Integration guide

Integrate Tracer with your authorization system: send complete payloads, handle ALLOW, DENY, and REVIEW decisions, and stay within an 80ms latency budget.

Integrating Tracer means deciding where in your authorization flow to make the validation call, what data to send, and how to handle the three possible decisions. The pattern is short: your system collects the full transaction context, calls POST /v1/validations, acts on ALLOW / DENY / REVIEW, and moves on. Tracer never reaches back into your stack. There are no webhooks or callbacks, and the integration ends with the response.

What changes in your operation: decisioning moves from in-process logic to an external call. The call is synchronous (request/response, no webhooks), so it sits on the critical path of the transaction. Done well, it adds under 80ms p99 and gives you a single point for policy and validation history. Done poorly (no timeout, no retry strategy, no fallback), it becomes a single point of failure.

Trade-off to be honest about: you’re adding a network hop. The good news is the contract is simple: idempotent by requestId, no callbacks, deterministic three-state response. The bad news is you must think about timeouts, retries, and what to do if Tracer is unreachable. Most of this guide is about that.

Who is this guide for? Integration engineers writing the request from your system to Tracer, and architects deciding where the call sits in the flow. Risk and fraud analysts who write rules can skip ahead to the Rules engine guide. Compliance can read the Audit and compliance guide instead.

This guide covers payload requirements, the integration flow, and practices that keep the validation call inside your latency budget.

Tracer sits outside your ledger: it never calls Midaz. Your application orchestrates the two. It calls Tracer to validate, and submits the transaction to Midaz only if the decision is ALLOW. Tracer evaluates your configured policies and limits against the context you send, not account balances. The ledger stays the source of truth for what an account holds.

Midaz can also drive Tracer through an optional per-ledger reservation seam. It remains one-way, Ledger → Tracer. The rest of this guide covers the application-orchestrated HTTP pattern. The seam contract appears below.

Integration overview

---

Tracer expects calls from authorization systems (payment gateways, workflow orchestrators, or transaction processors) that need real-time validation decisions. The integration follows a simple request-response pattern:

Figure 1. Integration overview with Tracer

Key principle: Tracer does not fetch external data during validation. Your system must provide all context needed for rule evaluation.

---

Midaz Ledger reservation seam

---

This opt-in seam belongs to Ledger HTTP v2, not to Tracer HTTP v2. Ledger HTTP v1 never invokes it. Tracer’s public HTTP API remains v1-only, including its reservation operations under /v1. The gRPC service lerian.midaz.reservation.v1.ReservationService is an internal service-to-service transport, not a public v2 API.

Set TRACER_BASE_URL to inject the seam. When it is unset, Ledger makes no reservation call. Ledger calls the injected client only when the per-ledger tracer.mode is advisory or enforce. An unset or off mode still skips reservations even when you set TRACER_BASE_URL. An honored per-call skip also bypasses the seam.

gRPC is the default transport: configure TRACER_GRPC_PORT on Tracer and point TRACER_BASE_URL at that gRPC listener. With TRACER_TRANSPORT=rest, point it at Tracer’s HTTP listener instead. Both transports use the same reservation service and the same five transitions:

Transition	Ledger action
Reserve	Hold limit capacity before the transaction commits. Retries are idempotent at the (transactionId, limitId, scopeKey, periodKey) grain.
ConfirmByTransaction	Commit every hold for a transaction; retrying after terminal state is a no-op and can return flipped=0.
ReleaseByTransaction	Return every hold for a transaction; retrying after terminal state is a no-op and can return flipped=0.
Confirm	Commit one hold; retrying a terminal reservation is a no-op.
Release	Return one hold; retrying a terminal reservation is a no-op.

For multi-tenant calls, REST forwards the tenant in the X-Tenant-Id header and gRPC forwards it as x-tenant-id metadata. Neither transport places it in the reservation message.

