Repository navigation
bug(deploy): three more path defects around the backend PYTHONPATH and module symlinks #15560
Description
Three findings from the B3/B4 work in PR #15559. All pre-existing, none fixed
there — that PR was scoped to the autobot_shared symlink and the
backend_install_dir meaning, and widening it would have made a deploy-path
change harder to review.
1. ai-stack builds backend-module symlink targets absolutely
autobot-slm-backend/ansible/roles/ai-stack/tasks/main.yml — the
"Create src/ symlinks to backend modules" task builds each src: as an absolute
path under the backend install directory.
Same defect family as B4, but wider: B4 was one well-known link, this is an
arbitrary set of backend modules. A link whose target does not depend on where
the link lives cannot survive the tree moving, which is the property #13539's
release scheme depends on.
- Targets resolve relative to the link's own directory
- A test resolves them with
realpathunder a relocated root, in the style ofrepo_tests/autobot_shared_symlink_relative_13539_test.py— asserting on the resolved destination, not on the absence of a string
2. fix-backend-environment.yml hand-writes a PYTHONPATH it just rendered
The playbook renders autobot-backend.service.j2 and then, a few lines later,
writes a PYTHONPATH again via lineinfile/shell. Two sources for one value,
which can drift — and the template is the canonical one.
This sits directly on the defect B3 exposed: the same playbook was producing
PYTHONPATH=…:/opt/autobot_shared, a directory that does not exist. #15559 fixed
the rendered value; the hand-written duplicate is still a second, independent
writer of the same setting.
- Establish whether the hand-written path is needed at all, or is a leftover from before the template covered it
- If it is needed, it derives from the same variables as the template rather than restating them
- A test asserts the two agree, or that only one writer exists
3. setup-user-backend.yml defines a variable nothing consumes
backend_shared_dir is defined and never referenced in that file or the role it
invokes. Dead configuration reads as meaningful to whoever edits next.
- Confirm it is genuinely unconsumed across every playbook and role that could inherit it
- Remove it, or wire it in if the omission is the actual bug
Evidence
Found while fixing #13539's B3/B4 blockers (PR #15559). The B4 sweep also turned
up three symlink sites the design document had not listed, so the population here
should be assumed larger than these three until swept.
Activity
fix(deploy): derive the extracted path defaults from base_dir and fol…
chore(ci): take the stranded baseline entries with the paths the extr…
test(guard): drop the pin whose command moved and fail on a dangling …
fix(ansible): stage the SLM frontend publish and repair three backend…
mrveiss commentedon Sep 4, 2026
Closed on code evidence, verified against merged
Dev_new_gui(PR #15602,e1d8b51d).srcspelled out../../{{ backend_install_dir | basename }}/{{ item }}; the basename comes from the variable, so the two sibling install dirs cannot drift apart. Proved with real symlinks undertmp_path+os.path.realpath, with the historical absolute form as the contrast caselineinfilewriters fighting the templategrep -c lineinfilereturns 2 on the file, and both hits are inside comments recording why they went — no live task. The play now slurps the rendered unit, derivesbackend_unit_pythonpath(4 uses), and fails if the unit carries nonebackend_shared_dirnames a variable no role readssetup-user-backend.ymlnow usesbackend_shared_standalone_dir(2 uses), which is the nameroles/backendandroles/autobot_sharedactually declarerepo_tests/ansible_backend_path_defects_15560_test.pypresent in baseDefect 2 is the one worth recording properly. The two
lineinfiletasks rewroteEnvironmentFile=andEnvironment="PYTHONPATH=..."a few lines after the template had already written both — one setting with two writers that disagreed. That is how this play came to produce aPYTHONPATHentry pointing at a directory that does not exist (#13539/B3). #15559 corrected the rendered value; the hand-written duplicate was a separate, independent writer and survived that fix untouched.Removing them is not deletion of working code: the template is confirmed the sole writer of both settings, its single-line output matches the regex the play reads back with, and every downstream consumer — the import test and the summary — was moved onto the derived value in the same change. Verified before removal rather than after.
Defect 3 was wired in rather than dropped: the play's
backend_deployed_rootnow actually reaches the standaloneautobot_sharedcopy, which the misnamed variable meant it never did.The guard generalises the third defect instead of pinning it — every variable
setup-user-backend.ymldefines must be referenced somewhere in the tree, so the next variable named for a reader that does not exist fails at authoring time.fix(ansible): the ai-stack manifest default must not be a Jinja templ…