Files
nucleic/ml/purpose-classifier/data/round2-10.jsonl
T

201 lines
79 KiB
JSON

{"prompt": "the agent stops reporting on some hosts after a kernel upgrade:\n\nlevel=info msg=\"loading eBPF programs\" kernel=6.11.0-19-generic btf=/sys/kernel/btf/vmlinux\nlevel=error msg=\"failed to load program\" prog=tcp_connect err=\"field Sport: can't resolve field: not found\"\nlevel=error msg=\"failed to load program\" prog=tcp_close err=\"field Sport: can't resolve field: not found\"\nlevel=warn msg=\"falling back to procfs polling\" interval=10s\nlevel=info msg=\"agent started\" mode=degraded programs_loaded=3/9\n\nsame binary, same config, works on 6.8 hosts. the struct sock layout changed and we compile with CO-RE, which was supposed to handle exactly this", "purpose": "debugging", "secondary": null, "mixed": false, "difficulty": 0.85, "slice": "pasted-context", "lang": "en"}
{"prompt": "agent CPU spikes to a full core on hosts with many containers and stays there:\n\nperf top -p 41221:\n 38.11% [kernel] [k] bpf_prog_a11c3f2_tcp_connect\n 21.44% agent [.] lumen/agent/proc.(*Resolver).containerFor\n 14.02% agent [.] runtime.mapaccess2_faststr\n 9.88% agent [.] os.ReadFile\n 4.11% [kernel] [k] __d_lookup\n\ncontainerFor reads /proc/<pid>/cgroup on every event and parses it; the host runs 400 containers and about 12,000 events a second", "purpose": "debugging", "secondary": null, "mixed": false, "difficulty": 0.7, "slice": "pasted-context", "lang": "en"}
{"prompt": "our go modules define three `Meta` types that are converted between at every boundary", "purpose": "refactor", "secondary": null, "mixed": false, "difficulty": 0.5, "slice": "boundary", "lang": "en"}
{"prompt": "listing search returns nothing for one region and the logs are unhelpful:\n\nElasticsearch::Transport::Transport::Errors::BadRequest ([400] {\"error\":{\"root_cause\":[{\"type\":\"query_shard_exception\",\"reason\":\"failed to create query: [geo_bounding_box] field [location] is not a geo_point field\"}],\"type\":\"search_phase_execution_exception\",\"phase\":\"query\",\"grouped\":true}})\n app/queries/listing_search.rb:88:in `execute'\n app/controllers/api/v3/listings_controller.rb:22:in `index'\n\nindex mapping for listings-2026-07:\n \"location\": { \"type\": \"object\", \"properties\": { \"lat\": {\"type\":\"float\"}, \"lon\": {\"type\":\"float\"} } }\n\nindex mapping for listings-2026-06:\n \"location\": { \"type\": \"geo_point\" }\n\nthe july index was created by a rollover after someone deleted the index template last month", "purpose": "debugging", "secondary": null, "mixed": false, "difficulty": 0.7, "slice": "pasted-context", "lang": "en"}
{"prompt": "agent's memory grows on busy hosts until the OOM killer takes it:\n\nrss: 180MB → 2.1GB over 4 hours\nbpf map usage:\n connections max_entries=65536 current=65536 (full)\n sockets max_entries=65536 current=65536 (full)\n events (ringbuf) 16MB, consumer lag 14MB\n\ngo heap profile:\n flat flat% cum cum%\n 1.4GB 62.1% 1.4GB 62.1% lumen/agent/enrich.(*Cache).Put\n 0.4GB 17.8% 0.4GB 17.8% encoding/json.Marshal\n\nCache is keyed by (pid, fd) with no eviction, and on this host processes churn thousands of short-lived connections a second", "purpose": "debugging", "secondary": null, "mixed": false, "difficulty": 0.75, "slice": "pasted-context", "lang": "en"}
{"prompt": "agent config across environments, and only prod fills its maps:\n\n# staging\nmap_sizes: { connections: 65536, sockets: 65536 }\nsample_rate: 1.0\nenrich: { container: true, k8s: false, process_tree: false }\ncpu_limit_percent: 10\nhosts: 40\n\n# prod\nmap_sizes: { connections: 65536, sockets: 65536 }\nsample_rate: 1.0\nenrich: { container: true, k8s: true, process_tree: true }\ncpu_limit_percent: 10\nhosts: 40000\n\nstaging hosts run 5 containers each, prod hosts run up to 400, and process_tree is the option that doubles memory", "purpose": "quickFix", "secondary": null, "mixed": false, "difficulty": 0.35, "slice": "pasted-context", "lang": "en"}
{"prompt": "the map screen", "purpose": "frontendImpl", "secondary": null, "mixed": false, "difficulty": 0.5, "slice": "vague-eval", "lang": "en"}
{"prompt": "our API changelog needs an entry for the fixed page size and what partners should do about it", "purpose": "writing", "secondary": null, "mixed": false, "difficulty": 0.4, "slice": "boundary", "lang": "en"}
{"prompt": "saved searches fire notifications for listings that don't match, here's one:\n\nsaved_search 4471:\n bbox: [-0.21, 51.48, -0.09, 51.53]\n min_beds: 2\n max_price_cents: 65000000\n property_type: [\"flat\", \"maisonette\"]\n\nlisting 88412 that triggered it:\n location: { lat: 51.61, lon: -0.19 }\n beds: 2\n price_cents: 62000000\n property_type: \"house\"\n\nthe matcher runs as a percolator query built from the saved search, and both the bbox and the type are wrong here", "purpose": "debugging", "secondary": null, "mixed": false, "difficulty": 0.7, "slice": "pasted-context", "lang": "en"}
{"prompt": "what does the ringbuffer do when full?", "purpose": "review", "secondary": null, "mixed": false, "difficulty": 0.6, "slice": "core", "lang": "en"}
{"prompt": "a customer needs the agent on 5.4 kernels, which we dropped in 2.4 because CO-RE and ring buffers made everything simpler. i'd like an honest view of what supporting them again costs — a compatibility layer, a separate build, or declining the business — and what it means for the next two years of kernel support the deal is worth about a fifth of our ARR, so \"no\" needs to be well argued.", "purpose": "planning", "secondary": null, "mixed": false, "difficulty": 0.85, "slice": "core", "lang": "en"}
{"prompt": "agent's configuration, which customers deploy across their fleets with no documentation beyond a sample file:\n\nagent:\n endpoint: https://ingest.lumen.io\n tenant_key: <secret>\n sample_rate: 1.0 # events per second per process, 0 disables\n programs: [tcp, dns, file, exec]\n map_sizes: { connections: 65536, sockets: 65536 }\n ringbuf_bytes: 16777216\n enrich: { container: true, k8s: true, process_tree: false }\n fallback_procfs: true\n cpu_limit_percent: 10\n\nthings only we know: sample_rate is per process, not per host, so a host with 500 processes is not sampled at all; cpu_limit_percent is advisory and enforced by our own scheduler, not cgroups; map_sizes above 65536 need a kernel with BPF_MAP_TYPE_LRU_HASH support; and turning on process_tree roughly doubles memory", "purpose": "writing", "secondary": null, "mixed": false, "difficulty": 0.65, "slice": "pasted-context", "lang": "en"}
{"prompt": "listing count says \"1 results\"", "purpose": "quickFix", "secondary": null, "mixed": false, "difficulty": 0.1, "slice": "core", "lang": "en"}
{"prompt": "compiler miscompiles one function under -O2 and i've reduced it this far:\n\nerror: internal compiler error: broken MIR in Item DefId(0:412 ~ lumenc[a11c]::opt::fold)\n --> src/opt/fold.rs:141:9\n |\n141| let folded = self.fold_const(expr)?;\n | ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^\n = note: value of type `ConstValue` has no in-memory representation\n\nthread 'rustc' panicked at compiler/rustc_middle/src/ty/consts.rs:88:22:\nassertion failed: !ty.has_infer()\nstack backtrace:\n 0: rust_begin_unwind\n 4: lumenc::opt::fold::ConstFolder::fold_binary\n 5: lumenc::opt::run_passes\n\nonly with our own const-folding pass enabled, and only when the input has a shift by a value we can't prove is in range", "purpose": "debugging", "secondary": null, "mixed": false, "difficulty": 0.85, "slice": "pasted-context", "lang": "en"}
{"prompt": "compiler's incremental builds are wrong about one in fifty times and it's terrifying:\n\n$ lumenc build\n Compiling app v0.4.1\n Finished in 2.1s\n$ ./target/app\nsegmentation fault\n\n$ lumenc build --no-incremental\n Compiling app v0.4.1\n Finished in 41.2s\n$ ./target/app\nok\n\nfingerprint debug output for the changed module:\n src/render.lm mtime=1753843201 size=8412 hash=9c1d0044 deps=[math, gfx]\n cached fingerprint: mtime=1753843201 size=8412 hash=9c1d0044 deps=[math]\n\nthe dependency list changed but the hash didn't, because the hash covers the source text and not the resolved imports", "purpose": "debugging", "secondary": null, "mixed": false, "difficulty": 0.85, "slice": "pasted-context", "lang": "en"}
{"prompt": "three services build the same \"is this listing visible\" condition and they disagree:\n\n# api/v3/listings_controller.rb\nscope.where(status: %w[live under_offer]).where(published_at: ..Time.current)\n\n# jobs/portal_feed_job.rb\nscope.where(status: \"live\").where.not(published_at: nil)\n\n# app/queries/saved_search_matcher.rb\nscope.where(\"status != 'withdrawn'\").where(\"published_at IS NOT NULL\")\n\nthe first includes under-offer listings, the second excludes them, and the third includes drafts that happen to have a published_at from a previous publish", "purpose": "refactor", "secondary": null, "mixed": false, "difficulty": 0.6, "slice": "pasted-context", "lang": "en"}
{"prompt": "pass manager has grown conditionals for every pass we've added. same output, better structure:\n\npub fn run_passes(module: &mut Module, opts: &Opts) {\n if opts.opt_level >= 1 { inline::run(module, opts.inline_threshold); }\n if opts.opt_level >= 1 && !opts.no_const_fold { fold::run(module); }\n if opts.opt_level >= 2 { \n loop_opt::run(module);\n if !opts.no_vectorize && module.target.has_simd() { vectorize::run(module); }\n }\n if opts.debug_assertions { assert_checks::run(module); }\n if opts.opt_level >= 2 && opts.lto { cross_module::run(module); }\n if opts.opt_level >= 1 { dce::run(module); }\n if opts.emit_ir { dump_ir(module); }\n if opts.opt_level >= 2 { fold::run(module); } // second fold, added later, nobody remembers why\n}", "purpose": "refactor", "secondary": null, "mixed": false, "difficulty": 0.65, "slice": "pasted-context", "lang": "en"}
