Skip to content

Evenfire is an intelligence layer platform for running LLM agents that take real actions, on your infrastructure

TypeScript
17.0 KB
Raw Blame

Manually verify password recovery with the development producer

Use this guide in a separate, authorized manual session after fetching PR #1056. Audience: a developer or platform owner with access to the intended development Member Registration producer, its test data and an approved recipient mailbox. This guide describes the configured integration lane and a separate email/UI journey. Neither lane was executed while implementing the structural redesign. Both remain NOT VERIFIED until their own receipts are recorded.

A. Obtain and build the published revision

From the intended main Evenfire checkout, inspect before changing anything:

pwd
git remote get-url origin
git branch --show-current
git status --short
git rev-parse HEAD
git fetch origin fix/issue-1055-password-admission-dos

Confirm the remote is evenfire-ai/evenfire. Preserve unrelated local work; if switching would overwrite it, stop and use another clean checkout or agree on its disposition. Do not hard reset, force checkout, or automatically stash it. In a clean checkout already on the PR branch, use:

git merge --ff-only origin/fix/issue-1055-password-admission-dos
git rev-parse HEAD

If the branch is not checked out and no conflicting local work exists, switch to the existing local branch with git switch fix/issue-1055-password-admission-dos (or create its tracking branch with git switch --track origin/fix/issue-1055-password-admission-dos). Then fast-forward as above. Divergence needs a decision; this guide does not authorize rebasing or rewriting history. Compare the actual SHA with GitHub's current PR head and record it in the receipt.

Use Node 24 and the checked-in lockfiles:

node --version
npm --prefix packages/workflow-runtime-core ci
npm --prefix packages/workflow-runtime-core run build
npm --prefix control-api ci
npm --prefix external-rest-api ci
npm --prefix control-api run build
npm --prefix external-rest-api run build
node scripts/ci/password-recovery-suite-contract.mjs verify

The composed local private suite needs the built External REST app. Each service owns its dependencies; another CI job's artifacts do not satisfy this prerequisite. For the optional later UI journey, install/build Profile UI according to its live local guidance. Do not treat a local UI attached to older backend images as verification of this revision; record the backend revisions separately.

B. Establish the real development producer

Obtain the actual development cluster, zone, project, account, Kubernetes user and context from the owner's approved private infrastructure record. Keep concrete infrastructure identifiers out of this public repository and sanitized receipts. The existing setup sequence is shown with verified-input variables below. It is reference context for a later authorized session, not proof that Member Registration is deployed in that cluster:

gcloud container clusters get-credentials "${VERIFIED_DEV_CLUSTER:?Confirm the development cluster}" \
  --zone="${VERIFIED_DEV_ZONE:?Confirm the development zone}" --project="${VERIFIED_DEV_PROJECT:?Confirm the development project}" \
  --account="${AUTHORIZED_DEV_ACCOUNT:?Confirm the authorized account}"
kubectl config set-credentials "${VERIFIED_DEV_KUBE_USER:?Confirm the Kubernetes user}" \
  --exec-arg="--account=${AUTHORIZED_DEV_ACCOUNT:?Confirm the authorized account}"
LOCAL_KUBE_CONTEXT="${VERIFIED_DEV_CONTEXT:?Confirm the development context}" npm run ui

Confirm that the Kubernetes user entry belongs to the intended context before changing its account binding. Required expansions fail before a command runs when a verified input is missing; no command falls back to the current context. Populate these variables locally from the approved record; do not invent values, commit them or paste them into a PR receipt.

Earlier read-only discovery found no registration namespace in the owner-selected development context. The Control API configuration referenced the external registration hub at registration.evenfire.ai; its deployment ownership, version, database and isolation were not established. Reconfirm these facts. Do not select the first similarly named service or infer the producer from the UI launch command.

After access is authorized, use explicit-context read-only discovery:

kubectl --context="${VERIFIED_DEV_CONTEXT:?Confirm the development context}" get namespaces
kubectl --context="${VERIFIED_DEV_CONTEXT:?Confirm the development context}" get services --all-namespaces
kubectl --context="${VERIFIED_DEV_CONTEXT:?Confirm the development context}" get deployments --all-namespaces \
  -o custom-columns=NAMESPACE:.metadata.namespace,NAME:.metadata.name,IMAGES:.spec.template.spec.containers[*].image

Confirm the actual namespace/service/deployment with its owner. Read only the specific non-secret endpoint field from the confirmed configuration object; do not dump Secrets, environment files or whole configuration objects. For a confirmed Deployment, inspect its image and the running Pod image ID with context-qualified commands and compare the immutable image provenance with the producer's source revision. Record the actual deployed revision. Ask the owner to identify the producer database, tenant/key binding, recipient policy and whether the target is isolated or approved for this synthetic write. An image tag, HTTP health response or UI screenshot alone does not establish version, database identity or permission to send email.