The listener may trust this value only over direct mTLS or behind a verified service-mesh sidecar. With TRACER_TLS_MODE=mtls, each side presents and verifies certificates. The mesh and empty TLS modes require a sidecar that enforces mTLS. Without one, the process-to-listener connection is plaintext and an untrusted caller can spoof the tenant. Tracer enables its gRPC listener only when you set TRACER_GRPC_PORT.

Failure semantics are explicit. A reservation denial is a successful response, not a transport error. The advisory mode records the outcome and proceeds. The enforce mode rejects the transaction before any balance movement, regardless of failPosture. The failPosture setting applies to any reserve-call error: closed rejects and open proceeds without a reservation.

Confirm and release failures are warning-level, non-blocking operations. The Tracer TTL reaper reconciles a missed terminal transition.

Payload-Complete Pattern

---

Tracer uses the Payload-Complete Pattern. Every request must carry all context required for validation. This design ensures:

Benefit	Description
Predictable latency	No external calls during validation; response time stays under 80ms
Simplicity	Single request contains everything needed for decision
Reliability	No dependency on external services during validation
Flexibility	Your system controls data freshness and enrichment logic

Your responsibilities

As the integrating system, you are responsible for:

1. Enriching the payload with account, segment, portfolio, and merchant data before calling Tracer
2. Providing accurate context for rule and limit evaluation. Tracer cannot fetch missing data
3. Handling the decision (ALLOW, DENY, or REVIEW) appropriately in your workflow
4. Implementing retry logic if Tracer is temporarily unavailable
5. Managing review workflows when Tracer returns REVIEW. Tracer does not include case management

Tracer validates what you send. If your payload lacks context (e.g., account status, segment membership), rules that depend on that data cannot evaluate correctly. Always ensure payloads are complete before submission.

Tracer’s responsibilities

Tracer is responsible for:

1. Evaluating rules against the provided context
2. Checking limits against current usage
3. Storing validation history for investigation and reporting
4. Returning decision with detailed information

---

Integration flow

---

Follow these steps to integrate your system with Tracer.

Step 1: Prepare the transaction context

Before calling Tracer, gather all relevant data from your systems:

Figure 2. Preparing transaction context

Step 2: Call Tracer API

Send a POST request to /v1/validations with the complete transaction context including:

• Transaction details (type, subType, amount, asset, timestamp)
• Account information (required)
• Optional: segment, portfolio, merchant, and custom

For complete payload structure and field details, see the API reference.

Step 3: Handle the response

Process the decision returned by Tracer:

Decision	Action
ALLOW	Proceed with the transaction
DENY	Reject the transaction; show reason to user if appropriate
REVIEW	Queue for manual review in your review system

The response includes the validationId for correlation with validation history, details about which rules matched, and current limit usage information.

Using metadata

Metadata allows you to pass custom fields that your rules can evaluate. Use this for context like channel, device information, customer tier, or any business-specific attributes.

Metadata keys must be alphanumeric with underscores only, maximum 64 characters. Maximum 50 entries per request.

---

Request idempotency

---

Validation requests are idempotent based on the requestId field. If you send the same requestId twice, Tracer returns the cached result from the first request instead of reprocessing.

Response code	Meaning
201 Created	New validation processed
200 OK	Duplicate request detected; cached result returned

The response body is identical in both cases. Your client should handle both status codes as success.

Why it matters: Network timeouts and retries can cause duplicate requests. Without idempotency, a retried request could double-count against limits or create duplicate validation records. The requestId ensures exactly-once processing semantics.

Idempotency contract:

• Same requestId → Same response (guaranteed)
• Different requestId → Independent processing (even if transaction data is identical)

Always generate a unique requestId (UUID) for each new transaction. Reusing a requestId from a previous transaction will return the old result, not process the new transaction.

---

Authentication

---

Tracer supports two authentication modes. You can use them independently or together.

API key authentication

The simplest option. Send your API key in the X-API-Key header with every request.

