security: auditoria completa 0.8 — vulnerabilitats, hardening i tests by jgoy-labs · Pull Request #2 · jgoy-labs/server-nexe

jgoy-labs · 2026-02-21T15:10:09Z

Resum

Implementació completa del pla d'auditoria de seguretat per Nexe Server 0.8. S'han analitzat ~11.000 fitxers Python i s'han corregit 21 problemes classificats per severitat.

🔴 Crítiques resoltes

C-1 — Secrets NEXE_PRIMARY_API_KEY i NEXE_CSRF_SECRET rotats
C-3 — Bare except: substituïts per except Exception as e: amb logging a 4 fitxers

🟠 Altes resoltes

A-1 — Upload path traversal: sanititzar filename + whitelist d'extensions ({.txt, .md, .pdf, .csv, .rst, .html})
A-2 — Qdrant: suport QDRANT_API_KEY per entorns remots
A-3 — /health/ready: retorna {status, timestamp} sense exposar llista de mòduls sense auth
A-4 — HTTPException 500: eliminar detail=str(e) (no filtrar internals als clients)
A-5 — Qdrant log file: flush() + = None al shutdown
A-6 — SessionManager.cleanup_inactive() implementat amb TTL real
A-7 — Streaming Ollama: gestionar asyncio.CancelledError (client desconnectat)

🟡 Mitges resoltes

M-2 — Docker healthcheck per al servei nexe a docker-compose.yml
M-3 — qdrant/qdrant:latest → qdrant/qdrant:v1.12.0
M-4 — .dockerignore creat (exclou .env, storage/, .git/, tests/)
M-5 — httpx eliminat de requirements-dev.txt (duplicat)
M-6 — pip-audit afegit al CI pipeline

🟢 Tests de seguretat (14/14 passing)

Nou fitxer tests/integration/test_security.py:

Path traversal i extensions no permeses a upload
Regressió _sanitize_rag_context (truncació + filtratge injeccions)
/health/ready no exposa informació de mòduls sense auth
Session TTL cleanup

Altres millores incloses a la branca

Dockerfile: usuari non-root nexe:nexe, permisos 750
runner.py: validate_production_config() per detectar secrets buits en producció
scripts/generate_secrets.sh: script per generar secrets segurs
knowledge/SECURITY.md, DEPLOYMENT.md: documentació de seguretat

Pla de tests

pytest tests/integration/test_security.py -v → 14/14 passing
pip-audit -r requirements.txt → sense CVEs crítics
docker compose up --build -d → healthcheck OK per ambdós serveis
Verificar que /health/ready retorna {status, timestamp} sense detalls de mòduls
Verificar que upload de .exe retorna HTTP 400
Verificar que upload de ../../etc/passwd retorna HTTP 400

Assisted by AI

…cepts - memory/rag/routers/endpoints.py: ALLOWED_UPLOAD_EXTENSIONS whitelist, filename sanitization (Path.name) to prevent path traversal, generic 500 error messages (no internal details exposed) - core/endpoints/root.py: /health/ready returns only {status, timestamp} to avoid exposing internal module list without auth - plugins/web_ui_module/session_manager.py: implement cleanup_inactive() with timedelta TTL to prevent session memory leak - memory/memory/engines/persistence.py: QDRANT_API_KEY env var support for authenticated Qdrant deployments - core/endpoints/chat.py: log JSONDecodeError in Ollama stream, handle asyncio.CancelledError on client disconnect - plugins/web_ui_module/memory_helper.py: replace bare except with except Exception + debug logging - .env.example: document all env vars including QDRANT_API_KEY - .dockerignore: exclude .env, storage/, venv/, .git/ from Docker image - .github/workflows/ci.yml: pip-audit CVE scan + unit tests - core/endpoints/tests/test_security.py: 14 security regression tests Assisted by AI

Issue 8 - {{NEXE_*}} templates substituïts per IDs literals: - memory/embeddings: {{NEXE_EMBEDDINGS_MODULE}} -> 'embeddings' - memory/memory: {{NEXE_MEMORY_MODULE}} -> 'memory', dependencies corregides - memory/rag: {{NEXE_RAG_MODULE}} -> 'rag' - plugins/security: {{NEXE_SECURITY_MODULE}} -> 'security' - core/cli/manifest.toml: comentari corregit (cli) Issues 9+10 - URLs hardcoded restants (lifespan.py + ollama_module): - lifespan.py:96 unificat NEXE_OLLAMA_HOST > OLLAMA_HOST (fallback) - lifespan.py:247,585 usen NEXE_OLLAMA_HOST - lifespan.py:482,564,565 usen NEXE_API_BASE_URL / config - ollama_module/module.py:420 usa NEXE_API_BASE_URL Issue 11 - datetime.now() -> datetime.now(timezone.utc): - core/endpoints/bootstrap.py, root.py - plugins/security/core/logger.py - plugins/web_ui_module/session_manager.py, memory_helper.py - memory_helper.py: handle naive datetimes antics (replace tzinfo) - test_security.py actualitzat per usar datetime.now(timezone.utc) Issue 12 - Variables d'entorn sense prefix NEXE_ unificades: - BOOTSTRAP_TTL -> NEXE_BOOTSTRAP_TTL (amb fallback a nom antic) - AUTO_CLEAN_ENABLED -> NEXE_AUTO_CLEAN_ENABLED (idem) - AUTO_CLEAN_DRY_RUN -> NEXE_AUTO_CLEAN_DRY_RUN (idem) - OLLAMA_HOST mantingut com a fallback de NEXE_OLLAMA_HOST Issue 13 - .env.example: variables que faltaven documentades: - NEXE_BOOTSTRAP_DISPLAY, NEXE_BOOTSTRAP_TTL - NEXE_AUTO_CLEAN_ENABLED, NEXE_AUTO_CLEAN_DRY_RUN - NEXE_QDRANT_HEALTH_TIMEOUT - OLLAMA_HOST (com a fallback comentat) Tests: 954 passed, 0 failures Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

