Vectorii corpusului k-NN persista in tabela embedding_cache (PK model+text_hash,
blob float32 LE); la warmup se vectorizeaza doar textele lipsa din cache, deci
restartul cu corpus neschimbat nu mai plateste ~1-2 min de embed (embed=0).
- app/embedding_cache.py: serializare array('f'), load/save/purge chunk 500 cu
BEGIN/COMMIT explicit (conexiuni autocommit), validare dimensiune la scriere
si citire, orchestrare sync_corpus_vectors cu embed_fn injectat
- index_corpus(vectors=): vectori precalculati cu validare aliniere; mismatch
-> fallback embed complet
- ensure_embeddings_corpus: warmup in thread la startup (block=True), calea de
request ne-blocanta (acquire non-blocking pe lock; warmup in curs -> return
imediat); purjare orfane + modele vechi doar dupa indexare reusita
- log warmup: cache=N embed=M in Xs
- 31 teste noi (cold/warm/incremental, model schimbat, concurenta, ranking
exact, echivalenta float32); suita completa 1596 passed
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Curatare globala a comentariilor si docstring-urilor (app, tools, teste,
scripturi): eliminate referintele la PRD-uri, US-xxx, task-uri istorice si
review-uri; pastrata doar informatia functionala, formulata scurt. Regula
adaugata in CLAUDE.md (sectiunea Stil). Fara modificari de cod sau comportament.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Findings security-review P1/P2 (2026-07-03):
- headere de securitate pe toate raspunsurile (nosniff, X-Frame-Options,
Referrer-Policy, HSTS doar pe HTTPS) + teste
- body-cap global 10MB ca middleware ASGI pur (413 inainte de parserul
multipart/JSON; verificarea per-endpoint ramane strat 2)
- imagine Docker non-root (uid 10001), port 8010 aliniat, loguri pe
volumul /data
- fail-fast la boot cu rar_env=prod fara AUTOPASS_REQUIRE_API_KEY sau
AUTOPASS_SESSION_SECRET
- compose: env-uri critice obligatorii (:?) ca api/worker sa nu diverga
tacit; FORWARDED_ALLOW_IPS ca rate-limit-ul sa vada IP-ul real dupa
Traefik
- signup fara PII in stdout: log_event in loc de print cu email (idem
notify degradat)
- ratelimit: sterge cheile fara timestamp-uri valide (crestere monotona
a memoriei pe IP-uri reale)
- backup criptat SQLite (backup online API, gpg AES256) + verificare
restore + docs/backup.md
Suita completa verde: 1557 passed, 1 skipped (live).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Comutator accounts.auto_send_enabled per cont: Auto OFF (default) tine randurile
la ingestie (submissions.held=1), worker-ul (claim_one AND held=0) le sare pana la
eliberare umana (per rand/bulk/auto-release OFF->ON). Snapshot held prin chokepoint
unic held_for_account pe toate caile de ingestie (API, import, reresolve, reactivare).
- schema/migrare: coloana held + index partial idx_submissions_held; auto_send_enabled
- API: echo onest held+motiv (US-010), ruta /prezentari/{id}/trimite-acum
- web: toggle header, modal confirmare tipata, buton Trimite per rand + Trimite toate,
banner coada tinuta imbatranita (L.142), contor "In asteptare (manual)"
- worker: expire_held (US-008, inchide gaura retentie PII), metrics held gauges
- ops: tools/carantina_held + runbook rollback (R4)
Nota review (/code-review high): re-snapshot held lipseste pe caile repune/corectie
(requeue_submission, post_corectie, bulk-fix) — de aliniat separat cu create_prezentari.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Eliminat zgomotul de trasabilitate (US-xxx, PRD x.x, Rn, OV-x, Tn, decizii/naratiune
istorica) din 41 fisiere app/ + template-uri. Pastrate comentariile care documenteaza
invarianti si logica ne-evidenta (idempotenta/hash, reconciliere anti-duplicat, RAR 500
esec definitiv, creds per cont, WAF User-Agent, 422 fara echo de parola, scope NULL->1),
curatate doar de tokeni.