{"prompt": "portal partner's feed specification, which we produce:\n\nnightly full feed plus a delta feed every 15 minutes\n full: gzipped NDJSON, one listing per line, uploaded to their SFTP by 05:00\n delta: same shape, only listings changed since the last delta, plus a `deleted` array of ids\n a listing that is withdrawn must appear in `deleted`, not merely be absent\n price changes must produce a delta entry, which today they do not because price lives in a separate table with its own timestamp\n every entry needs a stable `source_id` that survives a listing being unpublished and republished\n they process deltas in order and will reject a delta whose sequence number is not consecutive\n if we miss a delta window, the next one must include everything since the last successful one", "purpose": "backendImpl", "secondary": null, "mixed": false, "difficulty": 0.75, "slice": "pasted-context", "lang": "en"}
{"prompt": "one visibility scope for listings", "purpose": "refactor", "secondary": null, "mixed": false, "difficulty": 0.5, "slice": "core", "lang": "en"}
{"prompt": "is that transmute actually sound?", "purpose": "review", "secondary": null, "mixed": false, "difficulty": 0.85, "slice": "core", "lang": "en"}
{"prompt": "incremental compilation is wrong about one build in fifty and the cause is that fingerprints hash source text but not the resolved dependency set, which the original RFC explicitly rejected as expensive. i want the fix designed properly — what goes into a fingerprint, how we invalidate the existing caches on upgrade, and how we'd detect a recurrence in CI rather than in a customer's production binary two customers have shipped a binary built this way, which is the part that keeps me up.", "purpose": "planning", "secondary": null, "mixed": false, "difficulty": 0.9, "slice": "core", "lang": "en"}
{"prompt": "i'd like an honest read of whether a plugin can crash the compiler with an out-of-range span, and if so the validation that stops it — with a test that used to crash i'd like enough detail that i can hand it to someone else to finish.", "purpose": "review", "secondary": "backendImpl", "mixed": true, "difficulty": 0.7, "slice": "mixed", "lang": "en"}
{"prompt": "a prospect's security team has asked six pointed questions about what the agent collects, what happens when our endpoint is unreachable, and the blast radius of a bad eBPF program. answer each from the code and write it as a page we can publish, rather than an email that gets forwarded and misquoted they will forward whatever we write to their own security review, so it has to survive being read carefully.", "purpose": "writing", "secondary": null, "mixed": false, "difficulty": 0.7, "slice": "core", "lang": "en"}
{"prompt": "saved searches notify on listings that match neither the bounding box nor the property type, which suggests the percolator query is built from stale saved-search fields. diagnose it, then fix the builder", "purpose": "debugging", "secondary": "backendImpl", "mixed": true, "difficulty": 0.75, "slice": "mixed", "lang": "en"}
{"prompt": "what happens to buffered events when the agent is restarted mid-upload", "purpose": "review", "secondary": null, "mixed": false, "difficulty": 0.6, "slice": "core", "lang": "en"}
{"prompt": "react native app crashes when returning from the camera on android 15 only:\n\nFATAL EXCEPTION: main\nProcess: com.lumen.estate, PID: 8812\njava.lang.RuntimeException: Unable to resume activity {com.lumen.estate/com.lumen.estate.MainActivity}: java.lang.IllegalStateException: Can not perform this action after onSaveInstanceState\n\tat android.app.ActivityThread.performResumeActivity(ActivityThread.java:5412)\n\tat androidx.fragment.app.FragmentManager.checkStateLoss(FragmentManager.java:1882)\n\tat com.lumen.estate.photos.PhotoPickerModule.onActivityResult(PhotoPickerModule.java:141)\n\nonly when the system killed our activity while the camera was open, which android 15 does far more aggressively", "purpose": "debugging", "secondary": null, "mixed": false, "difficulty": 0.7, "slice": "pasted-context", "lang": "en"}
{"prompt": "before this goes in, is the unsafe block justified or are we fooling ourselves?\n\npub fn intern(&self, s: &str) -> Symbol {\n if let Some(&sym) = self.map.borrow().get(s) {\n return sym;\n }\n let owned = s.to_owned();\n let leaked: &'static str = unsafe { std::mem::transmute::<&str, &'static str>(&owned) };\n std::mem::forget(owned);\n let sym = Symbol(self.strings.borrow().len() as u32);\n self.strings.borrow_mut().push(leaked);\n self.map.borrow_mut().insert(leaked, sym);\n sym\n}\n\nthe interner is per-compilation-session, sessions are dropped between builds in the language server, and the language server is long-lived", "purpose": "review", "secondary": null, "mixed": false, "difficulty": 0.85, "slice": "pasted-context", "lang": "en"}
{"prompt": "eBPF program's map handling, which i'd like a second opinion on before we ship it widely:\n\nSEC(\"kprobe/tcp_connect\")\nint BPF_KPROBE(tcp_connect, struct sock *sk) {\n struct conn_key key = {};\n key.pid = bpf_get_current_pid_tgid() >> 32;\n BPF_CORE_READ_INTO(&key.sport, sk, __sk_common.skc_num);\n BPF_CORE_READ_INTO(&key.daddr, sk, __sk_common.skc_daddr);\n\n struct conn_val val = {};\n val.ts = bpf_ktime_get_ns();\n bpf_map_update_elem(&connections, &key, &val, BPF_ANY);\n\n struct event *e = bpf_ringbuf_reserve(&events, sizeof(*e), 0);\n if (!e) return 0;\n e->pid = key.pid;\n bpf_ringbuf_submit(e, 0);\n return 0;\n}\n\nthe connections map is a hash with 65536 entries and nothing deletes from it except a userspace sweep every 30 seconds", "purpose": "review", "secondary": null, "mixed": false, "difficulty": 0.8, "slice": "pasted-context", "lang": "en"}
{"prompt": "por favor revisa la consulta de búsqueda antes de que la pongamos en producción:\n\ndef execute\n Listing.search(\n query: {\n bool: {\n must: [{ match: { description: params[:q] } }],\n filter: [\n { geo_bounding_box: { location: bbox_from(params) } },\n { range: { price_cents: { lte: params[:max_price] } } },\n { terms: { property_type: params[:types] } }\n ]\n }\n },\n size: params.fetch(:size, 50),\n sort: [{ _score: :desc }, { listed_at: :desc }]\n )\nend\n\nparams[:types] llega directamente del cliente sin validar, size no tiene límite superior, y el índice tiene 4 millones de anuncios", "purpose": "review", "secondary": null, "mixed": false, "difficulty": 0.65, "slice": "pasted-context", "lang": "es"}
{"prompt": "photo pipeline, three lambdas that grew separately. worth a read before i extend it:\n\nupload-handler: validates content-type, writes to s3://raw/, publishes to SNS\nthumbnail-worker: sharp().resize(1200).jpeg({quality: 80}).toBuffer() → s3://thumbs/\nfloorplan-worker: detects floorplans by filename heuristic (\"plan\" in the name), applies OCR, writes JSON\n\nnone of them read EXIF; the thumbnail worker strips all metadata; the floorplan heuristic misfires on any listing whose address contains \"plan\"; and a failure in any of them leaves the listing in \"processing\" forever because the state is only advanced on success", "purpose": "review", "secondary": null, "mixed": false, "difficulty": 0.65, "slice": "pasted-context", "lang": "en"}
{"prompt": "mobile app's offline handling for saved listings, which i suspect is why favourites go missing:\n\nconst save = async (listing: Listing) => {\n const current = JSON.parse((await AsyncStorage.getItem('saved')) ?? '[]')\n const next = [...current, listing]\n await AsyncStorage.setItem('saved', JSON.stringify(next))\n try {\n await api.post('/saved', { listingId: listing.id })\n } catch {\n // will sync later\n }\n}\n\nthere is no \"sync later\"; the app also fetches the server list on launch and overwrites AsyncStorage with it", "purpose": "review", "secondary": null, "mixed": false, "difficulty": 0.6, "slice": "pasted-context", "lang": "en"}
{"prompt": "changelog for the agent, from the commits since 2.4:\n\n41c9e0b fix(ebpf): CO-RE relocation for struct sock on 6.11 kernels\n88f21c0 feat(maps): LRU hash for the connections map, evicting instead of filling\nc0aa774 perf(enrich): cache container lookups per cgroup id instead of per event\n2e91b45 fix(sample): sample_rate is now per host, not per process\naa30f19 feat(config): cpu_limit_percent enforced via cgroups where available\n9c1d004 chore: minimum kernel is now 5.10\n4410bb7 fix(ringbuf): drop oldest instead of blocking when the consumer is behind\nb77e910 feat(k8s): pod and namespace enrichment from the kubelet API\n\nour readers are platform engineers deploying this to fleets; two of these change behaviour they've built alerts around", "purpose": "writing", "secondary": null, "mixed": false, "difficulty": 0.5, "slice": "pasted-context", "lang": "en"}
{"prompt": "CI for the compiler takes 40 minutes and this is the whole config:\n\njobs:\n test:\n strategy:\n matrix:\n os: [ubuntu-24.04, macos-15, windows-2022]\n steps:\n - uses: actions/checkout@v4\n - uses: dtolnay/[email protected]\n - run: cargo test --all-features\n - run: cargo test --release --all-features\n - run: ./scripts/run_ui_tests.sh\n - run: cargo bench --no-run\n\nno caching of any kind, the release test run duplicates the debug one for most tests, and the ui test script rebuilds the compiler from scratch", "purpose": "quickFix", "secondary": null, "mixed": false, "difficulty": 0.35, "slice": "pasted-context", "lang": "en"}
{"prompt": "unsere React-Native-Screens holen ihre Daten jeweils selbst, mit drei verschiedenen Mustern:\n\n// ListingScreen.tsx\nuseEffect(() => { api.get(`/listings/${id}`).then(setListing) }, [id])\n\n// SavedScreen.tsx\nconst { data } = useQuery({ queryKey: ['saved'], queryFn: fetchSaved })\n\n// SearchScreen.tsx\nconst [state, dispatch] = useReducer(searchReducer, initial)\nuseEffect(() => { let cancelled = false; search(state.filters).then(r => !cancelled && dispatch({type:'ok', r})); return () => { cancelled = true } }, [state.filters])\n\ndrei Muster, drei Fehlerbehandlungen, und nur eines davon behandelt den Offline-Fall überhaupt", "purpose": "refactor", "secondary": null, "mixed": false, "difficulty": 0.6, "slice": "pasted-context", "lang": "de"}
