Repository navigation
bug(ci): verify-generated-types self-skips on autobot_shared-only changes that can alter both generated api.ts files #14885
Description
Found while delivering #12986 in #14883, and deliberately NOT fixed there: it is a behaviour change to a gate, not a deduplication, so it belongs in its own change with its own measurement.
The gap
.github/workflows/verify-generated-types.yml's changes job declares two filters:
types— watchesautobot-backend/**/*.pyplus the autobot frontend's generatedapi.tsand lockfilesslm_types— watchesautobot-slm-backend/**/*.pyplus the SLM frontend's equivalents
Neither watches autobot_shared/**/*.py.
Both products' OpenAPI schemas are dumped by importing the application, and both applications import from autobot_shared. A change confined to that tree can therefore alter the generated types while this workflow computes "nothing to verify" and self-skips — the same silent-coverage-shrink shape #12986 was filed about, one workflow over.
Why it was left out of #14883
The per-product split between types and slm_types is deliberate and must be preserved; #14883 records it explicitly in pipeline-scripts/check_workflow_path_filters.py's NOT_CONSUMERS table with this reasoning, so it reads as considered rather than missed. Adding autobot_shared/**/*.py to both widens when the gate runs, which is the safe direction, but it is still an unmeasured behaviour change and it did not belong bundled into a consolidation PR.
Ask
- Measure how often a pull request touches only
autobot_shared/**/*.pyand would newly trigger this workflow. - If the cost is acceptable, add
autobot_shared/**/*.pyto bothtypesandslm_types— keeping the per-product backend split intact — and move the workflow fromNOT_CONSUMERSinto the guard's declared-consumer table so it is policed from then on. - Evidence must be a run showing the workflow triggering on an
autobot_shared-only change, not a diff.
Acceptance criteria
- A measured answer to whether
autobot_sharedchanges can alter either generatedapi.ts - Either both filters watch it, or the reason they should not is recorded in the guard's table
-
pipeline-scripts/check_workflow_path_filters.pyreflects whichever answer is chosen -
actionlintpasses
Related
- refactor(ci): backend-Python path filter duplicated 7x across 6 workflows with no source of truth #12986 — the same defect class, fixed for six other workflows
- refactor(ci): consolidate duplicated, drifted and never-run CI guards (#14371, #13984, #12986, #14034) #14883 — where this was found and recorded
Activity
mrveiss commentedon Aug 24, 2026
Fixed in #14924, with one residual.
All four acceptance criteria are delivered there: the measurement (both
api.tsfiles depend onautobot_sharedthrough the user-schema shims andnormalize_pattern_anchors), both filters widened while keeping the per-product split,pipeline-scripts/check_workflow_path_filters.pyasserting it from now on, and the workflow parsing cleanly.The "Ask" section's bullet 3 — a run rather than a diff — is not delivered and cannot be pre-merge without pushing a throwaway commit for CI evidence. Tracked as #14925: the push trigger now names the shared tree, so the first
autobot_shared-only push toDev_new_guiafter the merge produces that run by itself.One adjacent defect was fixed in the same file rather than filed:
verify-generated-types-slmused== 'true'where its sibling uses!= 'false', so an empty filter output skipped it silently.fix(ci): close two gate escapes — a missing route and an accidental s…
mrveiss commentedon Aug 24, 2026
Delivered and merged as
1208b77ba(PR #14924).Measured before widening, rather than assumed.
autobot_shared/user_management/schemas/user.pyis the sole definition of the user schemas both backends re-export, and bothapi.tsfiles carry them;normalize_pattern_anchorsrewrites both specs. 46 of the last 300 first-parent commits touch onlyautobot_shared/**/*.py— that is the size of the blind spot, and the added CI cost.Widened, not un-skipped. The required context is already the separate reporting job, and un-skipping a path-filtered required job is exactly what deadlocks PRs on "Expected — Waiting for status" (#13388).
verify-types-run,verify-generated-types-slmand the requiredverify-generated-typesall ran green.All four acceptance criteria met. The one item not delivered is the "Ask" bullet requesting an actual CI run on an
autobot_shared-only change — that requires pushing a throwaway commit purely for CI evidence, which this repo forbids. Filed as #14925: the push trigger now watches the shared tree, so the first genuine such push after merge produces that evidence naturally.