Skip to content

CHANGELOG.md

Latest commit

 

History

History
82 lines (54 loc) · 6.61 KB

CHANGELOG.md

File metadata and controls

82 lines (54 loc) · 6.61 KB
Raw

Changelog

1.3.0 — 2026-07-28

Upgrade notes

No flags, environment variables or model names were removed, but three behaviours change:

  • A partly translated run now exits 1 instead of 0. Previously only a run with no successful translations at all failed, so a .po with untranslated strings still reported success. Check any pipeline that treats a non-zero exit as fatal.
  • The default model is now gpt-5.4-mini. Output and per-token rates both change: against this pricing table, input is 87.5% higher and output 181.25% higher. Pin the previous behaviour with --model gpt-4.1-mini or MODEL=gpt-4.1-mini.
  • GPT-5 models now work. A configuration naming one previously failed every request; it will now do real, billable work.

Added

  • --allow-unknown-model (or ALLOW_UNKNOWN_MODEL=true) accepts a model that is not in the pricing catalogue, costed at the fallback rate. Without it, a model missing from config/openai-pricing.json is rejected, which would otherwise mean waiting for a release to use a newly published model.

Fixed

  • GPT-5 models could not be used at all. Every request sent max_tokens, which the GPT-5 and o-series families reject. Requests now send max_completion_tokens for those models and keep max_tokens for the supported GPT-4.1 models.
  • temperature is now omitted for models that accept only the default value (the GPT-5.6 and GPT-5.5 families, the GPT-5 base family, and the o-series), which previously failed the request outright.
  • Reasoning models were starved of completion budget. The token limit was sized from estimated visible output, but on the max_completion_tokens path that same limit also pays for hidden reasoning, which is largely fixed rather than proportional to the batch. Small batches therefore truncated: in our tested configurations gpt-5-nano and gpt-5-mini failed every run at batch sizes 1 to 5 while batches of 10 or more succeeded. Those models now get a flat headroom allowance. Raising the limit is not prepaid, though any extra tokens a model does generate are billable.
  • A truncated response was parsed as zero translations and written out as blank strings while the run reported success. Truncation is now detected via finish_reason and fails the batch without retrying, since an unchanged retry is unlikely to help and incurs another charge.
  • Potomatic's parser required an exact <f0> tag, so a malformed opening tag such as <f0"> sent the entry down the singular path and blanked every plural form, shipping it untranslated. Form tags are now matched tolerantly while still requiring the tag name to end, so <f10> is not read as form 1 and <f0evil> is not read as form 0.
  • Token counting fell back to a rough length / 4 estimate for GPT-5 names that tiktoken does not recognise. It now falls back to the o200k_base encoding those models use, which matches the API's own token accounting exactly in our tests.
  • Token encoders were rebuilt after every successfully returned response, costing roughly 60 ms each. They are now created once per model and shared.
  • --temperature 0 and TEMPERATURE=0 were silently replaced by the default, because zero was treated as unset.
  • The A/B prompt tool crashed with translations.map is not a function on every run, because it treated the { translations, validationStats } object returned by parseXmlResponse() as an array.

Changed

  • Documentation no longer describes providers other than OpenAI. The README and .env.example advertised Gemini and Anthropic keys with automatic provider detection, but --provider gemini fails outright and a GEMINI_API_KEY was only ever used as an OpenAI key. The --provider flag still works for openai and is now hidden from --help until there is a second provider to choose.
  • OpenAI models and pricing updated to July 2026, adding the GPT-5.6 (Sol, Terra, Luna) and GPT-5.5 tiers.
  • The model catalogue lists only models the chat completions endpoint serves, which excludes the -pro tiers. None of them were listed in 1.2.0, so no existing configuration is affected.
  • tiktoken updated to 1.0.22, which recognises more model names directly. Counts are unchanged, since every name it knows resolves to the same o200k_base encoding used for the rest.
  • The ab-prompt-test npm script has been removed. tools/ is not published, so the script pointed at a file that was never in the package; run node tools/ab-prompt-test from a checkout instead.
  • A batch that failed after being billed now reports what it cost instead of zero, so --max-cost accounts for it, and the failure message states the number of attempts actually made.
  • The A/B prompt tool no longer hardcodes its two models, which were outside the supported catalogue. Models, temperatures and target language are set with AB_MODEL_A, AB_MODEL_B, AB_TEMP_A, AB_TEMP_B and AB_LANG.

1.2.0 — 2026-04-06

Fixed

  • Output directory was not being created when saving .po files if it didn't exist.

Added

  • Post-translation validation that catches placeholder mismatches, non-breaking space issues, and plural form errors - blanking bad translations for automatic retry.
  • New --extra-prompt-path CLI flag to append domain-specific prompt content without replacing the base prompt.
  • Language-specific plural form guidance in the default prompt (Arabic 6-form, Russian 3-form, French 2-form, Japanese 1-form).
  • XML source elements now include a placeholders attribute listing expected placeholders for LLM verification.

Changed

  • Update OpenAI models and pricing to April 2026.
  • Default model changed from gpt-4o-mini to gpt-4.1-mini.
  • Hardened default prompt with explicit negative examples for placeholder preservation.
  • Blanked translation details now logged at warn level instead of debug.

1.1.0 — 2025-11-12

Fixed

  • Fix plural-form translations in languages with complex rules (e.g., Arabic, Russian,).
  • Treat environment variables such as FORCE_TRANSLATE=false correctly instead of interpreting them as true.
  • Ensure configuration files are detected when running Potomatic from paths containing spaces (such as ~/Local Sites/) and on Windows systems.

Added

  • Support multiple API key formats (OPENAI_API_KEY, POTOMATIC_OPENAI_API_KEY, etc.) with automatic provider detection based on the key name.
  • Add CLI options --prompt-file-path and --po-header-template-path to customize the locations of translation prompt and header template files.

Changed

  • Update OpenAI pricing data to reflect current model costs (November 2025).

1.0.1 — 2025-06-27

1.0.0 — 2025-06-27

  • Initial release.