subtitle-studio

A self-contained multilingual subtitle translation workflow plugin for dsh (DeepSeek Harness), built around the "everything is a plugin" idea. It parses and writes SRT/VTT subtitles, translates them sentence-by-sentence through a configurable LLM, merges bilingual outputs, validates alignment, and processes whole directories — exposed both as dsh tools (ctx.tools) and as a dependency-free CLI.

subtitle-studio is a new, self-contained implementation. Its subtitle parsers are written from scratch (no heavy subtitle libraries), its LLM layer speaks OpenAI-compatible HTTP out of the box (DeepSeek by default) and can plug into the harness's ctx.llm seam, and everything is UTF-8 on both ends.

---

Table of contents

• Features
• Installation
• Using subtitle-studio as a dsh bundle
• CLI
• Configuration
• Glossary format
• Library & service API
• Testing
• Limitations
• License

Features

1. Parsing & writing — SRT and WebVTT with a tolerant state machine:

   • accepts missing sequence numbers, missing blank separators, CRLF, and a UTF BOM (UTF-8 / UTF-16LE / UTF-16BE);
   • keeps multi-line cue text and VTT identifiers/settings verbatim;
   • skips junk lines and malformed timecode lines, reporting them as issues instead of aborting;
   • output is always UTF-8 and round-trips preserve timing to the millisecond.

2. Sentence-by-sentence translation — batched requests over a character budget, JSON payloads keyed by cue index for reliable alignment, bounded retries (with a corrective nudge on malformed JSON), timeouts, and glossary injection. Works with any OpenAI-compatible endpoint (DeepSeek is default) or the dsh ctx.llm seam.

3. Bilingual subtitles — merge original + translation either stacked (translation lines under the original) or interleaved (alternating cues), with optional per-line tags and a separator. The timeline is copied verbatim from the source, so it is preserved by construction.

4. Alignment validation — checks translation count vs. source, missing and extra cues, empty translations, overlong cues that should be split, and timeline overlaps.

5. Batch processing — translate whole directories with a bounded concurrency pool, per-file retry with backoff, an atomic checkpoint file for pause/resume, and token/cost estimation (with approximate per-model rates that can be overridden).

6. Two entry points — five dsh tools (sub_parse, sub_translate, sub_merge, sub_export, sub_glossary) plus a full CLI.

Installation

Requirements: Node.js ≥ 18.18.

# 1. install dev dependencies and build
npm install
npm run build

# 2. run the CLI directly
node bin/subtitle-studio.js --help

# or install the bin globally (makes `subtitle-studio` available)
npm link

The package ships zero runtime dependencies. TypeScript is only a dev dependency (build to lib/ with tsc).

Using subtitle-studio as a dsh bundle

The package is a valid dsh bundle: package.json declares the bundle manifest, cordis.patch.yml inserts the plugin row into a profile, and lib/index.js exports the Cordis-style entry (name + apply(ctx, config)).

1. Add the bundle to a profile

dsh plugin --profile <name> add subtitle-studio

This installs the package and appends subtitle-studio to the profile's dsh.profile.bundles. The profile loader then applies our patch (the cordis.patch.yml shipped by the bundle), which inserts a row:

- insert:
    - id: subtitle-studio
      name: subtitle-studio
      config:
        targetLanguages: []
        sourceLanguage: ''

2. What the plugin registers

• Tools on ctx.tools (registered when the tools service is available):

  tool	purpose
  sub_parse	parse a subtitle file into structured cues (JSON output)
  sub_translate	translate a subtitle file into one or more languages (LLM call)
  sub_merge	combine a source subtitle with a translation payload (stacked/interleaved)
  sub_export	convert/export a subtitle to SRT or VTT (UTF-8), optionally bilingual
  sub_glossary	manage the JSON terminology glossary (list/add/remove/merge)

• Service subtitleStudio (when ctx.provide exists) exposing parse, translate, merge, validate, glossary, cost, stringify, convert — a convenient library surface for other plugins.

