Suite inputs and trusted results
e2e-tests.yaml covers application source, E2E specs, shared packages, database
fixtures, dependencies, Docker configuration, and runner scripts. A broad **
push filter reaches unknown future inputs; explicit negative paths omit known
non-runtime apps, documentation, and agent tooling. Its planner
hashes each suite’s tracked inputs before allocating Docker or Playwright jobs.
Only an exact, same-day success cache from protected main can omit a suite.
Keys include shared runtime/schema/auth helpers, the suite’s specs and satellite,
workflow/toolchain inputs, runner image, and E2E_RESULT_CACHE_EPOCH (default 1).
General shards conservatively share all web spec inputs because changing test
counts can redistribute Playwright shards. Shared API/package changes invalidate
all dependent suites; an unrelated suite’s cached proof can remain valid.
Missing, unavailable, or uncertain evidence runs tests. Failed, flaky, empty,
and malformed reports never publish proofs. Cached receipts contain counts and
source/run provenance, not raw reports, credentials, or customer data. Failure
artifacts contain sanitized text diagnostics only; raw traces, blob reports,
HTML reports, and test-result directories are not uploaded. Sanitize collected
server/probe output before printing console groups, and block artifact upload
if sanitization fails. Omit unstructured files containing credential-valued
collections or header records; reject symlinked diagnostics paths. Cancelled workflows do not allocate new E2E cohorts;
started jobs retain their cleanup steps. Branch
runs cannot publish trusted proofs. Manual dispatch always runs all suites;
there is no cron. Increment repository variable E2E_RESULT_CACHE_EPOCH to
invalidate all proofs. A new UTC day refreshes proofs on the next relevant push.
Release-only changes retain the existing skip. Validate with
node --test scripts/ci/e2e-result-plan.test.js scripts/ci/e2e-result-receipt.test.js.
Inactive database proposal
The default fingerprints include database migrations, seeds, real SQL tests, active helpers, and unknown database inputs in all eleven suites. The sole conditional exception isapps/database/tests/programming-hosted-typegen-review/,
which currently contains an inactive proposal and its review fixtures.
Before planning, a bounded, quiet Git scan checks the same immutable source ref
used to enumerate inputs. Any tracked reference to the directory basename outside
the proposal itself, documentation, and the planner’s own policy/regressions keeps
the directory in every fingerprint. This covers CI, scripts, applications,
packages, and root configuration without interpreting their dependency graphs.
A timeout, invalid ref, or uncertain Git result includes the inputs and executes
all suites without consulting cached proofs. The scan never logs source contents.
Only a definitive no-reference result excludes this directory. Other database
paths retain their existing conservative ownership; similarly named active paths
are included. Before activating the proposal, move its helpers into a runtime
location or remove this exception. Use explicit imports/callers; do not construct
hidden dynamic references into an excluded review directory. Review tests exercise
real Git trees with CI/script/app/package/root-config activation, and an older
immutable tree remains independent of later worktree or index changes.
Changing the planner itself changes its global input hash, so this policy update
requires one fresh baseline. Existing caches are retained; there is no epoch
change, cache flush, legacy-key fallback, or weakened main/platform/freshness
requirement. Subsequent proposal-only edits can reuse otherwise eligible results
for all ten web suites and Inventory/Storefront. Eligibility is not evidence of
an actual cache hit or measured time saving; verify the workflow summary first.