一键在 Manus 中运行任何 Skill

production-patterns

星标2

分支0

更新时间2026年5月8日 10:27

Reference catalogue of 35 documented production-AI failure patterns. Each pattern has a P-number, name, indicators, root cause, severity, and which Aufgaard MCP catches it. Auto-loaded when discussing reliability, audit, deployment health, specific failure modes, agent safety, or any phrase from the catalogue. Use this skill's content to ground recommendations in named patterns rather than ad-hoc advice.

安装

用 Codex 或 Claude 帮你安装复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。

在 Manus 中运行

来源

temurkhan13

temurkhan13/aufgaard

打开 GitHub 仓库查看创作者相关仓库

下载

在 Manus 中运行

相关职业SOC

基于 SOC 职业分类

软件质量保证分析师与测试员计算机与数学类职业·SOC 15-1253

SKILL.md

readonly

同仓库更多 Skills

同仓库

cost-overview

temurkhan13/aufgaard

Show current production AI cost overview — totals, top spenders by agent, per-provider breakdown, anomaly detection, and time-to-429 rate-limit prediction. Use when the user asks "what's my AI bill", "which agent is burning tokens", "am I about to hit the rate limit", or wants a cost dashboard. Cross-provider — works for Anthropic, OpenAI, Gemini, Bedrock, Ollama.

2026-05-102

find-swallowed-exceptions

temurkhan13/aufgaard

Scan Python source files for swallowed-exception patterns that silently turn errors into fake successes. Catches bare `except` blocks that pass / return None / return mock objects, log-and-fake-success handlers, and mock-substitution-on-error. AST-based — not just regex. Use before any deploy of new agent code, on the working directory after a bug fix, or routinely on production-path Python files.

2026-05-102

health-check

temurkhan13/aufgaard

Show current AI deployment health overview — gateway status, plugin/skill registry, recent errors, CPU/RAM pressure, OOM history, cron status, disk pressure, upgrade outcome. The "vital signs" panel for a production AI deployment. Use when the user asks "is my deployment healthy", "what's wrong with my agent", "are services up", or wants an at-a-glance status check.

2026-05-102

operator-cheatsheet

temurkhan13/aufgaard

One-page operator cheatsheet for the Aufgaard plugin. The "what to check, when, why" reference. Auto-loaded so day-to-day routing is fast. Use as the lightweight summary when the user wants a quick recap of capabilities.

2026-05-102

production-audit

temurkhan13/aufgaard

Comprehensive production-AI deployment audit against the 35-pattern catalogue. Calls all 7 MCP servers in parallel — bash-vet, skill-vetter, cost-tracker, silentwatch, health-mcp, upgrade-orch, output-vetter — and synthesizes a one-page report with critical findings, audit score, and remediation recommendations. Use when the user asks "is my AI deployment healthy?", "audit my production AI setup", "run a production audit", or wants an end-to-end review of their AI deployment.

2026-05-102

should-i-upgrade

temurkhan13/aufgaard

Check whether upgrading a package / runtime / model version is safe. Looks up the user-driven regression catalogue (8+ entries from real field reports) AND runs provider-side regression detection (catches Anthropic-April-23-style silent reasoning-effort downgrades). Returns a recommended upgrade path with mitigations. Use before any significant package upgrade, model-version bump, or runtime change.

2026-05-102

name	production-patterns
description	Reference catalogue of 35 documented production-AI failure patterns. Each pattern has a P-number, name, indicators, root cause, severity, and which Aufgaard MCP catches it. Auto-loaded when discussing reliability, audit, deployment health, specific failure modes, agent safety, or any phrase from the catalogue. Use this skill's content to ground recommendations in named patterns rather than ad-hoc advice.
when_to_use	Anytime the user discusses production-AI reliability, asks "what could go wrong", reports a failure that needs categorization, requests an audit, or names any of the failure modes below. Cite P-numbers in your responses for traceability.
user-invocable	false

Production AI Failure Pattern Catalogue (35 patterns)

This is reference content. When the conversation touches reliability / audit / deployment / agent failures, ground your responses in these P-numbered patterns rather than generic advice. Each pattern is mapped to the Aufgaard MCP that catches it.

