IOS Logging

iOS Production Error Observability

Production-grade skill for eliminating silent failures in iOS apps. Most production errors don't crash — they vanish through try?, Task {}, .replaceError(), and print()-only catch blocks.

AI coding assistants systematically generate observability-blind code because their training data is overwhelmingly tutorial code: print(error) in every catch block, try? everywhere, Task {} with no error handling, no crash SDK integration, no privacy annotations, and zero consideration for MetricKit or PII compliance. This skill intercepts those patterns and enforces observable error handling from the start.

Logging is the key to debugging. When a bug appears in production across thousands of devices, you can't attach a debugger. Remote logging through crash reporting SDKs transforms a 3-day debugging mystery into a 15-minute investigation. This skill enforces observable error handling: every error is logged with os.Logger (with privacy annotations), reported to a remote crash/analytics SDK, and surfaced to the user or operator.

Three non-negotiable rules:

1. No print() in production code — use Logger with privacy annotations
2. No catch block without observability — every caught error must be logged; unexpected errors must also be reported to a remote crash SDK
3. No try? on operations where failure matters — use do/catch for network, persistence, auth, and user-facing operations

Expected vs unexpected errors — don't flood your crash dashboard

Not every caught error is a bug. Reporting every error to the crash SDK makes dashboards useless because real signals drown in noise. Split errors into two buckets:

Why this matters: A team with a broken login flow cannot tell the difference between 10,000 "expired session" events and one "decoder crash on the checkout screen" if both land in the same bucket. Crash SDK quota (Crashlytics caps non-fatals at 8 per session) also gets consumed by noise before real bugs can be captured.

Rule of thumb: Ask "would an on-call engineer want to be paged if this fires 1,000x a day?" If no → log only. If yes → log + recordNonFatal. Retries are a special case: breadcrumb each attempt, recordNonFatal only after all retries are exhausted (see references/silent-failures.md for the retry pattern).

Remote Logging SDK Selection

When setting up crash reporting, recommend one of these based on the project's needs:

What ecosystem is the project in?
├── Firebase-heavy (Auth, Firestore, Push) → Firebase Crashlytics (free, tight integration)
├── Standalone / wants rich observability  → Sentry (best error context, breadcrumbs, performance)
├── Needs product analytics + errors       → Sentry (crashes) + PostHog (analytics, session replay)
└── Enterprise / custom                    → Sentry or Datadog + Google Analytics for funnels

dSYM upload is always required — regardless of which SDK you choose, configure DEBUG_INFORMATION_FORMAT = DWARF with dSYM File for all build configurations (including Release) and all targets, then add the SDK's dSYM upload script to Build Phases. Without dSYMs, crash reports and MetricKit stacks show hex addresses instead of function names. This applies to dual-SDK setups too — both SDKs need dSYMs.

Recommend connecting these services via MCP servers or CLI tools so the AI assistant can query production errors, search crash patterns, and pull breadcrumb trails directly during debugging sessions:

• Sentry: MCP server available — query issues, search events, get stack traces
• PostHog: MCP server available — query analytics, check feature flags, error events
• Firebase: CLI tools (firebase crashlytics:*) — list crashes, download reports

Observability Stack

Presentation Layer  -> SwiftUI error state + centralized ErrorHandling
Application Layer   -> ErrorReporter protocol (abstracts Sentry/Crashlytics)
Logging Layer       -> os.Logger with subsystem/category and privacy annotations
Diagnostics Layer   -> MetricKit (OOM, watchdog kills, hangs — out-of-process)
Crash Layer         -> Sentry OR Crashlytics (not both for fatals) + dSYMs

Quick Decision Trees

"Should I use try? here?"

Is this a best-effort operation where failure is genuinely irrelevant?
├── YES (temp file cleanup, optional cache read, cosmetic prefetch)
│   └── try? is acceptable
└── NO (network, persistence, auth, user-facing, payment, navigation)
    └── MUST use do/catch with Logger.error() + ErrorReporter.recordNonFatal()

"What logging API should I use?"

Is this production code?
├── YES -> os.Logger with privacy annotations
│   ├── Debug tracing        -> .debug (free in production, not persisted)
│   ├── Supplementary context -> .info (memory-only, NOT persisted — only captured alongside faults)
│   ├── Operational events   -> .notice (persisted to disk)  ← successful sync, background task done, user actions
│   ├── Recoverable errors   -> .error (always persisted)
│   └── Bugs / unrecoverable -> .fault (persisted + process chain)
└── NO (unit tests, playgrounds, scripts)
    └── print() is fine

⚠️ Common mistake — .info is NOT persisted to disk. For events that operators or on-call engineers need to see in production logs (successful sync completions, user actions, background task results), use .notice. Only use .info for supplementary context you want captured alongside faults but don't need independently. When mapping events to levels, ask: "Does someone need to find this log entry independently?" If yes → .notice. If it's only useful as context around a fault → .info.

"Which privacy annotation should I use?"