Environment variable	Description
API_KEY_ENABLED	Enable API key authentication (default: false)
API_KEY	The secret key value
API_KEY_ENABLED_ONLY_VALIDATION	Use API key only for the /v1/validations endpoint (default: false)

Plugin authentication (Access Manager)

For enterprise deployments, Tracer can delegate authentication to the Lerian Access Manager. This enables centralized authentication across all Lerian services.

Environment variable	Description
PLUGIN_AUTH_ENABLED	Enable plugin authentication (default: false)
PLUGIN_AUTH_ADDRESS	URL of the auth service (default: http://localhost:4000)

Authentication priority

When you enable both modes, Tracer uses this priority:

1. If PLUGIN_AUTH_ENABLED=true and the endpoint has no API-key-only flag → Plugin auth
2. If API_KEY_ENABLED=true or the endpoint carries the API-key-only flag → API key auth

Infrastructure endpoints (health checks, version probe, OpenAPI spec) bypass authentication and are not part of the public /v1/* API surface documented in this reference.

You can configure the /v1/validations endpoint for API-key-only authentication via API_KEY_ENABLED_ONLY_VALIDATION=true. This is useful in high-throughput scenarios where plugin auth adds unacceptable latency. This flag is incompatible with multi-tenant mode (MULTI_TENANT_ENABLED=true). The service fails to start with error code 0458.

Multi-tenant authentication

When MULTI_TENANT_ENABLED=true, Tracer runs in multi-tenant mode and the authentication model changes:

• Plugin auth is mandatory. The service fails to start with error code 0457 if PLUGIN_AUTH_ENABLED=false.
• Every /v1/* request must carry a JWT bearer token issued by Access Manager: Authorization: Bearer <jwt>.
• tenantId comes from the JWT claim, not from a header, path, body, metadata, or rule scope. There is no X-Tenant-ID header. The tenant identifier has no effect anywhere other than the token claim.
• Each tenant operates on its own PostgreSQL database. The multi-tenancy platform service resolves the tenant-specific connection at request time.
• Public endpoints (/health, /readyz, /metrics, /version) stay unauthenticated in multi-tenant mode too. The bearer-token requirement applies only to /v1/*.

If the JWT is absent, malformed, or expired, the request returns HTTP 401 with "code": "Unauthenticated". Missing API keys return the same code, with no separate TRC code.

One case is distinct. A token that parses but carries no sub claim returns HTTP 401 and error code 0474 up front. The sub claim is what the audit writer uses to attribute the action to a principal. Tracer fails loudly rather than recording the change against a generic system actor. Make sure your Access Manager tokens always carry it.

If the multi-tenant deployment hits its per-instance tenant cap, requests for cold tenants return HTTP 503 with error code 0466 and a Retry-After header. The client should back off and retry. The cap auto-resets as the LRU pool evicts cold tenants.

See Multi-tenancy for the platform-wide tenant model.

---

Performance considerations

---

Optimize your integration for low latency and high reliability.

Timeout budget

Tracer targets a response in under 80ms (p99). Configure your client timeout accordingly:

Configuration	Recommended value
Client timeout	100ms
Connection timeout	50ms
Read timeout	100ms

Retry strategy

Implement retry logic for transient failures:

On 5xx error or timeout:
  - Wait 10ms
  - Retry once
  - If still failing, apply fallback policy

Do not retry on 4xx errors. These indicate invalid requests that will fail again. For retries on 5xx/timeout, reuse the same requestId to take advantage of idempotency.

Fallback behavior

Decide what happens when Tracer is unavailable:

Strategy	When to use
Fail-open	Allow transaction if Tracer is down (prioritize availability)
Fail-closed	Deny transaction if Tracer is down (prioritize security)
Queue for review	Queue transaction for manual review

Your choice depends on your risk tolerance and business requirements.

