CheckResultsService — reuse any integration check's results

SkillFiles & storage

Shows your agent how to reuse existing integration check results, like 2FA status or AWS S3 findings, in app features.

Available today. Use it from your connected AI after setup.

Connect ahel once, and every AI you use reads what you have installed.

Then ask your AI: use the CheckResultsService — reuse any integration check's results skill

About this capability

How to reuse ANY integration check's results in a feature via the universal CheckResultsService (apps/api integration-platform). Use whenever a feature needs data produced by an integration check — "show 2FA status on People", "surface AWS S3 findings in X", "reuse a check's results", "per-user/per-

What this skill tells your AI

The instructions your AI receives, as published by trycompai/comp in .claude/skills/check-results-service/SKILL.md and read by ahel’s review.

The one idea

Integration checks (2FA, AWS S3 encryption, device posture, …) all write per-resource results to one table (IntegrationCheckResult). CheckResultsService is the single, universal, read-only way to get those results into any feature. It is deliberately feature-agnostic: it fetches results and hands them back in a stable envelope, and it does not know or care what the data means.

Service = fetch the results (generic). Feature = interpret + map + present.

If you're about to query IntegrationCheckResult or CheckRunRepository directly from a new feature — stop. Use this service instead.

  • Service: apps/api/src/integration-platform/services/check-results.service.ts
  • Exported from IntegrationPlatformModule — inject it into any feature module.
  • Reference consumer (copy this): the 2FA feature — two-factor-source.controller.ts.

When to use it

Use it whenever a feature needs the output of an integration check:

  • "Show each employee's 2FA status on the People tab" (the current consumer)
  • "Surface which S3 buckets failed encryption inside feature X"
  • "Show device compliance per user from the MDM check"
  • "List which connected integrations can supply data for task T"

The API (3 methods)

// 1. Discover sources: which connected integrations can feed a task, + connection state.
listSourcesBoundToTask(organizationId, taskTemplateId): Promise<CheckSourceInfo[]>

// 2. The primitive: full results of a check's latest REAL run for ONE connection.
getLatestResultsByCheck({ organizationId, connectionId, checkId, resourceType? }): Promise<CheckResultRow[]>

// 3. Convenience: results for a task-bound check from a chosen source (provider slug).
//    Resolves task -> check and slug -> connection for you.
getLatestResultsForTask({ organizationId, taskTemplateId, sourceSlug, resourceType? }): Promise<CheckResultRow[]>

Notes:

  • "Latest REAL run" = newest run that isn't inconclusive/held and whose connection isn't disconnected. Held runs (our-side self-heal) never leak to a feature.
  • No row cap — you get the full result set (unlike the 30-row task-history display).
  • resourceType filters rows (e.g. 'user', 'bucket'). Omit to get all.
  • Empty array = "no data" (source not bound/connected, or never really ran). Never throws for "no results".

Person-scoped results: the shape contract

Checks about PEOPLE (employee access, 2FA/MFA, training) follow a standard emission shape — this section is the canonical definition of it:

  • One row per person — never a single aggregate row with a roster buried in evidence.
  • resourceType: 'user' (exactly this string).
  • resourceId = the person's email, lowercased + trimmed. Fallback username || id only when the provider genuinely exposes no email (such rows won't join to members — acceptable, still visible in evidence views).
  • evidence carries what the provider knows: email, name, role, isAdmin, status, lastLogin, plus a checkedAt timestamp.
  • Access/inventory rows always emit as pass (having access is information, not a violation); compliance-gate checks (2FA, training) pass/fail per person; error paths (bad creds, missing scopes) stay org-level rows.

So a feature joining check results to org members does exactly this — no parsing, no AI:

const rows = await checkResults.getLatestResultsForTask({
  organizationId, taskTemplateId, sourceSlug, resourceType: 'user',
});
const forMember = rows.filter((r) => r.resourceId === member.email.toLowerCase());

If a source returns zero 'user' rows, its check hasn't been normalized to the standard yet (or genuinely has no per-person data) — render that as "no per-person data from this source", and fix the CHECK to emit the shape above. Never work around it by parsing aggregate evidence in the feature.