{"prompt": "design spec for the map search screen, react native, which is our most-used surface:\n\nMap search\n- Map fills the screen; a bottom sheet at three detents holds the result list, snapping to peek / half / full.\n- Clusters above 12 listings show a count; below that, individual pins priced in thousands (\"£625k\").\n- Panning refetches with a 400ms debounce and a \"Search this area\" button rather than automatic refetch — users hated automatic.\n- Selected pin lifts, turns accent, and scrolls the sheet to that listing without changing the detent.\n- Saved listings show a filled heart on the pin; toggling from the sheet updates the pin immediately and reconciles later.\n- Offline: last results and their photos are shown with a banner; the search button is disabled with an explanation.\n- Accessibility: every pin reachable by the list, which is the accessible path; the map itself is marked as decorative.\n- Must hold 60fps while panning with 300 pins on a mid-range android device from 2021.", "purpose": "frontendImpl", "secondary": null, "mixed": false, "difficulty": 0.8, "slice": "pasted-context", "lang": "en"}
{"prompt": "price arrows are colour-only", "purpose": "frontendImpl", "secondary": null, "mixed": false, "difficulty": 0.25, "slice": "boundary", "lang": "en"}
{"prompt": "why is incremental sometimes wrong?", "purpose": "review", "secondary": null, "mixed": false, "difficulty": 0.8, "slice": "boundary", "lang": "en"}
{"prompt": "two of the three compiler engineers leave in six weeks and the remaining list includes a soundness landmine, a correctness bug and a pile of ergonomics work. sequence it with the departures as the main constraint, and tell me plainly which item is the most dangerous thing to leave undocumented rather than merely undone the person joining in october has no context on any of this.", "purpose": "planning", "secondary": null, "mixed": false, "difficulty": 0.85, "slice": "core", "lang": "en"}
{"prompt": "listings stuck in processing generate about forty support calls a week and the fix is always the same re-upload. write the troubleshooting page support can follow, and separately tell me which of the six behaviours in my notes are product bugs that documentation would merely paper over support has no engineering access, so every step has to be something they can do themselves.", "purpose": "writing", "secondary": null, "mixed": false, "difficulty": 0.55, "slice": "boundary", "lang": "en"}
{"prompt": "fleet view has to stay usable when the control plane itself is degraded, which today means it shows a spinner forever. rebuild it to the spec with stale data marked by age, filters in the URL, and bulk actions that say how many hosts they will touch before they do it during an incident this screen is the only thing platform engineers trust, so stale-but-labelled beats empty.", "purpose": "frontendImpl", "secondary": null, "mixed": false, "difficulty": 0.7, "slice": "boundary", "lang": "en"}
{"prompt": "viewing booking service does eight things in one method and only the availability check needs to be transactional. split it, then document the new boundaries and what is guaranteed to have happened by the time the API returns this has come up in three separate reviews now and never gets done.", "purpose": "refactor", "secondary": "writing", "mixed": true, "difficulty": 0.65, "slice": "mixed", "lang": "en"}
{"prompt": "agents stopped loading three eBPF programs after a kernel upgrade despite CO-RE, which is exactly what CO-RE was meant to prevent. work out why the relocation failed, then fix it so the next kernel doesn't do this to us assume whoever picks this up next has no context at all. assume whoever picks it up next has no context beyond what you write.", "purpose": "debugging", "secondary": "backendImpl", "mixed": true, "difficulty": 0.85, "slice": "mixed", "lang": "en"}
{"prompt": "compiler's error messages are inconsistent and i want a house style before we add fifty more. current examples:\n\nerror: type mismatch\nerror: expected `Int`, found `String`\nerror[E0412]: cannot find type `Foo` in this scope\nERROR: unresolved import `std::collectionz`\nerror: the trait bound `T: Display` is not satisfied\nwarning: unused variable x\nnote: `Bar` is defined here but not exported\n\ninconsistent: capitalisation, whether codes are used, whether the primary message names the file, whether we suggest a fix, and whether we use backticks around identifiers\n\nwrite the error message style guide, with rules for wording, structure, spans, notes and suggestions, and rewrite these seven as examples", "purpose": "writing", "secondary": null, "mixed": false, "difficulty": 0.6, "slice": "pasted-context", "lang": "en"}
{"prompt": "compiler team's list, and one of them is leaving in six weeks:\n\n- incremental fingerprints must include the resolved dependency set, which is the correctness bug\n- error messages need a house style and a lint that enforces it\n- the interner's unsafe transmute is a landmine in the language server\n- plugin diagnostics can crash the compiler with an out-of-range span\n- the second const-fold pass exists for a reason nobody remembers and removing it changes output\n- ui tests take 20 of our 40 CI minutes and rebuild the compiler from scratch\n\nsequence these with the departure in mind, and say plainly which one is the highest risk to leave undone", "purpose": "planning", "secondary": null, "mixed": false, "difficulty": 0.85, "slice": "pasted-context", "lang": "en"}
{"prompt": "pass manager as an ordered list", "purpose": "refactor", "secondary": null, "mixed": false, "difficulty": 0.6, "slice": "core", "lang": "en"}
{"prompt": "one cache layer in the agent", "purpose": "refactor", "secondary": null, "mixed": false, "difficulty": 0.6, "slice": "core", "lang": "en"}
{"prompt": "portal feed shows withdrawn listings for hours and misses price changes entirely, because price lives in a separate table with its own timestamp and the feed keys on the listing's updated_at. before we patch it i'd like a view on whether the feed should be event-driven rather than diff-driven, and what that means for partners who process deltas in strict sequence partners process deltas strictly in sequence and reject anything with a gap.", "purpose": "planning", "secondary": null, "mixed": false, "difficulty": 0.85, "slice": "boundary", "lang": "en"}
{"prompt": "incremental fingerprint RFC rejected hashing the resolved dependency set as rarely different, which is exactly the case that has now miscompiled production code twice. read the RFC against the current implementation and tell me which of its other assumptions are similarly optimistic i'd rather find the rest of them now than after the next miscompile. flag anything you'd want to change before doing it rather than after.", "purpose": "review", "secondary": null, "mixed": false, "difficulty": 0.85, "slice": "boundary", "lang": "en"}
{"prompt": "three services each define what makes a listing visible and they disagree about under-offer, drafts with a stale published_at, and withdrawn listings. settle on one definition, apply it in all three, and give me the list of listings whose visibility changes so the content team can be warned before partners notice the portal feed is the one partners see, so it's the definition that matters most externally.", "purpose": "refactor", "secondary": null, "mixed": false, "difficulty": 0.65, "slice": "core", "lang": "en"}
{"prompt": "i'd like to understand what the second const-fold pass changes, since removing it alters output", "purpose": "review", "secondary": null, "mixed": false, "difficulty": 0.75, "slice": "core", "lang": "en"}
{"prompt": "the agent has three caches with three lifetimes, one unbounded, one re-read from procfs on every event, and one that isn't really a cache. consolidate them behind one layer with explicit bounds and eviction, and keep the enrichment output identical for a recorded hour of events from a busy host a recorded hour from the four-hundred-container host is in the fixtures bucket for comparison.", "purpose": "refactor", "secondary": null, "mixed": false, "difficulty": 0.75, "slice": "boundary", "lang": "en"}
{"prompt": "agent degrades after the kernel upgrade", "purpose": "debugging", "secondary": null, "mixed": false, "difficulty": 0.8, "slice": "core", "lang": "en"}
{"prompt": "our health endpoint reports the agent healthy while it's in procfs fallback with three programs unloaded", "purpose": "quickFix", "secondary": null, "mixed": false, "difficulty": 0.35, "slice": "core", "lang": "en"}
{"prompt": "listing thing", "purpose": "debugging", "secondary": null, "mixed": false, "difficulty": 0.5, "slice": "vague-eval", "lang": "en"}
{"prompt": "agent's enrichment cache and the process resolver both key on things that churn:\n\n// enrich/cache.go\ntype Cache struct { m map[connKey]*Meta } // keyed by (pid, fd), never evicted\n\n// proc/resolver.go\nfunc (r *Resolver) containerFor(pid uint32) string {\n b, _ := os.ReadFile(fmt.Sprintf(\"/proc/%d/cgroup\", pid)) // on every event\n return parseCgroup(b)\n}\n\n// k8s/enricher.go\nfunc (e *Enricher) podFor(containerID string) *Pod {\n return e.byContainer[containerID] // refreshed every 30s from the kubelet\n}\n\nthree caches with three lifetimes, one of which is unbounded and one of which isn't a cache at all", "purpose": "refactor", "secondary": null, "mixed": false, "difficulty": 0.65, "slice": "pasted-context", "lang": "en"}
{"prompt": "agent stuff again", "purpose": "debugging", "secondary": null, "mixed": false, "difficulty": 0.55, "slice": "vague-eval", "lang": "en"}
{"prompt": "the listing gallery should preload the next two images while the current one is showing", "purpose": "frontendImpl", "secondary": null, "mixed": false, "difficulty": 0.45, "slice": "core", "lang": "en"}
