proDreams/revu 2.4.1

on GitHub

Fixed

• The crash that restarted the container is gone, and it was a version. The
  parser core was tree-sitter 0.26 while every grammar wheel here is built against
  0.25, and the mismatch showed up where the grammar's field tables are read: the tree
  walk touched node fields through a stale view of the language and, under memory
  pressure, the process died with SIGSEGV — Error 139 in the container, which no
  Python except can catch and which arrives as a bare restart. Once the heap was
  poisoned on purpose it reproduced every time (PYTHONMALLOC=malloc MALLOC_PERTURB_=165): six crashes in six runs on 0.26.0, none in eight on 0.25.2,
  while plain parsing never crashed on either — which is what placed the fault in the
  walk and not in the grammar. The core is held below 0.26 until the wheels move, and
  run_prod sets PYTHONFAULTHANDLER=1, so a crash of any kind now leaves a Python
  stack in the container log instead of an exit code on its own.

• The image builds again. Three causes, in the order they surfaced. uv sync --locked ran in a layer holding only the metadata, and building the project needs
  its own source, so the install failed with Expected a Python module at: src/revu/__init__.py; the first sync now passes --no-install-project and the
  project is installed once its source has been copied in. The certificate step then
  failed the same way through uv run, which syncs — and therefore builds — before
  running certifi; it uses the environment's own interpreter now. And uv came from
  ghcr.io/astral-sh/uv:latest, so a new uv could change how this image builds
  without a line in the repository changing, which is exactly what had happened: the
  uv_build pin was written against an older uv. The image tag and the backend range
  are pinned together now.

• A version tag is released, and only once. The release workflow listened for
  v* while every tag in this repository is bare (1.0.0 through 1.5.0), so
  git push origin 2.4.0 started nothing at all; and the release step rebuilt the tag
  name from the version, so pushing 2.4.0 would have released v2.4.0, creating it
  if it did not exist. Both spellings are accepted now, and the tag that was pushed is
  the one released. The image was also built twice — once by the push to main, once
  by the tag that followed — so build.yaml is now one workflow for both, and the tag
  path verifies the image the main push published instead of rebuilding it. Its lint
  step ran nothing either: --hook-stage manual selects hooks declaring that stage,
  and none here do.

• Gitea's conversations are read through Gitea's own routes.
  GET /repos/{owner}/{repo}/pulls/{index}/comments does not exist in Gitea's API —
  it is GitHub's route, inherited by sharing the GitHub-compatible comment reader —
  and it answers 404. The review survived that by running without the history, which
  is what made it quiet: a repeated review on Gitea never saw its own earlier inline
  findings, so the feature that makes a second review build on the first never worked
  there. The routes are a per-host hook now, and Gitea reads its issue timeline (an
  inline review comment is an issue comment carrying a pull_request_url) plus
  /pulls/{index}/reviews for the review bodies.

• Deep context no longer drops a file without saying so. The tree walk was
  recursive, and a generated tree is as deep as its source is nested: a call chain
  a thousand calls long exhausted the interpreter's stack, parse swallowed the
  RecursionError and returned an empty result — indistinguishable, at the call
  site, from a file with nothing in it. A file with 900 nested calls parsed to zero
  symbols before this; the walk is iterative now and the same file yields its
  function. Anything that still fails is logged with its traceback instead of
  disappearing.

• The walk also holds its Parser in a local until the walk is over. Tree keeps
  its language alive, so this is belt and braces rather than a fix — but a parser is
  cheap to keep and a use-after-free in a C extension cannot be caught.

• A wrapper definition (a decorator, a template) is asked through the same rule
  lookup as every other node, so nested_inside_definition and requires_field are
  no longer skipped for it.

Changed

• A refusal by the Git host (401, 403, 404) logs what that host's own documentation
  requires of a token: Gitea's scope groups (read:repository for /repos/*,
  write:issue for issues/*, a scope required even for a public repository),
  GitHub's repo, GitLab's api. It goes to the log and not to the pull request —
  scopes are the operator's business, while the author only needs to know that the
  review failed.