agentic-cli 0.4.3__tar.gz → 0.5.0__tar.gz

{agentic_cli-0.4.3 → agentic_cli-0.5.0}/.gitignore

@@ -48,6 +48,7 @@ plans/
 docs/
 changes/
 findings/
+demo_dir/
 
 # Git worktrees
 .worktrees/

{agentic_cli-0.4.3 → agentic_cli-0.5.0}/CHANGELOG.md

@@ -5,6 +5,81 @@ All notable changes to this project will be documented in this file.
 The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.1.0/),
 and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
 
+## [0.5.0] - 2026-04-17
+
+### Added
+- **Knowledge Base concept pages** (PR #70): Agent-authored concept/summary pages via `kb_write_concept` and `kb_search_concepts`. Title-weighted case-insensitive search, slug-based addressing, merge-on-overwrite, UTC timestamps.
+- **Hybrid BM25 + vector KB search** (PR #69): `kb_search` fuses BM25 and vector results with Reciprocal Rank Fusion. New `create_bm25_index` abstraction, structure-aware document chunking.
+- **Per-document KB sidecars**: Markdown sidecars written at ingest, surfaced via async `kb_read`. Lazy-regenerated on first read if missing. Append-only `ingest_log.md` audit trail and auto-maintained `index.md`. New `backfill_sidecars()` with per-document locking and in-progress guard.
+- **KB tool bundles**: `KB_READER_TOOLS` and `KB_WRITER_TOOLS` for convenient agent wiring.
+- **MemoryStore lifecycle** (PR #66): Embedding-backed semantic search, contradiction detection (`store_with_similarity_check`), `ForgettingPolicy` + `apply_forgetting()`, archive filtering, `update_memory` and `delete_memory` tools.
+- **Tool Reflection Store**: Bounded per-tool heuristic memory (`save_reflection`) learned from failures, wired via session-end hook.
+- **OS-native sandboxing** (PR #63): macOS seatbelt and Linux namespace sandboxes for shell and Python execution.
+- **Dual thinking boxes** (PR #62): Persistent task-progress box alongside the ephemeral LLM-events box; task state persists across turns. New status icons (`[✓]`, `[▸]`).
+- **ADK `ConfirmationPlugin` + LangGraph tool wrapper**: Framework-level HITL for `DANGEROUS` tools.
+- **ADK `LLMLoggingPlugin`** replaces `LLMEventLogger` for structured LLM traffic logging. **`TaskProgressPlugin`** emits task progress events.
+- **Dual-backend state tools**: Native planning/task tools for ADK (`tools/adk/state_tools.py`) and LangGraph (`tools/langgraph/state_tools.py`) built on shared `tools/_core/` logic.
+- **`research_demo` is now a pip-installable console script**; new `/kb-backfill` demo command.
+
+### Changed
+- **KB tool naming**: `search_knowledge_base` → `kb_search`, `ingest_to_knowledge_base` → `kb_ingest`. Added `kb_read`, `kb_list`.
+- **Service registry** (PR #65): Replaced 8 per-concern ContextVars with a unified service registry; tools reach KB, memory, sandbox, reflection stores, etc. through it. Self-contained, backend-specific tool factories.
+- **ArXiv flow** (PR #67, #68): `ArxivSearchSource` is service-registered, async-safe, and handles PDF download + ID parsing. ArXiv ingestion moved to `arxiv_tools.ingest_arxiv_paper`. Search results expose `pdf_url`/`abs_url`/`src_url`. Id-indexed entry cache.
+- **Workflow manager simplification** (PR #64): Private internals, removed public `graph` property, inlined native `part.thought` check, simplified initialization.
+- **KB ingestion** uses async LLM summarizer; densified sidecar prompt with higher output budget; `embedding_device` autodetection.
+- **`persistence/_utils.py` → `file_utils.py`** at the package top level.
+- **Cross-KB search** uses RRF fusion; `PlanStore` decoupled from task progress display.
+
+### Removed
+- **`@requires` decorator** — replaced by service-registry detection.
+- **`open_document` tool** (CLI-app concern, not framework) and **`unified_search` tool** (buggy, redundant).
+- **`ArtifactManager`** and activity-logging dead code.
+- **`context.py`**, **`task_progress.py` module**, legacy `planning_tools.py` and `task_tools.py` — state tools auto-inject per backend.
+- **`ThinkingDetector`** (inlined into native `part.thought` check).
+- **Old HITL infrastructure** — superseded by `ConfirmationPlugin` and the LangGraph tool wrapper.
+
+### Fixed
+- Concept timestamps use UTC with trailing `Z`; `kb_search` rejects whitespace-only queries.
+- `kb_read` PDF lazy-gen fallback; drops frontmatter; adds scope.
+- `KnowledgeBase.clear()` removes sidecars and rebuilds index; `delete` drops lock entry.
+- KB audit log uses UTF-8 + UTC timestamps.
+- Tags sentinel handling, `load_all` filter, re-embedding on update, contradiction-detection edge cases in `MemoryStore`.
+- Eager detection for missing ML deps; `Content-Type` PDF detection in `web_fetch`.
+- Final task-progress emission at end of event stream; full-color styling for the task box.
+- Research demo: stale prompts, bundle/tool-name alignment, path-placeholder escaping, service-registry push in `/kb-backfill`.
+
+## [0.4.4] - 2026-03-11
+
+### Added
+- **Session Save/Resume**: Persistent conversations across CLI exits with `/save` and `/resume`
+- **Sandbox Executor**: Stateful multi-turn Python code execution with sandboxing and default deps
+- **Context Window Management**: Native ADK and LangGraph context trimming with source-level detection
+- **Context Window Visibility**: Trim detection, status bar token display, and `/status` breakdown
+- **Dynamic Model Registry**: Replace static model lists with live API discovery from Google/Anthropic
+- **Task Progress Display**: Rich-colored task progress in thinking box, persisted across turns
+
+### Changed
+- **Architecture**: Separated workflow+tools layer from UI/CLI layer; composable settings mixins
+- **DRY/SOLID Cleanup** (PR #59, #60): Deduplicated config paths, ArXiv parsing, session persistence, artifact loading; extracted `drain_trim_events()`, `format_detail_rows()`, `format_task_checklist()`; removed dead code (`get_help()`, `FinanceResearchState`, unused settings introspection); replaced dynamic `type()` with `_SessionEvent` class; moved token-drop heuristic into `UsageTracker`
+- **MessageProcessor**: Removed dead code, unused params, cleaner state init
+- **HITL**: Removed dead Future-based machinery, simplified callback path
+- **Ripgrep**: Cache availability check with `lru_cache`
+
+### Fixed
+- Executor import validation blocking allowed submodules
+- Settings not persisting (replaced `exclude_defaults` with explicit exclusion)
+- `verbose_thinking` not persisting across restarts
+- Thread safety, path traversal, and TOCTOU race conditions in persistence
+- VectorStore crash safety: write mappings before FAISS index
+- Embeddings: update in-memory state after `ingest_document`
+- `open_document`: extension allowlist, DANGEROUS permission, audit logging
+
+### Removed
+- `FinanceResearchState` (domain-specific code in shared library)
+- Dead `get_help()`, `get_ui_section`, `is_ui_hidden` methods
+- Dead `generate_tool_summary`, `clear_context`/`unbind_context`, model constant re-exports
+- Dead `cli/settings.py` re-export shim
+
 ## [0.4.3] - 2026-02-12
 
