Problem: the dashboard was Jev-centric; NLU/arbiter/fallback DeepSeek, STT
providers and the audio-enhancer were barely visible, and stored texts/voices
had no real access path.
Bot (additive; reply path untouched):
- nlu.result now records latency_ms + the DeepSeek usage tokens
- matching.arbitration/fallback.result fill the model + latency_ms columns
(payload unchanged; SQL coalesces column/payload so old rows still count)
- storage.py: MinIO dual-write layer — every voice (raw + denoised) and user
attachment is mirrored fire-and-forget into bucket bot-media; the blocking
minio SDK runs via asyncio.to_thread (senior-engineer review) so the event
loop is never stalled; empty MINIO_* env = layer off; bucket auto-created
- requirements: minio>=7.2
Backend (admin-gated, read-only):
- GET /bot-metrics/models — one record per model (jev, deepseek-nlu,
deepseek-arbiter, deepseek-fallback, audio-enhancer + dynamic per-provider
STT) with calls/ok-rate/p50/p95/tokens/last-seen + version-drift vs the pin
- GET /bot-metrics/models/{id}/daily — per-model daily calls/fails/tokens/p50
- GET /bot-metrics/stages — every stage tagged with its Persian pipeline
section (صوت/درک پیام/تطبیق/ابهام و انتخاب/پیشنویس)
- /traces now carries input_kind + a 120-char text excerpt (consensus →
input.text → first STT transcript)
- media serving: POST /media-url mints a short-lived HMAC URL (audio/a tags
cannot send Authorization); GET /media streams from MinIO bot-media with
the read-only ./bot/data mount as fallback for pre-MinIO files; hard ref
validation (voices|attachments only, no .., extension whitelist)
- router exempted from the global Bearer dependency (files.py pattern) —
every endpoint keeps its own require_role('admin')
Frontend /bot-monitor (RTL, recharts):
- four tabs: نمای کلی / مدلها (records table + per-model expandable daily
charts: calls/fails, tokens, p50) / بخشها (sections grouped) / پیامها
- message drawer rebuilt: inline <audio controls> for raw + denoised voice,
download links, every STT transcript + consensus side-by-side, user
attachments, then the event timeline
- terminology: «ترس» → «پروندهی پیام» with tooltips + a one-line legend
Grafana: bot-observability gains a «کارکرد مدلها» table (stage→model
mapping for legacy rows); the explorer gains a message text column.
Seed: real (tiny) ogg files written + mirrored to MinIO, NLU usage tokens,
column-level model/latency on matching events.
Verified locally: /models returns all 7 records incl. per-STT and NLU tokens;
media 200 via signed URL (MinIO) and 401 without; ../ and .py rejected (400);
disk fallback serves a legacy file; pm-role 403 on every new endpoint;
bucket auto-created with 14 mirrored objects; tsc + vite build clean;
Grafana re-provisions 24 panels.
Categories are kept only as the label each supplier carries (suppliers.category_id
+ the material_categories list it points at). Everything that treated category as
a shared key across materials, inventory, quotes and unit conversions is removed —
including two features that were never populated (0 rows on every volume).
Removed (backend):
- routers category_attributes, category_ref_units, unit_conversions + registration
- ?category_id= filters on /materials, /inventory and /quotes (the /suppliers one
stays — it filters suppliers by their own label)
- GET /inventory/materials/options: required a category and had no caller
- the material-payload category derivation (expand_service.category_by_material,
the inventory serializer lookup) and the material category on /pm/on-the-way;
the wire keys stay, now always null, so the payload shape is unchanged
- startup DDL: DROP the empty category_attributes / category_reference_units /
unit_conversions tables (idempotent) and their CREATE statements; seed.py
truncate list updated
Removed (frontend):
- the category-keyed conversion path: refUnitForCategory / findConversion /
refConverted, the categoryUnits + unitConversions store slices, api clients,
hooks and the "≈ converted quantity" line on quote detail and project pages
(it needed rows in those tables to render, and there were none)
- material.supplier_category and its read sites (the edit modal grew one
yesterday to restore labels; the label lives on the supplier now)
- the inventory page's category chips, the row filter and the «دسته» column,
plus the lookup's categories fetch and catName helper
Kept: supplier label + its list and picker, /suppliers?category_id=, the global
standard_conversions feature and the units page (material_units conversions).
