regression-testing

Forbidden invocations

NEVER invoke /sdd-* skills from this workflow. SDD is an optional user-installed ceremony; this skill ships self-contained and does not chain SDD under any condition. If you need to refactor KATA, fixtures, cli/, scripts/, or api/schemas/ pipeline, exit this skill first and invoke /framework-development — which itself runs Plan → Code → Verify → Archive natively (no SDD required).

This boundary is mechanical, not advisory: scripts/lint-skills.ts rejects any /sdd- mention outside this section. See: .agents/skills/agentic-qa-core/references/skill-composition-strategy.md §4 (governs users who manually install SDD).

Regression Testing — Execute, Analyze, Decide

Orchestrates the full release-readiness pipeline: trigger a CI suite, monitor it to completion, classify failures, score against release criteria, and emit a GO / CAUTION / NO-GO verdict plus a stakeholder report.

Three phases, always in this order: Execute → Analyze → Report. Do not skip analysis and jump to a report. Do not guess classification without reading failure logs.

Compact Rules

DO: run Execute → Analyze → Report in that order. Never skip analysis and jump to a report, and never classify a failure without reading its logs.
DO: clear the readiness preflight before triggering anything — gh authenticated, the suite's workflow file present, GitHub Actions secrets set, Allure resolvable, active env confirmed. A long run that 401s mid-way is the expensive failure.
DO: persist RUN_ID the moment the trigger returns, before anything else. Resume re-attaches to a live run instead of re-triggering CI; a trigger that landed without the id saved costs the whole run again.
DO NOT: mark a failure REGRESSION without checking its history first — the single most common misclassification. A first-ever failure with no history is NEW TEST, unverified, not a regression.
DO: classify every failure into exactly one of KNOWN-BLOCKED / KNOWN ISSUE / ENVIRONMENT / NEW TEST / FLAKY / REGRESSION, and assess severity on a separate axis — a FLAKY test on checkout is still CRITICAL. Classification is fail-closed: every class but REGRESSION cites its evidence (the @blocked:{BUG-KEY} marker; a ticket created BEFORE the run; the matching log line plus the other failures on that host; the empty history query; the ids and results of at least 5 runs on an unchanged build). No citation, or a ticket this session filed, = REGRESSION.
DO: exclude @blocked:{BUG-KEY} tests from the gating pass-rate and report their count with each blocking key. They are parked behind an already-filed bug: never REGRESSION, and never a new bug.
DO NOT: use ENVIRONMENT as a scapegoat. Many unrelated tests failing on one host is environment; one test failing on an endpoint other tests reach fine is more likely a REGRESSION.
DO NOT: call a test flaky on fewer than 5 runs of history — mark "insufficient history" and re-evaluate rather than guessing.
DO NOT: emit GO while any REGRESSION-class failure stands. Hard vetoes regardless of score: any @critical test failing, any HIGH/CRITICAL-severity regression, or a pass rate below 90%.
DO: file only CONFIRMED product failures — the REGRESSION class, plus a NEW TEST failure once manually confirmed to be a real defect. FLAKY, ENVIRONMENT and KNOWN ISSUE get no issue at all. Triage decides WHETHER to file; the defect-management doctrine decides the type and the fields.
DO NOT: open a GitHub issue for a quality failure. It is filed in the issue tracker, parented to the QA Defect Management process epic and linked to the source Story — never to a product or dev epic.
DO: create the RTR (Regression Test Results, a Test Execution: RTR: {scope-id}: Regression Testing, parent QA Test Artifacts, Test Environment set, assignee self, testPlan → RTP) BEFORE triggering CI, persist its key beside RUN_ID, and pass it as the execution_key dispatch input. One RTR per verdict. The STR is created or completed ONLY when the run is the sprint close (then it links to both the STP and the RTP).
DO NOT: import a regular regression run into the sprint STR, and never let smoke or sanity write into a regression execution: they import only when an execution key is passed explicitly.
DO: close the RTR (or the sprint-close STR) via complete only AFTER the verdict comment is posted on it, and leave the RTP at its ready status: a suite run never completes the plan it ran from.
DO NOT: invent a sprint number. The RTR needs none (its scope-id is {env}-{YYYY-MM-DD} or a release tag). N matters only for the sprint-close STR: take it from the user or from the STP's own scope-id, and ask before creating anything at sprint altitude.
DO NOT: skip the artifact download on a red build (evidence vanishes after the retention window), and never merge smoke and regression results into one pass-rate — their SLOs differ.
DO: inspect traces and screenshots with /playwright-cli ([AUTOMATION_TOOL]); it is the only browser path, no MCP backs it.

Read full SKILL.md when: driving the CI commands, applying the GO/CAUTION/NO-GO scoring table, resolving a borderline classification, wiring the TMS artifacts, or writing the report.

Inputs

.github/workflows/*.yml — workflow files for regression / smoke / sanity suites; defines triggers, inputs, and artifact uploads. LOAD /github-actions-docs before editing or diagnosing one: Actions syntax (matrix, needs, reusable workflows, artifact retention, permissions) is the part of this skill's surface that changes upstream without telling anyone, and a guessed key fails at runner start with a message that points nowhere. Reading one does not need it; changing one does.
.context/PBI/qa-artifacts/master-test-plan.md — regression Epic key + expected pass-rate SLOs per suite.
playwright.config.ts — reporter config, retry policy, project matrix; needed to interpret retry counts and shard splits.
Previous run's Allure report (artifact URL or local download under ./analysis/previous/) — baseline for trend computation.
kata-manifest.json — registry of tests and ATCs available; used to cross-reference failed test IDs.
.agents/jira-required.yaml — Jira refs (project key, work types, transitions) for filing regression issues.
agentic-qa-core/references/defect-management-doctrine.md — canonical authority for classifying (Bug/Defect/Improvement), the mandatory field matrix, QA-Assignee ownership, and the QA process epic when a confirmed regression is filed in Jira (Phase 3). Read BEFORE filing any defect.
RTP key (RTP: {{PROJECT_KEY}}: Regression Test Plan, a Test Plan item parented to QA Master Test Plan): found by title via [TMS_TOOL] in Phase 1, it is the RTR's testPlan target. Producer: /test-documentation (promotion of regression-candidate TCs); this skill never creates it and never writes into it. Modality jira-native: there is no Test Plan item, the promotion is the regression-candidate label on the Test issues, so there is no key to resolve.
RTR (RTR: <<SCOPE_ID>>: Regression Testing, a Test Execution parented to QA Test Artifacts): the run record this skill CREATES in Phase 1 before the trigger, CI imports into, and Phase 3 closes after the verdict. One per verdict. Its key is persisted beside RUN_ID in plan.md and passed to CI as the execution_key dispatch input.
agentic-qa-core/references/artifact-lifecycle.md — canonical authority for artifact statuses: the RTR (or the sprint-close STR) closes at {{jira.status.test_execution.close}} after the verdict, the RTP stays at {{jira.status.test_plan.ready}}, every created artifact carries assignee = self, and an unmapped transition slug goes through the §4 fallback instead of a silent skip. Read BEFORE firing any transition.

