Backend:
- /api/validation/ CRUD for validation tasks (policies)
- Run history with 6 filters + pagination
- Alembic migrations (provider_id, policy_id)
- 28 pytest tests
Frontend:
- /validation — task list page with status filters + CRUD
- /validation/[id] — task config (Config + History tabs)
- /validation/history — run history with metrics
- Sidebar nav entry under Operations
- i18n en/ru
Playwright (ScreenshotService):
- Removed 25s hardcoded sleeps → conditional waits
- CDP fallback to full_page screenshot
- Total timeout enforcement (120s)
- Debug screenshots → temp dir with cleanup
QA fixes:
- Debug screenshot accumulation in production (tempdir + rmtree)
- Frontend API payload field name mismatch
- History issues field reference fix
- delete_runs scoped by policy_id
Belief protocol:
- @RATIONALE/@REJECTED on 27 C4+ contracts
Closes: VALIDATION-REWORK-2026
- .env.example: единый источник истины для всех переменных окружения
- .gitignore: добавлено исключение !.env.example (правило .env* не применяется)
- Все ранее внесённые правки аудита (7 классов) уже закоммичены в 68740cd
- Итог: 33 QA-теста, 0 хардкод-секретов, 0 .bak, 0 except:pass
Why: tier rules forbidding @PRE/@POST on C2 cause agents to either
remove useful documentation or generate audit noise. Both outcomes are
worse than having slightly richer metadata.
Changed:
- Tier table in all 4 agent prompts: removed 'Forbidden' column,
replaced with 'descriptive signal, NOT gatekeeper'
- All tags (@PRE/@POST/@RATIONALE/@REJECTED) welcomed at any tier
- semantics-core: complexity table collapsed to descriptive
- qa-tester: removed tier-based reject rules from P1 and Phase 1
Hard rules preserved: CONTRACT-FIRST, ANCHOR SAFETY, FRACTAL LIMIT,
RESURRECTION BAN. Everything else is descriptive intent.
All 5 functions now have proper #region/#endregion contracts:
- _detector_cache_key [C:1] — private helper
- build_detector [C:2] — public API
- get_detector [C:2] — public API with caching
- detect_language [C:2] — public API, @RAISES for safety
- batch_detect [C:3] — public API with @RELATION edges
Clean audit: 0 warnings on _lang_detect.py
- Split sidebar into 3 groups: Resources / Operations / System
- Added Migration section (Overview + Mappings) - was hidden in routes
- Added Git section (Repository Status) - was hidden in routes
- Added Tools section (Mapper, Debug, Storage, Backups) - was hidden in routes
- Moved Settings gear icon to footer (quick link to /settings)
- Moved Profile to footer (avatar + username)
- Moved Collapse button to header (always visible)
- Organized Admin with all sub-items (Users, Roles, ADFS, LLM)
- Reduced visual noise: section labels collapse independently
- Kept existing features: RBAC filtering, health badges, mobile overlay, collapse animation
- Updated tests to match new structure
- Updated i18n (en/ru) with new keys
- TranslationRunGlobalIndicator: replace fromStore + 6x with
(() => store.subscribe()) to prevent accumulating render_effect
instances via createSubscriber that caused infinite Svelte reactive flush
- translate/[id]/+page.svelte: add runComplete flag to hide progress bar
after completion without touching the store; add collapsible run detail
cards with status badge, counts, and error_message; filter invalid runs
- translationRun.js: remove stale guard (no longer needed)
- Sidebar.svelte: fix mobile overlay close on route change
- docs/adr/ADR-0007-rejected-fromStore-derived.md: document REJECTED pattern
- frontend: add to reload job data on param change after goto
- frontend: set existingJob from createJob response for immediate tabs
- frontend: suppress error toasts for schedule 404 (normal state)
- frontend: conditional ScheduleConfig mount to avoid 'new' job loading
- backend: skip preview check for jobs with direct Superset datasource
- backend: add orthogonal test for datasource+no-preview success case
- Extract _fallback_cancel_request for row lock timeout recovery in orchestrator_cancel
- Add flush lock timeout fallback in cancel_run
- Set error_message when all records fail in executor
