CHANGELOG

[1.2.1] - 2026-08-18

Public release. License switched from proprietary to MIT.
README (EN/DE): new "Use Cases" section — including cleaning unintended AI-typical text markers combined with honest AI disclosure.
SECURITY.md: public reporting channel (GitHub Security Advisories), documented default host is localhost.
New docs/ai-act-note.md (scope and intended use under the EU AI Act).
Git history rewritten before publication to remove an internal infrastructure address from early commits.

disable_thinking option (default: on) — thinking-capable models (qwen3.x, deepseek-r1, ...) spend most of their latency on hidden reasoning tokens, pointless for mechanical restyling. Measured on qwen3.5:4b: the same file dropped from 647s to 7.9s (~80x). The client sends think: false and transparently falls back for models that reject the parameter. Toggle via pastapress config --thinking on|off. Note: on small models, disabling thinking can cost some phrasing quality — content fidelity was unaffected in testing, but grammar got rougher; use --thinking on or a larger model when polish matters more than speed.

Critical: llm_client.py contained an unterminated triple-quoted string (an orphaned legacy SYSTEM_PROMPT block swallowing the class definition), which made the entire package unimportable (SyntaxError). The CLI and all tests were broken as of 1.0.0.
Leading/trailing whitespace of text chunks (indentation, boundary spaces) is now preserved around the LLM call — previously lost via response stripping.
Directory mode no longer re-enqueues previously generated *_pasta-press.* output files (avoided double-pressing on repeated runs).
Removed duplicated output-path logic in core.process_file (now uses get_output_path) and a redundant second reassemble call.
Plain-text reading now handles UTF-8 BOM (utf-8-sig) and falls back to cp1252 for legacy Windows files instead of failing.

.doc support removed — Pandoc has no reader for legacy binary .doc; the tool now fails with a clear message asking for .docx instead of pretending support.
Default Ollama host is now http://localhost:11434; config.json is no longer tracked in git (auto-created on first run, see config.example.json).
LLMClient.process_text now raises LLMProcessingError after exhausting retries instead of silently returning the original text. PastaPressCore keeps the original chunk, counts failures (last_failed_chunks), and the CLI reports partial failures. If all chunks fail (e.g. Ollama unreachable), no output file is written and the command exits non-zero.
Oversized paragraphs (no blank-line delimiters) are now split at line and word boundaries (secondary split) so single huge paragraphs no longer exceed the chunk limit.
Model auto-selection now also recognizes qwen model families.
Directory mode reports how many files were actually enqueued.

Test suite expanded from 4 to 36 tests: chunker secondary-split and lossless roundtrip invariants, LLM client retry/error paths (mocked at HTTP level), core end-to-end file processing, document parser encodings, and CLI behavior (queue isolation, failure exit codes, config persistence).

Core LLM text refinement logic via TextChunker and LLMClient.
Seamless chunking preserving text structure and formatting delimiters.
Support for gleichwertig, wissenschaftlich, einfach, kurz, and original text styles.
Translation mode functionality.
Queue manager for batch directory processing.
pypandoc integration for automatic .docx, .odt, .rtf, .doc to .md parsing.
CLI application using click (process, process-queue, text, config).
Auto-detection of available Ollama models.
Bilingual documentation and GitHub repository setup.