Subagent Dispatch Strategy

Orchestration & Session contracts: this skill follows agentic-qa-core/references/orchestration-doctrine.md (mandatory subagent dispatch — main thread is command center) AND agentic-qa-core/references/session-management.md (Phase 0 resume check, plan-first persistence at .session/<skill-slug>/<scope>/, archive on completion). Phase 0 (resume check) and Phase 1 (plan write) are NOT optional. The orchestrator also applies the per-stage Definition-of-Done gates in agentic-qa-core/references/stage-gates.md: verify a stage's DoD BEFORE recording its progress checkpoint and advancing.

This skill is per-run scope: <scope> = <env>-<YYYY-MM-DD> (e.g. staging-2026-05-20). Session state lives at .session/regression-testing/<scope>/{plan.md, progress.md} per agentic-qa-core/references/session-management.md §3 + §9. The single highest-value resume case: if the Monitor subagent dies while watching a long CI run but RUN_ID was captured in plan.md, Phase 0 re-attaches via gh run view <RUN_ID> instead of re-triggering CI (saves the whole run's wall-clock; read the last gh run view duration).

This skill is compliant with the doctrine in AGENTS.md §3 "Orchestration Mode" and the session contract in .agents/skills/agentic-qa-core/references/session-management.md. Every dispatch follows the 7-component briefing format defined in .agents/skills/agentic-qa-core/references/briefing-template.md, and the pattern selected per stage matches the decision guide in .agents/skills/agentic-qa-core/references/dispatch-patterns.md. The two CI-bound stages (long-running watch, multi-artifact download) and the high-volume failure classification step are the hotspots — everything else stays inline because the dispatch overhead is not justified.

Stage	Pattern	Subagent role
Trigger workflow (`gh workflow run`)	Single	inline — no dispatch needed (one shell call)
Wait/monitor `gh run watch`	Background	one Monitor subagent runs the watch; main thread continues with prep work; subagent notifies on exit
Download the run's artifacts	Parallel	one subagent per artifact `gh run view --json artifacts` lists (no rate-limit risk)
Classify failures (chunks of ~10 tests each)	Parallel	N subagents based on failure volume; cap per `dispatch-patterns.md` to avoid context dilution
Compute metrics (pass-rate, trends)	Single	inline — needs aggregated state, low cost
Generate executive report	Single	inline — final synthesis, decisions live here
GO / CAUTION / NO-GO verdict	Single	inline — main thread owns release decisions

Error protocol: On any subagent failure: STOP, report full context to user, present retry / skip / abort options. Do NOT auto-fix. See .agents/skills/agentic-qa-core/references/orchestration-doctrine.md. A skill that itself broke (a wrong step, a missing verifier, a stale rule) is reported upstream per ../agentic-qa-core/references/upstream-feedback.md: drafted and redacted locally, filed only on explicit OK, verified with gh issue view.

Fleet seam (optional)

Triage itself is never parallelized across sessions: one conductor reads the run, classifies, and owns the verdict. The seam is for what comes AFTER Phase 2, when the classification leaves several independent failure clusters that each need code — and only when the user asks for them to be worked at once.

Topology: one worktree per failure cluster. A cluster's fix is code (a spec, a locator, a fixture), so each worker gets its own checkout and its own branch; two sessions in one checkout contend on the git index even on disjoint files. One cluster = one worker = one branch. Never two workers on one cluster.
The conductor writes launch.txt in .session/regression-testing/<scope>/ — one self-contained line per cluster — always, whether or not any orchestration transport exists on the machine. Launching, supervising and closing those sessions is orca-orchestration/SKILL.md ([ORCHESTRATION_TOOL]): supervised launch is the native path, and launch.txt is the payload for the human-paste fallback when nothing can launch it.
The verdict never moves. GO / CAUTION / NO-GO, the metrics, the RTR (or the sprint-close STR) and every Jira write stay with the conductor (Phase 3). A worker fixes its cluster and reports; it does not re-score the run, does not file the defect, and does not transition the RTR.
Each worker's fix is authored under /test-automation (Plan → Code → Review) on its own branch, and lands per git_strategy — a regression fix is not exempt from the automation gate.
Silence rule: the absence of an orchestration transport is never named to the user, never appears in the preflight gate, and never appears in the Environment block or the report.

Readiness Preflight Gate (MANDATORY — runs before Phase 0)

Full doctrine: agentic-qa-core/references/preflight-gate.md. Runs FIRST, before the resume check and any gh workflow run. Two laws: (1) args-as-answers — the suite (regression/smoke/sanity), env, and any grep/test_file are provided args; ask only the gaps. (2) probe, don't assume. Surface gaps + REDs as ONE AskUserQuestion checklist; self-fix with approval + explanation; STOP on any blocking RED. This generalizes the Phase 1 §Preflight (gh auth) to a full readiness check pulled to t=0. Generic baseline (env resolution, secret/restart handling, the two laws, output contract) is inherited from the reference §3.1 — not repeated here. Below is only this skill's specific capability delta (note: test-user creds, MCPs and browsers live inside the CI runner, not the orchestrator).

Capability	Need	Why here
GitHub CLI authenticated	REQUIRED	Every stage drives CI via `gh` (`gh auth status`, `gh workflow run`, `gh run watch`, `gh run download`). Not authed → user runs `gh auth login` (suggest the `!` prefix); do not proceed.
Workflow files present	REQUIRED	`.github/workflows/` must hold the regression/smoke/sanity workflow for the chosen suite, with the inputs this skill passes.
GitHub Actions Secrets/Variables	REQUIRED	The runner authenticates with env-prefixed creds (`secrets.<ENV>_USER_EMAIL` / `_PASSWORD`) + `XRAY_` / `ATLASSIAN_` as Repository/Environment Secrets — the suite 401s mid-run without them. `gh secret list` (add `--env <env>` for environment scope) shows them; missing → `gh secret set <NAME>` from `.env`. if `/test-framework-adaptation` did not push the secrets, this is the most common silent gap.
Allure local	REQUIRED	`bunx allure` resolves (devDep, no global install); `allurerc.mjs` present for `bun allure:agent` markdown triage.
Active env	REQUIRED	The suite runs against `<<ACTIVE_ENV>>` (default `{{DEFAULT_ENV}}`). Confirm it is the intended target before a long run.
`[TMS_TOOL]` (result sync)	OPTIONAL	Only when `.agents/project.yaml` `testing.tms_cli` is set: Phase 1 creates the RTR before the trigger, Phase 3 posts the verdict on it and closes it. jira-xray → `/xray-cli` + `XRAY_*`.
`[ISSUE_TRACKER_TOOL]` (file regression issues)	OPTIONAL	Only on NO-GO / CAUTION-with-regressions, to file issues. Load `/acli` then.