{"prompt": "our language server and compiler share a crate that assumes one compilation per process", "purpose": "planning", "secondary": null, "mixed": false, "difficulty": 0.8, "slice": "boundary", "lang": "en"}
{"prompt": "is our percolator matcher using the saved search as stored or as it was when created", "purpose": "review", "secondary": null, "mixed": false, "difficulty": 0.6, "slice": "core", "lang": "en"}
{"prompt": "what guarantees does the delta feed make if we miss a fifteen-minute window entirely", "purpose": "review", "secondary": null, "mixed": false, "difficulty": 0.65, "slice": "boundary", "lang": "en"}
{"prompt": "a walkthrough of how a viewing reaches the agent's calendar would help before i touch sync", "purpose": "review", "secondary": null, "mixed": false, "difficulty": 0.55, "slice": "boundary", "lang": "en"}
{"prompt": "agent's sampling is per process at a fixed rate, so a host with five hundred processes is effectively unsampled, and when the ringbuffer fills the kernel silently overwrites exactly the events we most want. work through what adaptive sampling with per-category floors would require, including the awkward fact that customers alert on raw event counts today one customer alerts on events per host per minute and would notice adaptive sampling immediately.", "purpose": "planning", "secondary": null, "mixed": false, "difficulty": 0.9, "slice": "core", "lang": "en"}
{"prompt": "listing page fails seven accessibility items including a carousel that only responds to swipe. fix them, and write the accessibility statement the portal partner asked for", "purpose": "frontendImpl", "secondary": "writing", "mixed": true, "difficulty": 0.7, "slice": "mixed", "lang": "en"}
{"prompt": "saved searches need an edit screen that shows what the search currently matches before saving", "purpose": "frontendImpl", "secondary": null, "mixed": false, "difficulty": 0.55, "slice": "boundary", "lang": "en"}
{"prompt": "les photos des annonces arrivent parfois pivotées de 90 degrés, mais seulement depuis l'app iOS :\n\nfichier reçu : IMG_4471.HEIC, 4032x3024\nEXIF Orientation: 6 (rotate 90 CW)\npipeline de traitement :\n 1. upload direct vers S3 (pas de transformation)\n 2. lambda de vignettes : sharp().resize(1200).jpeg().toBuffer()\n 3. écriture de la vignette, 1200x900\n 4. l'app web affiche la vignette\n\nles vignettes générées perdent l'orientation EXIF et sharp n'est pas configuré avec rotate(); le web affiche donc l'image couchée alors que l'original est correct", "purpose": "debugging", "secondary": null, "mixed": false, "difficulty": 0.6, "slice": "pasted-context", "lang": "fr"}
{"prompt": "booking-viewing flow double-books slots occasionally and the audit trail looks like this:\n\n11:02:14.101 POST /viewings agent=a_881 listing=88412 slot=2026-08-03T14:00Z → 201 viewing v_4471\n11:02:14.118 POST /viewings agent=a_902 listing=88412 slot=2026-08-03T14:00Z → 201 viewing v_4472\n11:02:14.140 slot_availability recalculated for listing 88412: 14:00 marked unavailable\n11:02:14.155 notification sent to vendor: 2 viewings booked at 14:00\n\nthe controller checks availability with a SELECT and then inserts, no unique index on (listing_id, slot), and availability is a materialised view refreshed after the fact", "purpose": "debugging", "secondary": null, "mixed": false, "difficulty": 0.75, "slice": "pasted-context", "lang": "en"}
{"prompt": "availability logic behind viewing bookings, which i inherited:\n\ndef available_slots(listing, from_date, to_date)\n slots = listing.vendor.availability_windows.flat_map { |w| w.slots_between(from_date, to_date) }\n booked = Viewing.where(listing: listing, starts_at: from_date..to_date).pluck(:starts_at)\n blocked = listing.blackouts.where(date: from_date..to_date).flat_map(&:slots)\n (slots - booked - blocked).sort\nend\n\ncalled from the API on every listing page view, from the notification job, and from the agent app's calendar sync; vendors edit their availability windows from a separate screen with no locking", "purpose": "review", "secondary": null, "mixed": false, "difficulty": 0.7, "slice": "pasted-context", "lang": "en"}
{"prompt": "RFC for our incremental compilation fingerprints. does it actually close the hole we hit?\n\n## Fingerprints\nEach module's fingerprint is the hash of its source text. A module is recompiled when its fingerprint changes or when any of its dependencies were recompiled.\n\n## Rationale\nSource text is cheap to hash and captures every change a user can make.\n\n## Known gaps\n- Changes to compiler flags are not captured; users are told to run a clean build.\n- The dependency graph is recorded from the previous build.\n\n## Rejected\nHashing the resolved dependency set was rejected as \"expensive and rarely different\".", "purpose": "review", "secondary": null, "mixed": false, "difficulty": 0.8, "slice": "pasted-context", "lang": "en"}
{"prompt": "query the agent's control plane runs every 30 seconds for every tenant:\n\nSELECT h.id, h.hostname, h.last_seen_at, h.agent_version,\n count(distinct c.id) FILTER (WHERE c.state = 'running') AS containers,\n (SELECT count(*) FROM events e WHERE e.host_id = h.id AND e.ts > now() - interval '5 minutes') AS recent_events,\n (SELECT max(ts) FROM events e2 WHERE e2.host_id = h.id) AS last_event\nFROM hosts h\nLEFT JOIN containers c ON c.host_id = h.id\nWHERE h.tenant_id = $1 AND h.deleted_at IS NULL\nGROUP BY h.id;\n\n40,000 hosts across tenants, events is 2.1 billion rows partitioned by day, and this runs per tenant per 30 seconds", "purpose": "review", "secondary": null, "mixed": false, "difficulty": 0.75, "slice": "pasted-context", "lang": "en"}
{"prompt": "support notes about listings stuck in processing, which need to become a real page:\n\n- a listing goes to \"processing\" when photos are uploaded and leaves it when all workers report success\n- if any worker fails, the listing stays in processing forever and the agent sees no error\n- the most common failure is a HEIC the thumbnail worker can't decode, about 40 a week\n- support fixes it by re-uploading the photo, which starts the pipeline again\n- there's a script that force-advances a listing, which two people know about\n- agents phone in because the listing isn't live and they have a viewing booked\n\nwrite the support-facing troubleshooting page, and separately note which of these are product bugs rather than documentation gaps", "purpose": "writing", "secondary": null, "mixed": false, "difficulty": 0.55, "slice": "pasted-context", "lang": "en"}
{"prompt": "die Notizen aus der Störung von gestern, daraus soll die Kundenmitteilung werden:\n\n14:02 Kunden melden, dass in einer Region keine Suchergebnisse erscheinen\n14:18 bestätigt: alle Abfragen mit Kartenausschnitt schlagen fehl, Textsuche funktioniert\n14:35 Ursache gefunden: der neue Index vom 1. Juli hat kein geo_point-Mapping\n14:50 Index-Template war im Juni versehentlich gelöscht worden\n15:20 Reindexierung gestartet, 4,1 Millionen Anzeigen\n17:40 Reindexierung abgeschlossen, Suche wieder normal\n\nBetroffen war eine Region, dreieinhalb Stunden lang; keine Daten verloren, aber Makler konnten ihre eigenen Anzeigen nicht finden und haben teilweise doppelt eingestellt", "purpose": "writing", "secondary": null, "mixed": false, "difficulty": 0.55, "slice": "pasted-context", "lang": "de"}
{"prompt": "vendor's questions before they sign, which we should answer as documentation rather than an email:\n\n\"What data does the agent collect from our hosts, exactly? Does it read file contents or only metadata? Are command-line arguments captured, and if so can we redact them? What happens to the agent if your ingest endpoint is unreachable for a day — does it buffer, drop, or fill our disk? Can we run it in a mode that collects nothing until we've reviewed the schema? What is the blast radius if one of your eBPF programs has a bug on a production kernel?\"\n\nanswer each from the code and write it as a page we can publish", "purpose": "writing", "secondary": null, "mixed": false, "difficulty": 0.7, "slice": "pasted-context", "lang": "en"}
{"prompt": "marketplace's public API, which three portal partners integrate against off a README from 2024:\n\nGET /api/v3/listings\n filters: bbox, min_price, max_price, min_beds, property_type[], status\n sort: price_asc | price_desc | listed_desc | relevance (default relevance, which needs `q`)\n pagination: cursor-based, `after` opaque, page size 50 fixed\n a listing under offer is returned with status=under_offer unless `status` excludes it\n withdrawn listings disappear entirely, which partners handle badly — they keep showing them\n the price on a listing can change without the listing changing its updated_at, because price lives in a separate table\n\nwrite the reference, and be explicit about the updated_at problem because it's why partners' caches go stale", "purpose": "writing", "secondary": null, "mixed": false, "difficulty": 0.6, "slice": "pasted-context", "lang": "en"}
{"prompt": "runbook for a stuck ingest pipeline is a slack thread. what we actually do:\n\n- symptom is agents reporting fine but events not appearing for a tenant\n- first check is the per-tenant kafka lag dashboard; over 5 minutes of lag means the consumer is behind\n- if lag is flat and high, the enrichment stage is stuck on a slow k8s API; restarting the enricher clears it\n- if lag is climbing steadily, it's usually one tenant sending 100x their normal volume\n- `lumenctl tenant throttle <id> --rate 5000` slows them without dropping, and we tell the account manager afterwards\n- never restart the ingest tier during business hours; it drops the ringbuffer contents on every agent connected to it\n\nwrite the runbook page, in the order someone paged at 3am would need it", "purpose": "writing", "secondary": null, "mixed": false, "difficulty": 0.55, "slice": "pasted-context", "lang": "en"}