#3 loader: fallback només considera classes definides al mòdul (afegit attr.__module__ == module.__name__ per evitar classes importades) #2 singletons: afegits reset_config() i reset_loader() per testing #1 core/app.py: import os mogut al bloc d'imports del principi #4 loader.py: docstrings traduïts de català a anglès (unificació open-source) Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

Fix-All BUS sobre 3 tracks paral·lels per resoldre tots els bugs del QA post-BUS de normalització abans del DMG v0.9.0. 8 commits dev consolidats en aquest sync. TRACK A — Memory/RAG/Sessions - Bug #1 (PID file canònic) — single source of truth a storage/run/server.pid - F5 — 3 col·leccions canòniques (nexe_web_ui, user_knowledge, nexe_documentation) creades a get_memory_api() en lloc de només la primera - F7 — ingest_knowledge defaulteja a nexe_documentation (era user_knowledge) i és idempotent (eliminada la sequence delete_collection + create_collection destructiva que esborrava docs ad-hoc dels usuaris a cada install) - F8 — root cause empíric Bug #4: MemoryModule obria un SEGON QdrantClient real a storage/vectors/qdrant_local/, divergent del singleton del pool. MEM_SAVE escrivia a una col·lecció, MEM_RECALL llegia d'una altra. Ara tots dos comparteixen storage/vectors/. - F1 — _check_duplicate retorna contracte honest (success=False, duplicate=True) enlloc de fingir success=True amb document_id=None. Era el segon root cause de Bug #4: el dedup bloquejava SAVEs amb fals positius silenciats. - F2 — typo cols list (nexe_web_ui duplicat) - F3 — list_memories scroll-based (sense semantic search amb query anglesa) - Bug #10 — collections= filter a list/save/delete (sidebar checks reals) - Bug #6 — frontend hydration document attached. Eren 2 bugs encadenats: l'endpoint /history no retornava attached_document, i removeFilePreview() feia POST /clear-document destructiu cada switch de sessió. - Bug #3 — MEM_SAVE-only response fallback. Quan el model emet només [MEM_SAVE: ...] sense text envoltant, ara genera 'Memòria desada: <fact>' perquè el bloc save s'executi i el frontend mostri confirmació. - auto_save crida eliminada per HOMAD memoria v1 (2026-04-01) — manual MEM_SAVE only fins a Part 2. TRACK B — Tray / Multi-instance / Packaging - Bug #1 (PID file) compartit amb Track A - Bug #2 — setproctitle a server i tray (server-nexe / nexe-tray a ps/Activity Monitor). Force Quit encara mostra Python perquè requereix CFBundleName via .app bundle real (deute v0.9.1). - Bug #9 — menu polish: server-nexe.com duplicat substituït per '📖 Documentació' al main level (3 idiomes), website_item es manté al submenú Configuració. TRACK C — UX cosmètic - Bug #5 — slow_request middleware exclou /ui/upload (uploads naturalment triguen >1s i el log apareixia duplicat amb l'access log d'uvicorn). - Bug #8 — 3 ⓘ visibles als checkboxes del sidebar de col·leccions amb tooltips als 3 idiomes (la infraestructura CSS/i18n ja existia). Pytest D-1 final: 4424 passed, 0 failed, 35 skipped, 1 xfailed, 86% coverage en 76.11s. Baseline pre-BUS era 4396. +28 tests nous, ZERO regressions. Tests nous: - tests/test_pid_file.py: 7 tests Bug #1 - tests/test_ingest_knowledge_idempotent.py: 8 tests F7 (3 classes) - plugins/web_ui_module/tests/test_memory_helper_async.py: 1 test F1 - plugins/web_ui_module/tests/test_memory_delete.py: 7 tests F3+Bug#10 - plugins/web_ui_module/tests/test_mem_save_injection.py: 5 tests Bug #3 Out of scope (deute v0.9.1+): - routes_chat.py 54KB decapitació general (deute formal P0) - Bundle .app real amb py2app per CFBundleName (deute v0.9.1) - Resums per capítol (Part 2 redisseny memory) - RDBMS font de veritat + vector store reconstruïble (HOMAD memoria v1, Part 2)