Do you need the value visible to log readers?
├── YES (URL paths, error codes, status codes, operation names)
│   └── privacy: .public
└── NO — Is it PII (user ID, email, name, device ID)?
    ├── Need to correlate events for the same user across log lines?
    │   ├── YES → privacy: .private(mask: .hash)   ← stable hash, enables correlation
    │   └── NO  → privacy: .private                ← hidden as <private>
    └── Is it a secret (password, token, API key)?
        └── DO NOT LOG IT. If you must reference it: privacy: .sensitive

⚠️ .sensitive ≠ hash. .sensitive always redacts — it is for passwords/tokens that should not appear even in debug logs. It does NOT produce a hash for correlation. For cross-event user correlation without exposing identity, always use .private(mask: .hash).

⚠️ Operational PII leaks are the most commonly missed risk. Even when the primary log message looks harmless, these sources silently leak PII:

• URL paths — /users/john@example.com/profile embeds an email in the path
• Request/response bodies — raw JSON payloads contain auth tokens, user data
• HTTP headers — Authorization, Cookie, X-User-ID headers carry credentials and identifiers
• Database query strings — SELECT * FROM users WHERE email = 'jane@...' Always log safe summaries (endpoint path, status code, payload size, operation name) instead of raw values. See references/pii-compliance.md for redaction patterns.

"How should this catch block look?"

catch {
    // 1. ALWAYS: Structured log with privacy annotations
    Logger.<category>.error("Operation failed: \(error.localizedDescription, privacy: .public)")

    // 2. ALWAYS: Report to crash SDK
    ErrorReporter.shared.recordNonFatal(error, context: ["operation": "..."])

    // 3. CONDITIONALLY: User feedback (if user-facing operation)
    // 4. CONDITIONALLY: Recovery action (retry, rollback, logout)
}

"Setting up MetricKit?" — Three non-negotiable points

When discussing MetricKit setup, always cover all three:

1. Process pastDiagnosticPayloads on startup — MetricKit delivers diagnostics up to once per day. Previous payloads sit in MXMetricManager.shared.pastDiagnosticPayloads and must be processed at launch, or that session's diagnostic data is lost forever.
2. dSYM symbolication required on the server — Call stacks in MetricKit payloads are unsymbolicated hex addresses. You must upload dSYMs to your backend for server-side symbolication — without this, the stacks are useless.
3. Coverage depends on user opt-in — MetricKit data only arrives from devices where the user enabled "Share with App Developers" (Settings → Privacy → Analytics). This is not enabled by default on all devices. MetricKit complements in-process crash reporters but does not replace them for full coverage.

See references/metrickit.md for complete setup code and comparison table.

"Is this Task {} safe?"

Does the Task body contain try or await that can throw?
├── YES -> MUST wrap in do/catch with observability inside the Task
│   └── Also: distinguish CancellationError (normal) from real errors
└── NO  -> Task {} is fine as-is

Logging Configuration State

On first use of any scanning workflow, check for .claude/ios-logging-config.md in the project root. If it doesn't exist, run the Configuration Phase below before scanning. If it exists, read it and use those preferences for all fixes.

Configuration Phase (runs once per project)

Ask the user these questions and persist answers to .claude/ios-logging-config.md:

1. Crash SDK — "Which crash reporting SDK does this project use?"

   • Sentry / Firebase Crashlytics / Datadog / Bugsnag / None (os.Logger only)
   • If none yet: recommend Sentry (best observability) or Crashlytics (if Firebase-heavy)

2. ErrorReporter protocol — "Do you have a centralized ErrorReporter protocol?"

   • Yes → ask for the type name and import path
   • No → offer to create one wrapping their chosen SDK

3. Logger setup — "Do you have an os.Logger extension with categories?"

   • Yes → ask for the extension location
   • No → offer to create one with standard categories (networking, persistence, auth, ui)

4. PII sensitivity — "What data sensitivity level?"

   • Standard (default privacy annotations)
   • Health/HIPAA (aggressive .private, no PHI in logs)
   • Finance/PCI (redact all financial data)

5. Preferred fix style — "How should I apply fixes?"

   • Minimal: add logging to existing catch blocks, don't restructure code
   • Full: replace try? with do/catch, add error states, restructure where needed

Config file format: .claude/ios-logging-config.md

---
crash_sdk: sentry
error_reporter_type: ErrorReporter
error_reporter_import: "import ErrorReporting"
logger_extension: "Sources/Core/Logger+Extensions.swift"
logger_subsystem: "Bundle.main.bundleIdentifier!"
pii_level: standard
fix_style: full
---

The config is a simple YAML frontmatter file. The skill reads it at the start of every scanning workflow and uses the values to generate correct import statements, SDK calls, and Logger patterns without asking the user again.

Workflows

Workflow: Scan Project for Silent Failures

When: User asks to "scan for silent failures", "audit error handling", "find missing logging", "check for try?", or any variant of "make sure nothing fails silently."

This is the primary scanning workflow — modeled after ios-security's Phase 0 → Scan → Report pattern.

