llm-testbench/EXPANSION.md

Expansion spec — ≥10 tests per subcategory

Orchestrator plan (fable). Authors: sonnet, one per category, isolated worktrees. Verification: opus, after integration. Baseline commit: 26d8c22.

Targets (existing → target)

Category	Subcategory	Have	Target
programming	golang	3	10
programming	python	2	10
programming	typescript	1	10
programming	c	1	10
operations	macos	2	10
operations	linux	2	10
operations	kubernetes	1	10
research	web	1	10
research	whitepapers	1	10
research	codebase	1	10
agents	tool-routing	1	10
agents	planning	1	10
agents	delegation	0	10 (new subcategory)

Total ≥130.

Non-negotiable authoring rules (lessons from two opus reviews + live run)

1. Self-contained prompts. Every fixture (code, logs, excerpts, robots.txt, rosters) inline in the prompt. No external references, no live web.
2. Deterministic. One canonical expected answer — and the evaluator must accept EVERY materially-correct form of it. Before writing the evaluator, list 3 plausible correct phrasings and 2 plausible wrong answers; assert all 5 in the unit test. If you cannot enumerate the correct forms, redesign the test (usually: force a JSON or number-only answer).
3. Answer-format forcing. Structured answers: "Respond with only a JSON object/array, no other text." Numeric: "Respond with only the number." This is the main determinism lever.
4. Never bare NotContains for anything a good answer might mention in negated form ("do not run kubectl edit"). Use the negation-aware pattern from kubernetes.go if needed.
5. Numbers: use eval.Numeric + word-boundary-safe extraction. Beware fixtures whose prose contains numbers (LP64, x86_64, 64-bit) — prompt must demand number-only replies.
6. Ground truth discipline: every constant carries // ground truth: derivation; recompute in the unit test where cheap (call the same function, compute the math, run the code).
7. Omit MaxTokens on all new tests (0 = runtime default; reasoning models burn 3k+ tokens thinking — small caps caused mass truncation zeros in the live run).
8. No Microsoft content. No hello-worlds. Ground tests in the operator's real work: Go vector-search/HNSW engines, golangci strictness (fieldalignment, gosec), semver/ conventional commits, goreleaser, Proxmox/LXC (pct), Talos/ArgoCD/Traefik/CNPG home cluster GitOps, macOS BSD-vs-GNU pain, launchd, 1Password SSH agent, Vue 3 + Vite + Tailwind frontends, Encore, ONNX rerankers, embeddings/recall metrics, agent orchestration.
9. Namespace helpers with a subcategory prefix (k8s…, py…, wp…) — four authors share package tests; collisions break the merge.
10. File ownership: touch ONLY your category's files in internal/tests/ (+ their _test.go). Add new files per subcategory if a file grows past ~600 lines (agents_delegation.go etc.). NEVER touch catalog_test.go (its exact-map assertion is being replaced centrally — expect it to fail; verify your work with go test ./internal/tests/ -run '<YourTests>' and full runs of the other packages).
11. Exec evaluators (GoRun/PyRun/CRun) for tests where correctness = behaviour. Reuse existing harness patterns from golang.go/python.go/c.go. Budget: exec tests are expensive to evaluate — aim for 2-3 per programming subcategory, rest static/numeric/JSON.
12. Register every test in your category's register function; keep IDs kebab-case, prefixed sensibly (go-, py-, ts-, c-, macos-, linux-, k8s-, web-, paper-, code-, agent-).
13. Gates you must pass in your worktree before committing: gofmt -l . empty, go build ./..., go vet ./..., go test ./internal/... ./cmd/... -run '.' green EXCEPT the known catalog_test exact-map failure, golangci-lint run clean, gosec -tests ./... clean.
14. Finish with ONE commit: test(catalog): expand <category> to 10 tests per subcategory. Conventional commit, body in active voice, sentences ≤25 words, no contractions.

Topic menus (author may substitute equivalents of the same difficulty if a topic cannot

be made deterministic; keep counts ≥10 per subcategory)

programming/golang (+7)

channel deadlock spot (inline code, "which line blocks forever, why" → JSON {line, reason-keyword}); defer/recover order trace (exact output); slice append aliasing trace (exact output, GoRun or derived constant); generics constraint fix (broken code inline → fixed code, GoRun); errgroup bounded fan-out (GoRun harness proving limit); method sets / pointer-vs-value receiver interface satisfaction (JSON); context cancellation propagation trace (exact); map iteration order question (contains-based, careful).

programming/python (+8)

softmax of inline vector (numeric, 4dp); asyncio.gather vs sequential await trace (exact); mutable default argument bug (inline code → "what does third call return" exact); dict/set comprehension trace (exact); regex log extraction script (PyRun, inline fixture); pathlib rewrite of os.path code (ContainsAll pathlib idioms + NotContains-style care); generator exhaustion trace (exact); JSON transform script (PyRun); GIL/threading question with single factual answer (JSON); f-string format spec output (exact).

programming/typescript (+9)

discriminated union narrowing trace (exact/JSON); generic utility type authoring (Contains structural checks); Promise.allSettled result shape trace (JSON); Vue 3 ref vs reactive unwrapping behaviour (JSON answer); computed vs watch choice scenario (JSON); ESLint v9 flat-config migration question (ContainsAll: export default, defineConfig/array, no .eslintrc); tsconfig strict family (JSON: which flag catches inline bug); array method chain trace (exact); optional chaining + nullish coalescing trace (exact); type-only import syntax (Contains).

