Run quick end-to-end smoke testing for {{COMPANY_SLUG_UPPER}} changes by tracing changed code paths through the real local stack and verifying behavior live.
Installation
Installer avec Codex ou Claude Copiez ce prompt, collez-le dans Codex, Claude ou un autre assistant, puis laissez-le vérifier la page du skill et l'installer pour vous.
Run quick end-to-end smoke testing for {{COMPANY_SLUG_UPPER}} changes by tracing changed code paths through the real local stack and verifying behavior live.
End-to-End Test of the Current Change
Run a full end-to-end test of whatever was implemented in this conversation. The goal is a comprehensive scenario matrix that leaves nothing questionable — exercise every layer the change touched, prove the change works through real (not mocked) interfaces, and surface anything that doesn't.
Scope: This drives the change against your primary service (started by /qspinuplocal). If your change spans multiple services, start the others yourself and add their triggers to the scenario matrix — /qspinuplocal is single-service by design.
Companion skills
/qspinuplocal owns the local stack lifecycle for your primary service (against a local DB). This skill always calls /qspinuplocal for spin-up — it handles .env overrides, the load_dotenv(override=True) footgun, port collisions, and worktree resolution.
/qmanualt owns the UI E2E flow (Playwright + Claude in Chrome). For any UI change, hand control to /qmanualt after the API/worker layer is verified — see "UI testing — delegate to /qmanualt" below.
This skill (/qe2etest) is the orchestrator: it audits the diff, traces every changed code path forward to its production trigger (HTTP endpoint, worker queue, cron, scheduled job), drives those triggers itself against the real stack, delegates UI to /qmanualt, and verifies in the DB. It runs /qspinuplocal to start the stack.
The cardinal rule
An E2E pass requires invoking the same entrypoint production uses. A unit test, a Python REPL call into the changed function, or "the migration applied cleanly" is NOT an E2E pass — it's a partial check. If the change is reachable from a worker, the worker MUST be the thing that fires the code path. If the change is reachable from an HTTP route, curl MUST be the thing that fires it. Always. No exceptions, no shortcuts to "save time."
The single most common failure mode of this skill is the agent verifying a code change at the wrong altitude — proving the function works in isolation while never demonstrating the production trigger reaches it. Section "Step 1.5" below exists to prevent exactly that.
Hard rules — non-negotiable
Production trigger or it didn't happen. Every changed code path MUST be fired by the same entrypoint production uses. Worker code → run the worker script. HTTP code → curl. Cron code → invoke the cron entry. If you can't fire the production trigger, the result is BLOCKED, not "passed via alternative verification."
/qspinuplocal is unconditional for any change touching API / Service / Repository / Worker / UI / Cross-repo. The only exemption is a change purely in alembic/versions/* with no runtime impact. Don't second-guess: call the skill first, then test.
DEV_MODE=true is the default for local stack runs. The DEV_MODE=false purity goal is a CI / staging concern. Locally, the external auth provider, Dash routes, Auth-gated wizard flows, and many internal endpoints don't have a non-DEV-MODE path that works without a real user session. Use DEV_MODE=true unless you have a specific reason to test the auth boundary itself.
DEV_MODE is env-var only — NEVER commit auth/middleware code changes. If you find yourself thinking about editing oauth_provider_auth.py, middleware, or any auth code on the feature branch to make a test pass, stop. That's a code change leaking into the testing surface; revert before commit. Precedent: feedback_devmode_badge_deadlock.md.
Local database only: local_demo_db (the canonical default) or whatever DB /qspinuplocal was invoked against. Never write to a staging or production DB. Read-only queries against staging are fine if the user explicitly provides them and the URL points at a DB provider branch — otherwise stay local.
Worktrees take priority. If the user mentioned worktree paths in this session, the spin-up MUST point at those worktrees (pass them to /qspinuplocal). Otherwise the test exercises develop, not the change.
Real interfaces over mocks. UI clicks via the Chrome extension, API calls via curl, worker runs via the actual processor scripts, DB checks via psql. If the runtime can't be brought up, mark BLOCKED — don't substitute a unit-test for the e2e check.
Hard fail on missing layers. If the change touched the worker but the worker can't be started, that's a finding. Surface it; don't quietly skip.
Before auditing the conversation, pull in the spec source-of-truth so the scenario matrix is grounded in stated acceptance criteria, not just the implementation diff. This is what differentiates a comprehensive E2E from a check-what-I-just-wrote review.
