foolproof-product

name

foolproof-product

description

Use this skill on ANY code change in codevira's product surface (mcp_server/, indexer/). Triggers on Edit/Write to product code, on phrases like "fix bug", "add feature", "handle error", or any modification to commands the end-user runs. Enforces 10 product- reliability principles: no silent failures, atomic state mutations, defensive parsing, bounded resources, predictable detection, reversible operations, helpful error messages, graceful degradation, observability, self-diagnosis. Refuses to ship code that violates any of them.

Foolproof product — non-skippable code invariants

The release-readiness skill prevents shipping a release without evidence. THIS skill prevents writing code that misbehaves once shipped. The two are complementary; both must be enforced.

v2.0.0 shipped with 23 bugs (A–O) because individual code changes satisfied "the function returns a value" but violated the product invariants below. This skill makes those invariants explicit and checkable.

The 10 principles (P1–P10)

For every code change you make in mcp_server/ or indexer/, walk through this checklist. Skipping any P is a discipline breach.

P1 — No silent failures, ever

For every code path that does work:

• Returns >0 results → succeeds with concrete counts in the response
• Returns 0 results → MUST emit a clear "no results because " message with a fix_command field

Anti-pattern (the v2.0 bug pattern):

def index() -> dict:
    matched = [f for f in files if matches(f)]
    if not matched:
        return {"chunks": 0}   # ← silent zero. NEVER.
    return {"chunks": len(matched)}

Correct:

def index() -> dict:
    matched = [f for f in files if matches(f)]
    if not matched:
        return {
            "chunks": 0,
            "warning": f"No files matched. watched_dirs={watched_dirs} "
                       f"file_extensions={extensions}",
            "fix_command": "codevira configure",
        }
    return {"chunks": len(matched), "files": [str(f) for f in matched]}

P2 — Self-diagnose on startup

Every long-running entry point (MCP server boot, watcher start, indexer process) must run startup health checks:

• Database files openable + expected tables present
• Schema version matches (migrate forward OR halt with error)
• Required dependencies importable (chromadb, sentence-transformers)
• File system permissions on data dir

On any failure: halt with a clear remediation hint. Never start in a degraded state without telling the user.

P3 — Atomic state mutations

Any operation that writes external state (files, DB rows, IDE configs) must be atomic. The temp-file → fsync → rename pattern is the rule:

def write_config_atomic(path: Path, content: dict) -> None:
    """Write JSON atomically — readers never see half-written state."""
    tmp = path.with_suffix(path.suffix + ".tmp")
    tmp.write_text(json.dumps(content, indent=2))
    os.fsync(tmp.open("rb").fileno())  # force kernel buffer to disk
    tmp.replace(path)                  # atomic rename on POSIX

Database mutations: SQLite transactions (begin → ... → commit OR rollback on exception). Multi-step operations: rollback on partial failure.

P4 — Defensive parsing

Every external input is parsed with try/except + fail-safe defaults:

def load_config(path: Path) -> dict:
    """Load config with safe fallback. Never crashes on bad input."""
    if not path.is_file():
        return DEFAULT_CONFIG  # log a warning
    try:
        return yaml.safe_load(path.read_text()) or DEFAULT_CONFIG
    except yaml.YAMLError as e:
        logger.warning("Bad config at %s: %s — using defaults", path, e)
        return DEFAULT_CONFIG

User input that arrives malformed should produce a log warning + safe default, NOT a crash. The crash is a P4 violation.

P5 — Bounded resources

No unbounded loops. No retry without a circuit breaker. No log line written without a rate limiter.

The 41-crash UDAP log was P5 violation: the watcher hit a corrupted Chroma store and retried every file in sequence with no halt condition.

Pattern:

class CircuitBreaker:
    def __init__(self, max_consecutive_failures: int = 5):
        self.failures = 0
        self.limit = max_consecutive_failures

    def call(self, fn, *args, **kwargs):
        try:
            result = fn(*args, **kwargs)
            self.failures = 0
            return result
        except Exception:
            self.failures += 1
            if self.failures >= self.limit:
                raise CircuitOpen(f"halted after {self.limit} consecutive failures")
            raise

Apply at every retry boundary.

P6 — Predictable detection

If configure says "8 files will be indexed" and index then matches 0 files, that's a bug. The same matcher must drive both. Single source of truth for "what counts as a source file."