programming/c (+9)

pointer arithmetic trace (numeric); bitmask set/clear/toggle result (numeric); union size LP64 (numeric); string function output (CRun); integer promotion/overflow trace (numeric); array decay sizeof trace (numeric — inline both sizeof forms); macro expansion pitfall (exact); undefined behaviour spot (JSON {line, kind}); struct bitfield packing size (numeric, derivation comment); const pointer vs pointer-to-const (JSON).

operations/macos (+8)

BSD sed -i in-place syntax (Regex accepting -i '' forms); launchctl bootstrap vs load (Contains); plutil/defaults plist read command (Regex); mdfind/Spotlight query (Contains); APFS snapshot create/list commands (Contains); ssh IdentityAgent config for a custom agent socket (Contains — 1Password-style agent); caffeinate usage (Contains); BSD vs GNU stat format flags (Regex/Contains); dscl or security keychain CLI single-answer task (Contains); zsh PATH precedence question (exact).

operations/linux (+8)

journalctl unit+time filter command (Regex); cgroup v2 OOM diagnosis from inline memory.events (JSON); nftables/iptables DNAT one-liner (Regex); LVM extend sequence ordering (JSON array of steps); ssh -L port-forward command from scenario (Regex); cron expression for given schedule (exact); systemd timer vs cron unit pair (ContainsAll OnCalendar, WantedBy=timers.target); /proc/meminfo available-memory interpretation (numeric from inline fixture); rsync archive+delete+exclude command (Regex); find+xargs -print0 pipeline (Regex).

operations/kubernetes (+9)

ImagePullBackOff from inline events (JSON {cause, fix-keyword}); Pending pod with node taints (JSON); PVC Pending with storageclass mismatch (JSON); Service selector mismatch debugging from inline YAML (JSON {field, value}); HPA not scaling — missing metrics-server (Contains); NetworkPolicy blocking DNS (ContainsAll port 53/UDP + label fix); ArgoCD OutOfSync causes from inline app status (JSON); Traefik IngressRoute host rule fix from inline YAML (Regex/JSON); QoS class of inline pod spec (exact: Guaranteed/Burstable/BestEffort); readiness vs liveness misuse scenario (JSON); CNPG-style StatefulSet PDB question (Contains).

research/web (+9)

extract product data from inline HTML to JSON (JSONField set); hreflang reciprocal-link error in inline snippet (JSON); sitemap.xml rules question with single factual answer (exact); HTTP status code semantics for scenarios (JSON mapping); CORS preflight scenario — which header missing (exact); DNS record type for scenario (exact); security headers audit of inline response headers → which missing (JSON array); URL parsing components of inline URL (JSON); canonical vs 301 choice scenario (exact); robots.txt crawl-delay/wildcard evaluation on inline file (JSON).

research/whitepapers (+9)

All excerpt-QA: author writes a technically-correct 120-180 word excerpt inline, asks 2-4 precise extraction/derivation questions, JSON answers. Topics: product quantization (m, nbits, memory math — compute compression ratio numeric); LSM-tree compaction (write amplification derivation numeric); Raft (election timeout ranges, quorum size for N=5 numeric); Bloom filter false-positive rate (numeric from formula with given m,n,k); transformer attention (d_k scaling reason, JSON); CAP theorem partition scenario (exact choice); TF-IDF compute for inline mini-corpus (numeric); B-tree order/fanout height derivation (numeric); recall@k computation from inline result lists (numeric — retrieval evaluation, on-brand); HNSW distance computations counting (numeric).

research/codebase (+9)

trace inline Python function output (exact); trace inline TS function output (exact); bug line in inline diff (JSON {line}); data race spot in inline Go snippet (JSON {variable}); dead code identification (JSON array of function names); big-O of inline function (exact); import cycle detection from inline import lists (JSON array); return-type inference of inline generic function (exact); loop iteration count with off-by-one subtlety (numeric); which commit broke it — git bisect reasoning over inline commit list + symptom (exact commit id).

agents/tool-routing (+9)

Variants of roster+tasks→JSON mapping, escalating difficulty: distractor tools with overlapping descriptions; task needing NO tool (answer from context); multi-step task → ordered tool sequence (JSON array); cheapest-adequate-tool choice (cost field in roster); parallel-vs-sequential tool dispatch decision (JSON); tool with missing required parameter — must answer "cannot" (JSON); ambiguous task → clarify vs act (JSON); tool output → next tool chaining (JSON); read-only vs mutating tool safety choice (JSON).

agents/planning (+9)

Dependency-graph orderings (varying graphs, JSON arrays); critical path length (numeric); which steps parallelize (JSON array of sets — careful: canonicalize ordering in evaluator); rollback trigger placement (JSON); idempotency — which steps safe to retry (JSON array); plan repair after step failure (JSON); milestone/dependency date math (numeric); resource-constrained ordering (JSON); detect the cycle in a task graph (JSON array); minimal replan after requirement change (JSON).

agents/delegation (10, new)

Roster mirrors real specialist agents (code-writer, code-reviewer, test-writer, security-reviewer, devops-debugger, docs-writer, performance-engineer, release-manager, web-researcher, orchestrator). Tasks → which specialist (JSON); which TWO in sequence for build-then-verify (JSON array); what context must the handoff include (JSON array of essentials from an enumerated list); when to verify vs trust (JSON); batch related fixes vs separate dispatches (JSON); main-thread vs delegate decision for a quick edit (JSON); reviewer independence — same agent must not verify its own work (JSON); escalation to human criteria (JSON); parallel dispatch safety — shared-file conflict spot (JSON); minimal-privilege agent choice (JSON).