For each ticket referenced in the session (Jira ID in commit messages, branch name, or chat):
Fetch the ticket via the atlassian MCP (mcp__atlassian__jira_get_issue) or by extracting the ID from git log and reading it. Capture: summary, description, every acceptance criterion (AC) verbatim, attached design notes / TRD links.
If the ticket links to a Confluence TRD or design doc, fetch that too (mcp__atlassian__confluence_get_page). Read the §3 / §4 / §AC blocks.
If the branch contains multiple commits closing multiple tickets, collect ACs across ALL of them.
Save the union of ACs to a working note so the brainstorm step can read them as input:
Skip this step only when the change has no Jira backing (true ad-hoc work). For epic-scale work or wave-of-tickets sessions, this step is non-negotiable — the matrix without ACs is the matrix of what the implementer remembers being asked to do, which is always a subset of what they were actually asked.
Step 1 — Audit the conversation
Re-read what was implemented in this session. Build a checklist of what changed and which runtime layers it touches.
Layer
What it covers
How to detect from the diff
DB / migrations
Alembic migrations, ORM-model changes, new indexes, listeners
New file in alembic/versions/, edits under data_models/orm/, new SQLAlchemy event listeners
API
FastAPI routers, request/response schemas
Edits under api/v1/*.py or api/routes.py; new endpoints in *_router
Service / Repository
Pure backend logic
Edits under services/ or repositories/ not exposed via HTTP
Worker
job-queue consumers, claim loops, scheduled jobs, OR any service/repo function reachable from a worker
Edits under your worker entrypoint(s) (services/*worker*.py, scripts/*-worker.py), OR a function whose call graph reaches one (grep the changed symbol there)
UI
React under {{CODEBASE_PATH_PREFIX}}/{{PRIMARY_REPO_NAME}}/ui/components/react/src/
Edits under that path, new endpoints called by fetchJson, new modals / state
Cross-repo contract
any boundary between two of your repos
Same symbol/string referenced in two repos in the same change set
Layer ambiguity → assume the wider scope. A change to a service function used by both an HTTP route and a worker is BOTH an API change AND a Worker change. Test both triggers.
Cover EVERY mode in scope — every scenario, every mode (Worker / UI / API / DB). Each mode this change touches must end up with ≥1 scenario that actually runs through its production trigger; a mode that's in scope but left untested is a gap to flag, not a silent skip. For a mode the change genuinely doesn't touch, record it explicitly as N/A — <one-line why> (e.g. "no worker claim loop reaches the diff") rather than omitting it — so the reader sees the full surface and your reasoning. A frontend-only change collapses to UI × data-states, with Worker N/A and API/DB as the read-only signals that back the UI.
Print the audit so the user sees what you're about to test:
Layers touched: DB (X), API (Y), Service (Z), Worker (W), UI (U)
Test plan:
- DB: <one line per scenario>
- API: <one line per scenario>
- ...
Step 1.5 — Trace the production trigger (MANDATORY)
For every changed symbol identified in Step 1, walk forward through the call graph until you reach a production entrypoint: an HTTP route, a worker claim loop, a cron entry, or a UI click handler. Until you can name the trigger, you do not know how to run the E2E test.
For each changed file/function, fill in:
Changed: <file>:<symbol>
↑ called by: <function/method>
↑ called by: <controller/worker entry>
↑ production trigger: <HTTP path | worker script | cron job | UI button>
↑ claim predicate: <e.g. "WHERE status='Ready' AND status_id IS NULL">
If multiple production triggers reach the changed code, list ALL of them. Each one is a separate test scenario.
If you can't construct the call chain in five minutes of grep / Read, stop and ask the user which trigger they want exercised — don't guess and end up running the wrong test.
Example — what this prevents
Changed record_classification/service.py:_engine_record_adapter (drops tax_rate before policy engine).
Without this step, the agent verifies via a unit test that the adapter now forwards tax_rate, declares pass, and ships. The user has to come back later and say "actually reprocess the record through the worker." That's a failure of this skill.
With this step:
Changed: record_classification/service.py:_engine_record_adapter
↑ called by: RecordClassificationService.classify_items
↑ called by: scripts/classify_worker.py (the WORKER)
↑ production trigger: classify_worker.py polling loop
↑ claim predicate: public.record.status='Ready' AND no classification rows exist
Now the test plan writes itself: spin up the stack, reset a record to Ready with classifications cleared, run python scripts/classify_worker.py --once, verify the resulting classification rows.
