uneven-ai 1.0.6 → 1.1.0

package/CHANGELOG.md

@@ -5,6 +5,79 @@ All notable changes to Uneven AI will be documented in this file.
 The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.0.0/),
 and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
 
+## [1.1.0] - 2026-04-20
+
+### Added
+
+- **Privacy Hardening**: Enhanced protection for development environment privacy. Absolute paths from the host machine are now automatically stripped from distribution binaries.
+- **Scalability**: New memory-optimized indexing architecture for massive codebases, allowing the engine to scale to millions of lines with a minimal RAM footprint.
+- **Project Integrity**: Real-time configuration schema validation to prevent malformed settings from affecting system stability.
+
+### Fixed
+
+- **Stability**: Improved sub-process management to ensure clean termination of development servers and dependent tasks, preventing orphan processes and resource leaks.
+- **Reliability**: Advanced loop protection and binary-header identification in the knowledge indexing engine for safer directory traversal.
+- **Error Propagation**: Enriched internal error system for more precise diagnostics and user feedback during AI inference.
+
+## [1.0.8] - 2026-04-19
+
+### Fixed
+
+- **watcher/terminal**: `start()` now returns `Ok(None)` for non-zero exit codes instead of silently returning `Ok(())`; `Send + 'static` bound removed from `stream_output` callback; `&mut self` corrected to `&self`; I/O errors on reader tasks now emit warning instead of silent EOF
+- **watcher/error_parser**: `message` now extracts the line containing the regex match instead of always returning the first line of output (fixes preamble cases like Rust panics)
+- **watcher/mod**: path traversal guard added to `watch_directory` — `../` paths are now rejected
+- **retrieval**: cosine similarity clamped to `max(0.0)` before `to_bits()` ordering — negative scores were ranking semantically opposite documents above similar ones
+- **retrieval**: `write_hnsw` index save now uses atomic tmp+rename — SIGKILL during save no longer corrupts the index permanently
+- **retrieval**: `stats()` now returns distinct values for documents and embeddings — both fields were identical before, making diagnostics useless
+- **retrieval**: `clear()` now uses `tokio::fs::remove_file` — was blocking the Tokio executor with synchronous `std::fs` calls
+- **retrieval**: lock poisoning in `hnsw_backend` handled via `unwrap_or_else(|e| e.into_inner())` in 6 locations
+- **pentester/static_analysis**: `scan_secrets` was a no-op — fully reimplemented with BFS file walk, static `Lazy<Regex>` and minimum-length filter on API key patterns
+- **pentester/port_scanner**: empty host now rejected with `PentesterError`; scan concurrency capped at 50 via batching; non-ECONNREFUSED errors propagated instead of swallowed
+- **pentester/mod**: `static_analysis()` and `active_testing()` were stubs returning empty results — now delegate to their submodules; `SecurityFinding.severity` promoted from unvalidated `String` to typed `enum Severity`; CIDR `/0` now rejected
+- **llm/inference**: stale `INTERRUPT` flag from a previous call no longer silences the next generation; race condition in `unload_model` between `LLAMA` and `LOADED_MODEL` closed; `is_model_loaded` uses `AtomicBool` — no longer returns `false` while writer holds the lock
+- **llm/inference**: `run_inference` now returns real token count and real model name instead of hardcoded values
+- **llm/embeddings**: double-init race fixed with `AtomicBool::compare_exchange`; lock poisoning handled via `unwrap_or_else(|e| e.into_inner())`
+- **knowledge/indexer**: `index_directory` was checking path type but never traversing the filesystem — BFS walk now implemented; symlink cycle protection via `entry.file_type()`; path traversal guard added
+- **knowledge/file_loader**: TOCTOU `exists()` check removed; extension check now runs before filesystem access
+- **agent/fix**: `generate_patch` now produces valid unified diff with headers; `apply_patch` uses atomic write-then-rename; path traversal guard and stale-snapshot guard added
+- **agent/mod**: `apply_fix` always returned `Ok(true)` without touching files — now delegates to `fix::apply_patch`; `diagnose` now delegates to `diagnose::diagnose_error`
+- **napi**: `store_documents_batch` validates that all five input slices have equal length — `zip` was silently truncating to the shortest; `load_file_content` path traversal guard added
+- **application/engine_service**: `initialize` now validates `brain.provider` and `brain.model`; concurrent-init race closed with `AtomicBool`; `UnevenEngine` singleton wired to NAPI via `OnceLock`
+- **utils/config-loader**: partial `brain` override no longer overwrites config defaults with `undefined`; `syncConfigFromTs` uses atomic write-then-rename; backslash-escaped quote detection corrected
+- **utils/migration**: `ENOTEMPTY` on already-migrated destination no longer stalls migration on every startup; per-file rename errors now tracked individually
+- **utils/process-lock**: `release()` removes SIGINT/SIGTERM/exit listeners — prevented `MaxListenersExceededWarning` accumulation in test suites
+- **utils/error-parser**: `extractStackTrace` now captures the exception message line preceding the first stack frame
+- **license/gate**: expired pro license (`active: false`) no longer granted access; `process.exit(0)` on blocked feature corrected to `process.exit(1)`
+- **license/manager**: `status()` now reflects real expiry; HTTP error responses no longer silently extend grace period; `writeLicense` failures inside `.catch()` no longer produce unhandled rejections
+- **license/storage**: `readLicense()` validates all required fields at runtime — partial/corrupt cache no longer propagates `undefined`; `writeLicense` uses atomic write-then-rename
+- **io/git-manager**: shell injection in `checkout`, `deleteBranch`, `revertCommit`, `commitFix` fixed — migrated to `execFileSync`/`execFile` with argument arrays
+- **io/web-scraper**: concurrent fetch capped at 10; `undici` Agent reused via singleton instead of instantiated per request
+- **io/process-watcher**: SIGKILL fallback added after 5s if process ignores SIGTERM; `off()` method added for listener removal
+- **io/file-watcher**: `start()` no longer hangs indefinitely — error handler and 30s timeout added
+- **io/logger**: UTC fields used consistently in `getTimestamp()`; `debug()` now writes to log file
+- **io/db-loader**: empty `tables: []` now warns instead of silently loading 0 documents; DB type detection migrated from substring match to `URL.protocol`; `sanitizeUrl` redacts username in addition to password
+- **adapters/external-providers**: Gemini safety-filter responses no longer throw TypeError; streaming readers closed in `finally` across all four providers; Gemini NDJSON parser uses proper `JSON.parse` instead of string-match heuristic
+- **domain/resource-guardian**: uncaught callback exceptions in `setInterval` no longer crash the process; CPU hot-plug length mismatch guarded
+- **domain/snapshot**: `restore()` validates that target path is inside project root before writing
+- **domain/session**: `lockTimeoutHandle` nullified before `forceReleaseLock` — prevented new lock from being cancelled by expired-timeout cleanup
+- **cli/analyze**: double-disconnect race in `runBatchDashboard` fixed; spinner interval leak on thrown error fixed
+- **cli/askf**: `isSafeFilename()` guard added — LLM-supplied filenames with path traversal or absolute paths now rejected
+- **cli/ci**: `__UNEVEN_GH_LINES` env var replaced with module-level array; external `Promise.race` deadline added to prevent indefinite CI hang
+- **cli/log**: `readFileSync` replaced with `createReadStream` — large log files no longer risk OOM
+- **orchestration/engine**: `unloadTimer`, `saveProgress` and `activeFixes` race conditions fixed; `promptUserConfirmFix` and `llmInfer` now have timeouts; `docType` inconsistency between index phases corrected
+- **orchestration/incremental-index**: `markIndexedWithHash` rejects empty hash, negative size, or non-array `chunkIds`
+- **orchestration/knowledge-retriever**: ReDoS in `extractSource` fixed (`\S+` → `\S{1,500}`)
+- **development/fix-engine**: snapshot-before-lock race fixed; prompt injection via `error.message` sanitized; path traversal guard added; `fixAttempts` memory leak fixed
+- **development/test-runner**: SIGKILL flag logic corrected; output buffer capped at 2 MB
+- **development/analyst-job-manager**: concurrent `saveJobState` race serialized per job-id; task ID collision on same millisecond fixed
+
+## [1.0.7] - 2026-04-17
+
+### Fixed
+
+- **license**: `parseKey` now correctly validates `UNVN-PRO-*` / `UNVN-TEAM-*` keys — 1.0.6 was published with a stale build that still rejected these keys as malformed (`TMR` prefix check)
+- **license**: gate message now shows correct command `uneven-ai license activate` — 1.0.6 stale build showed `uneven license activate`
+
 ## [1.0.6] - 2026-04-16
 