- Track insert_failed state in scheduler and frontend store
- Migrate TranslationRunProgress from local polling to centralized store
- Fix nginx resolver for Docker variable-based proxy_pass
- Add FRONTEND_HOST_PORT env var to docker-compose
- Sidebar.svelte: defer healthStore.refresh() until feature flags confirm
health_monitor is enabled; skip entirely when disabled
- health.js: remove throw error from catch block — log and return null
instead, preventing 'Uncaught (in promise)' on expected 404
- test_report_audit_immutability.py: fix mock assertions —
audit_service uses logger.reason/reflect/explore, not logger.info
- HealthStore and Sidebar now produce zero network noise when
FEATURES__HEALTH_MONITOR=false
- _compute_source_hash now accepts context_keys parameter
- source_data is filtered to only include context-relevant fields (not row_id, primary keys)
- If no context_columns configured, hash is based on source_text + dict_hash + config_hash
- This ensures identical text in different rows produces a cache hit
- Add source_hash column (SHA256 of source_text+source_data+dict_hash+config_hash) to TranslationRecord for cache dedup
- Add Alembic migration b1c2d3e4f5a6 for the new column and composite index (source_hash, status)
- Implement _compute_source_hash / _check_translation_cache in executor.py — before LLM call, check if same source+dict+config combo was already translated successfully
- Implement cache-aware routing in _process_batch: full cache hit → skip LLM, partial hit → route to LLM
- Store source_hash on all new TranslationRecord rows for future cache hits
- Strengthen dictionary prompt instructions from 'use when applicable' to 'MUST use — mandatory'
- Add _enforce_dictionary() post-processing: force-replace dictionary terms in LLM output if LLM ignored them
- Fix CRITICAL: use run.dict_snapshot_hash/config_hash (exist on TranslationRun, not TranslationJob)
- Fix MAJOR: verify cached languages cover ALL target_languages before accepting cache hit
- Fix MAJOR: prevent regex back-reference injection in dictionary enforcement (use lambda)
- Fix: add joinedload to cache lookup to avoid N+1 queries
- Fix: remove redundant single-column index (composite index is sufficient)
- Fix(frontend): parse paginated /translate/dictionaries response correctly (result.items)
- Add mask_api_key() and is_masked_or_placeholder() to llm_provider service
- Return masked keys in all provider CRUD endpoints
- Reject masked/placeholder keys in fetch_models and test_provider_config
- Show masked key with Change button in ProviderConfig.svelte edit form
- Exclude masked keys from fetch-models, test, and submit payloads on frontend
- Update semantics-core skill with clarified complexity tier rules
- Switch agent modes from subagent to all
- Show full stats grid (total/success/fail/skip) + progress bar + cancel button during active run
- Compact colored banner for terminal states (completed/partial/failed/cancelled) with auto-dismiss
- Import cancelRun from translate API for functional cancel button
- fromStore() reactivity confirmed working — no store changes needed
- Hide on /translate/[id] to avoid duplication (existing behavior preserved)
- Remove resetTranslationRun() from handleRunComplete so the store keeps its terminal state
- Global indicator (TranslationRunGlobalIndicator) now shows 'Completed' for 5s then auto-dismisses
- Page navigation while run is active keeps the indicator visible until terminal state
- Add RUN_CANCELLED event logging in orchestrator when executor returns CANCELLED (cancel flag path)
- Skip TranslationLanguage creation for languages matching detected source language (no ru in stats)
- Update tests for new behavior (3 entries instead of 4, no fr entry for source-match)
- Fix clickhouse insert test mocks for .options(joinedload()) chain
- All 208 translate tests pass
- Add LITELLM = "litellm" to LLMProviderType enum (already in HEAD from prev commit)
- Add comment in LLMClient.__init__ explaining standard Bearer auth for LiteLLM