…rence Symptom: after a restart the sidebar only listed .json sessions (.enc were invisible), new sessions were persisted unencrypted even with encryption enabled, and a reboot's .json->.enc migration could overwrite an existing .enc belonging to a different conversation with the same id (collision observed in the wild: "Hola Diana" overwritten by "Hola Anna"). Root cause is a three-bug chain: 1. Loader early-init (core/loader/manifest_base._get_module): calls `instance._init_router()` immediately after `__init__`, *before* `initialize()` runs. The plugin must have its routable resources ready at construction time. 2. Plugin double-create (plugins/web_ui_module/module.py): `__init__` created a SessionManager() without crypto, then `initialize()` replaced self.session_manager with a new SessionManager(crypto_provider=crypto). Two instances, divergent state: #1 loaded only .json, #2 loaded .enc + migrated .json->.enc. 3. Router local reference (plugins/web_ui_module/api/routes.py): `create_router()` captured session_mgr = module_instance.session_manager into a local at creation time. Because of bug #1, that snapshot was the crypto-less #1. When `initialize()` later replaced the attribute, the router closures kept pointing at the stale #1. Net effect: the UI saw #1 (only .json sessions), new sessions went through #1 and were written unencrypted, and every reboot migrated those .json files into .enc, overwriting whatever .enc already lived under the same id. Fix: - module.py: remove the premature SessionManager() from __init__; build the one real SessionManager(crypto_provider=crypto) in initialize() using whatever crypto_provider is available (may be None). One instance, one truth, one load-from-disk. - routes.py: introduce _SessionManagerProxy. Instead of capturing session_mgr = module_instance.session_manager, route closures hold the proxy and `__getattr__` re-reads module_instance.session_manager on every request. Late-binding, so the real SessionManager built in initialize() is always what the routes hit. Raises a clear RuntimeError if a route fires before initialize() has completed (rather than the opaque 'NoneType has no attribute list_sessions'). Verified: after restart, a single "Loaded N sessions from disk" per reboot (was 2 before), all .enc sessions visible in the sidebar, new sessions persisted as .enc. Data note: no encrypted sessions were lost in the wild. All existing .enc files decrypt cleanly with the current MEK; they were only made invisible by the stale reference. The only silent loss was the conversation overwritten by the colliding-id migration, which predates this fix.

…clause SQL for nosec coverage Onada 1 bandit audit — findings #2-#23 (22/23). Each remaining HIGH+MEDIUM bandit finding from MUTHUR baseline 701b4e5 was reviewed empirically by reading the actual call site, then either silenced as a false positive (19) or marked ACCEPT with explicit justification (3). Format: `# nosec B<rule>: <reason>` for FP, `# nosec B<rule>: ACCEPT — <reason>` for accepted residual risk (so bandit silences the finding while the ACCEPT prefix flags it as a conscious architectural decision rather than a tooling false positive). By rule: B104 ×2 (FP): host string comparison for allow-list construction in TrustedHostMiddleware / web_ui URL rewriting — not a network bind. Files: core/middleware.py, plugins/web_ui_module/module.py B108 ×2 FP + 1 ACCEPT: /tmp paths inside read-only allow-lists or detection arrays don't create files. The single ACCEPT is rag_logger.py last-resort fallback when both ~/Nexe-Logs and storage/logs are unwritable; mono-user local install threat model accepts the predictable-path race vs. silently disabling RAG telemetry. Files: installer/install.py, memory/memory/cli/rag_viewer.py, memory/memory/rag_logger.py B310 ×8 (FP): all are urlopen() against literal hardcoded http://localhost:{constant}/... URLs (Ollama daemon ps/tags, own server health, dev smoke-test). One additional case (core/cli/client.py) already validates the scheme against ALLOWED_URL_SCHEMES before urlopen. Files: core/cli/client.py, dev-tools/smoke_test.py (×2), installer/install.py, installer/tray.py (×2), installer/tray_monitor.py, plugins/web_ui_module/api/routes_auth.py B314 ×1 (FP): CI tooling parsing coverage.xml emitted by pytest-cov in the same workflow run. File: .github/scripts/generate_coverage_badge.py B608 ×6 (FP): the canonical IN-clause-with-dynamic-placeholders pattern, where placeholders = ",".join("?" for _ in n) and every value is bound as a parameter. The sqlite_store case additionally interpolates a table name validated by _validate_table() against a frozenset whitelist. Files: memory/memory/api/text_store.py, memory/memory/engines/persistence_sqlite.py, memory/memory/storage/sqlite_store.py, memory/memory/workers/{dreaming_cycle,gc_daemon,sync_worker}.py Note: 5 of the 6 SQL strings were also refactored from multi-line f"""...""" to single-line `sql = f"..."` + `conn.execute(sql, ...)` so the # nosec annotation lands on the line bandit reports (bandit doesn't follow # nosec across multi-line statements when the report points at an inner line). Functionally identical. B615 ×2 (ACCEPT): snapshot_download() without revision= pinning in `nexe pull` CLI path. Pinning infrastructure already exists at installer/installer_catalog_data.py:360 (MODEL_WEIGHT_SHA256 map, backlog item C19) but all entries are still None pending hash population. Realignment of CLI pull path scheduled for v1.0.5. File: core/cli/cli.py (×2) Verified empirically: - 0 HIGH+MEDIUM bandit findings remain (was 23) - Total bandit count 195 → 172 (-23 exact) - Suite: 5127 passed, 0 regressions - 22 new # nosec annotations in production (1 preexisting B110 in plugins/mlx_module/core/chat.py:617 unrelated to this audit) No behavioural changes.