How to find the claim predicate
For worker code, grep for SELECT against the entity table in the worker script. Example:
The first significant WHERE clause is the claim predicate. Copy it verbatim into the audit. The reset SQL in Step 4 (Worker testing) is just the inverse of that predicate applied to your test entity.
Step 2 — Spin up the local stack (call /qspinuplocal)
Always call /qspinuplocal. Don't hand-roll uvicorn / serve.py / processor commands — that path repeatedly hits the same pitfalls ({{PRIMARY_REPO_NAME}} load_dotenv(override=True) clobbering env vars, stale {{ENV_SERVICE_URL_KEY}}=:9000 in older .env files, port 8000 occupied by Docker, missing DEV_TENANT_ID). The skill handles all of these; reimplementing them in this step burns time and ships subtle bugs.
Pass through what the user gave you in this session:
Worktree paths — REQUIRED in qship context. Pass BOTH --worktree-core and --worktree-app whenever the change lives in a worktree (i.e. always, under /qship). /qspinuplocal Step 4.5 mutates .env and Step 0.5 now hard-refuses to run against the maintree without an explicit --allow-maintree opt-in — you must NOT pass --allow-maintree from qe2etest. Locating the worktrees:
If pwd is under {{STATE_ROOT}}/worktrees/<TICKET>/, derive --worktree-core $TICKET_DIR/{{PRIMARY_REPO_NAME}} and --worktree-app $TICKET_DIR/{{PRIMARY_REPO_NAME}}.
Otherwise scan git -C {{CODEBASE_ROOT}}/{{PRIMARY_REPO_NAME}} worktree list and ... {{PRIMARY_REPO_NAME}} worktree list for the branch the user is working on and pass those paths.
If only ONE worktree exists (e.g. {{PRIMARY_REPO_NAME}} only) and the other repo is unchanged in this ticket, you may pass --no-core or --no-app to skip the unaffected side — but never substitute the maintree path for the missing worktree.
Worker — if Step 1 marked the Worker layer in scope (including the wider sense from the call-graph trace in Step 1.5), pass --worker. The polling worker is then live and will pick up reset entities automatically.
DB name — usually local_demo_db (the canonical default). Use whatever the user told you to use; don't default to {{LOCAL_DEV_DB_NAME}} when they named a specific DB.
Maintree-protection contract (non-negotiable): the user's day-job runs against the maintree at {{CODEBASE_ROOT}}/{{COMPANY_SLUG}}-*. This skill MUST NOT mutate the maintree .env files — they carry the user's working tenant / DB / {{ENV_SERVICE_URL_KEY}} config and a rewrite would silently break their next IDE run. Verify by reading the .env file path printed in /qspinuplocal Step 4.5 output: it must start with {{STATE_ROOT}}/worktrees/ (or whatever explicit --worktree-app path you passed), NOT {{CODEBASE_ROOT}}/. If /qspinuplocal aborted with the "refusing to use maintree" guard, fix your invocation — do NOT bypass the guard with --allow-maintree.
Port detection + fallback (read this BEFORE any curl / Playwright call)
/qspinuplocal Step 1 picks ports deterministically: it tries the canonical port first (e.g. 8000), kills its OWN stale process if it finds one, and walks a fallback ladder (18000/18001/…) when a port is held by something else (the user's maintree IDE run, Docker, another tenant's stack, a sibling qship run). It writes the chosen ports to /tmp/{{COMPANY_SLUG}}-ports.env and prints them in the Step 9 report. The fallback is what lets a qship / qe2etest run coexist with the user's day-job stack on the canonical ports.
Source the port file at the top of this skill's session AND re-source it after every restart, because the fallback may have shifted between runs:
set -a; . /tmp/{{COMPANY_SLUG}}-ports.env; set +a
# now ${{ENV_SERVICE_URL_KEY}} / ${{ENV_SERVICE_URL_KEY}} / $SERVICE_HEALTH / $SERVICE_HEALTH# point at whatever the running stack actually bound to.
Hard rule for this skill: NEVER hardcode :8000 / :8001 / http://localhost:8000 / http://localhost:8001 in any curl command, Playwright base URL, evaluate_script call, or DOM-assertion fragment. Use the env vars. If a hardcoded literal lands in phase3-evidence.md, the qship enforcement hook treats the evidence as suspect.