 ### Changed

{agentic_cli-0.4.3 → agentic_cli-0.5.0}/CLAUDE.md

@@ -20,7 +20,6 @@ agentic-cli/
 │   ├── __init__.py           # Package exports, lazy imports
 │   ├── config.py             # BaseSettings (pydantic-settings)
 │   ├── constants.py          # Shared constants, truncate()
-│   ├── resolvers.py          # Model/path constants (GOOGLE_MODELS, etc.)
 │   ├── settings_persistence.py
 │   ├── logging.py
 │   ├── cli/
@@ -50,7 +49,7 @@ agentic-cli/
 │   ├── tools/
 │   │   ├── registry.py       # ToolRegistry, @register_tool, ToolCategory, PermissionLevel
 │   │   ├── executor.py       # SafePythonExecutor
-│   │   ├── knowledge_tools.py # search_knowledge_base, ingest_to_knowledge_base
+│   │   ├── knowledge_tools.py # kb_search, kb_ingest, kb_list, kb_read
 │   │   ├── arxiv_tools.py    # search_arxiv, fetch_arxiv_paper, analyze_arxiv_paper
 │   │   ├── execution_tools.py # execute_python
 │   │   ├── interaction_tools.py # ask_clarification
@@ -132,8 +131,8 @@ Workflow:
 ### Key Design Patterns
 - **Tool error handling**: All tools return `{"success": bool, ...}` dicts. Never raise `ToolError`.
 - **Tool registration**: Use `@register_tool(category=..., permission_level=..., description=...)` decorator. Tools are auto-discovered via the global `ToolRegistry`.
-- **Store consolidation**: Stores and managers (MemoryStore, PlanStore, TaskStore, ApprovalManager) live inside their tool files (e.g. `memory_tools.py`, `hitl_tools.py`), not in separate packages.
-- **Context access**: Tools use `get_context_*()` functions from `workflow.context` to access managers and stores via ContextVars.
+- **Service registry**: Tools access services and shared state via `get_service(key)` from `workflow.service_registry`. A single ContextVar holds a `dict[str, Any]` set by the workflow manager during processing. Complex services (KBManager, SandboxManager, MemoryStore) are lazily created; simple state (plan string, task list) lives directly in the registry dict.
+- **Manager detection**: Tools decorated with `@requires("kb_manager")` etc. are scanned by `BaseWorkflowManager._detect_required_managers()` which lazily creates only the needed services.
 - **Atomic writes**: Use `atomic_write_json`/`atomic_write_text` from `persistence/_utils.py` for file persistence.
 
 ### Console Output