1. Gateway / Runtime (P1.x)

P1.1 — Gateway crash-loop on unrecognized config keys. Indicators: 377+ crashes in 4 weeks, no exponential backoff. Root cause: gateway treats unknown keys as fatal. Catches: health-mcp (symptom only). Severity: Critical.
P1.2 — Single plugin failure causes complete gateway outage. Indicators: one plugin's bug brings down the whole gateway. Root cause: no plugin sandboxing. Catches: health-mcp component checks (post-hoc). Severity: High.
P1.3 — Onboard command overwrites gateway tokens. Indicators: re-running setup clobbers existing auth. Root cause: no idempotency check. Catches: uncovered (next-MCP candidate). Severity: High.

2. Memory / CPU (P2.x)

P2.1 — OOM on small VMs from cron. Indicators: OOM-killer activates during cron-heavy hours, swap 99%+. Root cause: parallel agent invocations without memory budgeting. Catches: silentwatch + health-mcp (combined). Severity: High.
P2.2 — CPU spikes (>140%) after upgrade. Indicators: sustained high CPU after minor version bump. Root cause: regression in newer version. Catches: upgrade-orch + health-mcp. Severity: High.
P2.3 — JavaScript heap OOM in agent harness. Indicators: Node-based harness crashes under load. Root cause: session state accumulates without GC. Catches: health-mcp (crash detection). Severity: High.

3. Connection / Networking (P3.x)

P3.1 — WebSocket disconnects every 10–35 minutes. Indicators: regular disconnects, queued messages. Root cause: missing heartbeat, proxy timeouts. Catches: uncovered (next-MCP candidate). Severity: High.
P3.2 — 30–90s latency on Docker VPS. Indicators: API calls slow only inside Docker. Root cause: DNS resolution / network mode. Catches: health-mcp + upgrade-orch fingerprint. Severity: Medium-High.
P3.3 — Discord/Telegram inbound silent drop. Indicators: outbound works, inbound silently dropped. Root cause: webhook signature failing silently. Catches: upgrade-orch catalog (if listed). Severity: High.
P3.4 — HERMES.md silent billing-routing. Indicators: API bill jumps overnight without usage spike. Root cause: vendor harness-detection bug routing Pro/Max users to API rates. Catches: cost-tracker predict_429_in_window (precursor). Severity: Critical. Real case: $6k overnight, vendor-acknowledged.

4. Configuration / Upgrade (P4.x)

P4.1 — Cron CLI doc/syntax mismatch. Indicators: jobs fail silently or run on wrong schedule. Catches: bash-vet (destructive only). Severity: Medium.
P4.2 — Web search silent fail in --session cron. Indicators: works interactively, fails in cron mode. Root cause: cron context lacks env vars. Catches: silentwatch (canonical fit). Severity: High.
P4.3 — Post-upgrade regression cascade. Indicators: version bump introduces 3+ new bugs. Catches: upgrade-orch (catalog + fingerprint). Severity: High.

5. Skill / Plugin / Supply Chain (P5.x)

P5.1 — Malicious skill payloads. Indicators: prompt injection, encoded payloads, HTTP exfil in community skills. Catches: skill-vetter vet_skill (41 rules). Severity: Critical.
P5.2 — Plugin dependency rot. Indicators: plugin worked at install, broke 3 months later. Catches: skill-vetter dependencies scanner (partial). Severity: Medium.
P5.y — LLM-emitted destructive bash (chain-buried). Indicators: agent proposes cd X && Y && rm -rf $VAR/* && Z; operator approves on quick read. Catches: bash-vet vet_command_chain (30 rules). Severity: Critical. Verbatim source: chiefofautism (158↑/11.5K views), xornullvoid apt remove '*nvidia*' wipeout.
P5.z — Agent-config trust-boundary attack. Indicators: PR adds AGENTS.md / .cursor/rules.md / .git/hooks with prompt-injection / hardcoded webhook / secret-ref. Catches: skill-vetter v1.1 vet_agent_config. Severity: Critical. Real CVEs: Cursor CVE-2026-26268, Gemini CLI CVSS 10.0.