Verificare: pentru cele 27 module .py curatate, structura de cod (tokeni non-comentariu/
non-string) e IDENTICA fata de HEAD -> doar comentarii/docstring-uri schimbate. Singura
schimbare de cod e in tests/test_web_responsive.py (scos 3 assert pe markeri US-006/007/008,
inlocuite de asertiunile structurale alaturate). 0 tokeni US/PRD reziduali in app/.
Regresie: 896 passed, 1 deselected.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
O cheie AUTOPASS_CREDS_KEY setata dar invalida (format Fernet gresit)
arunca ValueError abia la primul encrypt_creds -> 500 brut pe
POST /v1/prezentari, fara mesaj util (cazul reprodus din client VFP).
crypto.validate_creds_key() valideaza cheia, apelata in main.lifespan:
o cheie invalida opreste pornirea cu mesaj clar + comanda de generare,
in loc sa explodeze la prima cerere. Cheie nesetata = OK (model efemer).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Canalul web trece de la 100% deschis (hardcodat cont 1) la autentificat si
multi-tenant. Un service nou se inregistreaza din browser, primeste o cheie API
(o singura data) si o sesiune; contul se creeaza "in asteptare" (active=0) si nu
trimite la RAR pana la activarea de catre admin (tools/account.py activate).
- users + app/users.py: parole scrypt (salt per-user, eticheta parametri onorata
la verify pentru migrare cost), email unic case-insensitive
- sesiune: SessionMiddleware (same_site=strict, https_only config) + app/web/session.py
(current_account/web_account/require_login->LoginRequired, set_session clear-inainte)
- CSRF (app/web/csrf.py) enforce in prod inclusiv pe login/signup + rate-limit
in-proces (app/web/ratelimit.py) pe signup si login
- signup/login/logout (app/web/auth_routes.py): signup tranzactie atomica,
cheie-o-data, log SIGNUP pentru descoperire admin
- dashboard + import scoped pe contul sesiunii (regula NULL->cont 1); toate rutele
web care ating date sensibile sub require_login; nomenclator ramane global
- banner "cont in asteptare" pentru conturi active=0
- gate worker: claim_one LEFT JOIN accounts COALESCE(active,1)=1 (account_id NULL=activ)
VERIFY context curat (2 runde): leak cross-account /_fragments/mapari prins+reparat.
/code-review high: csrf_token lipsa pe re-randari de eroare, scrypt_params ignorat,
login fara rate-limit -- toate reparate. 361 teste pass (de la 313).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Dashboard-ul incarca htmx din unpkg.com. Gateway-ul ruleaza offline (LXC/VPS
+ Cloudflare Tunnel); cand CDN-ul e blocat/inaccesibil, htmx nu se incarca si
dashboard-ul ramane static: zero polling (banner alerta blocate, coada,
mapari), butoanele de mapare nu trimit. Aceeasi decizie offline-first ca fontul
UI (FINDING-002, fara dependinta CDN).
Fix: vendorizez htmx.org@1.9.12 minificat in app/web/static/, montez StaticFiles
la /static in main.py, base.html refera /static/htmx.min.js. Verificat live:
window.htmx.version=1.9.12, fragment polls 200, zero regresie vizuala.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Redactare:
- handler RequestValidationError dropeaza input/ctx din 422 (vectorul de
scurgere a rar_credentials.password pe /v1/prezentari); pastreaza type/loc/msg
- app/security.py: scrub/scrub_text + CredentialRedactingFilter pe root+uvicorn
- models.py: password cu repr=False
Auth API-key:
- app/auth.py: hash SHA-256 in api_keys (cheia in clar emisa o singura data),
header X-API-Key / Authorization: Bearer, dependency resolve_account_id
- enforcement pe flag AUTOPASS_require_api_key (prod on->401, dev off->cont
default id=1; cheie prezenta invalida->401 mereu)
- account_id real curge din cheie in ingestie + mapare
- tools/apikey.py: CLI create/rotate/revoke/list (fara endpoint HTTP admin)
16 teste noi (tests/test_security.py). 85 pass total.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>