 ### Fixed
@@ -501,7 +574,7 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 - `src/core/license/` module: machine fingerprinting (SHA256 of platform+arch+hostname+homedir), encrypted local cache (`~/.config/uneven-ai/license.json`), offline grace period (7-day extension, max 30 days cumulative)
 - License key format: `TMR-{TIER}-{payload_b64}-{hmac}` — payload parsed locally, HMAC verified server-side
-- `uneven-ai license activate <key>` — activates key against `api.uneven-ai.dev`, caches result locally
+- `uneven-ai license activate <key>` — activates key against `api.uneven.dev`, caches result locally
 - `uneven-ai license status` — shows tier, seats, expiry, machine ID
 - `uneven-ai license deactivate` — removes machine from seat count
 - Pro feature gate (`src/core/license/gate.ts`): commands blocked on free tier show upgrade message and exit cleanly

package/LICENSE

@@ -3,7 +3,7 @@ Business Source License 1.1
 Licensor:             KR Riley Soluções
 Licensed Work:        Uneven AI
 Additional Use Grant: Personal, educational and non-commercial use is free.
-Change Date:          2029-01-01
+Change Date:          2056-04-19
 Change License:       Apache License 2.0
 
 Use Limitation: Production commercial use — including SaaS products,
@@ -24,6 +24,10 @@ contato@rileysolucoes.com.br
 
 ---
 
+Uneven AI's "Maria" agent is powered by Llama 3.2. Llama 3.2 is licensed under the Llama 3.2 Community License, Copyright © Meta Platforms, Inc. All Rights Reserved.
+
+---
+
 BUSINESS SOURCE LICENSE 1.1 SUMMARY:
 
 What's free?
@@ -36,5 +40,5 @@ What requires a license?
 ❌ Commercial production use (SaaS, paid services, internal company use)
 
 Automatic conversion:
-- On 2029-01-01, this license converts to Apache 2.0
+- On 2056-04-19, this license converts to Apache 2.0
 - At that point, all commercial uses become freely allowed