Test-user creds, OpenAPI / API token, DBHub and Playwright browsers live inside the CI runner, not the orchestrator — this skill does not exercise them locally, so they are out of scope for this gate. After the gate clears (all REQUIRED GREEN), continue to Phase 0 below.

Phase 0 — Session resume check (MANDATORY, inline)

Before suite selection or any gh workflow run, run the resume contract from agentic-qa-core/references/session-management.md §4:

Compute prospective <scope> = <env>-<YYYY-MM-DD> from invocation context (env defaults to {{DEFAULT_ENV}}).
Check .session/regression-testing/<scope>/progress.md.
If it does NOT exist → proceed to suite selection + Phase 1 preflight + plan.md write.
If it DOES exist:
- Read plan.md (captured suite, env, workflow_file, RTR_KEY if Phase 1 already created the run record, RUN_ID if Phase 1 already triggered).
- Read tail of progress.md.
- If RUN_ID is present AND progress.md last entry is Phase 1 — Trigger — status: completed but Monitor entry is missing/failed: surface the option to re-attach to the existing RUN_ID via gh run view <RUN_ID> --json status,conclusion instead of re-triggering. This is the high-value resume case.
- Otherwise surface the standard offer resume / restart / abort. On restart, archive to .session/.archive/<YYYY-MM-DD>-regression-testing-<scope>-aborted/ first.

When to run each suite

Suite	Workflow file	Duration	Use when
`regression`	`regression.yml`	a long run (read the last `gh run view` duration)	Pre-release validation, nightly full run
`smoke`	`smoke.yml`	a short run (read the last `gh run view` duration)	Post-deploy health check, `@critical` only
`sanity`	`sanity.yml`	varies (read the last `gh run view` duration)	Validate one feature / one file / one grep pattern

If the user says "run regression" with no qualifier, default to regression on {{DEFAULT_ENV}}. If they say "smoke" or "critical only", use smoke. If they specify a file, grep, or single feature, use sanity.

Local reporting (Allure, no global install)

Allure (major pinned in package.json) is a devDep — bunx allure resolves to the local node_modules/.bin/allure, no brew install allure / scoop install allure required. Configuration lives at allurerc.mjs, single-plugin BY DESIGN: with only the Awesome plugin the generated index.html IS the report (no card-chooser landing), and its top-left mode dropdown covers everything — Report (drill-down, tag filters), Graphs (complete executive chart set: status, dynamics, severities, stability, testing pyramid, durations…), Timeline. Never add plugin-dashboard instances — they duplicate Graphs with fewer charts and bring back the landing screen (rationale in allurerc.mjs comments). Trend charts are fed by historyPath: ./.allure/history.jsonl and populate from the 2nd run onward.

Use case	Script	Underlying command
Run tests + auto-generate report (human review)	`bun allure:run`	`bunx allure run -- bun test`
Run tests + emit markdown for AI review	`bun allure:agent`	`bunx allure agent -- bunx playwright test`, through the env launcher like the `test*` scripts
Generate report from existing `./allure-results`	`bun allure:generate`	`bunx allure generate ./allure-results`
Serve last generated report locally	`bun allure:open`	`bunx allure open`
Live-refresh report during iterative dev	`bun allure:watch`	`bunx allure watch ./allure-results`

bun allure:agent is the AI-friendly entry point: it produces a markdown summary the orchestrator (or a Verifier subagent) can read directly without parsing HTML. Use it whenever you need a structured pass/fail breakdown after a local re-run while triaging a CI failure (Phase 2 step 1, before downloading the merged-allure-results artifact from CI).

CI artifacts (merged-allure-results-{env}) are still produced by the workflow and downloaded via gh run download as documented in Phase 2. The published GitHub Pages reports are generated by scripts/ci/publish-allure-pages.ts with the SAME allurerc.mjs and allure devDep as local runs — the /{env}/{suite}/ URL redirects straight into the latest run's Awesome report (Report | Graphs | Timeline), with per-suite trend history and the --keep retention scripts/ci/publish-allure-pages.ts applies.

Allure version-currency check (MANDATORY during any Allure/Pages setup)

The boilerplate pins allure / allure-playwright / allure-js-commons at scaffold time, so by the time someone installs the repo and runs this setup they are usually behind upstream. Whenever this skill performs Allure setup (first local report, preflight RED on the "Allure local" row) or GitHub Pages setup (references/github-pages-setup.md), run this check FIRST:

npm view allure version && npm view allure-playwright version → compare against package.json.
Same major behind → summarize the news for the user (release notes: gh api repos/allure-framework/allure3/releases), then offer bun update allure allure-playwright allure-js-commons (or bump the ^ ranges + bun install). Keep allure-js-commons in lockstep with allure-playwright (it is imported directly by tests/components/TestFixture.ts for the layer auto-label).
New major available → NEVER upgrade silently. Present breaking changes and wait for explicit user approval.
Config-currency (older scaffolds) — the updater (bun run up, read package.json) syncs skills and appends new devDeps, but it NEVER overwrites allurerc.mjs or tests/components/TestFixture.ts (project-adapted files). If the local allurerc.mjs predates the current template (no historyPath, no categories, or stale plugin-dashboard instances), OFFER to migrate it: fetch the boilerplate's current allurerc.mjs as reference (https://raw.githubusercontent.com/upex-galaxy/agentic-qa-boilerplate/main/allurerc.mjs), preserve the project's name, and port the config. Same for the _allureLayer auto-fixture in TestFixture.ts (feeds the testingPyramid + durations-by-layer charts in Awesome's Graphs tab) — without it those charts render empty. Never overwrite silently; show the diff and wait for approval.
After any bump: bun allure:generate from existing results (or a sandbox run) and confirm the report renders — the root index.html must open the Awesome report directly, with the Report | Graphs | Timeline mode dropdown working.

Known gotchas to preserve on upgrade (context in allurerc.mjs comments):

Dashboard chart type values must match ChartType in @allurereport/charts-api — the plugin README's trend/pie examples are stale and yield an empty dashboard (404 on widgets/charts.json).
historyPath must stay OUTSIDE allure-report/ (./.allure/history.jsonl) or test:clean erases trend history.
Multiple instances of one plugin need an explicit import: field — a custom key alone does not resolve (relevant only if a project deliberately adds extra report views).

Phase 1 — Execute

Preflight (always)

gh auth status
gh repo view --json name,owner
gh workflow list

If gh is not authenticated, stop and ask the user to run gh auth login. Do not proceed.

Write .session/regression-testing/<scope>/plan.md per agentic-qa-core/references/session-management.md §6 BEFORE the Trigger step below. Capture: Goal (suite + env + reason for run), Inputs (workflow file path, env vars, optional grep/test_file for sanity), Approach (subagent pattern per stage from the dispatch table above), Phase breakdown (Trigger → Monitor → Download → Classify → Compute → Report → Verdict), Risks, Verification checklist (all 3 artifacts download + verdict emitted), Cross-references (.context/reports/regression-<env>-<date>.md will hold the final verdict). RTR_KEY lands in plan.md §Inputs after the Create-the-RTR step below, and RUN_ID joins it AFTER the Trigger step captures it — append, do not rewrite the body.