{"prompt": "我们的编译器诊断文档只有一页示例,第三方插件作者一直在猜。当前实际行为:\n\n- 诊断分为 error / warning / note 三级,note 不能单独出现,必须挂在前两者下面\n- 错误码格式是 `E0412`,但只有大约三分之一的诊断有码,其余没有,也没有规则说明什么时候该有\n- span 支持多段,但渲染时只显示第一段所在的文件,跨文件的诊断会丢失上下文\n- 建议(suggestion)分为 machine-applicable 和 maybe-incorrect,前者会被 `--fix` 自动应用\n- 插件返回的诊断不做校验,span 越界会导致编译器崩溃而不是报错\n- 同一位置的多条诊断按插件注册顺序输出,没有去重\n\n示例代码:\n\nDiagnostic::error(\"type mismatch\")\n .with_span(span, \"expected `Int`, found `String`\")\n .with_note(\"the function signature is defined here\")\n .emit();\n\n请写成给插件作者的诊断 API 文档,把错误码规则和 span 越界这两点写清楚", "purpose": "writing", "secondary": null, "mixed": false, "difficulty": 0.65, "slice": "pasted-context", "lang": "zh"}
{"prompt": "rubocop and the type checker on the rails app, gate goes on friday:\n\napp/queries/listing_search.rb:88:5: C: Metrics/MethodLength: Method has too many lines. [31/15]\napp/queries/listing_search.rb:41:9: W: Lint/UselessAssignment: Useless assignment to variable - `sort`\napp/models/listing.rb:212:11: C: Style/SafeNavigation: Use safe navigation (&.) instead of checking if an object exists\napp/services/viewings/book.rb:22:3: C: Style/Documentation: Missing top-level class documentation comment\n\nsorbet:\napp/services/viewings/book.rb:66: Expected `Time` but found `T.nilable(Time)` for argument `starts_at`\napp/models/listing.rb:141: Method `price_cents` does not exist on `T.nilable(Price)`\n\n4 offenses, 2 type errors, and the two type errors look like actual nil bugs", "purpose": "quickFix", "secondary": null, "mixed": false, "difficulty": 0.35, "slice": "pasted-context", "lang": "en"}
{"prompt": "clippy on the compiler, mostly noise but check the last two:\n\nwarning: this `RefCell` is borrowed twice in the same expression\n --> src/intern.rs:41:9\nwarning: large enum variant\n --> src/ast.rs:88:1\n | enum Expr { Lit(i64), Call(Box<Call>), Block([Stmt; 64]) }\nwarning: this loop could be written as a `for` loop\n --> src/lex.rs:141:5\nwarning: casting `usize` to `u32` may truncate the value\n --> src/intern.rs:52:20\nwarning: `mem::forget` on a type with a `Drop` implementation\n --> src/intern.rs:49:5\n\n34 warnings total", "purpose": "quickFix", "secondary": null, "mixed": false, "difficulty": 0.4, "slice": "pasted-context", "lang": "en"}
{"prompt": "index template that got deleted, restored from a backup. does this match what the search code expects?\n\nPUT _index_template/listings\n{\n \"index_patterns\": [\"listings-*\"],\n \"template\": {\n \"mappings\": {\n \"properties\": {\n \"location\": { \"type\": \"geo_point\" },\n \"price_cents\": { \"type\": \"long\" },\n \"beds\": { \"type\": \"short\" },\n \"property_type\": { \"type\": \"keyword\" },\n \"description\": { \"type\": \"text\", \"analyzer\": \"english\" },\n \"listed_at\": { \"type\": \"date\" }\n }\n },\n \"settings\": { \"number_of_shards\": 3, \"number_of_replicas\": 1 }\n }\n}\n\nthe search code also filters on `status` and sorts on `updated_at`, neither of which appears here", "purpose": "quickFix", "secondary": null, "mixed": false, "difficulty": 0.3, "slice": "pasted-context", "lang": "en"}
{"prompt": "npm audit on the mobile app before the store submission:\n\nreact-native 0.76.0 - 0.76.9\nHigh: Improper URI validation in the Linking module\nfix available via `npm audit fix --force`\nWill install [email protected], which is a breaking change\n\n@react-native-async-storage/async-storage 1.21.0 - 1.23.1\nModerate: Data written without atomic replace, corruption possible on crash\nfix available via `npm audit fix`\n\nreact-native-image-picker 7.1.0 - 7.2.2\nModerate: Activity result handling can leak the file URI to other apps\nfix available via `npm audit fix`\n\nthe async-storage one is interesting given our saved-listings bug", "purpose": "quickFix", "secondary": null, "mixed": false, "difficulty": 0.4, "slice": "pasted-context", "lang": "en"}
{"prompt": "os limites de alerta do ingest, acordam-nos por nada:\n\n- alert: IngestLag\n expr: kafka_consumergroup_lag{group=\"ingest\"} > 1000\n for: 1m\n labels: { severity: page }\n\n- alert: AgentDown\n expr: up{job=\"agent\"} == 0\n for: 0m\n labels: { severity: page }\n\n- alert: EnrichErrors\n expr: rate(enrich_errors_total[5m]) > 0\n for: 1m\n labels: { severity: page }\n\ncomportamento normal: o lag passa de 1000 em todas as manhãs às 9h durante cerca de dez minutos; há sempre alguns agentes offline num universo de 40 mil hosts; e o incidente real do mês passado (uma região inteira sem eventos durante duas horas) não gerou nenhum alerta", "purpose": "quickFix", "secondary": null, "mixed": false, "difficulty": 0.4, "slice": "pasted-context", "lang": "pt"}
{"prompt": "diagnostic emission code is duplicated across passes with slight differences:\n\n// in type check\nsess.emit_error(format!(\"type mismatch: expected {}, found {}\", a, b), span);\n\n// in the borrow checker\nsess.diagnostic(Level::Error, \"cannot borrow as mutable\")\n .span_label(span, \"second mutable borrow\")\n .emit();\n\n// in the const folder\neprintln!(\"error: {}\", msg); // yes, really\nsess.error_count.fetch_add(1, Ordering::Relaxed);\n\n// in the plugin bridge\nDiagnostic::error(msg).with_span(span, label).emit();\n\nfour ways to report an error, one of which bypasses the diagnostic system entirely and breaks --json output", "purpose": "refactor", "secondary": null, "mixed": false, "difficulty": 0.6, "slice": "pasted-context", "lang": "en"}
{"prompt": "viewing-booking service does everything in one method and the tests need a database and a stubbed calendar:\n\nclass Viewings::Book\n def call\n raise SlotUnavailable unless available?\n viewing = Viewing.create!(listing:, agent:, starts_at:, source:)\n CalendarSync.push(agent, viewing)\n VendorMailer.viewing_booked(viewing).deliver_later\n BuyerMailer.viewing_confirmed(viewing).deliver_later\n Sms.send(buyer.phone, confirmation_text(viewing)) if buyer.phone?\n listing.touch(:last_activity_at)\n Analytics.track(\"viewing_booked\", agent_id: agent.id, listing_id: listing.id)\n AvailabilityCache.invalidate(listing)\n viewing\n end\nend\n\nsame behaviour, but the availability check and the booking need to be atomic and the rest needs to be out of the request", "purpose": "refactor", "secondary": null, "mixed": false, "difficulty": 0.65, "slice": "pasted-context", "lang": "en"}
{"prompt": "quarter planning, and this is what i have to sequence:\n\n- the double-booking bug is small in volume but every instance is a furious vendor and an angry agent\n- portal partners are complaining that our feed shows withdrawn listings for hours\n- the agent's memory growth means one large customer caps us at 200 hosts per cluster\n- the compiler's incremental correctness bug has produced two miscompiles in production code this quarter\n- a new customer needs the agent on 5.4 kernels, which we dropped support for in 2.4\n- two engineers are leaving the compiler team and one is joining in october\n- there's a store submission deadline for the mobile app in three weeks", "purpose": "planning", "secondary": null, "mixed": false, "difficulty": 0.9, "slice": "pasted-context", "lang": "en"}
{"prompt": "architecture ticket, thinking needed before code:\n\nOBS-410 — Agent event sampling and backpressure\nThe agent currently samples per process at a fixed rate and drops nothing until the ringbuffer fills, at which point the kernel side silently overwrites. On busy hosts this means we lose the events we most want. The proposal is adaptive sampling driven by observed volume with per-category floors, plus explicit backpressure signalling to the control plane. Concerns: sampling decisions in eBPF cost cycles per event; the control plane cannot push config faster than every 30 seconds; customers alert on event counts and adaptive sampling would make those alerts meaningless; and we have no way to tell a customer what we dropped.", "purpose": "planning", "secondary": null, "mixed": false, "difficulty": 0.9, "slice": "pasted-context", "lang": "en"}
{"prompt": "customer's security requirement for the agent, which sales has already half-agreed to:\n\n\"The agent shall not transmit file contents, command-line arguments or environment variables outside the host without explicit per-category opt-in. It shall operate in an audit-only mode in which it collects and displays locally but transmits nothing. The customer shall be able to review the exact schema of transmitted data. A failure of the agent shall not affect the host: no kernel panic, no CPU starvation, no disk exhaustion. Evidence of the last requirement shall include the results of fault injection testing.\"\n\nwe transmit command lines by default, have no local-only mode, no published schema, and have never done fault injection. i want the plan and an honest view of what we can claim by the deadline", "purpose": "planning", "secondary": null, "mixed": false, "difficulty": 0.9, "slice": "pasted-context", "lang": "en"}
{"prompt": "spec for the agent fleet screen in our console, which platform engineers live in:\n\nFleet view\n- Table of hosts: hostname, agent version, kernel, last seen, events/sec, degraded reason. Virtualised to 40,000 rows.\n- Degraded hosts sort first by default; \"degraded\" needs a plain-language reason, not an error code.\n- Filters: version, kernel, tenant, degraded state; filters are in the URL so they can be shared in an incident.\n- Bulk actions: restart agent, change sample rate, collect diagnostics. Each shows how many hosts it will affect before confirming.\n- A version rollout view: percentage on each version over time, with the ability to pause a rollout.\n- Host detail: loaded programs with their status, map utilisation, recent config changes, and the last ten errors.\n- Everything must remain usable when the control plane is degraded — show stale data with its age rather than a spinner.", "purpose": "frontendImpl", "secondary": null, "mixed": false, "difficulty": 0.75, "slice": "pasted-context", "lang": "en"}