Common integration pitfalls:

• “My retry created a duplicate validation.” Reuse the same requestId across retries. Tracer deduplicates by that field. The second call returns the cached result (HTTP 200) without creating another validation. If you generate a fresh UUID on every retry, you defeat idempotency.
• “My client times out after 30 seconds but Tracer keeps processing.” Tracer respects its own deadlines (~80ms p99 target). If your client gives up on the call, Tracer still spends that work on a response. Set client timeout aggressively (100ms) and trust the retry path.
• “Tracer rejects my validation with error code 0421 (timestamp too old) on legitimate transactions.” Default tolerance is 24 hours. Check your server clock and the transactionTimestamp you’re sending. If you batch-process late, set the timestamp to the actual transaction moment, not the moment you’re calling Tracer.
• “Test transactions show up in production validation history.” Tracer records every validation, including from test/staging environments calling the same Tracer instance. Use metadata.environment (or similar) to tag and filter test traffic if you share Tracer between environments.

---

Data freshness

---

Since you control the payload enrichment, data freshness is your responsibility. Tracer trusts the data you provide and cannot detect stale information.

Data type	Freshness recommendation	Risk if stale
Account status	Real-time or near real-time	Transactions on suspended accounts may be allowed
Segment membership	Can be cached (changes infrequently)	Wrong limits or rules may apply
Portfolio assignment	Can be cached (changes infrequently)	Incorrect scope matching
Merchant data	Can be cached with periodic refresh	Risk rules may not trigger correctly

Stale data leads to incorrect decisions. If you suspended an account but your cache shows it as active, Tracer will allow transactions that it should deny. Your enrichment layer is the source of truth for Tracer.

---

Date and time format

---

All datetime fields must use RFC3339 format with mandatory timezone:

Valid formats:

2026-01-30T10:30:00Z           (UTC)
2026-01-30T10:30:00-03:00      (São Paulo timezone)
2026-01-30T00:00:00+00:00      (UTC explicit)

Invalid formats:

2026-01-30                     (date only - rejected)
2026-01-30T10:30:00            (missing timezone - rejected)

---

Integration checklist

---

Before going to production, verify:

• Your API Key is in place and secure
• Each request includes a unique requestId (UUID)
• Client handles both 201 and 200 responses as success
• Your client timeout is 100ms
• Your retry logic covers 5xx errors
• You chose a fallback behavior
• Your payload carries all required fields
• Timestamps use RFC3339 format with timezone
• Asset codes are uppercase ISO 4217
• Your system handles each decision (ALLOW/DENY/REVIEW)
• Your system logs validation IDs for validation-history correlation

---

Example integration (pseudocode)

---

def validate_transaction(transaction):
    # Step 1: Enrich payload
    payload = {
        "requestId": generate_uuid(),
        "transactionType": transaction.type,
        "amount": transaction.amount,
        "asset": transaction.asset.upper(),
        "transactionTimestamp": now_rfc3339(),
        "account": get_account_context(transaction.account_id),
        "segment": get_segment_context(transaction.segment_id),
        "merchant": get_merchant_context(transaction.merchant_id),
        "metadata": transaction.custom_fields
    }

    # Step 2: Call Tracer
    try:
        response = http_post(
            url="https://tracer.example.com/v1/validations",
            headers={"X-API-Key": API_KEY},
            json=payload,
            timeout_ms=100
        )
    except Timeout:
        return apply_fallback_policy()
    except ServerError:
        return retry_once_or_fallback()

    # Step 3: Handle decision
    if response.decision == "ALLOW":
        return proceed_with_transaction()
    elif response.decision == "DENY":
        return reject_transaction(response.reason)
    elif response.decision == "REVIEW":
        return queue_for_manual_review(response.validationId)

---

Next steps

---

• Rules engine - Create rules that evaluate against the context you provide
• Spending limits - Configure limits that apply to your transaction scopes
• Validation history and compliance - Query validation history and use it in your compliance processes