cc-defensive-programming

Defensive code keeps small failures from compounding into silent, hard-to-diagnose ones. These four checks catch the violations that corrupt data and hide bugs rather than crashing loudly — each is almost always required; exceptions need explicit justification:

| Check | Why | |---|---| | No executable code in assertions | Assertions are compiled out of production builds; the code vanishes with them | | No empty catch blocks | Silently swallows bugs that cascade into harder failures downstream | | External input validated at entry | Unvalidated input is security vulnerabilities and data corruption | | Assertions for bugs only | Anticipated runtime errors need handling, not assertions (which are disabled in production) |

Shared thresholds and the information-hiding rationale: Read(${CLAUDE_PLUGIN_ROOT}/references/cc-foundations.md).

Two rules that override the textbook

• "Internal team API" is still external. Any data crossing a network or process boundary is external input — validate it, even from a service your own team owns.
• A barricade reduces redundant validation; it does not replace defense-in-depth. Inside the barricade you may assume data is validated — except for security-critical paths (auth, crypto, PII), where you validate again. Barricade validation has bugs too.

External input = any data not provably controlled by the current code path: user input, files, network/API/DB responses, environment variables, and data from any other service.

When NOT to apply

• Pure functions without side effects (but still validate inputs at system boundaries).
• Prototype/spike code, time-boxed before committing to an error strategy.
• Test code — test doubles intentionally violate production patterns.
• Performance-critical inner loops where assertion overhead is profiled at >5%.

Security-critical code (authentication, authorization, cryptographic operations, PII handling) is never exempt, regardless of the above. When in doubt, validate.

Modes

CHECKER

Execute the defensive-programming, assertion, and exception checklists against the code: Read(${CLAUDE_SKILL_DIR}/checklists.md). Output one row per item: | Item | Status | Evidence | Location |, status ∈ VIOLATION (missing validation, empty catch, side-effecting assertion) / WARNING (inconsistent handling, missing docs) / PASS.

APPLIER

Produce assertion placement, error-handling strategy, barricade architecture, and validation implementations. Search the codebase for the existing error-handling pattern (exception vs return code vs Result type, custom exception classes, logging conventions) and follow it — consistency in error handling is what makes failures debuggable. Full gate: Read(${CLAUDE_PLUGIN_ROOT}/references/pattern-reuse-gate.md).

Assertion vs error handling

| Condition | Assertion | Error handling | |---|---|---| | Should never occur (programmer bug) | Yes | No | | Can occur at runtime / anticipated bad input | No | Yes | | External input | No | Yes — always validate | | Internal interface, same module | Yes | No | | Internal interface crossing a trust boundary | Yes | Yes | | Public-API precondition violation | Yes | Yes | | Security-critical / highly robust system | Both | Both (defense in depth) |

When a condition is genuinely unclear ("is this a bug or expected?"), clarify requirements with a domain expert rather than guessing.

Correctness vs robustness

• Correctness — never return an inaccurate result; shut down rather than produce a wrong answer (safety-critical, financial, data pipelines).
• Robustness — keep operating even if results are sometimes inaccurate (consumer apps, internal tools).

| Domain | Lean toward | Key question | |---|---|---| | Safety-critical (medical, aviation) | Correctness | — | | Enterprise/B2B, financial, data pipelines | Correctness | Would wrong data drive a bad business decision? Does downstream assume integrity? | | SaaS platforms | Balanced | Blast radius of a wrong answer vs unavailability? | | Consumer apps, internal tools | Robustness | Can the user recover from a crash? | | Real-time systems | Context-dependent | Is stale data better or worse than no data? |

Barricade design

1. Identify external interfaces (user input, files, network, APIs, inter-service calls).
2. Place validation at the barricade boundary — all external data checked here.
3. Inside the barricade, use assertions for internal bugs (data assumed validated).
4. Strategy: external boundary = error handling; internal = assertions.

Class-level barricade: public methods validate and sanitize; private methods within that class may assume safe data.

Error-handling strategy options

Choose one and apply it consistently — error handling is an architectural decision.

| Strategy | Use for | |---|---| | Return neutral value | Display defaults, non-critical config | | Substitute next valid data | Streaming/sensor data with redundancy | | Return previous answer | Display refresh, non-critical caching (never financial data) | | Substitute closest legal value | Input clamping, slider bounds | | Log warning and continue | Non-critical degradation, feature flags | | Return error code | APIs with status conventions, C-style interfaces | | Centralized error handler | Consistent logging, monitoring integration | | Display error message | User-facing apps (don't leak security info) | | Shut down gracefully | Safety-critical, data-corrupting errors |

Modern error propagation

Synchronous exception propagation assumes a synchronous call stack; these patterns need explicit handling:

• async/await — an unhandled rejection crashes Node; wrap awaited calls and re-throw domain-level exceptions.
• Callbacks — errors don't propagate; use the error-first callback(error, result) pattern or wrap in promises.
• Promise.all — first rejection loses other results; use allSettled and decide fail-entirely vs partial.
• Event handlers / React — synchronous throws are silently swallowed; log explicitly and use error boundaries for render errors.

Offensive programming (development builds)

Make errors loud during development so they're found early; make them unobtrusive with graceful recovery in production. Techniques: make asserts abort, fill allocated memory and freed objects with junk, fail hard on default/else clauses, email error logs to yourself.

Evidence

• "Garbage in, garbage out" is obsolete — production software must validate or reject [McConnell p.188].
• Error handling is an architectural decision, enforced consistently [McConnell p.197].
• Bugs cost ~100x more to fix in production than at design time [IBM Systems Sciences Institute].
• Mars Climate Orbiter was lost to a unit mismatch [NASA 1999] — nine months of success does not mean code is safe.

Chain

| After | Next | |---|---| | Validation complete | Skill(code-foundations:cc-control-flow-quality) |