{"prompt": "accessibility findings on the listing detail page, from an audit a portal partner ran:\n\n1. The photo carousel is operable only by swipe; there are no next/previous controls and arrow keys do nothing.\n2. Photo alt text is the filename (\"IMG_4471.HEIC\") for every image.\n3. The floorplan opens in a lightbox that traps focus and has no close button reachable by keyboard.\n4. Price change history is conveyed by red and green arrows with no text.\n5. The \"book a viewing\" form's date picker is a custom control with no role and no keyboard support.\n6. Headings jump from h1 to h4 in the description section.\n7. The map embed has no accessible name and is included in the tab order with nothing to do.", "purpose": "frontendImpl", "secondary": null, "mixed": false, "difficulty": 0.7, "slice": "pasted-context", "lang": "en"}
{"prompt": "design tokens versus what the listing pages actually use:\n\ntokens:\n color.text.primary #101418 / #F2F5F7\n color.text.muted #5B6570 / #9AA6B2\n color.brand #0B6E4F\n color.price.up #B42318\n color.price.down #067647\n space 4/8/12/16/24/32/48, radius 6/10/16, shadow.sm/md/lg\n type: display 28/34, title 20/26, body 15/22, caption 13/18\n\nthe listing pages: nine hardcoded colours, three of which are the brand green at slightly different values; paddings of 5, 11, 13 and 22; two type sizes not in the scale; price direction shown only by colour; and a shadow defined inline in four components\n\nbring it onto the tokens and give price direction a non-colour indicator while you're in there", "purpose": "frontendImpl", "secondary": null, "mixed": false, "difficulty": 0.55, "slice": "pasted-context", "lang": "en"}
{"prompt": "schema we agreed for agent config versioning, needs implementing in the control plane:\n\nCREATE TABLE agent_configs (\n id uuid PRIMARY KEY,\n tenant_id uuid NOT NULL,\n scope text NOT NULL CHECK (scope IN ('tenant','group','host')),\n scope_id text,\n body jsonb NOT NULL,\n version int NOT NULL,\n created_by text NOT NULL,\n created_at timestamptz NOT NULL DEFAULT now(),\n UNIQUE (tenant_id, scope, scope_id, version)\n);\n\nagents poll every 30 seconds with their current version; the effective config is host over group over tenant; a rollout can be paused, which means agents keep their current version rather than reverting; and an agent that receives a config it can't apply must report why and keep running on the previous one", "purpose": "backendImpl", "secondary": null, "mixed": false, "difficulty": 0.8, "slice": "pasted-context", "lang": "en"}
{"prompt": "sample_rate should be per host", "purpose": "quickFix", "secondary": null, "mixed": false, "difficulty": 0.25, "slice": "core", "lang": "en"}
{"prompt": "async-storage bump before submission", "purpose": "quickFix", "secondary": null, "mixed": false, "difficulty": 0.2, "slice": "core", "lang": "en"}
{"prompt": "cache cargo registry in CI", "purpose": "quickFix", "secondary": null, "mixed": false, "difficulty": 0.25, "slice": "core", "lang": "en"}
{"prompt": "thumbnails need rotate() for EXIF", "purpose": "quickFix", "secondary": null, "mixed": false, "difficulty": 0.25, "slice": "core", "lang": "en"}
{"prompt": "el filtro de precio no acepta decimales", "purpose": "quickFix", "secondary": null, "mixed": false, "difficulty": 0.2, "slice": "core", "lang": "es"}
{"prompt": "cap the search page size at 100", "purpose": "quickFix", "secondary": null, "mixed": false, "difficulty": 0.25, "slice": "boundary", "lang": "en"}
{"prompt": "add `status` to the index template", "purpose": "quickFix", "secondary": null, "mixed": false, "difficulty": 0.25, "slice": "boundary", "lang": "en"}
{"prompt": "process_tree off by default", "purpose": "quickFix", "secondary": null, "mixed": false, "difficulty": 0.3, "slice": "boundary", "lang": "en"}
{"prompt": "IngestLag needs a 15 minute window", "purpose": "quickFix", "secondary": null, "mixed": false, "difficulty": 0.25, "slice": "boundary", "lang": "en"}
{"prompt": "Release-Tests im CI weglassen", "purpose": "quickFix", "secondary": null, "mixed": false, "difficulty": 0.3, "slice": "boundary", "lang": "de"}
{"prompt": "unique index on (listing_id, slot)", "purpose": "quickFix", "secondary": null, "mixed": false, "difficulty": 0.35, "slice": "boundary", "lang": "en"}
{"prompt": "the compiler's error output needs colour that survives being piped into a file", "purpose": "frontendImpl", "secondary": null, "mixed": false, "difficulty": 0.4, "slice": "boundary", "lang": "en"}
{"prompt": "agent config is parsed in three places with different defaults for the same keys", "purpose": "refactor", "secondary": null, "mixed": false, "difficulty": 0.55, "slice": "core", "lang": "en"}
{"prompt": "our internal wiki page on the photo pipeline describes two lambdas and there are three, one of which advances the listing state and is therefore the one everyone needs to understand during an incident. write the page properly, covering each stage, what failure looks like, and which state transitions are irreversible tell me if any of that is a bad idea before doing it.", "purpose": "writing", "secondary": null, "mixed": false, "difficulty": 0.55, "slice": "boundary", "lang": "en"}
{"prompt": "pins should show price in thousands", "purpose": "frontendImpl", "secondary": null, "mixed": false, "difficulty": 0.3, "slice": "core", "lang": "en"}
{"prompt": "\"search this area\" button on the map", "purpose": "frontendImpl", "secondary": null, "mixed": false, "difficulty": 0.4, "slice": "core", "lang": "en"}
{"prompt": "degraded hosts should sort first", "purpose": "frontendImpl", "secondary": null, "mixed": false, "difficulty": 0.3, "slice": "core", "lang": "en"}
{"prompt": "filters belong in the URL", "purpose": "frontendImpl", "secondary": null, "mixed": false, "difficulty": 0.35, "slice": "core", "lang": "en"}
{"prompt": "photo alt text is the filename", "purpose": "frontendImpl", "secondary": null, "mixed": false, "difficulty": 0.25, "slice": "core", "lang": "en"}
{"prompt": "間取り図のライトボックスが閉じられません", "purpose": "frontendImpl", "secondary": null, "mixed": false, "difficulty": 0.35, "slice": "core", "lang": "ja"}
{"prompt": "carousel has no keyboard controls", "purpose": "frontendImpl", "secondary": null, "mixed": false, "difficulty": 0.35, "slice": "boundary", "lang": "en"}
{"prompt": "bottom sheet snaps to the wrong detent", "purpose": "frontendImpl", "secondary": null, "mixed": false, "difficulty": 0.4, "slice": "boundary", "lang": "en"}
{"prompt": "stale fleet data looks live", "purpose": "frontendImpl", "secondary": null, "mixed": false, "difficulty": 0.35, "slice": "boundary", "lang": "en"}
{"prompt": "pull the emails out of Viewings::Book", "purpose": "refactor", "secondary": null, "mixed": false, "difficulty": 0.5, "slice": "core", "lang": "en"}
{"prompt": "un seul modèle de récupération de données", "purpose": "refactor", "secondary": null, "mixed": false, "difficulty": 0.55, "slice": "core", "lang": "fr"}
{"prompt": "`price_cents` naming, everywhere", "purpose": "refactor", "secondary": null, "mixed": false, "difficulty": 0.3, "slice": "boundary", "lang": "en"}
{"prompt": "extract the bbox parsing helper", "purpose": "refactor", "secondary": null, "mixed": false, "difficulty": 0.35, "slice": "boundary", "lang": "en"}
{"prompt": "inline `slots_between`, one caller", "purpose": "refactor", "secondary": null, "mixed": false, "difficulty": 0.2, "slice": "boundary", "lang": "en"}
{"prompt": "one diagnostic emission path", "purpose": "refactor", "secondary": null, "mixed": false, "difficulty": 0.55, "slice": "boundary", "lang": "en"}
{"prompt": "doc comments on the plugin diagnostic API", "purpose": "writing", "secondary": null, "mixed": false, "difficulty": 0.4, "slice": "core", "lang": "en"}
{"prompt": "changelog for agent 2.5", "purpose": "writing", "secondary": null, "mixed": false, "difficulty": 0.3, "slice": "core", "lang": "en"}
{"prompt": "nota para os parceiros sobre o feed", "purpose": "writing", "secondary": null, "mixed": false, "difficulty": 0.4, "slice": "core", "lang": "pt"}
{"prompt": "document what sample_rate really means", "purpose": "writing", "secondary": null, "mixed": false, "difficulty": 0.4, "slice": "boundary", "lang": "en"}
{"prompt": "summarise the sampling proposal", "purpose": "writing", "secondary": null, "mixed": false, "difficulty": 0.45, "slice": "boundary", "lang": "en"}
{"prompt": "PR body for the fingerprint fix", "purpose": "writing", "secondary": null, "mixed": false, "difficulty": 0.3, "slice": "boundary", "lang": "en"}
{"prompt": "¿por qué el feed muestra anuncios retirados?", "purpose": "review", "secondary": null, "mixed": false, "difficulty": 0.6, "slice": "core", "lang": "es"}
{"prompt": "can two agents book the same slot?", "purpose": "review", "secondary": null, "mixed": false, "difficulty": 0.7, "slice": "boundary", "lang": "en"}
{"prompt": "walk me through the photo pipeline", "purpose": "review", "secondary": null, "mixed": false, "difficulty": 0.5, "slice": "boundary", "lang": "en"}
{"prompt": "geo search fails in one region", "purpose": "debugging", "secondary": null, "mixed": false, "difficulty": 0.65, "slice": "core", "lang": "en"}
{"prompt": "favourites disappear after relaunch", "purpose": "debugging", "secondary": null, "mixed": false, "difficulty": 0.6, "slice": "boundary", "lang": "en"}
{"prompt": "warum belegt der Agent 2 GB RAM?", "purpose": "debugging", "secondary": null, "mixed": false, "difficulty": 0.7, "slice": "boundary", "lang": "de"}
{"prompt": "endpoint for a host's loaded programs", "purpose": "backendImpl", "secondary": null, "mixed": false, "difficulty": 0.45, "slice": "core", "lang": "en"}
{"prompt": "press on", "purpose": "backendImpl", "secondary": null, "mixed": false, "difficulty": 0.5, "slice": "vague-eval", "lang": "en"}
{"prompt": "leaner", "purpose": "refactor", "secondary": null, "mixed": false, "difficulty": 0.5, "slice": "vague-eval", "lang": "en"}
{"prompt": "you know the priorities", "purpose": "planning", "secondary": null, "mixed": false, "difficulty": 0.55, "slice": "vague-eval", "lang": "en"}
{"prompt": "lo del compilador, sigue", "purpose": "debugging", "secondary": null, "mixed": false, "difficulty": 0.6, "slice": "vague-eval", "lang": "es"}
