[piotrekno1]

For 0.3.0. Merging this into release/0.3.0 adds it to the release PR #98.

Problem

• The integrations cut error_message to its first 200 characters. The public handle.track_error() applied no limit, so apps that track errors by hand had to truncate the message themselves.
• I measured 37 typical failures across openai, anthropic, onnxruntime, torch, llama.cpp, faster-whisper, mlx-whisper and huggingface_hub. The median message is 123 characters, but 9 are over 200, and several put the cause at the end:
  • mlx-whisper audio errors (about 1,000 characters) are the ffmpeg banner followed by the actual error. The first 200 characters hold only the banner.
  • torch load_state_dict errors (301) end with the mismatched shapes.
  • Hugging Face Hub errors (209 to 507) start with about 100 characters of request ID.

Fix

• ErrorEvent.to_dict() applies the limits, so every error event gets them, from an integration or from a manual track_error() call.
• ERROR_MSG_MAX_LEN goes from 200 to 500. A longer message keeps its start and its end, with … between them.
• provider_error_code is capped at 64.
• The integrations and decorators no longer slice the message themselves, so the limit is applied in one place.
• The transmitter and attachment log lines still use ERROR_MSG_MAX_LEN to cap server responses in logs, so those can now be up to 500 characters.

Note

The Android and Swift SDKs still use 200 with a head-only cut. They should follow.

Testing

• New tests in tests/test_event_serialization.py cover the cap, start-and-end truncation and short messages.
• Full suite: 913 passed, 28 skipped.
• The 3 measured messages over 500 characters keep their cause after truncation.
• Against the local ingest, a 2,100-character manual error was accepted (202). It arrived as 500 characters ending with its cause, with a provider code of 64 characters.

Release notes from CHANGELOG.md

The changelog preview on #98, and the GitHub release notes the tag build creates, listed every commit subject since v0.2.2. That included merge commits ("Merge 6c3a54e… into c3896d9…") and release housekeeping such as changelog moves. scripts/build_changelog_comment.py now uses the version's section of CHANGELOG.md instead. A missing or empty section fails the build rather than publishing empty notes. A test checks that the current version has a section, and CONTRIBUTING.md says to move the Unreleased entries at release time.

Full suite after both changes: 917 passed, 28 skipped.

After the release

Merge main back into devel (as in bf806c9), so this and the changelog commit on release/0.3.0 reach devel.

🤖 Generated with Claude Code