EvoCLI v0.1.0

EvoCLI v0.1.0

Personal AI Engineering Runtime — Rust Host + Python Soul

Fixed — Quality & Integration Sweep (49 issues across 6 waves)

Dependencies

• Remove rocketry zombie dependency (was declared but never imported after scheduling moved to Rust Job Queue)
• Add keyring>=24.0 (required by agent.py API key lookup, was missing from pyproject.toml)
• Add tomli>=2.0; python_version < '3.11' for Python 3.10 compatibility (7 files used import tomllib which is 3.11+ stdlib only)

Rust Host

• Remove 5 dead_code functions: default_provider, default_fast_model, default_smart_model (config.rs), project_id() (event_bus.rs), default_tiers_for (init.rs), git_diff (git.rs)
• Remove unused Confirm import from init.rs (code used fully-qualified dialoguer::Confirm::new())
• Fix unnecessary mut at tool_dispatch.rs:1418 (let mut entries in write_tree)
• Add #[allow(dead_code)] to event_bus::global() with explanation comment
• Add #[allow(unused_mut)] to with_code_index! macro to suppress 17 spurious warnings
• Fix jobs.db schema migration order: POST_MIGRATION_INDEX now runs after MIGRATION so project_id column exists before index creation
• Add [dev-dependencies] tempfile = "3" for jobs.db migration tests

MCP Exposure

• Expose 8 previously missing tools via evocli_as_mcp_tools(): code_intel_communities, code_intel_processes, web_fetch, git_snapshot_list, git_snapshot_restore, fs_read_range, contracts_list, contracts_get_checkpoints
• Update EVOCLI_MCP_TOOL_COUNT: 62 → 70; unit test tool_count_matches_expected validates this

Python Soul — Core Logic

• Fix code_intel.ranked_context: was returning hardcoded score=1.0 for all symbols; now uses sqrt(caller_count) + mention_bonus scoring with descending sort
• Fix session_id in handle_agent_run: was cwd_<md5> (shared across conversations in same dir); now sess_<uuid4> per conversation
• Fix session_id in handle_agent_stream and all slash commands (/add, /btw, /compress, /undo, /plan): DRY'd into _derive_stream_session_id() helper with design rationale documented
• Fix TUI lib.rs: both agent.stream submit paths now explicitly pass session_id (FNV-1a hash of CWD for project continuity); override_session_id field added to App struct for evocli session resume
• Fix evolution/decay_detector.py: 2 bare except: pass → log.debug with context
• Fix handlers/agent.py::auto-continue: if False: guard documented with protocol limitation explanation and Options A/B/C roadmap

Python Soul — Memory System

• Add normalize_project_id() to state.py: maps None/"."/"global" → os.path.abspath(os.getcwd()); prevents ".", "global", and actual cwd paths from hashing to different memory buckets
• Fix ALL 11 memory handlers in handlers/memory.py to extract and pass project_id from params to state.get_memory(project_id=project_id) (previously all called state.get_memory() ignoring caller's project context)
• Fix metrics.py::handle_evolution_transfer(): was no-op params.get("project_id") (value discarded); now properly passed to get_memory()
• Fix evolution/failure_miner.py::mine(): ignored its project_id parameter; now passed to get_memory()
• Fix code_chunks.py::generate_community_summaries(): stored summaries with wrong project_id; now uses pid from outer loop
• Fix memory_client.py::add(): global-scoped memories now stored with project_id="global" instead of self.project_id, enabling cross-project retrieval via vector search filter
• Fix memory_client.py::search(): tool memory matching now also checks tags field as fallback for legacy entries without tool_id
• Fix evolution/knowledge_classifier.py::promote_if_transferable(): used get_memory(project_id="global") which normalized to cwd; now uses get_memory(project_id=None) and lets add(priority="global") set project_id="global" in the stored entry
• Fix state.py::reset_all(): was resetting _memory (old single-instance var); now also calls _memories.clear() to clear the per-project memory dict

Python Soul — Context Engine

• Fix context_engine.py: get_memory() replaced with get_memory_if_ready(project_id) — avoids blocking the asyncio event loop on fastembed/LanceDB model initialization (30+ seconds on first run)
• Fix context_engine.py: Superpowers Skill guidance injection guarded by _embedder_cache is not None AND _skill_engine is not None — prevents blocking SkillEngine(real_bridge) initialization during context build
• Fix context_engine.py: added skip_repomap guard — RepoMap skipped when no current_file anchor, preventing full tree-sitter codebase scan on empty requests
• Fix code_analysis.py::_resolve_symbol_id: bare except: pass → log.debug with context

Python Soul — Path Safety

• Fix state.py::_history_path(): now uses SHA-256 suffix to prevent both path traversal and filename collisions between session IDs that normalize to the same safe prefix
• Fix session.py::SessionManager._path(): same SHA-256 collision-resistant sanitization

Python Soul — agent.py

• Fix code_chunks.get_index(): changed from single _index global singleton ("first caller wins") to _index_cache: dict[str, CodeChunkIndex] keyed by os.path.abspath(project_id)
• Fix agent.py::code_semantic_search: was calling get_index(self._session_id) (session-keyed, bug: all sessions after first got wrong project index); now calls get_index(os.getcwd())
• Fix 6 critical bare except Exception: pass in agent.py → log.debug with error context (API key lookup, event recording, stream tool parsing, context injection)

Python Soul — BOM

• Remove UTF-8 BOM from prompt_manager.py and handlers/system.py (caused ast.parse() failures in tooling)

Tests

• Add 2 jobs.db migration tests: test_fresh_db_creation, test_legacy_db_migration (fresh DB + legacy schema upgrade)
• Add 4 code_analysis handler tests: response structure, risk_level field, list return, `_RANKED_CONTEXT_WE

[截断]

---

安装

Windows

# 解压 zip 后：
.\setup.ps1 && .\evocli.exe init && .\evocli.exe

Linux x86_64

tar -xzf evocli-v0.1.0-linux-x86_64.tar.gz
cd evocli-v0.1.0-linux-x86_64/
bash setup.sh && ./evocli init && ./evocli

平台

平台	文件
Linux x86_64	evocli-v0.1.0-linux-x86_64.tar.gz
Windows x86_64	evocli-v0.1.0-windows-x86_64.zip

macOS / Linux aarch64: 暂不提供预编译包（依赖 ONNX/tantivy 的 SIMD 检测与交叉编译不兼容）。
源码构建：cargo build --release --bin evocli

sha256sum --check checksums-sha256.txt