codevira 1.6.1__tar.gz → 1.7.1__tar.gz

{codevira-1.6.1 → codevira-1.7.1}/CHANGELOG.md

@@ -13,6 +13,182 @@ This project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.htm
 
 ---
 
+## [1.7.1] — 2026-04-22 — Search Timeout Fix & Version Display
+
+Two small but user-visible fixes on top of v1.7.0.
+
+### Fixed
+
+- **`search_codebase` timeout on first call** (reported by a user testing
+  on Antigravity). The embedding model (`all-MiniLM-L6-v2`) was being
+  instantiated fresh on every MCP tool call, which triggered a ~90MB
+  download + PyTorch init on first ever use (30-60s on slow networks)
+  and 1-3s of re-init overhead on every subsequent call. Antigravity's
+  ~30s MCP tool timeout killed the query before the model finished loading.
+
+  Three-layer fix:
+  1. Module-level cache for the chroma client + embedding function,
+     keyed by `db_dir`. Subsequent calls are now instant.
+  2. Background `prewarm_embedding_model()` spawned at server startup
+     (both stdio and HTTP transports). Model loads in parallel with
+     the MCP handshake window.
+  3. Cold-path timeout guard: if a query arrives while warmup is still
+     in progress, returns `{"status": "warming", ...}` within 10 seconds
+     instead of blocking until the MCP timeout fires. The agent gets a
+     clean retryable response.
+
+- **`codevira register` banner showed hardcoded `v1.6`** after upgrading
+  to v1.7.0. Now reads `mcp_server.__version__` dynamically. Same fix
+  applied to `metadata.json` version field written during auto-init and
+  migration.
+
+---
+
+## [1.7.0] — 2026-04-18 — Token Efficiency & AI-First Tool Design
+
+**The biggest release since v1.0.** We realized Codevira was dumping 15k-60k
+tokens per session into AI agent context windows — defeating the entire
+"92% token reduction" value prop. This release redesigns tool responses
+around what agents actually need, not what the database can return.
+
+### Changed — Dependency model
+- **`chromadb` + `sentence-transformers` now required** (was `[search]` extra).
+  `pip install codevira` installs all 36 MCP tools out of the box.
+  Trade-off: ~500MB install (ML runtime) vs. ~50MB. Eliminates the
+  "why doesn't semantic search work?" confusion.
+- **`[search]` extra kept as no-op alias** for backwards compatibility.
+
+### Changed — Token-efficient tool responses (the big one)
+
+Every high-traffic tool now returns a **summary by default**, with opt-in
+full data. On a 500-node project, a single agent session went from ~60k
+tokens to ~5k.
+
+- **`get_session_context`**: Compacted ~4k → ~800 tokens. Dropped
+  `global_intelligence`/`indexing_progress` (admin data, not session data).
+  Truncated decision/summary text. Nested `current_phase` at top level.
+- **`get_node(path)`**: Default returns counts (`rules_count`,
+  `dependencies_count`) + flags. Pass `full=True` for the full arrays.
+  Typical response: ~100 tokens (was 500-3000).
+- **`get_impact(path)`**: Default returns 10 affected files + protected/
+  high-stability counts. Pass `summary_only=True` for just counts
+  (~80 tokens — perfect for gate checks before modifying).
+- **`search_codebase(query)`**: Default returns file/symbol pointers only.
+  Pass `include_content=True` to inline chunk source (500-3000 tokens per
+  match). `limit` capped at 20.
+- **`search_decisions(query)`**: Default limit 5 (was 10), context truncated
+  to 150 chars. Pass `full=True` for untruncated text. `limit` capped at 20.
+- **`get_history(file)`**: Default limit 5 (was 20), text truncated.
+  Pass `full=True` for untruncated. `limit` capped at 50.
+- **`get_full_roadmap`**: Completed phases summarized (number + name + date
+  + decision_count) instead of inlining all `key_decisions`. Pass
+  `include_decisions=true` for the old behavior.
+- **`list_nodes`**: Paginated (50 per page, max 500) with `offset` support.
+  Response includes total count + per-layer distribution.
+
+### Changed — AI-facing MCP tool surface trimmed to 23 tools (was 36)
+
+12 tools moved to admin-only — they still work via `call_tool` dispatch
+but are **hidden from `list_tools()`**. AI agents only see tools they
+should use. The hidden tools are either:
+- Dashboard/reporting (human workflows): `get_full_roadmap`,
+  `get_project_maturity`, `find_hotspots`, `analyze_changes`, `get_graph_diff`
+- Bulk discovery (replaced by targeted queries): `list_nodes`, `add_node`
+- Background automation (self-managed): `refresh_graph`, `refresh_index`
+- Redundant with session_context: `get_preferences`, `get_learned_rules`
+- Dumps too many tokens: `export_graph` (can be 50k tokens)
+
+Admins can still call these via CLI. Prompts like `architecture_overview`
+still reference them server-side.
+
+### Added
+- **Non-blocking `refresh_index`**: Returns in <100ms with