{"prompt": "clean as you go", "purpose": "refactor", "secondary": null, "mixed": false, "difficulty": 0.4, "slice": "vague-eval", "lang": "en"}
{"prompt": "same again", "purpose": "backendImpl", "secondary": null, "mixed": false, "difficulty": 0.5, "slice": "vague-eval", "lang": "en"}
{"prompt": "partner doc", "purpose": "writing", "secondary": null, "mixed": false, "difficulty": 0.5, "slice": "vague-eval", "lang": "en"}
{"prompt": "anything small", "purpose": "quickFix", "secondary": null, "mixed": false, "difficulty": 0.4, "slice": "vague-eval", "lang": "en"}
{"prompt": "cast an eye over it", "purpose": "review", "secondary": null, "mixed": false, "difficulty": 0.45, "slice": "vague-eval", "lang": "en"}
{"prompt": "優先順位はお任せで", "purpose": "planning", "secondary": null, "mixed": false, "difficulty": 0.5, "slice": "vague-eval", "lang": "ja"}
{"prompt": "more of that", "purpose": "backendImpl", "secondary": null, "mixed": false, "difficulty": 0.5, "slice": "vague-eval", "lang": "en"}
{"prompt": "mobile app's offline story is three different patterns and a hope, and the saved-listings bug is just the visible part. i want a position on what offline should mean for this app — which surfaces work, what happens to writes, how conflicts resolve — before anyone touches the storage layer again the store submission is in three weeks, so anything shipped now has to be small.", "purpose": "planning", "secondary": null, "mixed": false, "difficulty": 0.8, "slice": "boundary", "lang": "en"}
{"prompt": "compiler's diagnostics are inconsistent in seven visible ways and plugin authors are copying whichever example they found first. write the error message style guide — wording, capitalisation, when a code is required, how spans and notes are used, when to suggest a fix — with our current worst examples rewritten as illustrations the guide should be enforceable by a lint, so wording rules need to be mechanical where possible.", "purpose": "writing", "secondary": null, "mixed": false, "difficulty": 0.6, "slice": "core", "lang": "en"}
{"prompt": "customers deploy the agent across whole fleets from a sample config file with no documentation, which is how one of them ran unsampled for a month. write the configuration reference covering what each option really does, the ones that interact badly, and the three that mean something different from what their names suggest assume the reader is a platform engineer rolling this to forty thousand hosts.", "purpose": "writing", "secondary": null, "mixed": false, "difficulty": 0.65, "slice": "core", "lang": "en"}
{"prompt": "public listings API is documented by a README from 2024 and the thing partners get wrong every time is that a price change doesn't move updated_at. write the reference properly, with the caching implications spelled out, and a worked example of a partner keeping a mirror in sync three partners integrate against it today and a fourth starts next month.", "purpose": "writing", "secondary": null, "mixed": false, "difficulty": 0.6, "slice": "boundary", "lang": "en"}
{"prompt": "string interner leaks every string deliberately and transmutes a borrow to 'static, which is fine for a one-shot compiler and possibly catastrophic in a long-lived language server. work through whether the sessions really are isolated, what happens when one is dropped, and whether the language server can end up with a dangling symbol the language server runs for days in an editor, which is where this would show up.", "purpose": "review", "secondary": null, "mixed": false, "difficulty": 0.9, "slice": "core", "lang": "en"}
{"prompt": "eBPF program updates a hash map on every connection and relies on a userspace sweep every thirty seconds to remove entries, which on a busy host cannot possibly keep up. read the program and the sweep together and tell me what happens once the map is full — silently dropped updates, or something worse on our busiest customer host that's about twelve thousand events a second.", "purpose": "review", "secondary": null, "mixed": false, "difficulty": 0.85, "slice": "boundary", "lang": "en"}
{"prompt": "pass manager is a wall of conditionals including a second const-fold pass nobody can explain, and removing it changes generated code. restructure it into an explicit ordered pipeline with each pass declaring its own preconditions, and keep the emitted binary identical for our benchmark suite before we discuss removing anything our benchmark suite is the only thing standing between us and silently changing generated code.", "purpose": "refactor", "secondary": null, "mixed": false, "difficulty": 0.8, "slice": "core", "lang": "en"}
{"prompt": "two agents booked the same viewing slot seventeen milliseconds apart and our availability check is a select followed by an insert with no constraint behind it. work out how often this has happened historically before we add the index, because the vendor-facing consequences of the ones we've already sent are the actual problem vendors get an email per booking, so the ones already sent are what we actually have to handle.", "purpose": "debugging", "secondary": null, "mixed": false, "difficulty": 0.75, "slice": "core", "lang": "en"}
{"prompt": "map screen is our most-used surface and it currently refetches on every pan, drops frames with three hundred pins, and has no offline behaviour at all. build it to the new spec, keeping the sheet's three detents and making sure the accessible path through the list actually works the target device is a mid-range android from 2021, which is what most of our users have.", "purpose": "frontendImpl", "secondary": null, "mixed": false, "difficulty": 0.8, "slice": "core", "lang": "en"}
{"prompt": "before we change sampling i want the design written down — adaptive rates, per-category floors, what we tell customers about what we dropped — and then the drop accounting implemented, because we currently can't answer that question at all that's the whole of it, but shout if the context is thin.", "purpose": "planning", "secondary": "backendImpl", "mixed": true, "difficulty": 0.9, "slice": "mixed", "lang": "en"}
{"prompt": "incremental correctness bug needs a proper fix and a regression test that would have caught it. design the fingerprint change with me first, including the cache invalidation on upgrade, then implement it we've been burned by guessing at this before, so evidence over instinct please.", "purpose": "planning", "secondary": "backendImpl", "mixed": true, "difficulty": 0.9, "slice": "mixed", "lang": "en"}
{"prompt": "portal feed should probably be event-driven rather than diffing on updated_at, but that's a bigger change than this quarter allows. give me the target design, then implement the price-change events so partners at least stop showing stale prices i'd rather have the reasoning than a quick answer here. there's no rush on this week specifically, but it keeps costing us time.", "purpose": "planning", "secondary": "backendImpl", "mixed": true, "difficulty": 0.85, "slice": "mixed", "lang": "en"}
{"prompt": "unsere Offline-Strategie in der App ist dreimal unterschiedlich gebaut und verliert dabei Favoriten. Ich hätte gern zuerst ein Konzept, was offline überhaupt funktionieren soll, und danach die Umsetzung für die gespeicherten Anzeigen", "purpose": "planning", "secondary": "frontendImpl", "mixed": true, "difficulty": 0.75, "slice": "mixed", "lang": "de"}
{"prompt": "supporting 5.4 kernels again needs a decision and, if we say yes, a compatibility path that doesn't fork the codebase. work through the options, then prototype whichever one you'd recommend against a 5.4 test host happy to be told this is the wrong shape entirely. nobody has trusted this code for about a year, which is part of the problem.", "purpose": "planning", "secondary": "backendImpl", "mixed": true, "difficulty": 0.85, "slice": "mixed", "lang": "en"}
{"prompt": "agent's configuration reference needs writing and i suspect at least two options don't do what their names say. produce the reference, and give me the list of options whose behaviour we should change rather than document i've already spent an afternoon on it and got nowhere useful.", "purpose": "writing", "secondary": "review", "mixed": true, "difficulty": 0.65, "slice": "mixed", "lang": "en"}
{"prompt": "listings API reference has to exist before the fourth partner integrates, and while writing it please confirm whether withdrawn listings really do vanish from the feed rather than appearing as deleted", "purpose": "writing", "secondary": "review", "mixed": true, "difficulty": 0.6, "slice": "mixed", "lang": "en"}
{"prompt": "escribe la guía de despliegue del agente para clientes que lo instalan en toda su flota, y comprueba en el código si el límite de CPU se aplica de verdad o solo es orientativo", "purpose": "writing", "secondary": "review", "mixed": true, "difficulty": 0.65, "slice": "mixed", "lang": "es"}
{"prompt": "ingest runbook should be a page rather than a slack thread, and the ingest tier restart that drops every connected agent's ringbuffer deserves a guard. write the runbook, then add the confirmation if the answer is that it's fine as it is, that's a useful answer too.", "purpose": "writing", "secondary": "quickFix", "mixed": true, "difficulty": 0.6, "slice": "mixed", "lang": "en"}
{"prompt": "compiler reports errors four different ways, one of which bypasses the diagnostic system and breaks JSON output entirely. unify them, and tell me which existing tools would break if the JSON output suddenly contained diagnostics it never saw before it doesn't have to be elegant, it has to be defensible in a review.", "purpose": "refactor", "secondary": "review", "mixed": true, "difficulty": 0.65, "slice": "mixed", "lang": "en"}
{"prompt": "react native screens fetch data three different ways with three error behaviours. bring them onto one pattern, and write the short note for the team explaining which one and why, because this is the third time we've had this conversation it's been like this long enough that nobody trusts it any more. the last person who touched this left, so there's nobody to ask.", "purpose": "refactor", "secondary": "writing", "mixed": true, "difficulty": 0.6, "slice": "mixed", "lang": "en"}