Create the RTR (before the trigger)

The run record is born BEFORE CI starts, so the workflow receives its key and imports into it instead of into a shared secret. <<SCOPE_ID>> = <<ACTIVE_ENV>>-<YYYY-MM-DD> (the same value as the session <scope>), or the release tag (e.g. v2.3.0-rc1) when the user says the run is a release candidate. Regression suite only: smoke and sanity get no RTR (see the Trigger block).

Resolve the RTP: find the Test Plan titled RTP: {{PROJECT_KEY}}: Regression Test Plan. Missing → do NOT create it here (the RTP is /test-documentation's promotion artifact): create the RTR without the testPlan edge, record the gap as a stated N/A in the verifier and the report, and recommend running the promotion first.
Find-or-create the RTR. A same-scope RTR still {{jira.status.test_execution.active}} (no verdict yet: an ENVIRONMENT re-run, a resumed session) is REUSED. A same-scope RTR already {{jira.status.test_execution.close}} (verdict written) means this is a NEW run: append #2 (then #3) to the scope-id. Never reopen a closed RTR (reactive stays unused).
Persist <<RTR_KEY>> in .session/regression-testing/<scope>/plan.md §Inputs, on the line RUN_ID will join after the trigger. Resume reads both.
Pass it to CI as -f execution_key=<<RTR_KEY>> on gh workflow run (Trigger block below). The workflow's execution_key input overrides the STP_EXECUTION_KEY secret (references/ci-cd-integration.md §4).

[TMS_TOOL] Find Test Plan:
  summary: RTP: {{PROJECT_KEY}}: Regression Test Plan
  -> <<RTP_KEY>>   (missing: RTR without testPlan + stated gap; never create the RTP here)

[TMS_TOOL] Find-or-create Test Execution:
  summary: RTR: <<SCOPE_ID>>: Regression Testing        # e.g. RTR: staging-2026-09-23: Regression Testing
  parent: {QA Test Artifacts epic: qa.qa_epics.test_artifacts_epic.key, resolved by .name when the key is null}
  testEnvironments: [<<ACTIVE_ENV>>]
  assignee: self
  testPlan: <<RTP_KEY>>
  -> <<RTR_KEY>>   (persist beside RUN_ID, pass as execution_key)

# Modality jira-native
RTR: n/a (jira-native has no Test Executions); results = per-Test status writes + [LOCAL] report
  -> no key: trigger WITHOUT execution_key; the CI jira-native leg writes per-Test status,
     the verdict lives in the [LOCAL] report (and a comment on the Release issue when one exists)

Sprint-close variant. When the user says this run IS the sprint close, the target is the STR instead of an RTR: find-or-create STR: Sprint#{N}: Regression Testing per the existing contract (§TMS sync: parent QA Test Artifacts, Test Environment, assignee self, testPlan → STP AND testPlan → RTP, dual membership), persist its key in the same plan.md slot, and pass it as execution_key the same way. N comes from the user or from the STP's scope-id, never invented.

Trigger

# Full regression: execution_key = the RTR (or sprint-close STR) created in the step above
gh workflow run regression.yml \
  -f environment=staging \
  -f execution_key=$RTR_KEY \
  -f video_record=false \
  -f generate_allure=true

# Smoke (no execution_key: smoke never writes into a regression execution;
# pass one only when its results deliberately belong in a specific Execution)
gh workflow run smoke.yml -f environment=staging -f generate_allure=true

# Sanity (grep OR test_file, never both; same execution_key rule as smoke)
gh workflow run sanity.yml -f environment=staging -f test_type=e2e -f grep="@auth"
gh workflow run sanity.yml -f environment=staging -f test_file="tests/e2e/auth/login.test.ts"

Modality jira-native: omit execution_key on every suite; there is no Execution to import into.

Capture run ID

# Wait 3-5 seconds for the run to register, then:
gh run list --workflow=regression.yml --limit=1 --json databaseId,status,createdAt -q '.[0].databaseId'

Store as RUN_ID. Every subsequent step uses it.

Progress checkpoint after Trigger: append RUN_ID to .session/regression-testing/<scope>/plan.md §Inputs (so resume can re-attach) AND append a phase entry ## Phase 1.Trigger — <ts> with status: completed, next: Phase 1.Monitor, notes: RUN_ID=<value> RTR_KEY=<value> to progress.md. This is the critical persistence point — Trigger landing without RUN_ID persisted means resume cannot re-attach, and without RTR_KEY a resumed session cannot tell which Execution CI imported into.

Monitor to completion

Use the dispatch defined in §Subagent Dispatch Strategy: Background. Delegate gh run watch <RUN_ID> to a Monitor subagent so the main thread is freed to prepare the report scaffold and load the classification rubric. See references/ci-cd-integration.md §"Monitoring the workflow run (Background dispatch)" for the full briefing.

Reference command (executed inside the subagent, not inline on the main thread):

gh run watch <RUN_ID> --exit-status
# Fallback polling (only if gh run watch is unavailable):
gh run view <RUN_ID> --json status,conclusion
# status: queued | in_progress | completed
# conclusion (only when completed): success | failure | cancelled | timed_out

Do not start Phase 2 until the Monitor returns status: completed.

Output of Phase 1

A short execution summary with: workflow name, run ID, environment, duration, conclusion, per-job status, artifact list, and the Allure URL pattern https://{owner}.github.io/{repo}/{environment}/{suite}/.

Read references/ci-cd-integration.md when configuring new workflows, debugging CI-only failures, tuning sharding / retries / timeouts, or wiring up secrets and variables.

Phase 2 — Analyze

Step 1: Collect data

Use the dispatch defined in §Subagent Dispatch Strategy: Parallel for the three artifact downloads (allure / evidence / playwright). Fan out three subagents in a single tool-call block — each owns one artifact, writes to its own directory, and reports back when its download is verified. The metadata reads (gh run view) stay inline because they are short.

Reference commands (the metadata reads run inline; the three gh run download calls live inside the parallel subagents):

# Inline (main thread): full run context
gh run view <RUN_ID> --json status,conclusion,jobs,createdAt,updatedAt,url,headBranch,event,actor

# Inline (main thread): failed logs only (much smaller than --log)
gh run view <RUN_ID> --log-failed

# Inline (main thread): list artifacts so the parallel dispatchers know what to fetch
gh run view <RUN_ID> --json artifacts --jq '.artifacts[].name'

# Parallel subagent A — allure results
gh run download <RUN_ID> -n merged-allure-results-staging -D ./analysis/

# Parallel subagent B — failure evidence (screenshots, traces, videos)
gh run download <RUN_ID> -n e2e-failure-evidence       -D ./analysis/evidence/

# Parallel subagent C — playwright HTML report
gh run download <RUN_ID> -n e2e-playwright-report      -D ./analysis/playwright/

Each subagent uses the briefing shape in agentic-qa-core/references/briefing-template.md §"Parallel — Download 3 CI artifacts in regression-testing". Fan out one subagent per artifact gh run view --json artifacts lists; GitHub's per-run rate limits are not a concern at that size.