After spin-up, verify with curl "${SERVICE_HEALTH}". If it fails, dump /tmp/{{COMPANY_SLUG}}-service.log to the user and STOP. Don't proceed with a half-up service. If the chosen port differs from the default, mention that in the test-run preamble of phase3-evidence.md.
If /qspinuplocal is genuinely unavailable in this environment, document the gap to the user and ask to install it — DON'T hand-roll the commands. The hand-rolled path is a known time sink.
Step 3 — Build the scenario matrix (via brainstorming, two-pass)
Don't hand-write the matrix from memory. Use the superpowers:brainstorming skill to design a comprehensive + robust matrix in two passes — one grounded in the ACs from Step 0.5, one independent from the implementation. Union the two.
3a — AC-grounded pass
Invoke brainstorming with the AC source-of-truth as the input. For each AC, generate the 5 canonical scenario classes:
Class
What it asks
happy_path
the AC's stated outcome works for the obvious input
negative
the AC's stated outcome FAILS correctly for invalid input
atypical-but-valid input the AC implicitly accepts (Unicode, leading-whitespace, concurrent caller, partial-state)
auth
the AC's behaviour at the auth boundary (no token, role-gated, or — if the app is multi-tenant — wrong tenant)
Output: matrix-1 with 5 × N rows where N = number of ACs.
3b — Implementation-independent pass
Invoke brainstorming a SECOND time, this time with the implementation diff as the input and the AC list HIDDEN. Ask: "what could break in this code path that the spec didn't think to require?" This catches the gaps between what was asked and what was built.
Specifically have the brainstormer enumerate:
Idempotency — same write twice. Same job claim twice. Same row insert with conflict.
Concurrency / dedup — two simultaneous triggers, two clicks on Run, two webhooks for the same event.
Async propagation — write triggers worker, worker actually runs and downstream state catches up. Heartbeats, retries, dead-letter queues.
Cross-wave / cross-repo interactions — if this change is part of a multi-wave epic, scenarios that exercise THIS change × adjacent waves' changes together (a write under the new RLS policy reading via the new schema field, etc.).
Migration rollback / replay — if alembic touched, the down() path runs cleanly AND data is recoverable.
Failure mode adversarial — for every guarantee the implementation claims, design a scenario that would expose its violation (deadlock, partial commit, racing writers, lost update).
Observability — log lines / metrics / traces emit at expected verbosity. Errors are diagnostic (not "an error occurred").
Output: matrix-2.
3c — Union + score
Merge the two matrices, dedupe overlapping scenarios, score each on:
AC coverage: which AC does this scenario prove (or "implementation-only" if pass-2 only)?
Production-trigger fidelity: is the scenario fired through the production entrypoint (Step 1.5), or is it an integration shortcut? (Latter is OK but flag it.)
DB requirement (per the table below).
Final matrix: 5 columns minimum — ID | Type (AC-grounded / impl-only / cross-wave / adversarial) | AC ref | Method (pytest / curl / psql / playwright / worker) | Concrete invocation | DB chosen.
Target size: 25-50 scenarios for a single-ticket change, 50-100 for an epic-wave-merged change. Smaller is fine if the change surface is tiny — but justify low count in the report rather than skipping.
Print the matrix before you run it. Save it as /tmp/qe2etest-matrix-<branch-or-ticket>.md so the report can reference it.
Per-scenario database selection (mandatory)
Different local Postgres DBs carry different data shapes. Pick the most-data-rich DB for each scenario, not a single default. The mapping below is the canonical default for the {{COMPANY_SLUG}} monorepo — overrideable per scenario if its needs cite a specific DB.
Scenario category
Default DB
Why
Validation policy templates / instances, catalogs, taxonomies — the canonical default for most scenarios
local_demo_db
richest item/policy/tag data locally — default for validation-policy scenarios
Procedure per scenario:
(a) Identify which entity tables / behaviours the scenario queries.
(b) Pick the DB from the table above; if a scenario explicitly cites a DB in the matrix row, that wins.
(c) Group scenarios by DB to amortize server-restart cost. uvicorn holds a connection pool to one DB — switching DBs requires killing and respawning the server. Aim for ≤3 server restarts across the entire matrix. Order the matrix so all local_demo_db rows run together (this is the default and will usually be the largest group), then all local_alt_db, then any local_acme_corp_db or {{LOCAL_DEV_DB_NAME}} rows.
(d) Capture the chosen DB in the scenario's evidence row (column "DB used") so reviewers can reproduce.