Concrete rule: discover_source_files() is THE matcher. Any code in the indexer pipeline that filters files must call it, not roll its own.

P7 — Reversible operations

Every operation that touches user state has a documented reverse:

Operation	Reverse
codevira setup writes IDE configs	codevira clean removes them
codevira init registers in global.db	codevira clean --project <key>
Hook installation writes to ~/.claude/hooks/	codevira hooks uninstall
DB migration	Rollback script in indexer/migrations/

For IDE configs specifically: ALWAYS merge, never overwrite. User keys must survive a codevira setup/refresh.

P8 — Helpful error messages

Every error visible to the user (CLI output, MCP tool response, log) follows the WHAT + WHY + FIX pattern:

Bad:
  ✗ Error: file not found

Good:
  ✗ Cannot read config at ~/.codevira/projects/foo/config.yaml:
    file does not exist. Project may not be initialized.
    Fix: run `codevira init` from the project root.

For MCP tool responses, include a fix_command field as the codevira convention.

P9 — Graceful degradation

Single subsystem failure must NOT cascade. Build with this matrix:

Subsystem	Down state	What still works
ChromaDB corrupted	Keyword search via SQLite	get_node, decisions, graph, roadmap
HuggingFace network	Cached model	Everything (until model needs re-download)
~/.codevira/global.db corrupt	Auto-reinitialize, warn	Per-project state preserved
Watcher crash	Manual codevira index --full	All MCP tools
Hook script broken	MCP server alone	Memory + retrieval

Every new feature: ask "if this subsystem fails, what continues to work?" Document the answer in the docstring.

P10 — Observability

Every operation logged with structured data (not bare strings):

logger.info(
    "indexed_files",
    extra={
        "project": project_name,
        "files": len(matched),
        "chunks": chunk_count,
        "elapsed_ms": int(elapsed * 1000),
    },
)

codevira doctor is the truth oracle: it reads actual state, never infers. Crash log auto-rotates (30 days OR on version change). Audit log for any sensitive operation (IDE config write, DB migrate, hook install).

The mandatory checklist (before any product code change)

Before calling Edit/Write on mcp_server/ or indexer/:

For this change, I am affecting:
  - Subsystem: <indexer | mcp tool | hook | CLI | watcher>
  - Files touched: <list>

Principle check:
  P1 No silent failures:       [ ] All zero-result paths emit reason + fix_command
  P2 Self-diagnose:            [ ] Startup checks updated if subsystem touched
  P3 Atomic mutations:         [ ] Any file/DB write uses atomic pattern
  P4 Defensive parsing:        [ ] External input has try/except + fallback
  P5 Bounded resources:        [ ] Loops bounded; retries circuit-broken
  P6 Predictable detection:    [ ] Single matcher; no parallel implementations
  P7 Reversible:               [ ] Uninstall counterpart exists/documented
  P8 Helpful errors:           [ ] WHAT + WHY + FIX in every error message
  P9 Graceful degradation:     [ ] If <this fails>, <these still work>
  P10 Observability:           [ ] Structured log; doctor check covers new state

Skipping a P without justification is a discipline breach. If a P genuinely doesn't apply (e.g. a refactor with no state mutation skips P3), say so explicitly.

Anti-patterns this skill refuses

• Returning 0 results without a reason field.
• Writing files non-atomically.
• Catching exceptions and silently continuing (the P4 dark form).
• Retry loops without a max count.
• Error messages that say "failed" without saying WHY.
• New subsystems with no doctor check.
• New subsystems without a clean uninstall path.
• "It works in the happy path, ship it." All sad paths must be walked too.

Why this exists

23 bugs shipped in v2.0.0 (A–O) all trace to one or more violated P-principles. The post-mortem:

Bug	Violated P
A: configure/index disagree	P6 (single source of truth)
B: index says "up to date" with 0 chunks	P1 (silent failure)
C: status doesn't warn on uninitialized	P1, P10
D: config split (in-repo + centralized)	P3 (atomic), P6
E: docs-only silent 0 chunks	P1
F: init drops top-level files	P6
Chroma HNSW crash storm	P2 (startup check), P5 (circuit breaker)
Claude Desktop disconnects	P9 (graceful degradation under timeout)
41-crash log spam	P5 (rate limit), P10 (auto-rotate)

If P1–P10 had been enforced at code-write time, none of these would have shipped. This skill makes the enforcement structural.