Skip to content

v2.0.23

Choose a tag to compare

View all tags
@berges99 berges99 released this 29 May 08:47
v2.0.23
de4827d
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.

What's Changed

  • fix Windows CRLF doubling on write→edit multiline flow (a94a735)
  • ci: bump actions to Node 24-compatible versions (3566ea1)
  • fix(write): hash fs_state from disk bytes to match edit/read (1ca7b1e)
  • fix(jsonl): only retry Windows file lock on contention errors (5525d12)
  • test(bash): use tmp_path for redirection test on Windows (889cd00)
  • fix: CRLF handling and Windows path edge cases in file tools (3e8e60d)
  • fix: Windows File paths and remaining cross-platform CI failures (d3c9b22)
  • test(codegen): use sys.executable for CLI subprocess invocations (8ab2496)
  • test: fix py3.11 File instantiation and flaky parallel-params assertion (fa11a69)
  • ci: add 'os' input to workflow_dispatch for targeted manual runs (1e7d1ff)
  • fix(tracing): make JsonlTracingProvider import on Windows (415883e)
  • fix(file-content): always send string filename to OpenAI Responses API _safe_file_name returns None for data URLs and anonymous BytesIO files, so for PDFs in those source schemes to_openai_responses_input was emitting {"filename": null, ...} — rejected by the OpenAI Responses API with a validation error. Falls back to f"document{source_extension or '.pdf'}" when no display name is known. Deterministic (no UUID per call, preserves cache invariants) and human-readable in OpenAI logs. Adds regression tests for both no-name code paths and tightens test_pdf_valid to actually assert the filename value. (21a56c5)
  • refactor(file-content): drop dead _display_name property The validator already sets self.name = _safe_file_name(self.file) at construction when name is None, so the property's fallback branch was unreachable in practice — it just recomputed the same None. Removes the property and uses self.name directly at the single call site (PDF branch of to_openai_responses_input). (af3911e)
  • feat(file-content): add explicit name field for filename metadata Adds FileContent.name: str | None so the original filename survives the full chat-history round-trip. Previously dump() serialized FileContent.file to a bare URL and the frontend had to guess the display name from the URL — fragile with URL encoding, sanitization, uuid paths, and bytes uploads defaulting to {uuid}.ext. Changes: - FileContent: new name field; validator defaults from File metadata without triggering the lazy fetcher (no disk read / network call at construction time). Moves _cached_* to PrivateAttr so they no longer leak into dumps. Fixes a stale _cached_openai_input typo. - content_factory: reads/restores name on deserialize; old traces without the field still work (defaults from URL basename). - serialization.dump(): explicit shape for FileContent: {"type": "file", "name": "...", "file": "..."} name is omitted when None. Stops leaking private cache attrs. - EML attachment paths propagate attachment['filename'] into the synthesized FileContent. Tests (offline + real-API): - 17 unit tests: name defaulting per source scheme, no I/O at construction, dump shape, round-trip via content_factory, backward compat for pre-existing traces. - 1 offline integration test: Agent + JSONL + session-chain reload using TestModel. - 3 real-API integration tests (gated by @pytest.mark.integration): OpenAI gpt-4o-mini, Anthropic claude-haiku-4-5, Google gemini-2.5-flash-lite — each verifies a real LLM accepts a PNG with display name, JSONL persists the name, and a second turn successfully reloads + re-sends the file. (7ddbdc5)
  • Add Jira tool tests with optional live integration suite (324966d)
  • Register Jira tools and use platform Integration (82a1d85)
  • Add Jira tools module (9ffdfd7)

Full Changelog: v2.0.22...v2.0.23

Assets 8
Loading