Given a set of flaky tests (full-window or partial), reconcile the MIND Linear backlog -- cluster by root cause, prioritize via the branch filter, and create/update/close flake tickets. The reusable core; detect-flakes (full CI sweep; may close stale tickets) and report-incidental-flakes (one flake an author hit; create/update only) both apply it. Usually invoked by those skills, not directly.
Instalar com Codex ou Claude Copie este prompt, cole no Codex, Claude ou outro assistente e deixe que ele revise a página da skill e instale para você.
Um comando direto ignora o prompt de revisão. Verifique a origem antes de executá-lo.
Given a set of flaky tests (full-window or partial), reconcile the MIND Linear backlog -- cluster by root cause, prioritize via the branch filter, and create/update/close flake tickets. The reusable core; detect-flakes (full CI sweep; may close stale tickets) and report-incidental-flakes (one flake an author hit; create/update only) both apply it. Usually invoked by those skills, not directly.
Manage flakes into MIND tickets
Take the set of flaky tests the caller hands you and make the MIND Linear backlog reflect it. You supply the judgment -- clustering, matching, and state decisions; the scripts/flake_reconcile.py CLI only moves data and never clusters. One ticket per root cause: never per test, never per symptom.
What the caller gives you
A list of flaky tests in the shape the CLI's list-flakes emits -- at least test, sample_failure_lines, is_marked_flaky, branches, last_seen. (A single-flake caller assembles one such record.)
Whether the set is full-window (every flake in a CI window, e.g. detect-flakes) or partial (e.g. one incidental flake from report-incidental-flakes). This flag gates closing -- see step 4.
Whether the run is autonomous -- stated by the caller, or passed to this skill directly as --autonomous. This gates the approval pause in step 5, and nothing else.
Precondition: Linear is reachable through latchkey -- check with uv run python scripts/flake_reconcile.py list-tickets, which is read-only. (Do not test this with latchkey services info linear: its credentialStatus field exists in latchkey 2.x but not in 3.x, so it reads as unset on a current CLI.)
The CLI
All ticket reads and writes go through uv run python scripts/flake_reconcile.py <cmd> from the repo root (--help for flags):
list-tickets -- existing flake tickets as JSON (read-only)
preferred-status --branch <b> ... -- the branch filter; prints ready/backlog/unknown
update-ticket --id <issue_id> --body-file ... [--title ...] -- replaces body (and optionally title); never touches state
set-status --id <issue_id> --status ready|backlog
comment-ticket --id <issue_id> --body-file ...
close-ticket --id <issue_id>
Write bodies and comments to temp files so multi-line markdown survives. Nothing is written until you run a write command -- and never before step 5's approval.
The loop
1. Read the existing tickets
list-tickets > /tmp/tickets.json, then read it. Each ticket has identifier, issue_id, is_open, and description (the full body, including markers written by previous runs).
2. Cluster the flakes -- for resolution, not description
Cluster with exactly one goal: each ticket is a handoff unit one future fixer can resolve as a single coherent piece of work. Before committing to a cluster, ask: could one fixer, with one capability, reproduce and fix all of this in one focused effort? If the fix for test A is unrelated to the fix for test B, they are different tickets even when the symptom matches.
Group by root cause and its fix, never by shared symptom. Read the failure lines and group by understanding, not string matching. (Illustration: three tests that all "time out" may be three tickets -- three unrelated slow paths -- while dozens of tests failing at the same infrastructure bring-up step are one ticket, because one fix hardens that step for all of them.)
Occurrence counts are severity signal for the body, never a clustering criterion. Do not bundle unrelated one-offs to keep the ticket count down; a one-off with a distinct cause is its own (low-priority) ticket, and staleness (step 4) retires it if it never recurs.
Refuse defensibility and inertia. Do not accept a grouping merely because it is plausible, and never keep a cluster as-is just because it is already filed -- that dumps several unrelated investigations on one fixer under one title. When in doubt, split: independently-actionable beats tidy.
For each cluster settle on: a title naming the cause, the affected tests, a one-line resolution hypothesis (what fixing it looks like, and which subsystem it lives in), severity from the occurrence counts, and a callout of any affected test lacking @pytest.mark.flaky (unmarked flakes turn CI red).