(e) Seed the data state if it doesn't exist — don't skip the scenario. If a scenario needs a fixture/state that's absent from the chosen DB, CREATE it: insert/flip the minimal source rows, run the check, then restore baseline (the restore-.sh discipline in Step 4 — local-clone writes are pre-authorised). Capture the seed + restore in the evidence row. Two rules that make this work:
Seed the source, not a phantom column. Many matches/derived states are computed server-side, not stored — a match endpoint may derive its rows from a mapping/lookup table via a content hash or a join, with no stored foreign-key column on the entity itself. Insert/adjust the rows the endpoint actually reads, then re-probe the live endpoint to confirm the state took before asserting.
Missing tenant/DB → clone it, don't skip. If a needed shape lives in a tenant you don't have locally, clone it with /qlocalclonedb rather than dropping the scenario.
Mark a scenario SKIP only as a last resort when seeding is genuinely infeasible (cross-system state you can't construct locally) — and say exactly why; drop a note to /tmp/qe2etest-db-fixture-gap-<ts>.md. Never silently skip a scenario for want of data you could have seeded.
If unsure, default to local_demo_db. The clean re-clone command is:
The matrix from Step 3 is sorted by DB. Between groups, you MUST kill + respawn the stack so uvicorn binds to the new DB. The {{PRIMARY_REPO_NAME}} load_dotenv(override=True) caveat means setting env vars at command line is insufficient — /qspinuplocal rewrites the .env files. Use:
# Switch from current DB to <new_db>:
pkill -f "serve.py" 2>/dev/null
pkill -f "uvicorn.*{{COMPANY_SLUG}}\." 2>/dev/null
sleep 2
/qspinuplocal --db-name <new_db> [--tenant-uuid <uuid>] [--worker] [--worktree-app <path>]
Verify the swap actually happened before running scenarios:
Capture the DB swap as an evidence row in the matrix: S<id> | (db-switch) | local_alt_db → local_demo_db | server-restart succeeded.
API testing
Use curl against the local stack. Always:
Capture the HTTP status, response body, and (if the call writes) follow-up psql queries to confirm DB state.
Diff DB state before vs after where it matters (count, sum, specific row).
If auth is required, walk through the real auth flow (the external auth provider dev login) — never substitute the bypass.
# Example shape (use sourced ${{ENV_SERVICE_URL_KEY}} — never hardcode :8001)
curl -s -X POST "${{{ENV_SERVICE_URL_KEY}}}/api/v1/.../endpoint" \
-H "Content-Type: application/json" \
-H "Authorization: Bearer $TOKEN" \
-d '{...}' | jq
psql 'postgresql://{{LOCAL_DB_USER}}@localhost:5432/local_acme_corp_db' -c \
"SELECT count(*) FROM public.<table> WHERE <predicate>;"
Worker testing — REQUIRED if Step 1.5 traced any change to a worker entrypoint
The trigger is NOT "did this change introduce a new job category" — the trigger is "does the production call graph for this change include scripts/*-worker.py or any worker claim loop." If yes, the worker MUST run. No exceptions.
The pattern is always the same four steps. Do them in order; don't skip the reset.
1. Pick a representative entity
Use the example the user mentioned in the conversation (e.g. record EXAMPLE-001) when available — the user picked it because it surfaces the bug or feature. If they didn't name one, pick one that satisfies the claim predicate AFTER your reset and exhibits the specific data condition the change targets (e.g. for the tax_rate fix: a record with a 0%-tax line and a category-specific policy applicable to its vendor).
2. Reset the entity into the worker's claim shape
This is the step that's most often skipped, and skipping it makes the worker silently no-op (the entity is already past the claim predicate). Use the claim predicate you captured in Step 1.5 as a recipe for the reset:
-- Example for processor (claim: status='Ready' AND no classifications):DELETEFROM public.record_classifications
WHERE record_id = :doc_id;
UPDATE public.record
SET status='Ready'WHERE record_id = :doc_id;
If the entity has dependent rows the worker writes to (step_log, audit_trace, run_log), wipe those too — otherwise the post-run verification confuses "old run output" with "this run's output."
Snapshot the BEFORE state to the conversation before running the worker. The user wants to see the broken state explicitly, not just the fixed state. One SELECT of the relevant rows is enough.
Capture the tail of stdout. If the worker reports Processed: 0 | Failed: 0, the claim predicate didn't match — go back to step 2 and fix the reset.
