postprocess-and-packaging

name

postprocess-and-packaging

description

Optional agent post-processing and deliverable assembly: `services/postprocess/` (refine.py, glossary_check.py, cover.py, date_research.py, _srt_guard.py and their prompt .md files), `services/finalize/` (SRT → styled ASS + Netflix-TC punctuation), and `services/package/` (burn-in, cover copy, noise/remix). Read this before changing refine, glossary check, cover generation, finalize punctuation or ASS styling, or packaging/burn-in behavior.

Post-processing, finalize & packaging

Post-processing (services/postprocess/)

Optional agent passes, each a thin orchestrator over run_inference where the agent reads/writes files in the project dir and we validate afterward. _srt_guard.py validates structure only (block count, indexes, timecodes, non-empty text) — semantic quality is the agent's responsibility via prompts. Agent-written SRTs may carry a UTF-8 BOM (Codex does this); every reader of video.cht.refined.srt / video.cht.glossary_checked.srt — including finalize — must read with utf-8-sig.

• refine.py — polish TC subtitles (AGENT_POSTPROCESS_MODEL).
• glossary_check.py — full-text terminology/factual consistency check after refine. Runs when enabled unless video.cht.glossary_checked.srt already exists, treats Latin/kana blocks only as priority hints, may use web search or on-demand frames, and may correct .pre_pass/pre_pass.json after preserving the original as .pre_pass/pre_pass.raw.json (backup happens once; the updated pre-pass must still validate against PrePassResult). Its prompt ends with a required name-form audit (honorific/name-span/subject parity vs video.ja.srt) as the final defense against dropped honorifics, full-name expansion, and invented subjects (refine also mandates the subject check per window). When video.official.ja.srt exists (platform CC, see project-architecture), the conditional official_subtitle_reference.md block is appended: CC wording outranks ASR as source evidence. If the SRT or pre-pass changed, glossary_check.md (the report) must exist; packaging copies reports only if present.
• cover.py — stylize the poster. Always Codex (image generation), regardless of the backend settings; effort comes from AGENT_COMMON_MODEL. Runs async from the workflow (see project-architecture for the ThreadPoolExecutor/join rules).
• date_research.py — broadcast-date web research fallback (AGENT_COMMON_MODEL, called with web_search=True and cwd=None so the agent gets a throwaway temp dir, never the project dir). Unlike the others the agent writes no files: it returns schema-validated JSON (DateResearchResult: status/date/trust tier/sources), Python persists it to .artifacts/date_research.json (fixed-filename cache — a parseable file skips the agent, a corrupt one counts as a miss and is overwritten; delete to re-run) and apply_date_research_result writes Project.broadcast_date on the main thread at join time (below-high trust is adopted with a warning). Runs async from the workflow like cover.

__init__.py uses lazy __getattr__ imports so importing the package doesn't drag in every backend.

Finalize (services/finalize/)

SRT → styled ASS + cleaned SRT. Two jobs, not one:

• Netflix-TC punctuation rules (strip terminal commas/periods, collapse ellipses, convert mid-line 。→，, etc.). The ASS style header lives here.
• Mixed-script name spacing: pulls Latin-containing names from pre_pass.json (proper nouns + characters) and from the fixed glossary (_curated_name_units, longest-first to avoid substring collisions) and applies deterministic spacing. So services/fixed_glossary/ is a runtime input here, not just prompt content.

Input precedence (enforced in workflow.py, not here): glossary-checked SRT → refined SRT → translated SRT — first that exists wins.

Packaging (services/package/)

Post-loop deliverable assembly (not a pipeline stage): burn ASS into the video (core.py), copy cover and analysis artifacts (pre_pass.json, optional refine.md/glossary_check.md — silently skipped when absent), plus a noise/remix packaging path (noise.py, remix.py). Burn-in itself lives in services/media.py (duration-validated; see project-architecture).