If the producer is outside this cluster, discover it through its actual hosting owner. The GKE UI workflow does not create a supported port-forward to it. If a confirmed in-cluster producer needs a forward, obtain its exact Service port and namespace first. Use a task-owned loopback forward in a dedicated terminal, without inventing names or ports:

kubectl --context="${VERIFIED_DEV_CONTEXT:?Confirm the development context}" -n "$VERIFIED_NAMESPACE" port-forward \
  --address=127.0.0.1 "service/$VERIFIED_SERVICE" ":$VERIFIED_SERVICE_PORT"

Use the allocated local port shown by that process only for this session, and stop the forward afterward. Confirm any network/proxy requirements with the owner; do not port-forward or mutate a shared environment without authorization.

C. Provide configuration securely

Use an approved credential store or secure interactive entry into process memory. In a local Bash shell, disable tracing and prompt without echo:

set +x
read -r -s -p 'Isolated Control API PostgreSQL admin DSN: ' CONTROL_API_REAL_PG_ADMIN_URL; printf '\n'
read -r -s -p 'Authorized producer PostgreSQL DSN: ' CONTROL_API_REAL_MEMBER_REGISTRATION_PG_URL; printf '\n'
read -r -s -p 'Verified producer API base URL: ' CONTROL_API_MEMBER_REGISTRATION_SERVICE_BASE_URL; printf '\n'
read -r -s -p 'Authorized producer HMAC secret (hidden)> ' CONTROL_API_MEMBER_REGISTRATION_HMAC_SECRET; printf '\n'
export CONTROL_API_REAL_PG_ADMIN_URL CONTROL_API_REAL_MEMBER_REGISTRATION_PG_URL
export CONTROL_API_MEMBER_REGISTRATION_SERVICE_BASE_URL CONTROL_API_MEMBER_REGISTRATION_HMAC_SECRET

The first DSN must target an authorized disposable PostgreSQL 16 instance with CREATE/DROP DATABASE permission. The second must identify the producer's explicitly authorized test database. They have different effects. Do not point the admin DSN at production or an unrelated shared database.

The runner requires these four variables. Successful authentication also requires the matching producer tenant and HMAC key ID; confirm whether CONTROL_API_MEMBER_REGISTRATION_TENANT_ID and CONTROL_API_MEMBER_REGISTRATION_HMAC_KID need explicit values. Verify that CONTROL_API_MEMBER_REGISTRATION_MODE=remote matches the intended producer configuration. Do not accept test defaults as evidence of the deployed binding. Confirm the configured redirect/app metadata against the producer's allowed URLs; the standalone configured suite is not a deployment compatibility test.

Never echo these variables, persist them in a receipt, paste literal values into shell commands, or enable set -x. Do not print .env contents or URLs containing credentials. Test runners may report operational errors: inspect locally, sanitize before sharing and do not publish raw producer/database logs. Unset all supplied secrets after the session, including optional credential inputs, and stop any task-owned port-forward or test process.

D. Authorize the actual data and email effects

Before running, record owner approval for these concrete effects:

  • A random bug192_flow_* Control API database is created, migrated and filled with synthetic users, invitations and admission state; teardown drops it. DELETE statements in setup operate inside that temporary database.
  • The producer receives real authenticated registration/validation requests and stores a synthetic invitation_flow_registrations row for a random issue1055-* address under example.invalid.
  • Producer ensureApp/ensureLatestAppConfig may create apps/app_configs rows. Existing teardown deletes flow rows by the generated test email but does not remove these possible app/config records.
  • Registration also inserts invite_send_log and invite_redirect_tokens before sending email. Both records can persist when SMTP fails; configured teardown does not delete them. For an external tenant binding, send-log entries count toward the destination domain's daily quota. Confirm the actual deployed revision, tenant binding and applicable quota before running.
  • A redirect record's destination_url contains signed invitation proof. The inspected producer assigns a 30-day redirect expiry; expiry does not physically delete the record. Confirm retention and narrowly scoped cleanup with the owner. Never include the redirect token or destination URL in a receipt.
  • Registration calls the real invitation email path. Checked-in producer deployment configuration uses Postmark SMTP. A non-delivering configuration has not been established; outbound email may be attempted, and an invalid recipient can cause rejection/bounce rather than delivery.

The configured suite's generated address is not a configurable permitted mailbox. Confirm that this exact synthetic recipient/write scope is approved and compatible with the development email service before execution. Use a separate permitted mailbox for the delivery/UI journey below. Do not silently change the runner to skip email, change recipients or omit cleanup.

Clearly identifiable synthetic development records may remain when explicitly accepted by the owner. Obtain separate approval for retained app/config records, send-log entries, proof-bearing redirect records and any applicable quota effects; willingness to retain app/config rows alone does not authorize the other effects. Record sanitized counts or non-secret identifiers and their retention or cleanup disposition. Do not assume that deleting a flow row cascades to redirect/send-log records, or that expiration proves deletion. This acceptance does not disable the runner's existing flow-row DELETE or temporary-database DROP. If teardown fails, confirm the exact generated resources before any manual deletion; never broadly delete producer data.