…nada 4.1 Cluster 12) 13 cosmetic mypy findings closed via documented type-ignores or minimal annotations, no behavioural change: import-untyped (toml, yaml — D3 director category) - core/config.py:17 — toml lacks stubs; kept for write path (#3) - core/cli/config.py:21 — yaml stubs deferred to v1.1 (#8) - installer/install_headless.py:447 — toml lacks stubs (#15) - installer/install.py:378 — toml lacks stubs (#16) - core/cli/cli.py:325 — toml lacks stubs (#63) no-redef (tomli fallback for Python <3.11) - core/version.py:9 — # type: ignore[no-redef] (#11) - core/cli/router.py:25 — # type: ignore[no-redef] (#12) annotation/var-annotated/has-type - core/cli/i18n.py:43 — FP or-narrowing of Optional[str] or str (Director Onada 4.1 D2: # type: ignore[return-value]) (#2) - core/bootstrap_tokens.py:35 — declare _db_path / _initialized at class body and add -> 'BootstrapTokenManager' to __new__ (#6, #7) - core/paths/detection.py:44 — _detection_history: list[dict[str, Any]] (#10) - core/ingest/ingest_knowledge.py:243 — files: list[Path] = [] (#32) method-assign - scripts/bench_ingest_bug16.py:102 — benchmark monkey-patch documented with # type: ignore[method-assign] (#39)

main() had 90 NLOC (over MUTHUR 80 limit) due to orgànic creixement during the v0.9 release work. Extract 4 single-responsibility helpers preserving exact initialization order: - _set_process_title() — setproctitle call (Bug #2) - _setup_file_logging() — TimedRotatingFileHandler setup - _log_quick_commands_banner(host, port) — CLI banner - _run_uvicorn_server(...) — uvicorn.run + KeyboardInterrupt/Exception handling main() now reads as a linear sequence of well-named steps. 40/40 runner-related tests pass (server, sidecar_port_guard, pid_file, tray_helpers).

Fix-All BUS sobre 3 tracks paral·lels per resoldre tots els bugs del QA post-BUS de normalització abans del DMG v0.9.0. 8 commits dev consolidats en aquest sync. TRACK A — Memory/RAG/Sessions - Bug #1 (PID file canònic) — single source of truth a storage/run/server.pid - F5 — 3 col·leccions canòniques (nexe_web_ui, user_knowledge, nexe_documentation) creades a get_memory_api() en lloc de només la primera - F7 — ingest_knowledge defaulteja a nexe_documentation (era user_knowledge) i és idempotent (eliminada la sequence delete_collection + create_collection destructiva que esborrava docs ad-hoc dels usuaris a cada install) - F8 — root cause empíric Bug #4: MemoryModule obria un SEGON QdrantClient real a storage/vectors/qdrant_local/, divergent del singleton del pool. MEM_SAVE escrivia a una col·lecció, MEM_RECALL llegia d'una altra. Ara tots dos comparteixen storage/vectors/. - F1 — _check_duplicate retorna contracte honest (success=False, duplicate=True) enlloc de fingir success=True amb document_id=None. Era el segon root cause de Bug #4: el dedup bloquejava SAVEs amb fals positius silenciats. - F2 — typo cols list (nexe_web_ui duplicat) - F3 — list_memories scroll-based (sense semantic search amb query anglesa) - Bug #10 — collections= filter a list/save/delete (sidebar checks reals) - Bug #6 — frontend hydration document attached. Eren 2 bugs encadenats: l'endpoint /history no retornava attached_document, i removeFilePreview() feia POST /clear-document destructiu cada switch de sessió. - Bug #3 — MEM_SAVE-only response fallback. Quan el model emet només [MEM_SAVE: ...] sense text envoltant, ara genera 'Memòria desada: <fact>' perquè el bloc save s'executi i el frontend mostri confirmació. - auto_save crida eliminada per HOMAD memoria v1 (2026-04-01) — manual MEM_SAVE only fins a Part 2. TRACK B — Tray / Multi-instance / Packaging - Bug #1 (PID file) compartit amb Track A - Bug #2 — setproctitle a server i tray (server-nexe / nexe-tray a ps/Activity Monitor). Force Quit encara mostra Python perquè requereix CFBundleName via .app bundle real (deute v0.9.1). - Bug #9 — menu polish: server-nexe.com duplicat substituït per '📖 Documentació' al main level (3 idiomes), website_item es manté al submenú Configuració. TRACK C — UX cosmètic - Bug #5 — slow_request middleware exclou /ui/upload (uploads naturalment triguen >1s i el log apareixia duplicat amb l'access log d'uvicorn). - Bug #8 — 3 ⓘ visibles als checkboxes del sidebar de col·leccions amb tooltips als 3 idiomes (la infraestructura CSS/i18n ja existia). Pytest D-1 final: 4424 passed, 0 failed, 35 skipped, 1 xfailed, 86% coverage en 76.11s. Baseline pre-BUS era 4396. +28 tests nous, ZERO regressions. Tests nous: - tests/test_pid_file.py: 7 tests Bug #1 - tests/test_ingest_knowledge_idempotent.py: 8 tests F7 (3 classes) - plugins/web_ui_module/tests/test_memory_helper_async.py: 1 test F1 - plugins/web_ui_module/tests/test_memory_delete.py: 7 tests F3+Bug#10 - plugins/web_ui_module/tests/test_mem_save_injection.py: 5 tests Bug #3 Out of scope (deute v0.9.1+): - routes_chat.py 54KB decapitació general (deute formal P0) - Bundle .app real amb py2app per CFBundleName (deute v0.9.1) - Resums per capítol (Part 2 redisseny memory) - RDBMS font de veritat + vector store reconstruïble (HOMAD memoria v1, Part 2)

