Retest Bug (discovery stub)

Thin entry for agent skill discovery. Full procedure: WORKFLOW.md.

When invoked:

1. Announce once: Using **retest-bug-workflow** to retest the bug fix.
2. Settle and strategize first — before any tool call, follow settle-and-strategize.md: invoke `engineering:testing-strategy` to plan the retest approach, no guessing an expected/spec, no blind retry loops.
3. Design (Figma) comparison is mandatory on every UI retest — figma-design-comparison.md. A screen with no design reference is reported back to the person or channel that assigned the retest and its visual points stay BLOCKED; never assume a label, order, or layout.
4. Retest deeply — customer-escape-prevention.md: cover the whole surface the fix touched, measure overflow/overlap at every in-scope width, evidence on passed rows too, and never PASSED over a "cannot verify" actual result.
5. Run in parallel lanes by default — parallel-test-lanes.md: split the case list across subagents, one leased account per lane (never two lanes on one login), barrier for anything every lane sees; serial only with a named predicate.

5a. Login by command, never a login window — when an account must log in or its session has expired, give the user ready-to-run commands as one combined bash block per step that loops over the accounts (each isolated, prints OK / FAILED with a log path): relink every needed account, then setup/mint only the FAILED ones. Then probe and report. Never collect or fill a password (parallel-test-lanes.md §3).

1. Result quality gate before any verdict — result-quality-gate.md: all 20 layers on every case before a verdict is reported or written; the main thread re-runs it on the lanes' files.

6a. No-wait decision gate before any BLOCKED / "needs decision" / held post — no-wait-decision-gate.md: all 20 layers — read every source, apply precedent, create missing data, decide and act; never park a case or a finished post waiting for someone.

1. Read and follow WORKFLOW.md end-to-end — every step, gate, and reference.

Claude Code shortcut: /retest-bug → commands/retest-bug.md.

Contents

Every file this workflow uses, linked directly from here. Only WORKFLOW.md is read up front; load each other file at the step or moment its line names. References already linked above are not repeated.

• WORKFLOW.md — always: the full procedure, read end to end.
• execute-and-evidence.md — before Step 4 — read and follow Steps 4–5 and the result quality gate, end to end.
• comment-and-post.md — before drafting the comment — read and follow Steps 6–7, end to end.
• closeout-and-notify.md — right after the comment is posted — read and follow Steps 8–10, end to end.
• bug-priority-matrix.md — whenever a Priority/Severity is named or judged.
• comment-template-v2.md — before drafting the v2 comment — render it from run.json per this template; never type the body.
• defect-candidates.md — before reporting any defect found in the round — follow this exact table and the fixed owner question; never file without the owner.
• defect-report-completeness.md — for any verdict that is not a clean PASSED — what the write-up must contain.
• intake-one-pager.md — at intake, to show the one-page list of inputs the run needs.
• jira-comment-post-review.md — right after posting a Jira comment — full review checklist.
• jira-wiki-vs-markdown.md — before and after posting any Jira body (v2 wiki vs v3 ADF).
• non-pass-challenge-gate.md — before recording any non-PASSED verdict.
• portable-content.md — before writing any notify text or helper — no hardcoded webhooks, tokens, or machine paths.
• qa-closing-shared.md — before reporting done — shared QA closing gates, including the cross-ticket conflict check.
• qa-debug-discipline.md — when results conflict with expectations, with dev claims, or look flaky.
• qa-evidence-gates.md — read when recording or judging evidence clips and the AC/EC coverage gate.
• retest-fix-intake.md — when building the retest plan, before executing tests.
• root-cause-investigation.md — before calling any case FAILED or BLOCKED — investigate to a verified root cause.
• round-time-contract.md — before logging in or reusing a session for the round.
• shared-must-never.md — before any write or post — rules shared by every workflow.
• shared-preamble.md — at the start of every run, before the first reply (preamble, constraints, tone).
• test-through-real-steps.md — before executing any case — drive the real screen, never a backend shortcut.
• workspace-guide-discovery.md — on the first response, to locate the workspace project guide for this workflow.
• agent-rationalizations.md — read when tempted to skip a step; each excuse has its correct action.
• helix-session-constraints.md — read once at session start.
• jira-linebreak-conversion.md — read before posting a Jira table with multi-line cells.
• long-workflow-todos.md — read when the workflow has many steps; keep the todo list.
• playwright-preflight.md — read before any Playwright run.
• proactive-qa-triggers.md — read before the round report, to suggest (never start) at most one follow-up Helix workflow.
• resilient-selectors.md — read when writing or fixing Playwright selectors.
• session-closing.md — read at session end.
• skill-invocation-discipline.md — read before invoking or switching skills.
• skill-routing.md — read when choosing which skill handles a request.
• skill-rules-style.md — read before asking the user or stopping on a missing input.
• user-communication.md — read before any message to the user.
• project-config-template.md — read when the project has no retest config yet.
• worked-example.md — read for a full worked retest.
• debug-discipline.md — read when a case is still failing, to triage FAILED vs BLOCKED.
• gotchas.md — read when a Jira body hits encoding or v2 rendering issues.
• handoff-template.md — read when handing the session to another person or agent mid-round.
• post-mortem-template.md — read when writing up a mistake made in the round.
• jira-fast-publish.md — read at Step 7a when the comment has more than 3 table rows (ADF-direct publish and error recovery).
• handoff-file-template.md — read at Step 8 close-out when writing the round's helix-handoff-{KEY}.md.