E. Run the configured integration lane

The final checked-in command and physical suite remain:

bash scripts/ci/run-password-recovery-private-postgres.sh configured

It selects exactly control-api/test/passwordRecovery.memberRegistration.realPostgres.integration.test.ts through the strict suite manifest. Expect one physical suite and one passing scenario, containing admission-denial, real stored signed proof, preview/row binding, wrong-email/row negatives, reset session and later-login assertions. Require process exit zero, nonzero passing collection and zero skipped tests. The runner rejects missing files/environment and skipped/zero-test results. The configured suite does not start or replace the producer, and does not prove public External REST cookies or actual mailbox delivery.

If collection skips, configuration is missing, authentication fails or the producer/database cannot be identified, record NOT VERIFIED. For a test failure, record its phase and sanitized error category: configuration/tenant, producer registration, persisted proof, validation, Control API mutation/session, or teardown. Do not substitute mocked receipts. Record the exact source SHA, producer revision, permitted data effects and cleanup result.

local is a separate real-producer lane requiring the Member Registration source checkout, installed tsx, two disposable databases and the built HTTP child; see private runner reference. It is not configured-producer or email-delivery evidence. Do not run all merely to obtain a green status without both lanes' explicit authorization.

F. Verify actual email delivery and the public UI journey

Use the confirmed development invitation email service and a dedicated synthetic account with an owner-approved mailbox. Do not add a test mail system for this journey. Establish the intended Profile UI and External REST origins and backend revisions, and ensure no unrelated account uses the synthetic address.

  1. Request recovery through the intended Profile UI, which calls public POST /api/v1/auth/password-reset/request. Check its non-enumerating success without assuming that response proves delivery. Observe the actual message arriving in the permitted mailbox; record a redacted timestamp/message receipt.
  2. Inspect the destination origin/path privately before opening the link. The producer proof must bind to the intended invitation token/email/row; the row ID and invitation token are distinct. Do not publish the URL or proof.
  3. Use a valid fresh link and confirm the reset view shows the intended account. Separately exercise invalid proof and mismatched email/row input using only synthetic records: they must not update credentials, create a cookie or expose another account. Avoid modifying or probing an unrelated invitation.
  4. Submit an approved valid replacement password once. Confirm the correct account identity and authenticated session. Inspect cookie attributes privately: profile_session, HttpOnly, expected Path/SameSite and Secure for the intended HTTPS environment. Do not record cookie or token values.
  5. Confirm session hydration and navigation to the authenticated view, including a refresh. The UI must not automatically replay the reset after a hydration failure. Record the displayed account identity in sanitized form.
  6. Reuse the consumed proof once deliberately: completion must be rejected, credentials must stay unchanged and no new authenticated session may be published from that replay. Confirm sibling reset invalidation using only owner-approved synthetic links where applicable.
  7. Where normal admission permits, log in with the new password and verify the same account. Inspect only the authorized synthetic account/invitation state to confirm unrelated records were unaffected.

Do not deliberately saturate shared dev login capacity, consume another user's budget or inject wire/database/email failures into shared services. Busy-owner, unknown COMMIT, FATAL/lost acknowledgement, storage outage and connection cleanup faults belong in a separately authorized isolated environment. A normal admission 429 must remain bounded and must not trigger automatic reset replay.

Configured SQL/producer integration and this delivery/UI journey are separate evidence lanes. Passing either does not certify the other, and source/local synthetic-browser results do not certify deployed versions.

G. Record and share a sanitized receipt

Copy this checklist into the authorized review record:

Date / tester:
PR SHA fetched and tested:
Control API / External REST / Profile revisions actually running:
Member Registration immutable image / source revision:
Verified dev context and confirmed producer owner/service:
Authorized test database identity (no DSN):
Authorization for synthetic SQL, producer effects and email:
Configured command / exit code:
Physical suites / passed / failed / skipped counts:
Manual email receipt and timestamp (redacted):
Link origin and trusted account/token binding (no URL/proof):
Reset / invalid / mismatch / consumed-proof results:
Cookie attributes / session identity / hydration / navigation:
New-password login under ordinary admission:
Unrelated synthetic account/invitation state:
Flow-row cleanup / temporary database drop:
Remaining synthetic app/config records and owner disposition:
Send-log records / applicable quota effects / owner disposition:
Redirect records / expiry / retention or exact cleanup (no token or destination URL):
Stopped task-owned processes / forwards / unset credentials:
Unverified or failed paths:
Reviewer / owner merge-sign-off decision:

A later authorized PR update should state the exact tested SHA and producer revision, separate the two evidence lanes, include counts and sanitized outcomes, and retain any unresolved limitation. Never attach credentials, full logs, mailbox content, link tokens, database URLs or session cookies. Missing or skipped integration remains NOT VERIFIED; it is not a merge certification.