Step 2: Parse results

Source of truth priority: Allure results JSON > Playwright report.json > raw logs. Each Allure result has status, statusDetails.message, statusDetails.trace, and labels[] (look for testId = ATC ID, suite, and severity).

The suite label is tag-derived — single source of truth. Allure suite/grouping labels are NOT a separate taxonomy: they derive from the Playwright tag (@smoke / @regression / @e2e / @integration / @critical) that also drives CI scope selection. A test tagged @integration reports suite: integration automatically. So the suite you read here is exactly the scope CI ran — never reconcile it against a parallel Allure label set. Convention owner: test-automation/references/ci-integration.md §3.2.1.

Step 3: Compute metrics

Metric	Formula
Total	count of results
Passed / Failed / Skipped / Broken	count by `status`
Pass Rate	`Passed / Total * 100`
Duration	`max(stop) - min(start)`
Trend	current pass rate − previous run pass rate

Exclude KNOWN-BLOCKED from the gating pass-rate. Tests classified KNOWN-BLOCKED (tagged @blocked:{BUG-KEY}, see Step 4) are parked behind an already-filed bug — they are NOT regression failures and must not depress the pass-rate that drives the GO/NO-GO score. Compute the gating Pass Rate over Total − KNOWN-BLOCKED, and report the blocked count separately (with each {BUG-KEY}) so the release decision is not gamed in either direction.

Previous-run comparison requires downloading artifacts of the previous run:

PREV=$(gh run list --workflow=regression.yml --limit=2 --json databaseId -q '.[1].databaseId')
gh run download $PREV -n merged-allure-results-staging -D ./analysis/previous/

Step 4: Classify every failure

Use the dispatch defined in §Subagent Dispatch Strategy: Parallel when the failure list has more than 10 entries. Shard the failures into chunks of ~10 (cap at 10 subagents) and fan out one classification subagent per chunk; merge their JSON reports in the main thread. For ≤10 failures, classify inline (the dispatch overhead is not justified). See references/failure-classification.md §"Parallel classification (default for >10 failures)" for the full briefing and merge protocol.

Apply this decision tree to each failed test (whether classified inline or inside a parallel subagent). Never mark a test REGRESSION without checking history first — that is the single most common misclassification.

Failed test
  │
  ├── Tagged @blocked:{BUG-KEY}? ────────────► KNOWN-BLOCKED
  │   (test asserts test.fail('Blocked by {BUG-KEY}') — a deliberately
  │    parked test, not a fresh regression; excluded from gating pass-rate)
  │
  ├── Linked to a known-issue ticket? ───────► KNOWN ISSUE
  │
  ├── Error matches environment pattern? ────► ENVIRONMENT ISSUE
  │   (ECONNREFUSED, ETIMEDOUT, net::ERR_, Navigation timeout,
  │    browserType.launch, 502/503, context deadline exceeded)
  │
  ├── No history (first-ever run)? ──────────► NEW TEST FAILURE
  │
  ├── Failure rate > 20% over last 10 runs? ─► FLAKY
  │
  └── Passed in last ≤ 5 runs, now fails? ───► REGRESSION   (release blocker)

Category	Impact	Action	Evidence the classification must cite
KNOWN-BLOCKED	LOW	Already tracked by `{BUG-KEY}` — exclude from gating pass-rate, list in report with the blocking bug key. No new Jira bug (the marker already names the open bug)	the `@blocked:{BUG-KEY}` tag in the spec and its bug key
REGRESSION	HIGH	Block release, file Jira Bug/Defect (Phase 3 §File defects in Jira, doctrine Part 1), assign	none: it is the closed value
FLAKY	MEDIUM	Schedule stabilization, do not block — no Jira bug	the run ids and results of at least 5 runs, and the build each ran on (unchanged across them)
KNOWN ISSUE	LOW	Document against existing ticket, do not block — no new Jira bug	the ticket key, created before this run started, that names the ATC ID or test title
ENVIRONMENT	MEDIUM	Re-run after infra check — no Jira bug	the log line matching an environment pattern, plus the other tests that failed on the same host in the same run
NEW TEST	LOW	Manual verification → if a genuine product defect, file Jira Bug/Defect; else accept or fix	the history query (Allure history, TMS runs) that returned no earlier execution

Classification is a fail-closed gate (agentic-qa-core/references/orchestration-doctrine.md, FAIL-CLOSED GATES). Every class except REGRESSION lets a failure past the release gate, and the classifier is the one who picks it, so a non-blocking class stands only with its citation from the last column. No citation, an empty one, or a ticket this session filed to make the failure "known" → the failure counts as REGRESSION. The same holds for a parallel chunk: a classification returned without its evidence is merged as REGRESSION (references/failure-classification.md §Parallel classification).

KNOWN-BLOCKED — consuming the blocked-test marker. The @blocked:{BUG-KEY} tag + test.fail('Blocked by {BUG-KEY}') marker is defined in test-automation (references/automation-standards.md §7 Stability; the PROGRESS.md blocked-tests note lives in references/planning-playbook.md) — this skill only consumes it. The GO/NO-GO gate MUST recognize @blocked:{BUG-KEY} tests and classify them as KNOWN-BLOCKED, never REGRESSION: they are deliberately parked behind an already-filed bug, not a fresh failure. Exclude them from the pass-rate that gates the release (see §Compute metrics), and list each in the report under its own heading with the blocking {BUG-KEY}. Do NOT file a new Jira bug — the marker already names the open one.

sdet CI-fallback clause (integration-trunk suites only): an ENVIRONMENT-class red on a Sanity-CI run for a ticket branch may authorize merging into the integration trunk — never the final trunk → main PR — when proven by BOTH (a) the change passing locally on local AND staging, and (b) the same red being present independent of the change (nightly already red, or the failing line is shared pre-existing code). File a separate infra/flake ticket and reference it in the PR. This is NOT a relaxation of the GO bar: a REGRESSION-class failure is never eligible, and the final PR to main still requires a genuinely green test step. See .agents/skills/git-flow-master/references/sdet-integration-trunk.md §CI-fallback clause.

Read references/failure-classification.md when: the decision tree is ambiguous, you need the full error-pattern catalogue, you are classifying a borderline case, or you are computing flakiness over historical runs.

Step 5: Assess severity per failure

Severity is independent of classification. A FLAKY test on the checkout flow is still CRITICAL severity.

Severity	Criteria
CRITICAL	Core user journey (login, checkout, payment). Any `@critical` tagged test.
HIGH	Major feature (search, profile, dashboard)
MEDIUM	Secondary feature (filters, preferences)
LOW	Edge case or admin-only path

Output of Phase 2

An analysis block with: metrics table, trend delta, one section per failure category (Regressions first, then Flaky, Known, Environment, New), per-failed-test detail (name, ATC ID, suite, error, last-pass date, screenshot link), job summary, and a preliminary verdict.