…rence Symptom: after a restart the sidebar only listed .json sessions (.enc were invisible), new sessions were persisted unencrypted even with encryption enabled, and a reboot's .json->.enc migration could overwrite an existing .enc belonging to a different conversation with the same id (collision observed in the wild: "Hola Diana" overwritten by "Hola Anna"). Root cause is a three-bug chain: 1. Loader early-init (core/loader/manifest_base._get_module): calls `instance._init_router()` immediately after `__init__`, *before* `initialize()` runs. The plugin must have its routable resources ready at construction time. 2. Plugin double-create (plugins/web_ui_module/module.py): `__init__` created a SessionManager() without crypto, then `initialize()` replaced self.session_manager with a new SessionManager(crypto_provider=crypto). Two instances, divergent state: #1 loaded only .json, #2 loaded .enc + migrated .json->.enc. 3. Router local reference (plugins/web_ui_module/api/routes.py): `create_router()` captured session_mgr = module_instance.session_manager into a local at creation time. Because of bug #1, that snapshot was the crypto-less #1. When `initialize()` later replaced the attribute, the router closures kept pointing at the stale #1. Net effect: the UI saw #1 (only .json sessions), new sessions went through #1 and were written unencrypted, and every reboot migrated those .json files into .enc, overwriting whatever .enc already lived under the same id. Fix: - module.py: remove the premature SessionManager() from __init__; build the one real SessionManager(crypto_provider=crypto) in initialize() using whatever crypto_provider is available (may be None). One instance, one truth, one load-from-disk. - routes.py: introduce _SessionManagerProxy. Instead of capturing session_mgr = module_instance.session_manager, route closures hold the proxy and `__getattr__` re-reads module_instance.session_manager on every request. Late-binding, so the real SessionManager built in initialize() is always what the routes hit. Raises a clear RuntimeError if a route fires before initialize() has completed (rather than the opaque 'NoneType has no attribute list_sessions'). Verified: after restart, a single "Loaded N sessions from disk" per reboot (was 2 before), all .enc sessions visible in the sidebar, new sessions persisted as .enc. Data note: no encrypted sessions were lost in the wild. All existing .enc files decrypt cleanly with the current MEK; they were only made invisible by the stale reference. The only silent loss was the conversation overwritten by the colliding-id migration, which predates this fix.

