> ## Documentation Index > Fetch the complete documentation index at: https://docs.decktalk.ai/llms.txt > Use this file to discover all available pages before exploring further. # The stability promise > What DeckTalk promises a script, a GitHub Action or a service that builds on it, and what it keeps as a private cache. A caller that builds on DeckTalk reads four things and nothing else: the `--json` result of each command, the events file of each run, the files in `build/final/`, and the exit code. Those four keep their shape within a `schema` version, and a change to one of them is a new `schema`. Everything else under `build/` is DeckTalk's private cache. Its files and their shapes are what [Build artifacts](/reference/artifacts) describes, which is how the engine works today and not a promise. A newer release that cannot read a cache file builds it again, except the paid records below. ## The result of a command Every command prints one flat JSON object with `--json`, and every library call returns the same object as a frozen model. | Key | What it promises | | - | - | | `schema` | The shape version of the object, which is `1`. A reader checks it before it reads anything else. | | `ok` | True exactly when the command exits 0, under the threshold `--fail-on` and `--allow` set. | | `findings` | Every judgement the command made, in the order it made them. | | `error` | Filled only when the command could not run, with `code`, `message`, `hint` and `docs`. | | `run` | The id of the run the command opened, on every command that opens one. It names the events file. | | `written` | Every file the run wrote, project-relative, on every command that writes one. | `decktalk schema NAME` prints the JSON Schema of one command's result, and `decktalk schema` with no name prints every command with the result it answers with, the exit codes and the error codes. The same schemas are committed in the repository under `schemas/v1/results/`, one file per command, in the folder the `schema` number names. ```bash theme={null} uv run decktalk build --json | python -c "import json,sys; r=json.load(sys.stdin); print(r['schema'], r['ok'], r['run'])" ``` ## The events file Every run of a project appends one JSON object per line to `build/events/.jsonl`, where `` is the `run` of the result. `--events` prints the same lines on stderr as they happen. Every line carries `event`, which names its kind, and `time`, `seq` and `run`. `seq` counts within the run, so a reader can tell a gap from a lost line. The last line of a run is `run.done`, with the run's `outcome` and, when it failed, the same `error` the result carries. `decktalk schema event` prints every kind and every field, and the same schema is committed in the repository as `schemas/v1/events.json`. ## `build/final/` The deliverables, named after `[project] name`, which is `` below. | File | What it is | | - | - | | `.mp4` | The film. It is renamed into place only once it is whole. | | `.srt` | The captions, as SubRip. | | `.vtt` | The captions, as WebVTT. | | `.chapters.txt` | The chapter markers, as ffmetadata. | | `-transcript.html` | The whole film as one page, which is the media alternative. | | `-poster.png` | The opening slide, as a lossless PNG. | | `placements.json` | Where every section sits in the finished film. | ## Exit codes | Exit | Meaning | | - | - | | 0 | The command ran and found nothing at the threshold. `ok` is true. | | 1 | The command ran and found something at or above the threshold `--fail-on` set. | | 2 | The command was refused, which is `USAGE` or `APPROVAL`, and a retry as written is refused again. | | 3 | The command could not run, which is every other error code. | | 130 | The caller stopped the run, which is `CANCELLED`. | ## The paid records Four kinds of file record money spent, so they are never treated as a cache. Each stays where it is and is never deleted by DeckTalk. | Kind | File | Folder | | - | - | - | | Take audio | `.mp3`, one per voiced take, named by the digest of what was sent | The takes directory, `takes/` unless `[narration] takes_dir` names another, with a second copy in the take store, `[narration] store_dir` | | Provider words | `.words.json`, the words the speech provider sent back with its take | Beside its take, in the takes directory, `takes/`, and the take store, `[narration] store_dir` | | Score audio | `.mp3` for the ambience bed and each effect, and `music-part.mp3` for each part of the music, named by the item the score stage bought | The score directory, `score/` unless `[score] dir` names another, or the path a sound's own `out` names | | Ledger | `ledger.json`, what the score stage has bought, keyed by the request that made it | Beside the score audio, in the score directory, `score/` | The take store sits outside every project, and a take is written there first, once, when it is bought. A take's suffix is the one its voice declares for the format it asks for, which is `.mp3` for every ElevenLabs format. A newer release reads the words files and the ledger or refuses them with a sentence. It never builds one again and never deletes one, because counting a paid record as absent would buy what it records again. The refusal says so, and leaves the file where it is for the person whose money it is. A take no section plays any more stays too, and `decktalk status` lists it so its author can remove it with `git rm`. The take index, `takes.json`, is not one of them. It is a cache over the takes on disk, which `narrate` builds again from them when it does not read, so a broken index buys nothing. ## Related * **Every command, its flags and its output:** [CLI](/reference/cli) * **Every field of every event line:** [Python API](/reference/python-api) * **What the cache holds today:** [Build artifacts](/reference/artifacts) This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.