clew-code 0.4.6 → 0.4.7

package/README.md

@@ -107,7 +107,7 @@ Provider adapters map query payloads to their corresponding external APIs. The m
 
 Clew Code uses several advanced subsystems to enable autonomous, safe, and context-optimized agent operations:
 
-*   **Durable SQLite-backed Memory:** Memory is stored in a local SQLite database with tables for `memories` (tracking importance, confidence, access counts, and type rankings) and `memory_timeline` (event tracking). Context is injected dynamically into system prompts using an `importance * recency * confidence` score. Durable facts are automatically extracted during context compaction and synced across peers. Global project guidance is maintained under `.clew/memory/` in `MEMORY.md` (project overview), `DECISIONS.md` (architectural decisions), and `TASTE.md` (style/formatting preferences).
+*   **Durable SQLite-backed Memory:** Memory is stored in a local SQLite database with tables for `memories` (tracking importance, confidence, access counts, and type rankings) and `memory_timeline` (event tracking). Context is injected dynamically into system prompts using an `importance * recency * confidence` score. Durable facts are automatically extracted during context compaction and synced across peers. Global project guidance is maintained under `.clew/memory/` in `MEMORY.md` (project overview), `DECISIONS.md` (architectural decisions), and `TASTE.md` (style/formatting preferences).
 *   **Context Compaction and Checkpoints:** When the prompt token window fills up, Clew Code runs a multi-pass context compaction loop that recursively compresses history into factual summaries. To prevent loss of active work during compaction, the supervisor captures progress snapshots as checkpoints at 20%, 45%, and 70% milestones, along with a `notes.md` scratchpad for recovery.
 *   **Autonomous Task Queue and Daemon Mode:** Features a persistent task queue that operates with lease-based concurrency limits (allowing a maximum of 3 concurrent active tasks), exponential backoff retries, and a dead-letter queue (DLQ). The CLI can run in a background Daemon Mode to consume the task queue, run scheduled natural-language cron jobs, and perform memory distillation while the terminal is closed.
 *   **Model Context Protocol (MCP):** A unified client interface supporting stdio, Server-Sent Events (SSE), and direct in-process WebSocket connection adapters. It dynamically fetches tool declarations and resources, extending the agent's capabilities with external server tools.
@@ -239,6 +239,7 @@ Type these slash commands directly inside the interactive REPL:
 *   `/memory init` - Initialize a clean SQLite memory database for the project.
 *   `/memory scan` - Scan the active codebase to build code hierarchy memory.
 *   `/memory dashboard` - View memory statistics, confidence scores, and facts.
+*   `/rule` - Show project-specific behavioral rules auto-observed by the model. `/rule off` to disable, `/rule on` to re-enable.
 *   `/skills` - List and manage active custom skills defined in `.clew/skills/`.
 
 ### Swarm & LAN Peers
@@ -273,6 +274,7 @@ The agent has access to the following built-in tools based on active permission
 *   `LSPTool` - Query local Language Server Protocol diagnostics.
 *   `ProcessDelegateTool` - Delegate prompts to external CLI models or Codex.
 *   `MemoryFeedbackTool` - Update project memories, decisions, and preferences.
+*   `ProjectRuleTool` - Save, list, or remove project-specific behavioral rules auto-observed from user patterns (stored in `.clew/rules.json`).
 
 ---