4. Verify in the DB (the AFTER state)
Re-run the same SELECT from step 2's snapshot. The diff between BEFORE and AFTER is the proof. Be explicit in the report:
Pre-worker: line 5 had only `fixed` department; category + tax_code missing
Post-worker: line 5 has policy_auto for department, category, tax_code (category rule fired)
If the change affects worker-emitted log tables, also check those — step_log.processed_line_count is much more diagnostic than counting classification rows.
Cover scenarios beyond the golden path when relevant: claim → success, dedup when multiple jobs for the same key are queued, retry-on-transient-error, file-less vs file-backed shapes. But the golden-path "reset → run → verify" is non-negotiable.
UI testing — delegate to /qmanualt
For any UI change, hand control to the /qmanualt skill. It owns the Playwright + Claude Chrome extension flow. Brief it with:
What the change is (recapping from the audit).
Which page(s) the change is visible on.
The exact scenarios to walk (golden path + edge cases derived from Step 3).
The DB-side assertions it should run after each click via psql / a worker tick.
/qmanualt will:
Open the page in Chrome via the extension.
Drive the click sequence.
Capture screenshots of each meaningful state.
Read DOM + console for assertions.
Surface anything that doesn't render / behave correctly.
Do not try to drive the UI with raw mcp__Claude_in_Chrome__* tool calls in this command — that's what /qmanualt is for. Re-driving it here would duplicate logic that already lives in the skill.
UI Testing Anti-Mock Contract — UUID resolution
When the diff renders a UUID-typed field via any resolve*Name / resolve*Value hook, the Playwright run MUST hit the LIVE lookup endpoint, not a mocked hook.
This rule exists because the {{COMPANY_SLUG}} React UI handles lookup-pending rendering via a specific canonical pattern — and mocked-hook tests cannot prove that pattern is applied correctly because mocks return synchronously (no loading window). Mocks guarantee the lookup never fails in test → silently mask any production environment where it would.
Canonical pattern (the only acceptable shape — DO NOT reinvent):
The {{COMPANY_SLUG}} React UI gates UUID→name resolution on resolver.ready and renders a Mantine <Skeleton> while the lookup fetch is in flight. Two canonical reference implementations to grep against:
The resolve* hook family in src/components/shared/{entity,node,organization,attribute,entityLookup}.js exposes a .ready attribute on the returned function (see entityLookup.js:121 — resolve.ready = state.ready). Any new UI surface rendering a UUID-typed field MUST follow this <Skeleton> + .ready pattern. Truncated-UUID placeholders (Entity 00000000…) are NOT the codebase convention — flag them as a finding, do not introduce them.
When .ready === true && resolver(id) === null (lookup loaded but the specific id genuinely isn't cached), the convention is to render the raw UUID. That's intentional and consistent across the file — but only acceptable after the Skeleton gate has been in place for the loading window.
Detection — flag a UI change as "lookup-dependent" if its diff includes:
Any resolveEntityName, resolveRefName, resolveOwnerName, resolveAttributeValue, resolveAttributeName, resolveGroupName, or any other resolve*Name / resolve*Value call site.
Direct rendering of a column whose value matches the UUID regex \b[0-9a-f]{8}-[0-9a-f]{4}-[0-9a-f]{4}-[0-9a-f]{4}-[0-9a-f]{12}\b.
renderResolvedValue, renderRuleValue, or similar wrapper.
Quick detection command for the worker to run against its diff:
If any line matches, this Anti-Mock Contract applies.
For every lookup-dependent UI surface added in the diff, the Phase 3 test MUST:
Spin up your service via /qspinuplocal. If your change depends on another service (the UI fetches an endpoint a sibling repo owns), start that one too — otherwise React fetches 404 → resolver returns null → UI either flashes UUIDs (no Skeleton gate) or renders Skeletons forever. After /qspinuplocal returns, set -a; . /tmp/{{COMPANY_SLUG}}-ports.env; set +a and verify with curl -fs "${SERVICE_HEALTH}" BEFORE running scenarios. The hook gate cares that the service(s) your change touches are reachable — not that they're on a canonical port.
Drive Playwright (or Claude in Chrome) against either:
(a) The LIVE Dash UI at ${{{ENV_SERVICE_URL_KEY}}}/... — preferred. The Dash app already mounts the production React bundle with real hooks pointing at {{PRIMARY_REPO_NAME}} (${{{ENV_SERVICE_URL_KEY}}}); nothing to reconfigure. Source /tmp/{{COMPANY_SLUG}}-ports.env before invoking Playwright so baseURL reads ${{{ENV_SERVICE_URL_KEY}}}.