…clause SQL for nosec coverage Onada 1 bandit audit — findings #2-#23 (22/23). Each remaining HIGH+MEDIUM bandit finding from quality checks baseline 701b4e5 was reviewed empirically by reading the actual call site, then either silenced as a false positive (19) or marked ACCEPT with explicit justification (3). Format: `# nosec B<rule>: <reason>` for FP, `# nosec B<rule>: ACCEPT — <reason>` for accepted residual risk (so bandit silences the finding while the ACCEPT prefix flags it as a conscious architectural decision rather than a tooling false positive). By rule: B104 ×2 (FP): host string comparison for allow-list construction in TrustedHostMiddleware / web_ui URL rewriting — not a network bind. Files: core/middleware.py, plugins/web_ui_module/module.py B108 ×2 FP + 1 ACCEPT: /tmp paths inside read-only allow-lists or detection arrays don't create files. The single ACCEPT is rag_logger.py last-resort fallback when both ~/Nexe-Logs and storage/logs are unwritable; mono-user local install threat model accepts the predictable-path race vs. silently disabling RAG telemetry. Files: installer/install.py, memory/memory/cli/rag_viewer.py, memory/memory/rag_logger.py B310 ×8 (FP): all are urlopen() against literal hardcoded http://localhost:{constant}/... URLs (Ollama daemon ps/tags, own server health, dev smoke-test). One additional case (core/cli/client.py) already validates the scheme against ALLOWED_URL_SCHEMES before urlopen. Files: core/cli/client.py, dev-tools/smoke_test.py (×2), installer/install.py, installer/tray.py (×2), installer/tray_monitor.py, plugins/web_ui_module/api/routes_auth.py B314 ×1 (FP): CI tooling parsing coverage.xml emitted by pytest-cov in the same workflow run. File: .github/scripts/generate_coverage_badge.py B608 ×6 (FP): the canonical IN-clause-with-dynamic-placeholders pattern, where placeholders = ",".join("?" for _ in n) and every value is bound as a parameter. The sqlite_store case additionally interpolates a table name validated by _validate_table() against a frozenset whitelist. Files: memory/memory/api/text_store.py, memory/memory/engines/persistence_sqlite.py, memory/memory/storage/sqlite_store.py, memory/memory/workers/{dreaming_cycle,gc_daemon,sync_worker}.py Note: 5 of the 6 SQL strings were also refactored from multi-line f"""...""" to single-line `sql = f"..."` + `conn.execute(sql, ...)` so the # nosec annotation lands on the line bandit reports (bandit doesn't follow # nosec across multi-line statements when the report points at an inner line). Functionally identical. B615 ×2 (ACCEPT): snapshot_download() without revision= pinning in `nexe pull` CLI path. Pinning infrastructure already exists at installer/installer_catalog_data.py:360 (MODEL_WEIGHT_SHA256 map, backlog item C19) but all entries are still None pending hash population. Realignment of CLI pull path scheduled for v1.0.5. File: core/cli/cli.py (×2) Verified empirically: - 0 HIGH+MEDIUM bandit findings remain (was 23) - Total bandit count 195 → 172 (-23 exact) - Suite: 5127 passed, 0 regressions - 22 new # nosec annotations in production (1 preexisting B110 in plugins/mlx_module/core/chat.py:617 unrelated to this audit) No behavioural changes.

…nada 4.1 Cluster 12) 13 cosmetic mypy findings closed via documented type-ignores or minimal annotations, no behavioural change: import-untyped (toml, yaml — D3 director category) - core/config.py:17 — toml lacks stubs; kept for write path (#3) - core/cli/config.py:21 — yaml stubs deferred to v1.1 (#8) - installer/install_headless.py:447 — toml lacks stubs (#15) - installer/install.py:378 — toml lacks stubs (#16) - core/cli/cli.py:325 — toml lacks stubs (#63) no-redef (tomli fallback for Python <3.11) - core/version.py:9 — # type: ignore[no-redef] (#11) - core/cli/router.py:25 — # type: ignore[no-redef] (#12) annotation/var-annotated/has-type - core/cli/i18n.py:43 — FP or-narrowing of Optional[str] or str (Director Onada 4.1 D2: # type: ignore[return-value]) (#2) - core/bootstrap_tokens.py:35 — declare _db_path / _initialized at class body and add -> 'BootstrapTokenManager' to __new__ (#6, #7) - core/paths/detection.py:44 — _detection_history: list[dict[str, Any]] (#10) - core/ingest/ingest_knowledge.py:243 — files: list[Path] = [] (#32) method-assign - scripts/bench_ingest_bug16.py:102 — benchmark monkey-patch documented with # type: ignore[method-assign] (#39)

main() had 90 NLOC (over quality checks 80 limit) due to orgànic creixement during the v0.9 release work. Extract 4 single-responsibility helpers preserving exact initialization order: - _set_process_title() — setproctitle call (Bug #2) - _setup_file_logging() — TimedRotatingFileHandler setup - _log_quick_commands_banner(host, port) — CLI banner - _run_uvicorn_server(...) — uvicorn.run + KeyboardInterrupt/Exception handling main() now reads as a linear sequence of well-named steps. 40/40 runner-related tests pass (server, sidecar_port_guard, pid_file, tray_helpers).

