Build: final Python release #62

Activity

1. blooop

   added

   wayfinder:build

   on Aug 7, 2026
2. blooop

   added a parent issue on Aug 7, 2026

   • Wayfinder: perf/correctness/maintainability review #51
3. blooop

   marked this as blocked by 7 issues on Aug 7, 2026

   • Build: harden metadata storage (guarded load, atomic save) #56
   • Build: clear error when devpod binary is missing #57
   • Build: TTL + single spawn for background cache refresh #58
   • Build: repo hygiene sweep (template leftovers, dead code) #59
   • Build: cut redundant devpod shell-outs on hot paths #60
   • Build: parse workspace ids at one boundary (injective scheme) #64
   • Build: migrate clone dirs and workspaces to the new id scheme #65
4. blooop

   closed this as completedin #92on Aug 8, 2026
5. blooop commented on Aug 8, 2026

   

   blooop

   on Aug 8, 2026

   OwnerAuthor

   More actions

   Answer

   0.0.11 is released and both artifacts verified. PR #92, tagged v0.0.11.

   Acceptance was "both published artifacts install and run dl --version in a clean environment":

   $ python -m venv /tmp/dlvenv && /tmp/dlvenv/bin/pip install devlaunch==0.0.11
   $ /tmp/dlvenv/bin/dl --version          -> dl 0.0.11
   $ /tmp/dlvenv/bin/aid --version         -> aid 0.0.11
   $ pixi exec -c conda-forge -c https://prefix.dev/blooop --spec devlaunch==0.0.11 dl --version
                                           -> dl 0.0.11
   

   The venv also confirms the dependency set is right: iterfzf, tomli, and no tomli-w.

   The changelog

   [Unreleased] became [0.0.11], carrying #65's cache migration — the last of the seven review fixes. The other six had already gone out in 0.0.10 while this work was in flight, which is also why the 0.0.10 entry gained the two user-visible fixes it shipped without documenting (#56's corrupt-metadata guard and #57's missing-devpod error). They belong under the version that contains them; listing them under 0.0.11 would tell a user they arrive now. #59 is not backfilled — it changed nothing a user of the tool can observe.

   The entry names this as intended to be the last release of the Python implementation and points at the Rust successor sharing a core crate with blooop/wayfinder, per #53.

   tomli_w

   Dropped in #91, separately, because regenerating pixi.lock also rewrites it from format v6 to v7 — ~830 of 848 lines — and that would have buried the release diff. Its only caller was the save_config that #59 deleted. Both pyproject.toml and the conda run dep are now consistent with the code.

   One thing this ticket assumed and should not have

   The ticket said to "publish through both existing channels unchanged". PyPI is genuinely automatic. Conda was not, and never had been: conda-publish.yml's release: published trigger has never once fired, because Auto-publish creates the release with the default GITHUB_TOKEN and GitHub does not start workflow runs from events raised that way. The run history is workflow_dispatch all the way back. 0.0.11's conda package needed the same hand-dispatch as 0.0.9 and 0.0.10.

   Fixed in #102 by chaining off Auto-publish with workflow_run. That ordering is required rather than cosmetic: the recipe builds from tag: v${{ version }}, so the tag has to exist before rattler-build can fetch the source — which is presumably why the release trigger was chosen. A version-compare gate keeps it from firing on every push, since Auto-publish succeeds whether or not it published anything, and force_build — declared as an input and read nowhere — now does something. It can only be proven by the next release, by construction.

   Migration verified end to end against the real package

   Not just unit tests: a scratch cache with two old-scheme clones, each holding uncommitted work, migrated with the installed 0.0.11.

   main, feature-x  ->  testrepo-main-gahazofe, testrepo-feature-x-pazekoka
   

   dl --ls left it at version 1; the first command touching a workspace migrated it and printed both notices; uncommitted files, HEADs and branches all survived; .bare was never renamed; metadata agreed with disk at version 2; and the second run was a silent no-op.

   Worth recording: the first attempt at this test was invalid. Run from the repo directory, python -c imported the stale working tree instead of the installed package, so the migration appeared not to fire at all. Anything testing a released artifact has to run from outside the checkout.