(b) The worktree e2e-harness with the lookup-hook MOCK ALIASES REMOVED — i.e. webpack does NOT alias shared/entityLookup etc. to mocks/*Lookup.js files; the real hooks ship in the bundle and fetch at runtime from {{PRIMARY_REPO_NAME}}. For path (b), launch Chromium with --disable-web-security so the cross-origin fetch from the static harness server to {{PRIMARY_REPO_NAME}} doesn't get blocked by CORS preflight. Document the alias-removal + CORS-bypass in the evidence file.
A mocked-hook harness run is NEVER an acceptable substitute for either path.
Read the rendered DOM with evaluate_script and assert TWO things:
const uuidRegex = /\b[0-9a-f]{8}-[0-9a-f]{4}-[0-9a-f]{4}-[0-9a-f]{4}-[0-9a-f]{12}\b/i;
const root = document.querySelector('[data-rule-group-children-row]');
const visibleText = root.innerText;
// (i) NONE of the visible text matches the UUID regexconst leaks = visibleText.match(newRegExp(uuidRegex.source, 'gi')) || [];
expect(leaks).toEqual([]);
// (ii) The expected resolved name DOES appear in the visible textexpect(visibleText).toContain('Acme Corp Ltd');
Both assertions are required. (i) alone catches leaks but misses the case where the resolver hangs forever and the UI shows only Skeletons (no UUID visible, but no name either). (ii) alone catches missing names but misses leaks. Together they prove resolution actually completed.
For every UUID-bearing field, the test must capture (a) the raw UUID from the API/DB, (b) the expected resolved name from the lookup endpoint, (c) the rendered DOM text. The evidence file must record all three.
Verify the codebase pattern is applied — the new UI surface MUST gate its UUID-typed render on resolver.ready and show a Mantine <Skeleton> while not ready. Reference the canonical examples above. Truncated-UUID placeholders (Entity 00000000…) are NOT the convention — flag them as a finding. This is a code-shape check in addition to the runtime DOM assertion: grep the new surface for Skeleton or \.ready\b in proximity to the resolve* call.
If {{PRIMARY_REPO_NAME}} cannot be brought up, the test result is BLOCKED — not PASS. Write the BLOCKED rationale to the evidence file. A mock-backed-only run does NOT close the gap.
Bug-fix mandate: if a UUID leak is observed (rendered DOM contains a raw UUID where .ready === false should have shown a Skeleton):
DO NOT introduce a truncated-UUID fallback (Entity 00000000… / formatTruncatedUuid(...)). That's not the codebase convention.
DO apply the <Skeleton> + .ready gate per ResolvedRefCell.jsx and ResolvedSummaryView.jsx. Cite those files in the fix commit so future readers see the canonical reference.
If resolver.ready === true but an entity genuinely isn't in the cache (resolver returns null by design, e.g. soft-deleted entity), accept the raw UUID render as the codebase convention — but only after the Skeleton gate is in place for the loading window.
Matrix tagging: for every UI scenario, tag the "DB" column with one of:
Live UI (demo) — drove the real Dash UI against the named DB, with real lookup hooks → real {{PRIMARY_REPO_NAME}}.
Harness (real hooks + CORS-bypass) — drove the worktree e2e-harness with mock aliases REMOVED, Chromium --disable-web-security set, {{PRIMARY_REPO_NAME}} live.
Harness (mocked hooks) — mocked-hook harness only. This tag is ONLY acceptable for UI scenarios that DO NOT render any UUID-typed field. If the scenario touches lookup-dependent rendering, this tag means BLOCKED, not PASS.
Only the first two prove resolution; the third cannot.
Cross-repo contract checks
If the change crossed repos (e.g. {{PRIMARY_REPO_NAME}} publishes a string that {{PRIMARY_REPO_NAME}} consumes), grep the consumer repo for the old literal in addition to running the e2e — broken contracts are easier to catch with a static check than at runtime.
Step 5 — Verify in the database
For every write the test triggered, end with a psql SELECT that proves the expected row exists / is missing / has the right state. Don't assume the API response is the source of truth — it isn't. The DB is.
Specifically check, when applicable:
public.record_mappings: the is_active=true rows are what the matcher will use.
task_queue: status, result_payload, attempts, last_failure — and job_metadata for the input contract.
public.record_items.record_hash_id: linkage from line → hash.
Any audit / log table the change writes to.
