owarai-grillmaster

The unified model-inference layer under `services/inference/`. Read this before touching `run_inference`, any backend file (`gemini_api.py`, `gemini_cli.py`, `gemini_agy.py`, `codex.py`, `claude_sdk.py`), capability gating in `base.py`, the schema validate-and-repair loop in `schema_enforce.py`, or the agent frame tools under `services/inference/tools/` (`get_frames*.py`). Also read it when adding or renaming a `Backend`, or changing how audio/images/schema/prompts reach a model.

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.

Orchestration-level architecture of the Owarai GrillMaster pipeline: the resumable stage machine in `workflow/`, project state and path layout in `project.py`, settings/`.env`/ModelSpec in `settings.py`, the Typer CLI in `main.py`, and the supporting services (`services/srt/`, `services/media.py`, `services/ytdlp/`, `services/elevenlabs/`, `services/fixed_glossary/`, `services/progress.py`). Read this before adding/reordering a pipeline stage, changing resumability or cost accounting, adding a setting, adding a source platform, or any task that spans more than one service module. Deep dives live in the sibling skills inference-layer, translate-pipeline, and postprocess-and-packaging.

The two-stage translation package under `services/translate/` — pre-pass whole-film analysis (`pre_pass/`), concurrent chunk translation (`chunk/chunk_worker.py`), chunking (`chunker.py`, `facade.py`), structural validation and agent repair (`validation.py`, `structural_fix.py`), chunk caches under `.chunks/`, `pre_pass.json`, and the prompt `.md` templates. Read this before changing chunk boundaries, cache behavior, PrePassResult schema, chunk validation, or any prompt under `services/translate/`.

Parse and audit Codex rollout/session JSONL files. Use when the user asks whether Codex really called tools such as shell_command, web.run, image/file tools, MCP tools, or browser tools; wants to distinguish prompt mentions from real response_item function_call execution; wants local image/input summaries; or wants to inspect visible assistant messages, encrypted reasoning boundaries, final outputs, and tool-impact evidence from large Codex logs.

Parse and audit Gemini CLI session JSONL files and tmp session directories. Use when the user asks whether Gemini CLI really called tools such as get_frames, run_shell_command, read_file, or google_web_search; wants to distinguish prompt mentions from real tool execution; wants command/timestamp/image/search summaries; or wants to inspect thoughts/final output from large Gemini CLI chat logs safely.