- Add regression test test_provider_type_enum_values — guard against value drift
- Add test_litellm_client_uses_default_bearer_auth — verify no extra headers
- Clean up duplicate @RELATION lines in models.py
- All 20 backend + 4 frontend tests pass
- Add _normalize_timestamp_value to key column values in orchestrator.py and executor.py before SQL generation
- Fix test mocks for .options(joinedload()) chain, explicit None attrs for mock job
- Add global translation run progress indicator (TranslationRunGlobalIndicator + store)
- Fix translate page import missing translationRunStore
- All 208 translate tests pass
- Enable response_format (json_object) for all providers including Kilo/OpenRouter
- Skip reasoning_effort for Kilo/OpenRouter (returns 400 — mandatory reasoning)
- Add structured_outputs fallback: retry once without response_format if upstream
provider (e.g. StepFun) rejects it with 'structured_outputs is not supported'
- Handle model refusal field (refusal) instead of treating as empty content
- Fix None-handling guards for message/finish_reason/content fields
- Apply same fixes to both preview.py and executor.py _call_openai_compatible
- Connect orphaned TermCorrectionPopup, BulkReplaceModal, BulkCorrectionSidebar
The job config page was capped at 896px (max-w-4xl), making the
10×4 translation preview table extremely squished. Changed to
max-w-full xl:max-w-7xl — full width on all screens, 1280px max
on ultrawide.
- Removed max-w-[200px] compression on source text column → now
gets proportional space
- Scrollable containers (max-h-32 overflow-y-auto) for long text cells
- Language columns: min-w-[200px]→[250px] with whitespace-pre-wrap
- Source language badge and status columns remain narrow (w-28, w-24)
DeepSeek v4 Flash returns finish_reason=length when reasoning uses
all output tokens. Partial row extraction via regex recovers complete
rows from truncated JSON instead of failing with 400 error.
Also: extra_body reasoning_effort for Kilo/OpenRouter proxied providers
Added disable_reasoning toggle that works via:
1. API parameter: reasoning_effort: 'none' (DeepSeek, Qwen)
2. System prompt instruction: 'Respond directly with ONLY the JSON.
Do NOT include any reasoning, thinking, or chain-of-thought'
(works for ALL models universally)
Backend: model column, schema, preview+executor payload
Frontend: checkbox in LLM settings, i18n en+ru
New _token_budget.py calculates safe batch_size and max_tokens based on source text length, target languages, dictionary size, and model context window (64K DeepSeek v4 Flash). preview.py: auto-reduces sample_size when texts are long. executor.py: per-batch dynamic max_tokens. 12 new tests.
finish_reason=length from DeepSeek v4 Flash — model runs out of
output tokens due to reasoning_content (CoT) + multi-language (4 langs).
8192 tokens ensures complete JSON response for preview and execution.
Preview + executor now log finish_reason, message_keys, and raw
response structure when LLM returns empty content, enabling
diagnosis of DeepSeek API issues (empty content with finish_reason=length)
Preview + executor now log the raw LLM response (first 1000 chars)
when JSON parsing fails, enabling debugging of malformed LLM output.
Previously ValueErrors from JSON parse failures were silently returned
as 400 without any diagnostic data in logs.
Multi-language save fix:
- MultiSelect.svelte: selected prop changed to $bindable([])
(without $bindable, bind:selected never propagated to parent component,
so target_languages always stayed as default ['en'])
Target table column mapping (INSERT enhancement):
- target_language_column — stores language code (e.g. 'ru', 'en')
- target_source_column — stores original source text
- target_source_language_column — stores detected BCP-47 source language
- SQL generator includes these columns in INSERT when configured
- Frontend: 4 input fields under 'Target Column Mapping' section
with clear labels, placeholders, and i18n hints
Backend: model, schema, service, orchestrator
Frontend: job config page, i18n en+ru
DB: 3 new columns added to production
