ss-tools

Files

busya 58ac89c21e fix(translate): token budget overhaul + truncation retry + smart batch sizing + LLM provider fixes

Token Budget (_token_budget.py):
- DEFAULT_MAX_OUTPUT_TOKENS 8192→16384 (adequate for 50 rows×4 langs)
- Add PROVIDER_DEFAULTS with per-model context_window/max_output_tokens
- OUTPUT_PER_ROW_PER_LANG 60→120, add JSON_OVERHEAD_PER_ROW=50
- PROMPT_BASE_TOKENS 300→600, MAX_OUTPUT_HEADROOM 1000→3000
- CJK-aware token estimator (_estimate_tokens_for_text: 1.5 chars/tok for CJK)
- Output-aware batch sizing (_apply_output_aware_batch_sizing)
- Warn on dictionary cap mismatch

Executor (executor.py):
- Return finish_reason from _call_openai_compatible→_call_llm
- Truncation detection + batch splitting on finish_reason=length
- Smart batch sizing (_auto_size_batches): greedily split by row token budget
- Fix disable_reasoning hardcoded 8192→use calculated max_tokens
- TypeError guard on choices[0] access, base_url validation
- Broadened response_format fallback (matches response_format|structured|json_object)
- Remove invalid extra_body, fix partial-recovery regex for integer row_id
- 18 new tests for estimate_row_tokens and _auto_size_batches

Preview (preview.py):
- Align token constants with _token_budget.py (CHARS_PER_TOKEN=2.2, OUTPUT=120)
- Return finish_reason, broadened response_format fallback
- Fix hardcoded 8192, base_url validation

LLM Services:
- render_prompt: detect unfilled {placeholders}, log WARNING
- llm_provider: distinguish crypto exceptions (InvalidTag vs ValueError vs generic)
- Respect Retry-After header on HTTP 429

2026-05-16 09:58:31 +03:00