diff --git a/PROJECT.md b/PROJECT.md index 1db6f2a..468cc10 100644 --- a/PROJECT.md +++ b/PROJECT.md @@ -794,6 +794,10 @@ AI получает JSON-массив: 2026-07-30 один пост (`id=2348`) был платно отправлен в Sonnet 4.6 983 раза, потому что failed снова попадал в claim и цикл сразу продолжал работу. Failed посты нужно возвращать в `pending` только явным ручным действием. +Sonnet 4.6 в этом кейсе возвращал `category_id=18` (`Не целевой контент`) и +пустой `text`, объясняя отказ в `notes`; это валидный AI-отказ, а не ошибка +рерайта. Проверка минимальной длины текста применяется только к целевым +категориям. После failed AI batch worker помечает конкретные посты как `failed` и продолжает обрабатывать другие `pending` посты. Ошибка показывается на `/workers` из heartbeat meta и отправляется Telegram-уведомлением получателям diff --git a/STATE.md b/STATE.md index ea03f7d..4f4e46a 100644 --- a/STATE.md +++ b/STATE.md @@ -103,6 +103,7 @@ As of 2026-07-29: - Current live writer model setting checked on 2026-07-29: `ai_writer_model = anthropic/claude-sonnet-4-6`. - Sonnet 4.6 can return `rewrites` as a JSON string instead of a list. `validate_rewrites()` in `src/vk_parser_app/workers/ai_writer.py` now parses stringified `rewrites` before validation. - 2026-07-30 incident: post `2348` was retried 983 times by AI writer because automatic claim included `rewrite_status='failed'`; estimated internal cost was `$16.465977`. Live setting `ai_writer_enabled` was switched to `false` manually to stop spend. Automatic AI qualifier/writer claims must not include `failed`; failed posts require an explicit manual reset to `pending`. AI writer/qualifier keep running after a failed batch, show the error on `/workers`, and send a Telegram alert via daily-report/TG-poster bot recipients. +- Sonnet 4.6 may return `category_id=18` (`Не целевой контент`) with empty `text` and the explanation in `notes`. This is valid AI rejection, not a writer error; `validate_rewrites()` allows short/empty text only for non-target category. ## Prompt Architecture diff --git a/src/vk_parser_app/workers/ai_writer.py b/src/vk_parser_app/workers/ai_writer.py index 39c29f0..54b0bec 100644 --- a/src/vk_parser_app/workers/ai_writer.py +++ b/src/vk_parser_app/workers/ai_writer.py @@ -152,8 +152,6 @@ def validate_rewrites( if post_id not in expected_ids: raise ValueError(f"AI returned unexpected id={post_id}") text = strip_leading_emoji_markers(str(item.get("text") or "").strip()) - if len(text) < 40: - raise ValueError(f"AI returned too short rewrite for id={post_id}") by_id = {int(c["id"]): c for c in categories if c.get("id") is not None} by_name = {str(c["name"]).lower(): c for c in categories if c.get("name")} if item.get("category_id") is not None: @@ -170,6 +168,8 @@ def validate_rewrites( seen.add(post_id) category_tag = normalize_hash_tag(category_row.get("tag") or category_row.get("name") or "", "category") reject_by_category = int(category_row["id"]) == NON_TARGET_CATEGORY_ID or category_tag.lower() in NON_TARGET_CATEGORY_TAGS + if not reject_by_category and len(text) < 40: + raise ValueError(f"AI returned too short rewrite for id={post_id}") out.append( { "id": post_id,