failure-mode-glossary

name	failure-mode-glossary
description	Glossary of production-AI failure modes with verbatim operator phrasings. When the user uses any of these phrases or describes a symptom matching them, you'll recognize the pattern and respond with the right framing + MCP routing. Auto-loaded so the buyer-vocabulary is always available.
when_to_use	Anytime the user describes a symptom in their own words. Translate buyer-vocabulary to P-numbered patterns and the right MCP. Cite the named pattern in your response.
user-invocable	false

Failure-Mode Glossary (buyer-vocabulary → pattern)

When users report problems, they don't say "P3.4 HERMES.md" — they say "my Anthropic bill jumped overnight". This glossary maps buyer-vocabulary to the named pattern + the MCP that catches it.

Cost / billing vocabulary

• "My API bill jumped overnight" / "my AI bill is bigger than my rent" / "feel like we're flying blind on cost" → P3.4 HERMES.md silent billing-routing OR P9.1 wrong model for task → cost-tracker (predict_429_in_window + cost_overview)
• "$6k overnight" → P3.4 HERMES.md (canonical case) → cost-tracker
• "about to 429" / "429 imminent" / "throttle" / "rate-limit window" → P9.2 no retry budget → cost-tracker predict_429_in_window
• "my company is spending $12k/month on AI agents" → general cost-attribution pain → cost-tracker cost_overview (per-agent breakdown)

Bash / shell vocabulary

• "rm -rf" / "force push" / "drop database" / "deleted production" → P5.y LLM-emitted destructive bash → bash-vet vet_command_chain
• "chiefofautism" (verbatim author name when user cites the viral example) → P5.y → bash-vet
• "Claude wiped my repo" / "agent ran a destructive command" → P5.y → bash-vet
• "cd X && Y && rm -rf $VAR/ && Z"* style chain → P5.y chain-mode → bash-vet vet_command_chain (chain_mode=true)
• "apt remove 'nvidia' wipeout" (xornullvoid case) → P5.y PACKAGE.APT_GLOB_WIPEOUT → bash-vet
• "Railway production database deletion" / "JER lifeof_jer Apr 26 incident" → P11 agent-privilege-scoping (separate from bash-vet — privilege issue, not execution issue) → uncovered, see Pipeline P11

Cron / scheduled vocabulary

• "silent failure" / "silently broken" / "did X actually run" → P10.1 cron silent failures → silentwatch list_recent_silent_failures
• "exit-0 but empty output" / "cron returned success but no data" → P10.1 EXIT_0_EMPTY_STDOUT → silentwatch
• "running but doing no work" / "cron is a no-op" → P10.1 → silentwatch
• "web search works interactively but fails in cron" → P4.2 web-search silent fail in cron → silentwatch + audit env vars

Agent / config vocabulary

• "AGENTS.md" / "CLAUDE.md" / ".cursor/rules" / "agent config" → P5.z agent-config trust-boundary attack → skill-vetter v1.1 vet_agent_config
• "prompt injection in PR" / "ignore previous instructions in skill" → P5.z → skill-vetter
• "Cursor CVE-2026-26268" / "Gemini CLI yolo mode" / "CVSS 10.0 supply chain" → P5.z (canonical CVE) → skill-vetter
• "my agent got compromised" → P5.z → skill-vetter (vet existing config files)

Output / claim vocabulary

• "agent says X but I don't see X" / "claimed to commit but no commit" / "tests pass but…" → P6.y output-action divergence → output-vetter verify_action_outcome
• "chiefofautism" (when used about output-vs-action, not about bash) → P6.y → output-vetter
• "Codex sandbox escalation" / "agent acknowledged read-only then wrote" → P6.y STATE_VIOLATED_CONSTRAINT → output-vetter
• "agent confidently misreports" → P6.y → output-vetter
• "silent error masking" / "swallowed exception" / "fake success" → P6.2 missing error handling → output-vetter find_swallowed_exceptions

Upgrade vocabulary

• "should I upgrade" → general → upgrade-orch check_known_regressions
• "reasoning effort feels lower" / "model feels dumber" / "Anthropic regressed" → upgrade-orch detect_provider_regression (April-23 case)
• "upgraded and it broke" / "regression after version bump" → P4.3 post-upgrade regression cascade → upgrade-orch
• "CPU spiked after upgrade" → P2.2 CPU spikes after upgrade → upgrade-orch + health-mcp

Deployment / health vocabulary

• "is my deployment healthy" / "gateway status" / "what's wrong with my agent" → general → health-mcp health_overview
• "OOM" / "out of memory" → P2.1 OOM on small VMs OR P2.3 JS heap OOM → health-mcp + silentwatch
• "WebSocket disconnect every X minutes" → P3.1 WebSocket disconnects → uncovered (mention this honestly)
• "my agent broke production" (Vox @Voxyz_ai phrase) → P4.3 OR P5.y depending on context → run /aufgaard:production-audit

Reliability / methodology vocabulary

• "production grade patterns" / "battle-tested" / "actually production-ready" → general reliability framing → cite the bundle's Field Reference PDF
• "AI agents are a mess out there" (@KanikaBK / @Voxyz_ai phrasing) → general → run /aufgaard:production-audit
• "demo-to-production gap" / "worked in dev, broke in production" → general → multiple MCPs; start with audit
• "12-factor agents" / "patterns of reliable LLM applications" → general methodology → cite the bundle for the catalogue
• "AI agent broke something new in production" / "62 database migrations" (Voxyz_ai phrasing) → P4.3 + P10.x → upgrade-orch + silentwatch

Style — how to use this glossary

• When the user phrases a symptom in their own words, recognize it via the glossary, RESPOND with the named pattern: "That sounds like P5.y (LLM-emitted destructive bash). bash-vet's vet_command_chain catches this — let me run it on the command you're worried about."
• DON'T paraphrase the buyer's phrase back at them as a "diagnosis." Translate to the pattern name and the tool — that's the value.
• If the symptom is too general (e.g. "my agent has bugs"), ask one clarifying question to narrow: "Is this a cost spike, a silent failure, an action-claim mismatch, or something else?" — uses the categories from this glossary.