Verified against a restore of the production dump: startup DDL drops exactly the
three tables, suppliers keep their label (3/3), material_categories still serves
its 8 rows, the removed endpoints 404, and materials/inventory/quotes/units/
standard-conversions all return 200 with quote line items intact.
Downloads require an Authorization header, which <iframe>/<img> cannot
send. The previous blob: workaround broke under download managers (IDM
hooks blob URLs, fails to resolve them, and leaves a 0-byte file) and in
Chrome's mobile emulation, which has no inline PDF plugin.
Backend:
- preview_token.py: HMAC-signed token bound to one (quote, kind, file)
with a 5-minute TTL, keyed off an existing secret
- GET /quotes/{id}/preview-url/{kind}/{attachment} mints the URL
- the two file GET routes accept a Bearer session OR the signed ?t= token
(they leave the global auth dep, so the other six file routes now carry
it explicitly — nothing is open by omission)
Frontend:
- api.files.previewUrl() replaces the blob helper
- shared FilePreviewModal + useFilePreview; FileList and the dashboard
quote card both use it
- PDFs render with the browser's native viewer on desktop and pdf.js on
canvas where no inline viewer exists (mobile); pdf.js is fed the bytes
in memory so no PDF URL is exposed for a download manager to intercept
- removed the per-attachment download buttons on the dashboard card and
the modal header
Ops:
- nginx: explicit MIME types (.mjs as text/javascript) — a module worker
served as octet-stream is refused and crashed the page
- frontend service bypasses the nginx entrypoint, whose apk manifest call
hangs when the package mirror is unreachable (502s)
- slowapi limiter on POST /auth/login (10/min per IP, Persian error msg);
edge-level coarse limits stay in Traefik
- mem_limit on every main-compose service (~4.4GB worst case, protects the
8GB shared host during rebuilds)
- monitoring images pinned to verified tags (prom v3.4.1, grafana 11.6.3,
loki/promtail 3.5.0, node-exporter v1.9.1, postgres-exporter v0.17.1,
cadvisor v0.49.1); exporter image name fixed to postgres-exporter
- db-backup sidecar: pg_dump -Fc of procurement+casdoor every
BACKUP_INTERVAL_SECONDS into BACKUP_DIR, retention via BACKUP_KEEP_DAYS
- audit_log table (migration.sql + runtime ensure_audit_table) with
model + transactional helper (utils/audit.record) — wired into user
create/role-change and inventory transaction creation
- get_actor: quote_events.actor now stores the VERIFIED user's display
name instead of the raw bearer token
Wizard:
- failed saves now route to the exact step+field: a new data-field attribute on every required section (supplier_id, line_items, termsheets, project, quote_date, expiry_date, notes, proforma) plus resolveErrorTarget() that maps ApiError.field / message keywords (supplier, line item/material/unit/quantity, termsheet/schedule/milestone, project, dates, notes, file) to the offending step — the «رفع خطا» toast action scrolls/focuses the right control instead of guessing
- material records are no longer created on the fly while typing (rememberMaterial no longer fires addMaterial per blur); materials/specs/units are written once during the final save, avoiding half-created catalog rows
Accept flow:
- quote_service no longer rewrites LineItemProjects on accept — those candidate-project rows stay 100% untouched and only item_decisions.project_id is written, matching the per-item target-project model and the UI payload (item_projects: [])
Dates/time:
- JalaliDatePicker: hour/minute are now editable text fields (Persian digits, clamping, 3-digit retype handling) with a controlled useEffect that only rewrites the inputs when the numeric value actually changes — no more wiping in-progress typing; date-only detection tightened to a full YYYY-MM-DD match so datetime strings are not misread
Also: /api validation logging (RequestValidationError) now includes the path and error list for faster diagnosis.
Verified: api + frontend image builds (python, tsc, vite) clean on top of origin/production (incl. the team's mobile layout merges).
whqreza
ZCode