v3.1.3

New Features

Agent Mode (-r, --structured agent)

A two-pass agentic translation mode built on timestamp.

Pass 1 - Planning: Scans the full subtitle file in max-batch-size chunks. For each chunk the model produces a batch summary (character names, locations, events, tone, dialect) and decides a natural batch boundary. Summaries accumulate and are consolidated when they exceed the token budget. At the end of the scan, a refined system instruction is generated that filters the glossary and stylistic notes down to only what was observed in the file.

Pass 2 - Translation: Runs identically to timestamp mode using the enriched instruction and the agent-determined batch boundaries from Pass 1.

Best suited for content with recurring characters, specialized vocabulary, or stylistic consistency requirements. Costs additional API calls for the planning pass. Progress file resumption is not supported.

cli/translator.mjs -i subtitles.srt -r agent --from Japanese --to English

Other Changes

OPENAI_DEFAULT_MODEL environment variable

A new optional env var that sets the default model instead of hardcoding gpt-4o-mini. Useful when you always use a different model without passing -m on every invocation.

OPENAI_DEFAULT_MODEL=gpt-4o

OPENAI_API_RPM default raised to 500

The default requests-per-minute limit was raised from 60 to 500 to better match typical API tier limits. The .env.example value is now commented out (the built-in default applies unless you override it).

The fallback order for the moderator RPM was also corrected: OPENAI_API_MODERATOR_RPM is now checked before OPENAI_API_RPM (previously the two were swapped).

Context selection now uses real token counts

Context history chunking (-c, --context) switched from using model-reported completion token counts as a proxy to counting tokens directly with gpt-tokenizer. This makes the budget more accurate and independent of whether a prior response recorded token usage.

Full Changelog: v3.0.2...v3.1.3