Phase 3 — Report & Decide

GO / CAUTION / NO-GO scoring

Compute a weighted score from the analysis. Maximum is 10.

Factor	+3	+1	0	-1	-2	-3
Pass Rate	≥ 95%	90–95%			< 90%
Regressions	0	1-2 Low		1+ Medium		Any High/Critical
Critical tests	All pass					Any fail
Flaky tests		≤ 3	4-5	> 5

Verdict thresholds:

Score ≥ 7 → GO — release approved
Score 4-6 → CAUTION — manual review required, document accepted risks
Score < 4 → NO-GO — block release, fix regressions, re-run

Never auto-GO if: any @critical test fails, any REGRESSION with HIGH/CRITICAL severity exists, or pass rate < 90%. These are hard vetoes regardless of score.

File defects in Jira (when decision = NO-GO or CAUTION with regressions)

Quality issues go to Jira, not GitHub. A regression-discovered product failure is a defect-management artifact and follows agentic-qa-core/references/defect-management-doctrine.md — the same authority /sprint-testing uses. This skill files the issue IN JIRA with the full mandatory field matrix; it does NOT open a GitHub issue.

Only CONFIRMED real product failures become Jira issues. Use the Phase 2 Step 4 triage as the gate: file in Jira only for the REGRESSION class and for a NEW TEST failure once it is manually confirmed to be a genuine product defect (not a bad assertion). FLAKY, ENVIRONMENT, and KNOWN ISSUE do NOT get a Jira bug — they route to stabilization / infra / the existing ticket as the classification table already prescribes. The failure-triage classification and the defect issue-type are separate axes: triage decides whether to file; the doctrine decides what type and what fields.

For each issue that clears the gate:

Classify Bug vs Defect by the affected feature's lifecycle stage, NOT by where the failure ran (doctrine Part 1): the regressed feature is already live above Staging (production / superior env) → Bug; the feature is still pre-release (Staging or below) → Defect. A genuinely new, desirable behavior surfaced beyond the AC → Improvement (Part 1).
File it in Jira with the full mandatory field matrix (doctrine Part 5): severity (impact-based) → priority auto-derived (Part 5.1), native components = affected product module (Part 3, mandatory & pre-existing), root_cause + error_type + test_environment, qa_assignee = the authenticated session user (self; never-overwrite, Part 2), and evidence (Allure link + failure screenshots/traces/logs from ./analysis/evidence/).
Parent to the QA Defect Management epic — the QA process epic (qa.qa_epics.defect_epic.name), found-or-created; NEVER a product/dev epic (Part 4).
Link to the source Story/feature for traceability via the causal link (Part 4) — the regressed ATC's covering Story.
Write via acli/REST (doctrine Part 6): create with acli workitem create --from-json (create-time customfields under additionalAttributes.customfield_*, native components:[{name}]); set customfields/components on an existing issue via REST PUT /rest/api/3/issue/{KEY}; qa_assignee is read-before-write. Because this stage may run from CI, load /acli first (it owns auth, syntax, and the REST-PUT pattern in references/acli-integration.md).

Run the doctrine's filing gate (Part 9) before submitting each issue. Save the returned Jira key to reference in the report.

TMS sync (optional, when `[TMS_TOOL]` is configured via `.agents/project.yaml` `testing.tms_cli`)

Prerequisite: Load /xray-cli skill (Modality jira-xray) before executing the [TMS_TOOL] commands below. In Modality jira-native, load /acli instead and map test-execution operations to native Jira issues (see test-documentation/references/jira-setup.md).

A regression run maps to Jira items (items-first by excellence: no Story custom field exists at this altitude). The default pair is RTP + RTR. The STP + STR pair applies only to the sprint-close run.

RTP (Regression Test Plan): a Test Plan item titled RTP: {{PROJECT_KEY}}: Regression Test Plan, parent QA Master Test Plan (qa.qa_epics.master_test_plan_epic.name). Producer: /test-documentation (promotion of regression-candidate TCs). This skill only CONSUMES it as the RTR's testPlan target: it is never written into and never created here. An Xray Test Plan aggregates the LATEST status of each of its Tests across all Executions, so the RTP answers "is the regression suite green" on its own as RTRs accumulate (test-documentation/references/xray-platform.md §4).
RTR (Regression Test Results): a Test Execution item titled RTR: <<SCOPE_ID>>: Regression Testing (e.g. RTR: staging-2026-09-23: Regression Testing, or RTR: v2.3.0-rc1: Regression Testing for a release candidate), parent QA Test Artifacts (qa.qa_epics.test_artifacts_epic.name), testPlan → RTP, Test Environment + assignee = self at create. Producer: THIS skill, Phase 1, before the trigger (§Create the RTR). It is the CI import target: the execution_key dispatch input carries its key, and the STP_EXECUTION_KEY secret (name kept for downstream repos) is only the fallback for a scheduled run that no dispatcher minted a key for. One RTR per verdict: an ENVIRONMENT re-run before the verdict imports into the same RTR; a re-run after a verdict (a NO-GO fixed and retried) is a NEW RTR with a #2 scope-id; never reactive.
STP + STR (sprint close ONLY): STP: Sprint#{N}: {sprint objective} (Test Plan, parent QA Master Test Plan, relates to the Sprint; producer /sprint-testing, whose Session Start find-or-creates it on the sprint's first ticket; this skill consumes it and find-or-creates it only as a fallback, never writes results into it) and STR: Sprint#{N}: Regression Testing (Test Execution, parent QA Test Artifacts, relates to the Sprint, testPlan → STP AND testPlan → RTP, dual membership). The STR is the sprint-close recap of all sprint results: whoever arrives first creates it, the other completes it (/sprint-testing's batch close, or this skill when it runs the closing regression). A regular regression run during the sprint is an RTR, never the STR. The run's term is Regression Testing: "Sprint" already comes from the Sprint#{N} scope-id.

N is a sprint-close concern only. The RTR needs no sprint number. When the run IS the sprint close, N comes from the user or from the Sprint#{N} scope-id of the STP this skill finds. Never invent it: a guessed N forks a duplicate STP/STR pair. No STP found and no N given → ASK before creating anything at sprint altitude.

Environment gate: every Test Execution this skill creates (the RTR, and the STR at sprint close) carries the Test Environment taken from active_env in .agents/project.yaml, set at create time. An Execution without its environment fails the checklist: do not write results into it until the environment is set.

Ownership gate: every artifact this skill CREATES (the RTR; at sprint close the STR, and the STP in the fallback case) carries assignee = the authenticated session user, set at create time — agentic-qa-core/references/artifact-lifecycle.md §2. Xray refuses membership edits on a Test Plan the caller does not own, so an unassigned Plan turns into a blocker the moment tests must be added to it. If the find returns an artifact someone ELSE owns, do not reassign it silently: ask first.

