Version: 1.0
Date: 2026-09-16
Status: Trial complete, TFP-052 landed — desktop proposal implemented (rollback marker through UpdateChecker.interpret()’s existing seam, 9 new tests) and toil-classified both builds. Findings live in test-assist/docs/RELEASE_ROLLBACK_TRIAL.md, merged to test-assist’s main via test-assist#24 (merge commit b91c79b). Gate A below is satisfied.
Reviewing skills/proposed/’s overlap table raised a specific question:
verified-release-pipeline (the existing account skill) and sre-release
(Martin’s proposed candidate) were flagged as overlapping on ch. 8, with
sre-release also carrying “launch coordination and progressive rollout”
that the existing skill doesn’t cover. Checked precisely, not by
inference — read both skills’ full text. The line is exact:
verified-release-pipeline proves a built artifact runs correctly before
it’s attached to a release and stops there; sre-release continues past
that point into staged rollout, canarying, feature flags, and rollback.
Checked against test-assist’s real setup (not hypothetically) because
Emil described this as an actual, felt pain point — “frustrating trying to
get a proper update or rollback working” — not a theoretical gap:
release.yml (desktop/PySide6 build) already implements
verified-release-pipeline properly — tag-triggered, workflow_dispatch
for rehearsal, and a real verify step (--version/--selftest probes
written to file, since the windowed build has no usable stdout). No gap
on the “does the build actually work” front.python/update_check.py is deliberately manual: no on-launch poll, no
telemetry, no auto-download, stated explicitly in its own docstring as a
constraint (corporate-machine QA tool; an app that phones out gets
banned from the estate). This means sre-release’s thundering-herd/
jitter material doesn’t apply — there’s no automatic polling to
stampede..zip sits on the GitHub Releases page; UpdateChecker only ever
checks releases/latest, with no way to mark a release bad or point
users back. Browser build: pages.yml deploys to every visitor
instantly on every push to main, no staging, no canary; reverting
means reverting the commit and waiting for a full CI run.sre-release’s rollback/staged-rollout material and sre-toil’s
toil-ledger framework to test-assist’s real update/rollback
situation — both builds, findings printed and recorded, same pattern
as TFP-039’s CALIBRATION.md and TFP-036’s
INTROSPECT_CRAWL_SPIKE.md.update_check.py’s manual-only design constraint — that
was a deliberate decision, not an oversight, and isn’t being
second-guessed here.sre-release or sre-toil into skills/testing/ or
skills/engineering/ — this is a trial, same as the earlier
sre-code/sre-architecture read-only pass, not a promotion decision.investigation, skills, test-assistverified-release-pipeline doesn’t
cover it by design, and nothing else in this repo currently addresses
it either.sre-toil’s six-property test, applied to the actual
current process for handling a bad release: what happens today,
step by step, when v1.4.0 turns out to be broken, for each build.
Classify it — this is very likely toil (manual, repetitive if it
recurs, automatable, no enduring value), but do the actual
classification rather than assume it.UpdateChecker
already has a clean seam (interpret() is pure, already tested with
literal payloads per its own docstring) — propose the minimal
addition that lets a release be marked bad without violating the
manual-only constraint (e.g., checking a small denylist/known-good
marker alongside releases/latest, still only on explicit user
action, still no telemetry). Print the proposal; only implement it if
it’s genuinely minimal and doesn’t expand the update checker’s
footprint beyond what its own docstring commits to.pages.yml has no
staging today. Propose the smallest real fix — this could be as
simple as documenting a fast revert-and-redeploy runbook (time it for
real, don’t estimate), or, if genuinely cheap, a manual approval gate
before deploy-pages for anyone who wants a pause-point. Don’t
propose a canary/staged-rollout system sized for a team when this is
a single-maintainer static site — match the fix to the actual
scale.docs/RELEASE_ROLLBACK_TRIAL.md in test-assist (not testforge —
this is about test-assist’s real release process, the same pattern
TFP-036 used for a testforge-side finding about testforge’s own
code).test-assist/docs/RELEASE_ROLLBACK_TRIAL.md — the toil
classification, the current-process walkthrough, and the two scoped
proposals (desktop, browser).testforge/skills/ROADMAP.md’s sre-release/
sre-toil overlap notes to this finding, so it isn’t lost the way
TFP-036 sat unresolved for a full engagement’s length.update_check.py’s existing pure functions already are.Single ticket.
RELEASE_ROLLBACK_TRIAL.md exists with real findings, not a still-open
question.TFP-052 ran for real, not deferred: both builds toil-classified against
sre-toil’s six-property test (desktop 5-6/6, browser 5/6), zero rollback
mechanism confirmed by reading update_check.py and pages.yml directly,
and two proposals scoped to what a manual-update, no-telemetry,
solo-maintainer project actually supports. The desktop proposal was
implemented — a KNOWN_GOOD: marker read through UpdateChecker.interpret()’s
existing pure seam, with 9 new tests. skills/proposed/README.md already
points its verified-release-pipeline/sre-release overlap note at this
result (commit 97e3ef2).
Done (2026-09-17): test-assist PR #24 (branch
docs/ta-250-release-rollback-trial, commit 9479425) reviewed and merged
to main there (merge commit b91c79b). Reviewed for real before merging:
the 9 new tests were confirmed by name in the diff and their pass count
(405 passed, 2 skipped, 3 deselected, 1 xfailed) verified directly against
the PR’s own CI log, not just its description; the update_check.py diff
is a pure-function addition through interpret()’s existing seam with no
new network call; the browser-build proposal’s file list confirmed
doc-only (no pages.yml change). The PR’s “visual” check showed as
failing at merge time — confirmed to be a pre-existing flake also present
on main’s own tip (7f0c284), unrelated to this PR’s changes, before
merging past it. The rollback marker and RELEASE_ROLLBACK_TRIAL.md are
now on test-assist’s main. This backlog’s Gate A is satisfied.