+  `{"status": "Refresh started in background"}`. Heavy work (graph regen +
+  semantic embedding) runs in a daemon thread. Previously, this hung AI
+  agents for minutes on 500+ file projects.
+- **`codevira clean` command**: One-shot removal of all Codevira data, IDE
+  configs, and services. `--all`, `--dry-run`, `-y` supported.
+- **Google Antigravity global mode**: `codevira register` now includes
+  Antigravity with a single global entry (was missing + wrong config path).
+- **Browser-friendly landing page**: `GET /` on HTTP server returns helpful
+  HTML with setup instructions for browsers. API clients still get JSON.
+
+### Fixed
+- **`refresh_graph` ignored its `file_paths` parameter** — dead code that
+  always regenerated the entire graph. Cleaned up.
+- **`generate_graph_sqlite` crashed on macOS system paths**: Now skips
+  `Library`, `System`, `Applications`, `node_modules`, `.venv`, etc.,
+  and catches `OSError`/`ValueError` per-entry so one bad symlink doesn't
+  abort indexing.
+- **Crash log test isolation**: `crash_logger._get_log_dir()` now uses
+  `get_global_home()`. Tests no longer pollute the real user's crash log.
+- **`_get_embedding_fn` ValueError not caught**: When chromadb is installed
+  but sentence-transformers isn't, chromadb raises `ValueError`. Now caught.
+- **Playbook `add_route` → `add_tool`**: The valid task type was renamed
+  in code but the description still said `add_route`. Fixed.
+- **Antigravity config path**: Was wrong (`~/.gemini/settings/`). Now uses
+  the correct `~/.gemini/antigravity/mcp_config.json`.
+
+### Added — Post-release enhancements (merged into 1.7.0)
+
+- **`codevira status` is now fast** (~200ms for uninitialized projects,
+  ~1s for initialized). Was ~5-6s because it was loading the ~90MB
+  sentence-transformers embedding model just to count chunks. Now uses
+  `collection.count()` which doesn't need the embedding function, and
+  short-circuits entirely when there's no graph DB yet.
+- **`codevira status --global`** flag shows launchd service state +
+  cross-project memory stats in a dedicated panel. Works on both
+  initialized and uninitialized projects.
+- **`codevira status --check-stale`** flag opt-in for the slow SHA256
+  file-walk (was always-on, made status take 5s+).
+- **`codevira clean --legacy`** — remove `.codevira.migrated/` backup
+  directories accumulating across all initialized projects. Shows size
+  and confirms before deletion.
+- **`logs.retention_days` actually works now** (was dead config in earlier
+  versions). Opt-in only — default 0 keeps sessions/decisions forever.
+  Set > 0 for privacy-driven time-bounded history. Runs at most once
+  per 24h at server startup.
+- **HTTP/HTTPS transport marked as PREVIEW** (single-project only). The
+  server binds to one project at startup and cannot switch contexts per
+  request. Multi-project HTTPS routing via MCP `initialize` `rootUri` is
+  the top v1.8 priority. `codevira serve` prints a preview warning on
+  startup. README / FAQ / PROTOCOL updated to position stdio as the
+  clear recommendation for multi-project work.
+- **Dead-code audit** — removed 4 unused functions (`find_project_by_remote`
+  in global_db, `get_file_outcome_summary`, `add_open_changeset`,
+  `remove_open_changeset`), renamed `get_changeset` → `_get_changeset`
+  (was module-private usage only). Wired up 3 unused-but-useful functions
+  (`launchd_status`, `cleanup_legacy_dir`, `get_global_stats`) into the
+  CLI where they belong.
+- **Open-source readiness pass** — removed stray test-playground files
+  from git, fixed PR template typo (`mcp-server` → `mcp_server`), replaced
+  hardcoded author username in docstring examples (`/Users/sachin/...`
+  → `/Users/alice/...`), added `__all__` + `__version__` to
+  `mcp_server/__init__.py`, removed duplicate `requirements.txt`.
+
+### Tests
+- 1,306 tests passing (added 15 new tests for `log_retention.py`)
+
+---
+
+## [1.6.2] — 2026-04-16 — Crash Log Isolation & Browser UX
+
+### Fixed
+- **Crash log test isolation**: `crash_logger._get_log_dir()` now uses
+  `get_global_home()` instead of hardcoding `~/.codevira/logs/`. Tests
+  no longer pollute the real user's crash log with pytest mock tracebacks.
+- **`_get_embedding_fn` ValueError not caught**: When chromadb is installed
+  but sentence-transformers isn't, chromadb wraps the ImportError as a
+  ValueError. `_get_embedding_fn` now catches both and re-raises as
+  ImportError for consistent handling by callers.
+
+### Added
+- **Browser-friendly landing page**: `GET /` on the HTTP server now returns
+  a helpful HTML page for browsers (with setup instructions) instead of
+  just JSON. API clients with `Accept: application/json` still get JSON.
+
+---
+
 ## [1.6.1] — 2026-04-16 — Stability, Graceful Degradation & Cleanup
 
 ### Added