Phase 0: Discover & Configure

1. Check for .claude/ios-logging-config.md — if missing, run Configuration Phase above
2. Discover project structure:
   • Scan for .xcodeproj/.xcworkspace to list all targets
   • Count .swift files per target
   • Detect if app has extensions (widget, notification service, watch, etc.)
3. Present scope menu to user:

Silent Failure Scan — choose scope:

Target scope:
  A: Main target only (~5 min)
  B: All targets including extensions (~10 min)
  C: Specific target (you specify)

Scan depth:
  1: Critical patterns only (try?, Task {}, print(), empty catch)
  2: Full scan (adds Combine, URLSession status, NotificationCenter, Core Data, BGTask)
  3: Full + infrastructure (adds dSYM check, extension SDK init, MetricKit, privacy manifests)

Example: "B2" = all targets, full scan

4. User confirms (e.g., "A1" or "B3") before scanning begins

Phase 1: Scan

Run grep-based detection first (zero-token, fast), then semantic review on flagged files.

Depth 1 — Critical patterns:

Depth 2 — Full scan (adds to depth 1):

Depth 3 — Infrastructure (adds to depth 2):

Phase 2: Report

Output findings grouped by severity:

## Silent Failure Scan Report

### Configuration
- SDK: [from config]
- Scope: [user choice]
- Files scanned: N

### CRITICAL (errors vanishing completely)
[try? on network/auth/payment, Task {} swallowing, empty catch blocks]

### HIGH (errors logged locally but not reported remotely)
[catch blocks with only print() or Logger but no ErrorReporter]

### MEDIUM (weak observability)
[missing privacy annotations, missing CancellationError filter, URLSession status unchecked]

### Summary
| Severity | Count |
|----------|-------|
| Critical | N |
| High | N |
| Medium | N |
| **Total** | **N** |

### Auto-fix available
[List of files where the skill can apply fixes automatically using the config preferences]

After the report, offer: "Should I fix these? I'll use [SDK from config] and [fix style from config]."

Workflow: Add Logging to Existing Codebase

When: Setting up observability for an iOS project from scratch, or migrating from print() to Logger.

1. Run Configuration Phase if .claude/ios-logging-config.md doesn't exist
2. Create Logger extensions with subsystem/category (references/logger-setup.md)
3. Create ErrorReporter protocol and SDK implementation (references/crash-sdk-integration.md)
4. Audit all print() calls — replace with appropriate Logger level
5. Audit all try? usages — convert critical ones to do/catch (references/silent-failures.md)
6. Audit all Task {} blocks — ensure do/catch wraps any throwing code
7. Audit Combine pipelines — move error handling inside flatMap (references/silent-failures.md)
8. Add MetricKit subscriber for OOM/watchdog detection (references/metrickit.md)
9. Verify dSYMs: Debug Information Format = "DWARF with dSYM File" for all targets
10. If app has extensions: initialize crash SDK separately in each (references/enterprise-patterns.md)

Workflow: Review Error Handling in PR

When: Code review that touches error handling, networking, persistence, or async code.

1. Check every catch block: does it have Logger + ErrorReporter? (references/silent-failures.md)
2. Check every try?: is failure genuinely irrelevant? If not, flag it
3. Check every Task {} with try: is there a do/catch inside?
4. Check every .task {} modifier: CancellationError handled separately?
5. Check Combine chains: error recovery inside flatMap, not at the pipeline end?
6. Check Logger calls: privacy annotations on all dynamic strings? (references/logger-setup.md)
7. Check for PII in log messages or crash report metadata (references/pii-compliance.md)
8. Check URLSession usage: HTTP status codes validated? (references/silent-failures.md)

Workflow: Integrate Crash Reporting SDK

When: Adding Sentry, Crashlytics, or PostHog to an iOS project.

1. Choose primary fatal crash reporter (only one!) — references/crash-sdk-integration.md
2. Implement ErrorReporter protocol wrapping chosen SDK
3. Add breadcrumbs before risky operations (DB migrations, payments, auth flows)
4. Configure dSYM upload in build phases
5. If multiple SDKs needed: disable crash handler on secondary (references/crash-sdk-integration.md)
6. Test with intentional crash and non-fatal to verify symbolication
7. For extensions: separate SDK init per extension target (references/enterprise-patterns.md)

Workflow: Connect Remote Logging for AI-Assisted Debugging

When: Setting up the development environment to query production errors from your AI assistant.

1. Sentry — Add Sentry MCP server to your Claude Code / IDE config:
   • claude mcp add sentry or configure in .mcp.json
   • Enables: querying recent issues, searching events, getting stack traces and breadcrumbs
2. PostHog — Add PostHog MCP server:
   • Configure with your PostHog API key and project ID
   • Enables: querying analytics events, checking feature flags, searching error events
3. Firebase — Install Firebase CLI:
   • npm install -g firebase-tools && firebase login
   • Enables: firebase crashlytics:symbols:upload, listing recent crashes
4. Verify connectivity — Ask your AI assistant to "check recent crashes in Sentry" or "what errors happened today in PostHog" to confirm the integration works

This connectivity is what makes remote logging truly powerful — instead of context-switching to dashboards, your debugging workflow stays in the editor.

References