Lifecycle gate (agentic-qa-core/references/artifact-lifecycle.md §1):

The RTR is born {{jira.status.test_execution.active}} in Phase 1 and MUST be transitioned to {{jira.status.test_execution.close}} via {{jira.transition.test_execution.complete}} after the GO / CAUTION / NO-GO verdict comment is posted on it: never before the verdict, never left open, never reopened (reactive stays unused; a later run is a new RTR).
The STR (sprint close only) follows the same rule after the sprint-close verdict.
The RTP (and any Test Plan this skill only consumed) stays at {{jira.status.test_plan.ready}} and is never completed by a regression run: the RTP is long-lived, and a suite execution does not finish the plan it ran from. Do NOT fire {{jira.transition.test_plan.complete}} here.
The STP is closed by whoever owns sprint close, not by this skill — unless this skill IS the sprint close (see the sprint-close DoD in stage-gates.md), in which case {{jira.transition.test_plan.complete}} moves it to {{jira.status.test_plan.completed}} after the STR is closed.
Unmapped slug → artifact-lifecycle.md §4 fallback: list the LIVE transitions, propose the closest synonym in ONE AskUserQuestion, fire the live id on yes, recommend bun run jira:sync-workflows. Never skip silently, never guess an id.

The run record already exists when Phase 3 starts: the RTR was created in Phase 1 and CI imported into it through execution_key, so Phase 3 only writes what CI did not, posts the verdict and closes it. At sprint close, never assume another producer already created the STR; if /sprint-testing's batch close got there first, the find returns its item and this skill only completes it:

# Default: the RTR created in Phase 1 (§Create the RTR)
[TMS_TOOL] Update Test Execution:
  executionKey: <<RTR_KEY>>
  results: {per-ATC status + failure comments from Phase 2, only what the CI import did not land}

[ISSUE_TRACKER_TOOL] Add Comment:
  issue: <<RTR_KEY>>
  body: {GO / CAUTION / NO-GO verdict, score, blockers, workflow-run + Allure links}

# After the verdict comment is posted: close the run, never leave it ACTIVE
[ISSUE_TRACKER_TOOL] Transition: {{jira.transition.test_execution.complete}}   # active -> close
  issue: <<RTR_KEY>>

# Sprint close ONLY: the STR is the target instead, with dual plan membership
[TMS_TOOL] Find-or-create Test Execution:
  summary: STR: Sprint#{N}: Regression Testing
  parent: {QA Test Artifacts epic: qa.qa_epics.test_artifacts_epic.name}
  links: {relates to → Sprint; testPlan → STP key; testPlan → RTP key}
  environment: {active_env from .agents/project.yaml}
  assignee: self
  -> then the same Update / Add Comment / Transition sequence on the STR key

# Modality jira-native
RTR: n/a (jira-native has no Test Executions); results = per-Test status writes + [LOCAL] report
[ISSUE_TRACKER_TOOL] Update Test status: {per-Test status field write, one per executed Test;
  the CI leg already did this when AUTO_SYNC is on, so write only what it did not land}
verdict: `.context/reports/regression-{env}-{date}.md` ([LOCAL]) + a comment on the Release issue when one exists

Write the report

Save to .context/reports/regression-{env}-{date}.md. Use references/failure-classification.md only if you need the pattern catalogue; the report template itself is inline below.

Report template

# Regression Quality Report — {env} — {date}

## Executive Summary
**Verdict: {GO / CAUTION / NO-GO}**
Score: {score}/10. {one-line rationale}

| Metric | Value | Threshold | Status |
|--------|-------|-----------|--------|
| Pass Rate | {x}% | >= 95% | {ok/warn/fail} |
| Regressions | {n} | 0 | {ok/warn/fail} |
| Critical failures | {n} | 0 | {ok/warn/fail} |
| Flaky | {n} | <= 3 | {ok/warn/fail} |
| Duration | {d} | - | - |

## Release Blockers
{if NO-GO, enumerate regressions with severity, owner, ETA. Otherwise: "None."}

## Failure Details
### Regressions ({n})
  - {test} | {atc_id} | last passed {date} | [issue]({url}) | probable cause: {...}

### Flaky ({n}) — schedule stabilization
### Known Issues ({n}) — accepted
### Known-Blocked ({n}) — excluded from gating pass-rate
  - {test} | {atc_id} | blocked by [{BUG-KEY}]({url})
### Environment ({n}) — re-run after infra check

## Trend (last 5 runs)
{ASCII sparkline or pass-rate table}

## Links
- Workflow run: {url}
- Allure: {url}
- Created issues: {list}
- TMS execution: RTR: {key} (testPlan → RTP {key}) | STR: {key} (sprint close, testPlan → STP + RTP) | n/a (jira-native, stated skip)

## Recommendations
1. Immediate (pre-release): {...}
2. Short-term (this sprint): {...}
3. Long-term (tech debt): {...}

Post-decision actions

Decision	Actions
GO	Mark release candidate approved; schedule post-deploy smoke
CAUTION	Review with team lead; document accepted risks; proceed deliberately
NO-GO	Block release; assign regression issues; schedule fix verification; plan re-run

Whatever the verdict, close the run record: post the GO / CAUTION / NO-GO comment ON the RTR (on the STR only when the run is the sprint close), transition it to {{jira.status.test_execution.close}} via {{jira.transition.test_execution.complete}}, leave the RTP at {{jira.status.test_plan.ready}}, then run the light stage verifier (agentic-qa-core/references/artifact-lifecycle.md §5). Stage-specific lines:

[ ] RTR exists by KEY and existed BEFORE the CI trigger (its key went out as execution_key)
[ ] RTR carries its Test Environment (active_env), assignee = self, parent QA Test Artifacts
[ ] RTR -> RTP linked via the `testPlan` edge (a missing RTP is a STATED N/A naming the promotion gap)
[ ] Verdict comment posted ON the RTR (the durable record, not the local report file)
[ ] RTR at {{jira.status.test_execution.close}} via complete, AFTER the verdict; never reactive
[ ] Sprint close only: STR exists by KEY, testPlan -> STP AND -> RTP, closed after the sprint-close verdict
[ ] RTP untouched at {{jira.status.test_plan.ready}} (a regression run never completes it)
[ ] Modality jira-native: the RTR line is the stated skip note; per-Test status writes landed; [LOCAL] report written
[ ] Any unmapped slug went through the §4 fallback (asked), never a silent skip

Per-phase progress + Archive

After Phase 1 Monitor returns, after each Phase 2 step (Collect / Parse / Compute / Classify / Severity), and after Phase 3 Verdict, the orchestrator appends a phase entry to .session/regression-testing/<scope>/progress.md per agentic-qa-core/references/session-management.md §7. artifacts_touched records the downloaded CI artifacts (allure / evidence / playwright dirs) + the final .context/reports/regression-<env>-<date>.md.