The envelope you get back

interface CheckResultRow {
  resourceId: string;      // provider-native id: email (2FA), bucket ARN (S3), repo, …
  resourceType: string;    // 'user' | 'bucket' | …
  passed: boolean;         // did this resource pass the check
  title: string;
  description: string | null;
  evidence: Prisma.JsonValue;  // ← check-SPECIFIC payload. The service does NOT interpret it.
  collectedAt: Date;
  runId: string;
  connectionId: string;
}

The evidence rule (important): the envelope is universal; the check-specific data lives in evidence as raw JSON. The service never types or interprets it. Your feature validates evidence at its own edge with a zod schema and reads only the fields it understands:

const TwoFaEvidence = z.object({ isEnrolledIn2Sv: z.boolean() });
const parsed = TwoFaEvidence.safeParse(row.evidence);
// use parsed.data.isEnrolledIn2Sv — never `as any`

For a simple pass/fail feature you often don't even need evidence — just use row.passed and row.resourceId (that's all 2FA needs).

How to add a NEW consumer (step by step)

Say a feature wants "which AWS S3 buckets failed encryption":

  1. Inject the service: add CheckResultsService to your feature's constructor (the module already exports it; import IntegrationPlatformModule if your module doesn't already).
  2. Get results with the primitive (S3 usually spans many AWS connections, so you may loop connections from listSourcesBoundToTask or your own connection lookup):
    const rows = await this.checkResults.getLatestResultsByCheck({
      organizationId,
      connectionId,
      checkId: 'aws-s3-encryption',   // the check's manifest id
      resourceType: 'bucket',
    });
    
    …or, if your feature lets the user PICK a source for a task (like 2FA does), use getLatestResultsForTask with the task template id and the chosen sourceSlug.
  3. Interpret in YOUR feature: map resourceId/passed, validate evidence with zod, shape the response your UI needs. Do not add feature logic to the service.
  4. Test with the service mocked (see two-factor-source.controller.spec.ts).

Worked example — the 2FA feature (reference implementation)

The People-tab 2FA column is consumer #1. Study it end to end:

  • two-factor-source.controller.ts
    • available-2fa-sourceslistSourcesBoundToTask(org, TASK_TEMPLATES.twoFactorAuth)
    • two-factor-statusesgetLatestResultsForTask({ org, taskTemplateId: twoFactorAuth, sourceSlug: org.twoFactorSource, resourceType: 'user' }), then maps resourceId→email (lowercased) and passedenabled/missing. Emails with no row are resolved to "Not provided" on the client, never a false "missing".
  • The controller owns the 2FA-specific bits (the Organization.twoFactorSource column, the enabled/missing interpretation). The generic fetch is all in the service.

Rules / do & don't

  • ✅ DO put every "read a check's results" path through this service.
  • ✅ DO validate evidence with zod in your feature. No as any.
  • ✅ DO treat an empty array as "no data / not provided" — never a failure.
  • ❌ DON'T query IntegrationCheckResult / CheckRunRepository directly from a feature.
  • ❌ DON'T add feature-specific interpretation (enabled/missing, domain mapping) to the service — it must stay universal and read-only.
  • ❌ DON'T assume a source produces per-user (or email-keyed) rows — that's per-check. If a chosen source's check emits a different resourceType or a non-mappable resourceId, your feature should degrade gracefully (e.g. "Not provided"), not crash.

Source selection is NOT part of this service

Which integration an org uses for a purpose (e.g. Organization.twoFactorSource) is feature-owned config, mirroring employeeSyncProvider / deviceSyncProvider. Add a column per feature. (If a 3rd/4th feature needs it, consider generalizing selection then — not before.)

Signals

GitHub stars
2k
Forks
412
Last commit
Sep 2026
Advanced
Catalog kind
skill
Gateway key
check-results-service
Source
github.com/trycompai/comp