Auto PR dev to main/master

Version updated for https://github.com/yuri-val/auto-pr-action to version v1.4.0.

  • This action is used across all versions by ? repositories.

Action Type

This is a Composite action.

Go to the GitHub Marketplace to find the latest changes.

What’s Changed

Changes in this Release:

  • feat: default to gpt-5.6-luna and tune prompt for GPT-5.6

  • Switch the default model from gpt-5.4-mini to gpt-5.6-luna

  • Rewrite the system prompt per the GPT-5.6 prompting guide: lean instructions, real newlines (previously literal \n sequences), clear output contract (summary + emoji sections, no title/preamble)

  • Set reasoning_effort “low” and raise max_completion_tokens to 4096 so reasoning tokens don’t starve the visible answer on large releases

  • Expose pr_number as a composite action output via outputs.value (previously declared but never propagated)

Co-Authored-By: Claude Fable 5 noreply@anthropic.com

  • fix: bound the model payload so large releases don’t 400

A big release (the dev→master diff after a 256-commit feature branch merge) produced a ~2.3MB diff. Capped at 1MB it was still ~300k+ tokens of dense code+Cyrillic, overflowing the model’s context window → OpenAI returned HTTP 400 (context_length_exceeded), which the action reported only as the opaque “OpenAI API request failed”.

Changes:

  • Build a compact, high-signal payload: commit log + per-file diffstat first, then the unified diff. The log/stat (~50KB here) always survive truncation.
  • Lower the cap to a context-safe default (max_diff_bytes input, 200000) and truncate UTF-8-safely (head -c | iconv -f UTF-8 -t UTF-8 -c) so a multibyte char split mid-sequence can’t yield invalid UTF-8 (another 400 cause).
  • Drop the explicit temperature: 0.7 — the gpt-5 family only accepts the default, so a non-default value is itself a 400 risk; the default is fine for description generation.
  • Capture the HTTP status + API error body (no more –fail-with-body swallow) so failures like context_length_exceeded are visible in the logs.

Co-Authored-By: Claude Opus 4.8 (1M context) noreply@anthropic.com