After the Verdict emits, the orchestrator runs Archive per agentic-qa-core/references/session-management.md §8: moves .session/regression-testing/<scope>/ to .session/.archive/<YYYY-MM-DD>-regression-testing-<scope>/ (two-file dir preserved) and calls mem_session_summary with the archive path. .context/reports/regression-<env>-<date>.md stays in the reports dir as a local generated report — that directory is gitignored [LOCAL] output (.context/reports/README.md), so the file exists only on the machine that ran the suite and nothing downstream may depend on it. The durable record is the RTR in the TMS (the STR only at sprint close) plus the GO / CAUTION / NO-GO comment posted on it. Modality jira-native: the per-Test status writes plus this [LOCAL] report, by stated skip.

On Verdict = NO-GO with regressions still being filed as issues, archive WAITS until the issue-creation step completes (so the session state still references the open issue list at archive time).

Gotchas

Allure URL is predictable but only live after the "Build & Deploy Allure Report" job succeeds. If that job failed, the URL 404s — analyze from downloaded artifacts instead.
gh run watch can time out on long suites. Fall back to polling gh run view <RUN_ID> --json status every 60-90 seconds.
gh run view --log dumps every step's output and is often >50MB on large suites. Always prefer --log-failed during analysis; use --job=<JOB_ID> --log for targeted drilldown.
This repo ships retries: 0 everywhere (playwright.config.ts) — tests must be deterministic, and a retry would only mask the flake. A flaky test therefore surfaces as a plain intermittent failure and is caught by the >20% history rule, never by a retry-pass signal. If a downstream project has consciously enabled retries, a test that passes on retry is still flaky — see the "Conscious divergence: enabling retries" box in references/ci-cd-integration.md for how to read retry counts in Allure.
ENVIRONMENT is not a scapegoat. ECONNREFUSED to your app's own API probably means the app crashed, not "infra glitch". Check if the same run has many unrelated tests failing on the same host — that is environment. One test failing with a network error on an endpoint that other tests hit successfully is more likely a REGRESSION.
Never mark NEW TEST as REGRESSION. A first-ever failure with no history is not a regression — it is unverified. Manually confirm once before classifying.
Flakiness needs 5 runs of history minimum before you can call it at all (below that, mark "insufficient history" and re-evaluate next sprint — do not guess). The failure-rate itself is computed over a wider window: the last N = min(10, available) runs (see references/failure-classification.md). 5 is the floor to have any signal; 10 is the window the percentage is actually computed over.
Sanity + grep and test_file are mutually exclusive. Passing both makes the workflow ignore one silently. Pick one.
Video recording inflates artifact size by 5-10x. Only enable video_record=true when debugging flakiness or capturing bug evidence. Never enable it for nightly regression.
CI credentials come from GitHub secrets, not .env. Do not copy values from local .env into workflow YAML — reference ${{ secrets.NAME }} only.
Session-footer contract (mandatory at close). The final phase is not done until the two chat-facing blocks from ../agentic-qa-core/references/session-footer-contract.md are printed: (1) consolidated screenshot list — repo-relative paths, verified on disk, bug annotations first — plus in-flow surfacing of every capture's path the instant it lands; (2) Session Footer listing skills/MCPs/CLIs actually used + testing levels touched, with explicit "none" entries for expected-but-untouched levels. Framing for this skill: execution. Multi-subagent sessions: each stage report carries the five footer fields (skills_loaded, mcps_used, clis_used, testing_levels_touched, screenshots_captured); the orchestrator compiles the footer ONCE at close. Chat only — never in a Jira comment or ATR body. Lessons noticed during the session are PROPOSED to .session/<skill-slug>/<scope>/refinements.md and never applied to a live skill, per ../agentic-qa-core/references/skill-refinement-protocol.md; the footer's Refinements proposed: line counts them.

Specific tasks

Configuring or debugging GitHub Actions workflows — read references/ci-cd-integration.md
Enabling GitHub Pages so the published Allure reports are browsable ("set up GitHub Pages", "report URL is 404", "publish the reports site") — read references/github-pages-setup.md (enable via gh api, first-build stuck/errored gotcha + manual rebuild, gh-pages history squash job). Run the §Allure version-currency check first.
Making CI reports PRIVATE ("reports must be login-protected", "no publiques evidencia pública", "protege los reportes") — read references/private-hosting-setup.md (Test Report Portal: Vercel + Supabase + private R2, work-email login, portal-side retention, history round-trip replacing gh-pages). The publish step in all three suite workflows is already dual-mode — you only wire secrets. GitHub Enterprise orgs have a zero-infra shortcut (Pages visibility → Private); offer it first.
Setting up Allure locally for the first time, or the user asks "is Allure up to date?" — run the §Allure version-currency check under §Local reporting.
Classifying a borderline failure (REGRESSION vs FLAKY vs ENVIRONMENT) — read references/failure-classification.md
TMS / Xray result import — load /xray-cli skill
Downloading traces or screenshots for a failure — use [AUTOMATION_TOOL] per AGENTS.md Tool Resolution; for Playwright trace inspection load /playwright-cli
Session contract (Phase 0 resume, plan.md/progress.md schemas, archive policy, Engram per-phase checkpoint, RUN_ID re-attach mechanism) — read ../agentic-qa-core/references/session-management.md. This skill is a producer of session/regression-testing/<scope>/... topic keys.

Anti-patterns — NEVER do these

R1. NEVER classify a failure as FLAKY without the ≥5-run history citation (§Step 4 table) — masks real regressions.
R2. NEVER emit GO when known REGRESSION class > 0 — quality gate is binary: regressions block.
R3. NEVER auto-retry failing tests in CI without surfacing the retry count in the report.
R4. NEVER skip Allure artifact download on red builds — evidence vanishes after the retention window.
R5. Record the tested commit (gh run view <RUN_ID> --json headSha) in the report; compare trends only between runs whose commits are known — different commit = different baseline.
R6. NEVER mix smoke + regression suite results into one pass-rate number — different SLOs.
R7. NEVER mark a test KNOWN-failure without a Jira ticket linking the suppression to a tracking issue.

Quick reference

# Trigger + get run ID in one shot (RTR_KEY = the Execution created in Phase 1 §Create the RTR)
gh workflow run regression.yml -f environment=staging -f execution_key=$RTR_KEY && sleep 5 && \
  RUN_ID=$(gh run list --workflow=regression.yml --limit=1 --json databaseId -q '.[0].databaseId') && \
  echo "RUN_ID=$RUN_ID"

# Wait for completion
gh run watch $RUN_ID

# Failed logs only
gh run view $RUN_ID --log-failed

# All failure evidence
gh run download $RUN_ID -n e2e-failure-evidence -D ./analysis/evidence/

# Previous run for trend
PREV=$(gh run list --workflow=regression.yml --limit=2 --json databaseId -q '.[1].databaseId')
gh run download $PREV -n merged-allure-results-staging -D ./analysis/previous/