…rence Symptom: after a restart the sidebar only listed .json sessions (.enc were invisible), new sessions were persisted unencrypted even with encryption enabled, and a reboot's .json->.enc migration could overwrite an existing .enc belonging to a different conversation with the same id (collision observed in the wild: "Hola Diana" overwritten by "Hola Anna"). Root cause is a three-bug chain: 1. Loader early-init (core/loader/manifest_base._get_module): calls `instance._init_router()` immediately after `__init__`, *before* `initialize()` runs. The plugin must have its routable resources ready at construction time. 2. Plugin double-create (plugins/web_ui_module/module.py): `__init__` created a SessionManager() without crypto, then `initialize()` replaced self.session_manager with a new SessionManager(crypto_provider=crypto). Two instances, divergent state: #1 loaded only .json, #2 loaded .enc + migrated .json->.enc. 3. Router local reference (plugins/web_ui_module/api/routes.py): `create_router()` captured session_mgr = module_instance.session_manager into a local at creation time. Because of bug #1, that snapshot was the crypto-less #1. When `initialize()` later replaced the attribute, the router closures kept pointing at the stale #1. Net effect: the UI saw #1 (only .json sessions), new sessions went through #1 and were written unencrypted, and every reboot migrated those .json files into .enc, overwriting whatever .enc already lived under the same id. Fix: - module.py: remove the premature SessionManager() from __init__; build the one real SessionManager(crypto_provider=crypto) in initialize() using whatever crypto_provider is available (may be None). One instance, one truth, one load-from-disk. - routes.py: introduce _SessionManagerProxy. Instead of capturing session_mgr = module_instance.session_manager, route closures hold the proxy and `__getattr__` re-reads module_instance.session_manager on every request. Late-binding, so the real SessionManager built in initialize() is always what the routes hit. Raises a clear RuntimeError if a route fires before initialize() has completed (rather than the opaque 'NoneType has no attribute list_sessions'). Verified: after restart, a single "Loaded N sessions from disk" per reboot (was 2 before), all .enc sessions visible in the sidebar, new sessions persisted as .enc. Data note: no encrypted sessions were lost in the wild. All existing .enc files decrypt cleanly with the current MEK; they were only made invisible by the stale reference. The only silent loss was the conversation overwritten by the colliding-id migration, which predates this fix.

…clause SQL for nosec coverage Onada 1 bandit audit — findings #2-#23 (22/23). Each remaining HIGH+MEDIUM bandit finding from quality checks baseline 701b4e5 was reviewed empirically by reading the actual call site, then either silenced as a false positive (19) or marked ACCEPT with explicit justification (3). Format: `# nosec B<rule>: <reason>` for FP, `# nosec B<rule>: ACCEPT — <reason>` for accepted residual risk (so bandit silences the finding while the ACCEPT prefix flags it as a conscious architectural decision rather than a tooling false positive). By rule: B104 ×2 (FP): host string comparison for allow-list construction in TrustedHostMiddleware / web_ui URL rewriting — not a network bind. Files: core/middleware.py, plugins/web_ui_module/module.py B108 ×2 FP + 1 ACCEPT: /tmp paths inside read-only allow-lists or detection arrays don't create files. The single ACCEPT is rag_logger.py last-resort fallback when both ~/Nexe-Logs and storage/logs are unwritable; mono-user local install threat model accepts the predictable-path race vs. silently disabling RAG telemetry. Files: installer/install.py, memory/memory/cli/rag_viewer.py, memory/memory/rag_logger.py B310 ×8 (FP): all are urlopen() against literal hardcoded http://localhost:{constant}/... URLs (Ollama daemon ps/tags, own server health, dev smoke-test). One additional case (core/cli/client.py) already validates the scheme against ALLOWED_URL_SCHEMES before urlopen. Files: core/cli/client.py, dev-tools/smoke_test.py (×2), installer/install.py, installer/tray.py (×2), installer/tray_monitor.py, plugins/web_ui_module/api/routes_auth.py B314 ×1 (FP): CI tooling parsing coverage.xml emitted by pytest-cov in the same workflow run. File: .github/scripts/generate_coverage_badge.py B608 ×6 (FP): the canonical IN-clause-with-dynamic-placeholders pattern, where placeholders = ",".join("?" for _ in n) and every value is bound as a parameter. The sqlite_store case additionally interpolates a table name validated by _validate_table() against a frozenset whitelist. Files: memory/memory/api/text_store.py, memory/memory/engines/persistence_sqlite.py, memory/memory/storage/sqlite_store.py, memory/memory/workers/{dreaming_cycle,gc_daemon,sync_worker}.py Note: 5 of the 6 SQL strings were also refactored from multi-line f"""...""" to single-line `sql = f"..."` + `conn.execute(sql, ...)` so the # nosec annotation lands on the line bandit reports (bandit doesn't follow # nosec across multi-line statements when the report points at an inner line). Functionally identical. B615 ×2 (ACCEPT): snapshot_download() without revision= pinning in `nexe pull` CLI path. Pinning infrastructure already exists at installer/installer_catalog_data.py:360 (MODEL_WEIGHT_SHA256 map, backlog item C19) but all entries are still None pending hash population. Realignment of CLI pull path scheduled for v1.0.5. File: core/cli/cli.py (×2) Verified empirically: - 0 HIGH+MEDIUM bandit findings remain (was 23) - Total bandit count 195 → 172 (-23 exact) - Suite: 5127 passed, 0 regressions - 22 new # nosec annotations in production (1 preexisting B110 in plugins/mlx_module/core/chat.py:617 unrelated to this audit) No behavioural changes.