6. LLM-Specific (P6.x)

P6.1 — Keyword-list anti-pattern in scoring/classification. Indicators: KEYWORDS = [...] lists used to gate LLM output. Root cause: semantic-as-lexical translation. Catches: uncovered (methodology / code-review concern). Severity: Critical for scoring code.
P6.2 — Missing error handling around LLM calls. Indicators: no try/except, no retry, no timeout. Catches: output-vetter find_swallowed_exceptions. Severity: High.
P6.3 — Context-window overflow ungraceful. Indicators: sessions grow until model rejects. Catches: uncovered. Severity: Medium-High.
P6.4 — Untracked prompt versions. Indicators: prompts shipped with no versioning. Catches: uncovered (next-MCP candidate). Severity: Medium.
P6.5 — Missing eval harness. Catches: uncovered* (output-vetter is one, philosophical). Severity: High.
P6.y — Output-action divergence. Indicators: agent reports "I did X" but state didn't change. Catches: output-vetter v1.1+ verify_action_outcome (8 ACTION_OUTCOME.* rules). Severity: Critical (when production data involved). Real cases: chiefofautism, Codex sandbox-escalation, 698 CLTR scheming incidents.

7. Observability / Audit (P7.x)

P7.1 — No logs on critical paths. Catches: uncovered (next-MCP candidate). Severity: High.
P7.2 — No cost monitoring per agent / per user. Catches: cost-tracker (full). Severity: Medium-High.
P7.3 — Missing audit trail. Catches: uncovered (next-MCP candidate). Severity: Critical for regulated industries.

8. Security (P8.x)

P8.1 — Gateway exposed on 0.0.0.0 by default. Catches: uncovered (next-MCP candidate: deploy-config-vetter). Severity: Critical.
P8.2 — Secrets in plaintext config. Catches: skill-vetter v1.1 SECRET_REF rules (partial). Severity: Critical.
P8.3 — Session transcript persistence missing redaction. Catches: uncovered. Severity: High.
P8.4 — Prompt injection via untrusted input. Catches: skill-vetter (config) + output-vetter (response) (partial). Severity: High to Critical.

9. Cost / Resource (P9.x)

P9.1 — Wrong model for task (over-paying). Catches: cost-tracker cost-attribution + cheaper-routing. Severity: Medium.
P9.2 — No retry budget. Catches: cost-tracker predict_429_in_window (precursor). Severity: Medium.
P9.x — Cross-model deception (Pass 7 single-source). Indicators: one frontier model rewrites prompt to bypass another model's safety filter. Catches: uncovered. Severity: High to Critical. Source: Nav Toor / CLTR thread (~10K views).

10. Cron / Scheduled (P10.x)

P10.1 — Cron silent failures. Indicators: cron job appears to run but didn't do work. Catches: silentwatch (canonical fit). Severity: High.
P10.2 — Cron job concurrency conflict. Catches: uncovered (silentwatch extension candidate). Severity: Medium-High.

Aggregate coverage

12 ✅ FULL · 11 🟡 PARTIAL · 12 ❌ UNCOVERED. The Aufgaard suite addresses 23 of 35 patterns in MCP-native form. The uncovered 12 cluster into observability-completeness, deploy-config security, runtime monitoring extensions, and cross-model — these are the next-build frontier.

How to cite in conversations

When discussing a failure or making a recommendation:

✅ "This looks like P3.4 (HERMES.md silent billing-routing) — cost-tracker predict_429_in_window catches the precursor."
❌ "There may be a billing issue. You could try monitoring tokens."

Specificity is the value. The buyer didn't pay for "monitor things"; they paid for "you have this exact pattern, here's the exact MCP that catches it, here's what to do."

Coverage caveats — when to disclose

If the user reports a failure that's UNCOVERED in the catalogue, say so plainly: "This looks like P3.1 (WebSocket disconnect lifecycle) — currently uncovered by Aufgaard's MCPs. Recommended manual mitigations: heartbeat ping, proxy-timeout config, idempotency keys on resume." Don't invent coverage we don't have.