Refusal-first (precondition gate)

All preconditions and refusal rules are in WORKFLOW.md. MUST NOT start without a Jira issue key/URL — a Bug, or a Task / Story whose fix is being re-verified — and reachable environment config.

QA closing (mandatory before "done")

All close-out gates are in WORKFLOW.md and verify-closing-checklist.md (Retest section). MUST NOT claim PASSED/FAILED without evidence in the comment, and MUST NOT post a retest comment — bug or task — without its case list — ONE table only, headed and filled per the Table headers and Evidence per row rules in comment-and-post.md (a row per ER/AC with its covering cases as • TC_nn title (role) inside the row; never a separate Test cases run table). No Scope: FULL line; multi-point header fields and cells as bullets; after posting set column widths and centre every clip/image with tools/retest-guard/adf_colwidth.js. The design reference goes on the header's Design ref: line, never as a per-row column.

Before any comment draft is written to disk or shown as final — including "draft only" requests — run node tools/retest-guard/retest_guard.js --body <draft> --format v2 --bug-type FE (for other bug types use the --format/--bug-type values the WORKFLOW comment step names) and keep fixing the draft until it exits 0. If the guard fails, return to the comment step in WORKFLOW.md; never hand over a draft the guard has not passed. When the workspace has no tools/retest-guard/ (dependencies.md), apply the same rules by hand from the comment step and state guard not installed — checked by hand with the draft; exit 2 (guard present but could not run) is not a pass — fix the run.

<a id="tool-not-available"></a>Which tools count (every `tools/…` step in this workflow). "The workspace" is the repository of the project being retested — the one its project guide lives in. A tool counts only when it is in that workspace and ran in this round. A copy found in any other checkout or repository on the machine is not this project's tool: never run it against this ticket and never cite it. An offline or mock run (the input says so) has no tools: apply each tool's checks by hand and write <tool> not installed — checked by hand beside the result. Report an exit code or "passed" only by quoting the real output line of a run made in this round; a missing tool is never a lighter standard.

Dependencies

Run the check for each of these before step 1 — check commands and install hints are in dependencies.md:

• Node.js + in-repo guard tools (tools/retest-guard, tools/lane-lease-check, tools/run-time, tools/adoption-check)
• curl + Jira REST credentials
• Atlassian MCP
• Figma
• browser automation MCP
• Playwright
• QA capture and verify tools ({QA_TOOLS_ROOT})
• Python 3
• ffmpeg / ffprobe
• macOS-only tools (sips)
• notify channel webhook / bot
• engineering:testing-strategy
• superpowers:systematic-debugging