{"prompt": "photos come out rotated from iOS uploads and i think the thumbnail worker drops EXIF orientation. confirm it, then fix the pipeline and tell me how many existing thumbnails need regenerating", "purpose": "debugging", "secondary": "quickFix", "mixed": true, "difficulty": 0.6, "slice": "mixed", "lang": "en"}
{"prompt": "fleet view needs a rollout pause control, and we should agree what pausing means for agents mid-download before it exists. decide that with me, then build it", "purpose": "frontendImpl", "secondary": "planning", "mixed": true, "difficulty": 0.65, "slice": "mixed", "lang": "en"}
{"prompt": "CI duplicates the debug test run in release mode and rebuilds the compiler for the UI tests. cut it down, and tell me what coverage we lose by doing so", "purpose": "quickFix", "secondary": "review", "mixed": true, "difficulty": 0.5, "slice": "mixed", "lang": "en"}
{"prompt": "config versioning for agents needs the effective-config resolution agreed before it's built — host over group over tenant, and what a paused rollout means. settle that, then implement", "purpose": "backendImpl", "secondary": "planning", "mixed": true, "difficulty": 0.75, "slice": "mixed", "lang": "en"}
{"prompt": "percolator index is rebuilt nightly from saved searches, so an edit doesn't take effect until the next day", "purpose": "refactor", "secondary": null, "mixed": false, "difficulty": 0.55, "slice": "core", "lang": "en"}
{"prompt": "rails app has three ways to express money and the mobile API returns two of them in the same payload", "purpose": "refactor", "secondary": null, "mixed": false, "difficulty": 0.55, "slice": "core", "lang": "en"}
{"prompt": "lexer and the language server's lexer are separate implementations that have drifted on string escapes", "purpose": "refactor", "secondary": null, "mixed": false, "difficulty": 0.7, "slice": "core", "lang": "en"}
{"prompt": "rename `Listing#status` to something that admits it's a state machine with six values and two of them unused", "purpose": "refactor", "secondary": null, "mixed": false, "difficulty": 0.5, "slice": "boundary", "lang": "en"}
{"prompt": "could you explain how a listing gets into \"processing\" and what takes it out again", "purpose": "review", "secondary": null, "mixed": false, "difficulty": 0.5, "slice": "core", "lang": "en"}
{"prompt": "why does the fleet query run per tenant every thirty seconds rather than once for everyone", "purpose": "review", "secondary": null, "mixed": false, "difficulty": 0.6, "slice": "boundary", "lang": "en"}
{"prompt": "someone should check whether our search endpoint can be made to scan every shard by a crafted filter", "purpose": "review", "secondary": null, "mixed": false, "difficulty": 0.7, "slice": "boundary", "lang": "en"}
{"prompt": "is it expected that the thumbnail worker strips all metadata including copyright", "purpose": "review", "secondary": null, "mixed": false, "difficulty": 0.5, "slice": "boundary", "lang": "en"}
{"prompt": "pouvez-vous m'expliquer comment le cache de disponibilité est invalidé après une réservation ?", "purpose": "review", "secondary": null, "mixed": false, "difficulty": 0.55, "slice": "boundary", "lang": "fr"}
{"prompt": "docs/search.md describes the geo filter as optional, which stopped being true two versions ago", "purpose": "writing", "secondary": null, "mixed": false, "difficulty": 0.4, "slice": "core", "lang": "en"}
{"prompt": "a short note on why we're moving to an LRU map in the agent, for the decision log", "purpose": "writing", "secondary": null, "mixed": false, "difficulty": 0.35, "slice": "core", "lang": "en"}
{"prompt": "rustdoc on our plugin API promises span validation that doesn't exist", "purpose": "writing", "secondary": null, "mixed": false, "difficulty": 0.45, "slice": "boundary", "lang": "en"}
{"prompt": "write the customer note about dropping 5.4 kernel support, for the two customers still on it", "purpose": "writing", "secondary": null, "mixed": false, "difficulty": 0.45, "slice": "boundary", "lang": "en"}
{"prompt": "listing URL includes the address, so a withdrawn listing is still findable by search engines", "purpose": "quickFix", "secondary": null, "mixed": false, "difficulty": 0.35, "slice": "core", "lang": "en"}
{"prompt": "staging hosts run five containers and prod runs four hundred, with identical map sizes", "purpose": "quickFix", "secondary": null, "mixed": false, "difficulty": 0.3, "slice": "boundary", "lang": "en"}
{"prompt": "floorplan detector fires on any listing whose address contains the word plan", "purpose": "quickFix", "secondary": null, "mixed": false, "difficulty": 0.3, "slice": "boundary", "lang": "en"}
{"prompt": "how should we version the agent's wire protocol now that fleets run three versions at once", "purpose": "planning", "secondary": null, "mixed": false, "difficulty": 0.75, "slice": "core", "lang": "en"}
{"prompt": "what's the right way to test a compiler optimisation pass when the failure mode is a wrong program", "purpose": "planning", "secondary": null, "mixed": false, "difficulty": 0.85, "slice": "core", "lang": "en"}
{"prompt": "i want a position on whether saved searches should be percolated or evaluated on write", "purpose": "planning", "secondary": null, "mixed": false, "difficulty": 0.75, "slice": "core", "lang": "en"}
{"prompt": "three partners want real-time listing updates rather than a fifteen-minute feed, what's our story", "purpose": "planning", "secondary": null, "mixed": false, "difficulty": 0.75, "slice": "boundary", "lang": "en"}
{"prompt": "we need a plan for running the agent on hosts where we can't load eBPF at all", "purpose": "planning", "secondary": null, "mixed": false, "difficulty": 0.8, "slice": "boundary", "lang": "en"}
{"prompt": "what should happen to a tenant's events when they exceed their contracted volume mid-month", "purpose": "planning", "secondary": null, "mixed": false, "difficulty": 0.65, "slice": "boundary", "lang": "en"}
{"prompt": "an endpoint returning a listing's price history with the source of each change, for the detail page", "purpose": "backendImpl", "secondary": null, "mixed": false, "difficulty": 0.5, "slice": "core", "lang": "en"}
{"prompt": "delta feed needs a sequence number that survives a redeploy and a gap-recovery path", "purpose": "backendImpl", "secondary": null, "mixed": false, "difficulty": 0.65, "slice": "core", "lang": "en"}
{"prompt": "agents should buffer to disk with a size cap when ingest is unreachable, rather than dropping silently", "purpose": "backendImpl", "secondary": null, "mixed": false, "difficulty": 0.7, "slice": "boundary", "lang": "en"}
{"prompt": "viewing availability should be computed from a single source rather than three set operations at request time", "purpose": "backendImpl", "secondary": null, "mixed": false, "difficulty": 0.6, "slice": "core", "lang": "en"}
{"prompt": "host detail screen needs a map utilisation chart so we can see which map fills first", "purpose": "frontendImpl", "secondary": null, "mixed": false, "difficulty": 0.5, "slice": "core", "lang": "en"}
{"prompt": "whatever gets us through the submission", "purpose": "quickFix", "secondary": null, "mixed": false, "difficulty": 0.45, "slice": "vague-eval", "lang": "en"}
{"prompt": "next bit of the feed work", "purpose": "backendImpl", "secondary": null, "mixed": false, "difficulty": 0.55, "slice": "vague-eval", "lang": "en"}
{"prompt": "i'd like an honest read of whether our language server can end up holding a symbol from a dropped compilation session, given the interner leaks and transmutes, and if it can, the fix — ideally one that doesn't require rewriting every pass that holds a Symbol the sooner we know the size of it, the better.", "purpose": "review", "secondary": "refactor", "mixed": true, "difficulty": 0.85, "slice": "mixed", "lang": "en"}
{"prompt": "the lexer exists twice, once in the compiler and once in the language server, and they have drifted on string escapes badly enough that the editor highlights code the compiler rejects. merge them onto one implementation, and tell me which existing editor behaviours change as a result i've spent an afternoon on it already and got nowhere useful.", "purpose": "refactor", "secondary": "review", "mixed": true, "difficulty": 0.7, "slice": "mixed", "lang": "en"}
{"prompt": "agent configuration is parsed in three places with different defaults for the same keys, which is why a documented default is sometimes not the effective one. consolidate the parsing, then write down the resulting defaults so the reference we publish is generated rather than hand-maintained this has come up in three separate reviews now. the sooner we know roughly how big this is, the better for planning.", "purpose": "refactor", "secondary": "writing", "mixed": true, "difficulty": 0.6, "slice": "mixed", "lang": "en"}
{"prompt": "saved searches are percolated against an index rebuilt nightly, so an edit doesn't take effect for a day and users assume it's broken. decide with me whether to percolate live or evaluate on write, then implement whichever we land on for new saved searches first it doesn't have to be perfect, it has to be defensible.", "purpose": "backendImpl", "secondary": "planning", "mixed": true, "difficulty": 0.75, "slice": "mixed", "lang": "en"}
{"prompt": "our listing URLs embed the address, so a withdrawn property remains findable through search engines and vendors have complained about it twice this month. change the URL scheme for new listings, keep the old ones resolving with a redirect, and make sure withdrawn listings return the right status code rather than a soft 404 page that's the whole of it, but shout if the context is thin.", "purpose": "quickFix", "secondary": null, "mixed": false, "difficulty": 0.45, "slice": "boundary", "lang": "en"}