Step 6 — Report
Give the user:
Scenario matrix with results. Table: layer | scenario | result | evidence (file:line, db query, screenshot path). Pass/fail per row.
What broke and why — for any failure, a short root-cause hypothesis with the relevant log/grep.
Coverage gaps — anything in the audit you didn't get to and why.
Whether the change is safe to ship — explicit verdict: ✅ ready / ⚠ ready with caveats / ❌ blocked. Don't soften.
If everything passes, say so plainly. Don't pad with hedging. If something failed, lead with that, not with what passed.
Self-check before declaring pass
Before writing the report, run this check against your own work in the conversation:
Step 0.5: did you fetch every relevant Jira ticket's ACs (and any linked TRD §AC blocks) and save them as the brainstorm input?
Step 3a: did you invoke superpowers:brainstorming with the ACs visible to generate matrix-1 (5 × N rows, one per AC × scenario class)?
Step 3b: did you invoke superpowers:brainstorming a SECOND time with the implementation diff and ACs HIDDEN to generate matrix-2 (implementation-only scenarios — idempotency, concurrency, cross-wave, adversarial, observability)?
Step 3c: did you UNION the two matrices, dedupe, and score each row with Type / AC ref / Method / DB chosen?
Step 3 DB routing: is every matrix row tagged with a specific DB from the routing table (not blanket local_acme_corp_db)? Are the rows sorted by DB to minimize server restarts?
Step 4 DB swaps: between DB groups, did you pkill + re-/qspinuplocal and verify the swap with current_database() before running the next group's scenarios?
For every changed file in Step 1, can you point to a tool call where the production trigger for that file fired? (Not a unit test, not a REPL call — the actual HTTP request / worker run / cron invocation.)
If the Worker layer was in scope, did you reset the entity, run the worker --once, and SELECT the resulting rows from the DB?
Per-mode coverage: was every mode in scope for this change — Worker / UI / API / DB — actually exercised by ≥1 scenario that RAN (each with a concrete artifact), or explicitly recorded as N/A — <why>? A needed mode left untested is a gap to flag, not a silent skip.
Seed-don't-skip: for every scenario that needed a data state which didn't exist, did you SEED it (source rows, re-probed live) and RESTORE baseline — rather than skipping it? Any remaining SKIP must name why seeding was infeasible.
Did the stack run the fix worktree's code, not develop? (Quick check: ps aux | grep uvicorn should show the worktree path. Or: edit a log line in the changed file, restart, run the trigger, confirm the log line appears.)
Did you run against the right DB? (psql -l if uncertain.)
Did at least one scenario per Type bucket (happy_path / negative / boundary / edge / auth) actually run and report PASS or FAIL with a concrete artifact (curl response, pytest function, psql output, Playwright screenshot)?
UUID anti-mock: did every UI scenario that renders a UUID-typed field run against a LIVE lookup endpoint — either the LIVE Dash UI at ${{{ENV_SERVICE_URL_KEY}}} or the harness with real hooks + CORS-bypass — NOT mocked hooks? (See §"UI Testing Anti-Mock Contract".)
UUID anti-mock: did you evaluate_script against the rendered DOM and confirm BOTH (i) zero UUID-regex matches in user-visible text AND (ii) the expected resolved name IS present? (Both required — (i) alone misses Skeletons-forever; (ii) alone misses leaks.)
UUID anti-mock: is the lookup service actually reachable (curl -fs "${SERVICE_HEALTH}" returns 200) BEFORE every Playwright / Chrome assertion that depends on a live lookup? Source /tmp/{{COMPANY_SLUG}}-ports.env first; don't assume a canonical port is bound.
UUID anti-mock: if the diff added any new resolve*Name / resolve*Value call site, does the surface gate the render on .ready and render a Mantine <Skeleton> while not ready, per the ResolvedRefCell.jsx / ResolvedSummaryView.jsx convention? (Truncated-UUID placeholders like Entity 00000000… are NOT the convention — flag them.)
If any of those is "no", you have not done an E2E. Go back and finish, or mark BLOCKED — don't write "✅ ready" anyway.
When to push back instead of running
If the change is too small to need e2e (typo fix, isolated unit-tested helper), say so and skip.
If the runtime can't be brought up after a real attempt, surface the bring-up error and ask the user to fix it before proceeding.
If the user asked for e2e but the change is incomplete (e.g. backend ships, UI wiring is still TODO), test what's there and call out what's still untested.
