* refactor(memory): replace GLiNER model with heuristic extraction
- Removed GLiNER-related code and dependencies from the memory ingestion pipeline, transitioning to a heuristic-only extraction approach.
- Updated documentation and comments to reflect changes in extraction methods.
- Adjusted tests to ensure compatibility with the new heuristic extraction configuration.
- Bumped version of the tokenizers dependency and updated Cargo.lock accordingly.
* chore: apply formatting and tauri lockfile sync
* Enhance autocomplete functionality and logging. Increased debounce time for autocomplete suggestions and added minimum context character requirement. Improved inline suggestion handling with new cleanup logic for tab acceptance. Introduced a new logging option for autocomplete-only logs in CLI. Updated various components to support these changes, including sanitization and error handling in the autocomplete engine.
* Add autocomplete CLI adapter for improved argument handling
This commit introduces a new module, , which encapsulates the argument parsing and logging logic specific to the autocomplete namespace in the CLI. Key features include extraction of leading verbose flags, handling of the flag, and improved help message printing. The existing CLI command handling has been refactored to utilize this new adapter, enhancing code organization and maintainability.
* Refactor inline completion sanitization and enhance context handling
* feat(memory): add CLI support for memory commands
- Introduced a new `memory` subcommand in the CLI for memory ingestion, graph inspection, and debugging.
- Implemented various subcommands including `ingest`, `docs`, `graph`, `query`, and `namespaces` for comprehensive memory management.
- Updated the CLI entry point to route the `memory` command appropriately, enhancing the command-line interface functionality.
* refactor(memory_cli): streamline memory command ingestion and improve error handling
- Simplified the ingestion process by removing unnecessary workspace directory creation and embedding logic.
- Updated the ingestion function to utilize the `create_memory_client` for better client management.
- Changed the limit parameter type from `usize` to `u32` for consistency and improved error handling in command arguments.
- Enhanced logging for ingestion start to focus on model name only, removing redundant extraction mode information.
* refactor(memory): implement global memory client singleton for improved resource management
- Introduced a new `global.rs` module to manage a process-global memory client singleton, ensuring consistent access across subsystems.
- Updated `create_memory_client` to utilize the global client, enhancing memory management and reducing resource contention.
- Refactored various modules to replace local memory client instances with the global singleton, improving performance and reliability.
- Adjusted CLI and screen intelligence components to leverage the global memory client for document persistence and ingestion operations.
This refactor enhances the architecture by centralizing memory client management, leading to better resource utilization and simplified code structure.
* feat(memory): add put_doc_light for screen-intelligence, skip vectors/graph
Screen-intelligence captures are too frequent and ephemeral to justify
vector embedding and GLiNER graph extraction per frame. Adds a lightweight
storage path (put_doc_light) that persists the document row and markdown
file without chunking, embedding, or graph extraction.
Three-tier storage:
- put_doc_light: DB + markdown only (screen-intelligence)
- put_doc: DB + markdown + vectors + background graph (skill sync)
- ingest_doc: full synchronous pipeline (CLI, debugging)
* style: apply cargo fmt formatting
* fix: add missing window_id field in AppContext test helper
* fix(test): fall back to per-call MemoryClient when global not initialized
In tests with isolated OPENHUMAN_WORKSPACE, the process-global singleton
may not be initialized or may point at the wrong directory. Fall back to
creating a client from Config (which respects env vars) when the global
is not ready.
* fix(test): use per-call MemoryClient for screen-intelligence persistence
put_doc_light does no background work (no vectors, no graph), so a
per-call client created from Config is safe and avoids the global
singleton which may point at a different workspace in test suites.
