Deep LLM fine-tuning operational intuition — LoRA/QLoRA merge pitfalls, preference optimization (DPO/ORPO/SimPO/GRPO), chat-template footguns, completion-only loss masking, sequence packing, eval contamination, toolchain (TRL/axolotl/unsloth/torchtune). Load when post-training open-weights models, picking preference-optimization methods, configuring LoRA/QLoRA, suspecting eval contamination, or choosing a toolchain. Skip for inference/serving (use `llm-inference-serving`), prompt engineering, RAG, or pretraining. Triggers on: "QLoRA", "merge_and_unload", "DPO", "ORPO", "SimPO", "GRPO", "GRPOTrainer", "axolotl", "unsloth", "torchtune", "nf4", "apply_chat_template", "DataCollatorForCompletionOnlyLM", "cu_seqlens", "MMLU contamination", "catastrophic forgetting".
Deep Hetzner operational intuition — Cloud vs Robot vs Storage Box split, server-type taxonomy, region availability, IPv4 cost model, snapshot vs backup vs image, firewall default-deny, hcloud CLI, Floating-IP aliasing, cloud-init, Terraform provider drift. Load when sizing Hetzner resources, migrating to/from Hetzner, automating with hcloud/Terraform, or diagnosing Hetzner-specific gotchas. Skip for generic IaaS, Linux admin, or non-Hetzner clouds. Triggers on: "hcloud cli", "hcloud_server", "cax11", "cpx21", "fsn1", "hetzner robot", "primary ip cost", "without-ipv4", "vswitch", "floating ip hetzner", "storage box".
Deep LLM-inference-serving operational intuition — vLLM, SGLang RadixAttention, TensorRT-LLM tradeoffs; PagedAttention and KV-cache math, chunked prefill, prefix caching, disaggregated serving (DistServe, Mooncake), speculative decoding, quantization, MoE parallelism, TTFT/TPOT. Load when tuning serving-engine knobs, KV-cache sizing, batching design, quantization-format selection, speculative-decoding choice, or MoE layout. Skip for fine-tuning, embeddings serving, or hosted-API integration. Triggers on: "vLLM", "SGLang", "TensorRT-LLM", "PagedAttention", "RadixAttention", "chunked prefill", "prefix caching", "kv-cache-dtype", "EAGLE-3", "Medusa", "MLA", "AWQ", "NVFP4", "DistServe", "Mooncake", "EPLB", "TTFT", "TPOT", "goodput".
Deep RAG application-layer operational intuition — chunking (late chunking, contextual retrieval), embedding-model landscape, reranker economics (ColBERT, ColPali), hybrid retrieval (BM25 + dense + RRF, SPLADE-v3), evaluation (RAGAS, judge bias), grounding, long-context-vs-RAG. Load when designing chunking, embedding-model selection, reranker integration, eval design, hallucination audits, or long-context-vs-RAG sizing. Skip for ANN index choice (use `vector-search`), embedding fine-tuning, or prompt engineering. Pairs with `vector-search`. Triggers on: "late chunking", "contextual retrieval", "MTEB", "BGE-M3", "Matryoshka embeddings", "SPLADE-v3", "RRF", "ColBERTv2", "ColPali", "HyDE", "RAGAS", "groundedness", "lost in the middle", "Self-RAG", "CRAG".
Deep ArgoCD operational intuition — sync waves and hook lifecycle, ApplicationSet generators, Sync vs Health, drift root causes, sync options, AppProject constraints, rollback, multi-cluster. Load for sync ordering, hook design, drift diagnosis, generator selection, rollback planning, or multi-cluster registration. Skip for "what is GitOps", install walkthroughs, or basic Application YAML. Triggers on: "sync wave", "presync hook", "syncfail", "applicationset generator", "matrix generator", "hpa replicas drift", "ignoreDifferences", "ServerSideApply", "Replace=true", "self-heal", "app-of-apps cascade".
Deep bash correctness — `set -euo pipefail` traps, IFS/word-splitting, trap inheritance, process substitution, array gotchas, version-gated features, ShellCheck "looks right but isn't" patterns. Load when writing or hardening non-trivial bash scripts (installers, CI, unattended ops). Skip for basic if/while/for, one-shot interactive commands, or zsh/fish/POSIX-sh work. Triggers on: "set -e not triggering", "pipefail", "trap ERR", "PIPESTATUS", "macOS bash 3.2", "shellcheck SC2086", "harden bash script", "robust bash".
Deep C operational intuition — undefined behavior, sanitizers (ASan/UBSan/TSan/MSan), hardening flags, atomics/memory ordering, fuzzing, ABI quirks. Load when diagnosing UB, configuring sanitizers or hardened builds, investigating data races, or debugging alignment/atomics. Skip for ordinary C coding, malloc/free, or syntax help. Triggers on: "undefined behavior", "strict aliasing", "AddressSanitizer", "UBSan", "Valgrind", "use-after-free", "_FORTIFY_SOURCE", "memory_order", "libFuzzer", "ABI".
Diagnose a failing GitHub Actions run. Fetch the failing job's step logs, correlate with recent commits, and produce a ranked top-3 root-cause hypothesis plus a minimal fix proposal. Use when the user says "CI is failing", "why did the build break", "triage this GHA failure", "workflow failed", "what broke CI", or pastes a `gh run` URL / run ID. Read-only — does NOT apply fixes, restart the run, or tail logs. For workflow-file security audits use gha-security-review; for local test failures use fix-loop.