering) but never rendered.
**Loop Prevention Guardrail** (`app/src/ai/blocklist/controller.rs`):
Detects when the agent gets stuck in a recursive failure loop (same tool + same input failing repeatedly) and injects a corrective system instruction to break the cycle. Key details:
- Tracks last 10 failed action results per conversation via `LoopDetectionState`
- Uses tool discriminant + Display hash to identify repeated patterns
- Threshold: 3 identical failures triggers intervention
- Intervention: injects a `[SYSTEM] Loop detected` user message instructing the model to take a different approach
- Clears on any successful action result (progress = not looping)
- Resets when user sends a new query
### Key Architectural Patterns
1. **Entity-Handle System**: Views reference other views via handles, not direct ownership