3. Prioritize each cluster -- the branch filter
A flake seen only on one unmerged feature branch is probably that branch's own bug; filing it as ready sends a fixer chasing something they cannot reproduce on main. Take the union of branches across the cluster's tests. The rule is deterministic -- preferred-status computes it:
any flake on main -> ready (a live problem on main)
never on main, but on more than one feature branch -> ready (branch-independent / systemic; it will reach main)
only ever on a single unmerged feature branch -> backlog (likely branch-local; the branch's subject matching the failing test's domain is near-proof)
no branch evidence -> unknown: leave status alone
4. Reconcile against existing tickets
Every body carries a stable flake-cluster slug and a last-seen timestamp (see "Ticket body" below). Match each cluster to an existing open ticket by slug -- or by clear semantic overlap when a ticket predates the markers -- then decide:
Cluster with no open ticket -> CREATE, at its preferred status.
Cluster with an open ticket -> UPDATE the body (refresh the affected-tests table and last-seen).
Open ticket whose cluster is absent from the current set and whose last-seen is more than 21 days before today -> CLOSE as stale. Compute the cutoff by plain date math against today.
Close only on a full-window sweep. A partial caller cannot know that anything stopped flaking: it may CREATE and UPDATE, never CLOSE. A repeated full sweep with no new flakes should produce no changes.
Never clobber a human or agent move. Change an existing ticket's state only while it is still in a pre-work state (Todo/Backlog) and its preferred status differs -- then set-status. In Progress / In Review / Done / Canceled means someone moved it intentionally: leave its state alone -- never demote an In Progress ticket, never reopen a Done/Canceled one. On every state change you make, comment-ticket explaining why, citing the branch evidence. If a backlog-preferred cluster is already In Progress, do not demote it -- comment that it has only flaked on one unmerged branch, so the assignee knows it may not reproduce on main.
Recluster freely as new information arrives. Clusters are hypotheses, not commitments; do not leave a wrong grouping standing out of inertia.
Split:update-ticket (with --title) to re-scope the existing ticket down to one cause, then create-ticket for each part that split off.
Merge:update-ticket the survivor to cover the whole cause, then close-ticket the now-redundant ticket(s) (full-window callers only).
When you re-scope a ticket, give it a new flake-cluster slug matching its new shape so future runs dedup correctly.
5. Show the plan, then apply
Summarize every intended CREATE / UPDATE / CLOSE / state change.
Interactively:get approval before the first write. On approval, execute and report what changed (each command prints the affected ticket).
Autonomously (--autonomous): do not pause -- but do not skip the narration either. Write the same plan you would have shown to flake-sweep-summary.md at the repo root, then apply it, then append what actually changed, including anything that failed. Stating the plan before acting is load-bearing rather than ceremony: it is the only record of intent if a write goes wrong, and narrating a plan first measurably improves how faithfully it gets followed. Every other rule still applies -- closes remain restricted to full-window sweeps (step 4), and a human's or agent's state move is still never clobbered.
Run summary (autonomous runs)
flake-sweep-summary.md is read verbatim into a GitHub job summary and a Slack post, so write it
for someone who sees nothing else. Markdown, in this order:
one line: the window swept, how many flaky tests, how many clusters;
the planned CREATE / UPDATE / CLOSE / state changes -- written before applying them;
what actually changed, with ticket identifiers, plus anything that failed;
what a human should look at: unmarked flakes that can turn CI red, clusters you were unsure of,
and open tickets you deliberately left alone.
Ticket body
Sibling skills (the fixer and the reporters) parse this shape -- keep it. At the top of every body:
<!-- flake-cluster: <short-stable-slug> -->
<!-- last-seen: <ISO-8601 of the cluster's max last_seen> -->
Then, skimmable: the root cause and resolution hypothesis; an affected-tests table (test | flaking commits | hard-fails | @flaky); the representative failure line(s); an unmarked-flake callout if any; and how many commits/days the cluster spans. Footer it as auto-filed by manage-flakes.