3. Plugin configuration

The plugin reads the config object of its patch row (and any profile override). Keys:

key	type	default	meaning
llm.provider	"openai"|"dsh"	"openai"	HTTP backend vs. harness ctx.llm seam
llm.baseUrl	string	https://api.deepseek.com/v1	OpenAI-compatible endpoint
llm.apiKey	string	—	API key; supports ${ENV_VAR} expansion
llm.model	string	deepseek-chat	model name
llm.timeoutMs	number	120000	per-request timeout
llm.maxRetries	number	2	retries on transient failures
llm.jsonMode	boolean	true	request a JSON object response
llm.chunkChars	number	3500	cue-character budget per request
sourceLanguage	string	"" (auto)	source language tag
targetLanguages	string[]	[]	target languages for translation
glossary.paths	string[]	[]	glossary JSON files (merged)
output.layout	stacked|interleaved	stacked	bilingual merge layout
output.separator	string	""	line between original & translated blocks
output.tagTarget	string	—	per-line tag prefix for translated lines
output.format	srt|vtt	keep source	output container format
output.utf8Bom	boolean	false	write a UTF-8 BOM
batch.concurrency	number	2	parallel files during batch
batch.maxRetries	number	2	per-file retries
batch.checkpoint	string	subtitle-studio.checkpoint.json	checkpoint path
validation.maxChars	number	160	CJK over-long cue threshold
validation.maxWords	number	40	Latin over-long cue threshold

Example patch override (in the profile's own cordis.patch.yml, whole-row config replacement — see the dsh docs on patch layer semantics):

- id: subtitle-studio
  config:
    llm:
      provider: openai
      baseUrl: https://api.deepseek.com/v1
      apiKey: ${DEEPSEEK_API_KEY}
      model: deepseek-chat
    sourceLanguage: en
    targetLanguages: [zh, ja]
    output:
      layout: stacked

To use the harness's own LLM provider (e.g. the DeepSeek adapter configured in the harness), set llm.provider: dsh. The plugin then calls ctx.llm.stream.

4. Notes for harness authors

• The entry never imports @deepseek-ai/cordis; the context is consumed structurally, so the same source compiles with or without the harness installed. Type-augmentation fans can add declare module '@deepseek-ai/cordis' { interface Context { subtitleStudio: ... } } themselves.
• Tools are registered through ctx.inject(['tools'], sub => …) so the plugin starts even in a minimal profile; nothing hard-requires ctx.llm unless provider: dsh is set.

CLI

subtitle-studio <command> [options]

All commands accept --config <file.json> to load the plugin-shaped config, plus per-command flags (which override config).

parse

subtitle-studio parse movie.srt
subtitle-studio parse movie.vtt --pretty       # full JSON

translate

# one target language -> bilingual stacked SRT
subtitle-studio translate movie.srt \
  --target zh --source en \
  --glossary glossary.json \
  --output movie.zh.srt

# multiple targets -> fmt: <stem>.<target>.bilingual.<ext>
subtitle-studio translate movie.vtt --target zh --target fr --output movie.bilingual.vtt

# interleaved layout + shownote line tags
subtitle-studio translate movie.srt --target zh --layout interleaved --tag "[zh] " --output out.srt

# API key from environment
subtitle-studio translate movie.srt --target zh --api-key ${DEEPSEEK_API_KEY} --output out.srt

# persist a mid-run partial checkpoint
subtitle-studio translate movie.srt --target zh --output out.srt --save-partial partial.json

Flow: parse → validate timeline → print cost estimate → translate → validate alignment → merge → write bilingual subtitle + <name>.translation.json.

merge

# from a translation JSON file
subtitle-studio merge movie.srt translation.json --layout interleaved --output merged.srt

# or inline JSON
subtitle-studio merge movie.srt '{"entries":[{"index":1,"text":"你好"}]}' --layout stacked

export

# convert srt -> vtt preserving timestamps
subtitle-studio export movie.srt --output movie.vtt

# bilingual export
subtitle-studio export movie.srt --output movie.vtt --translation translation.json --layout interleaved

validate

subtitle-studio validate movie.srt                    # timeline sanity + overlaps
subtitle-studio validate movie.srt --compare translation.json  # alignment against a translation
subtitle-studio validate movie.srt --overlong         # over-length cue report
subtitle-studio validate movie.srt --no-overlap       # skip overlap checks