…nada 4.1 Cluster 12) 13 cosmetic mypy findings closed via documented type-ignores or minimal annotations, no behavioural change: import-untyped (toml, yaml — D3 director category) - core/config.py:17 — toml lacks stubs; kept for write path (#3) - core/cli/config.py:21 — yaml stubs deferred to v1.1 (#8) - installer/install_headless.py:447 — toml lacks stubs (#15) - installer/install.py:378 — toml lacks stubs (#16) - core/cli/cli.py:325 — toml lacks stubs (#63) no-redef (tomli fallback for Python <3.11) - core/version.py:9 — # type: ignore[no-redef] (#11) - core/cli/router.py:25 — # type: ignore[no-redef] (#12) annotation/var-annotated/has-type - core/cli/i18n.py:43 — FP or-narrowing of Optional[str] or str (Director Onada 4.1 D2: # type: ignore[return-value]) (#2) - core/bootstrap_tokens.py:35 — declare _db_path / _initialized at class body and add -> 'BootstrapTokenManager' to __new__ (#6, #7) - core/paths/detection.py:44 — _detection_history: list[dict[str, Any]] (#10) - core/ingest/ingest_knowledge.py:243 — files: list[Path] = [] (#32) method-assign - scripts/bench_ingest_bug16.py:102 — benchmark monkey-patch documented with # type: ignore[method-assign] (#39)

main() had 90 NLOC (over quality checks 80 limit) due to orgànic creixement during the v0.9 release work. Extract 4 single-responsibility helpers preserving exact initialization order: - _set_process_title() — setproctitle call (Bug #2) - _setup_file_logging() — TimedRotatingFileHandler setup - _log_quick_commands_banner(host, port) — CLI banner - _run_uvicorn_server(...) — uvicorn.run + KeyboardInterrupt/Exception handling main() now reads as a linear sequence of well-named steps. 40/40 runner-related tests pass (server, sidecar_port_guard, pid_file, tray_helpers).

Synced from dev since sync-20260519. Highlights: Security (Tier S — security audit 96/100 PUSH OK): - S2 XSS api_key UI sanitization (web_ui_module/ui/app.js) - S3 /installer/finalize legacy gate via idempotency marker - S9 path traversal reject in _resolve_model_path - S10 SQLiteStore thread-safe via check_same_thread + RLock - S1 health URL boot fix - S6 fsync before close on atomic writes - S7 CancelledError catch + mid-startup signal logging - TOCTOU atomic on /installer/finalize idempotency marker - basename guard applied to all model_id download pipelines Bug fixes vespre (2026-05-21 sessions): - Bug #5 fix(qdrant): retry-with-backoff lock acquisition at startup - Bug #4 test(live): order slow LLM tests after fast ones + 10s cooldown - Bug #2+#3 test(security): replace fragile inspect-based assertions - Bug #1 chore(installer): document accepted CVEs + portable OSV checker - Bug A fix(chat): remove data:[DONE] sentinel from text/plain stream - Bug B iter-2 fix(chat): natural-language date phrase replaces "Now:" tech UI productiva (F5.5 revert + post-revert): - web_ui_module serves full UI in sidecar mode again - footer i18n + version persistence - frontend reads nexe_api_key from query string on first launch F5.3 + F5.4 onboarding wizard: - HTTP endpoints for wizard (installer/finalize, preflight, progress) - NEXE_VERSION UI injection - onboarding_state.py + optional HF token via macOS Keychain (never on disk) - installer_constants.py + installer_progress.py + check_cves_osv.py Hygiene: - 113 hardcoded "Nexe 0.9" → version centralized (UI) - /v1/system/* blocklist on sidecar URL guard - ruff/mypy/pyright/typos cleanup (multiple commits) 54 files changed total. Nous tests: 17 (security regression sentinels + onboarding wizard + qdrant + installer + sqlite concurrent + sqlite store singleton).

jgoy-labs force-pushed the pre-production branch from aa487bd to 65c88c2 Compare February 21, 2026 15:19

jgoy-labs force-pushed the pre-production branch from 65c88c2 to d80343a Compare February 21, 2026 15:34

jgoy-labs merged commit 6574f63 into main Feb 21, 2026
0 of 2 checks passed

jgoy-labs deleted the pre-production branch February 21, 2026 15:34

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

security: auditoria completa 0.8 — vulnerabilitats, hardening i tests#2

security: auditoria completa 0.8 — vulnerabilitats, hardening i tests#2
jgoy-labs merged 1 commit into
mainfrom
pre-production

jgoy-labs commented Feb 21, 2026 •

edited

Loading

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

Conversation

jgoy-labs commented Feb 21, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Resum

🔴 Crítiques resoltes

🟠 Altes resoltes

🟡 Mitges resoltes

🟢 Tests de seguretat (14/14 passing)

Altres millores incloses a la branca

Pla de tests

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

jgoy-labs commented Feb 21, 2026 •

edited

Loading