glossary

subtitle-studio glossary list --path glossary.json
subtitle-studio glossary add --path glossary.json --source "DeepSeek Harness" --target "深度求索工具链" --scope zh
subtitle-studio glossary remove --path glossary.json --source "DeepSeek Harness"
subtitle-studio glossary merge --path glossary.json --with other.json

batch

# dry-run cost estimate
subtitle-studio batch ./movies --output-dir ./out --target zh --estimate

# real run with 3 parallel workers and resume-capable checkpoint
subtitle-studio batch ./movies --output-dir ./out \
  --target zh --glossary glossary.json \
  --concurrency 3 --checkpoint cp.json --layout interleaved

# resume after an interruption (skips done files, retries failed)
subtitle-studio batch ./movies --output-dir ./out --target zh --checkpoint cp.json --resume

Batch notes:

• The output directory mirrors the input tree, so two files that share a basename (e.g. a/clip.srt and b/clip.srt) never collide (out/a/clip.bilingual.srt, out/b/clip.bilingual.srt).
• The output directory itself is excluded from scanning, so a batch never re-translates its own artifacts (out/x.bilingual.bilingual.srt can't happen).
• --resume re-queues files left processing by a crash, skips done files, and (with --retry-failed, the default) retries failed files with a fresh attempt budget. The checkpoint is written atomically after every file.

cost

subtitle-studio cost movie.srt --target zh
subtitle-studio cost ./movies --target zh --target fr --model deepseek-chat --rate-in 0.27 --rate-out 1.10

Configuration

The CLI mirrors the dsh plugin configuration (see the table above) through --config or flags. API keys can reference environment variables:

node bin/subtitle-studio.js translate a.srt --target zh --api-key "${env:DEEPSEEK_API_KEY}"

Prices in the cost estimator are approximate list prices keyed by model (deepseek-chat, deepseek-reasoner) and are always overridable with --rate-in/--rate-out (or llm config equivalents).

Glossary format

A glossary is a JSON document:

{
  "name": "sample-glossary",
  "entries": [
    { "source": "DeepSeek Harness", "target": "深度求索工具链", "scope": "zh", "note": "official product name" },
    { "source": "hello", "target": "bonjour", "scope": "fr" },
    { "source": "bilingual", "target": "双语" }
  ]
}

• source → target map a term to its mandated translation.
• scope restricts an entry to one target language; entries without a scope apply to every language. This is how one glossary serves multiple targets.
• During translation the applicable entries are injected into the system prompt as mandatory terminology. The entry identity is (source, target, scope) — upserting the same key replaces it, different keys coexist.
• See examples/glossary.json and run subtitle-studio glossary list.

Library & service API

The public entry lib/index.js re-exports the engine pieces directly as functions:

import {
  parseSubtitle, stringifySubtitle, convertSubtitle, detectFormat,
  translateCues, translateDocument,
  mergeBilingual, mergeWithEntries,
  validateSubtitle, validateTranslationAlignment,
  loadGlossaryFile, mergeGlossaries, buildGlossaryPrompt,
  createLlmClient, estimateCost,
} from 'subtitle-studio'

All timestamps are integer milliseconds; all file I/O is UTF-8.

Testing

npm test

Runs tsc then the Node built-in test runner over test/*.test.js (122 tests covering time parsing, SRT/VTT tolerance, encoding/BOM, merge, validation, glossary, translation with a mocked HTTP backend, batch/checkpoint, and the CLI + dsh tool/plugin surface).

Limitations

• The token counter is a heuristic; use --rate-in/--rate-out for accurate billing, and treat cost prints as estimates.
• Interleaved bilingual output intentionally reuses the exact source timings for both cues, so the overlap validator flags them — validate the source document for overlaps, or exclude bilingual docs from overlap checks.
• jsonMode requires an endpoint that supports response_format; the client automatically retries without it if the provider rejects it with HTTP 400.
• The dsh LLM provider path assumes the harness ctx.llm.stream chunk shape; other chunk layouts degrade to generic text extraction.

License

MIT — see LICENSE.