vlasenkoalexey/tpu_performance_autoresearch_wiki
Agent Claude Code
Launch a single TPU training workload on a GKE cluster via XPK, poll until completion or hang, capture xprof + HLO dumps to GCS, and report structured verdict signals back to the master agent. Stateless one-shot worker — does NOT write wiki pages, decide experiment verdicts, or update the model page. Use for every…
vlasenkoalexey/tpu_performance_autoresearch_wiki
Agent Claude Code
Independent verifier for kernel-family experiments (the Roles section's verifier for the pallas lane). Given a final candidate kernel + the naive baseline, it independently re-benchmarks both in a fresh process, re-runs numerical parity, captures traces/LLO dumps with the canonical flag set, runs the hypothesis-firing…
vlasenkoalexey/tpu_performance_autoresearch_wiki
Agent Claude Code
Delta-audits an autoresearch run (kernel family OR model lane) since the last audit and returns corrective findings to the runner's context. Runs as a self-rescheduling watcher armed ONCE at launch (start-experiment Step 9·0) — never re-dispatched per iteration by the runner. Read-only over pages, branches, receipts…
vlasenkoalexey/tpu_performance_autoresearch_wiki
Agent Claude Code
Analyze a single completed experiment's xprof trace + HLO dump and return structured ## Profile + ## HLO Dump markdown sections that slot directly into the experiment page. Phases: Phase 1 walks xprof (bucket attribution, dominant ops, memory profile); Phase 2 walks HLO (module sizes, fusion verification, regression…
vlasenkoalexey/tpu_performance_autoresearch_wiki
Agent Claude Code
Explore an ingested repo silo-first. Use for orientation-scale questions ("how does X work in ", "where is Y implemented", pre-engineering surveys) against any repo with a wiki/codebases/ / silo. NOT for single known-symbol lookups or following a live stack trace — grep those directly.
vlasenkoalexey/tpu_performance_autoresearch_wiki
Skill Claude CodeCodex
Authoring discipline for kernel-family candidates (the author role of kernelexperiments/program.md step K4). Load BEFORE writing any kernel candidate — whether you are a solo agent running the whole K0–K9 loop (production: Gemini / Claude / Codex from a bare prompt) or a dispatched author process in the trial…
vlasenkoalexey/tpu_performance_autoresearch_wiki
Skill Claude CodeCodex
Bootstrap a new model family for the autoresearch loop. Creates the folder structure under wiki/experiments/, scaffolds the model-level program.md from a template (asking the user about lanes, sizes, hardware targets, sequence length, and architecture invariants), optionally creates lane-level program.md stubs, and…
vlasenkoalexey/tpu_performance_autoresearch_wiki
Skill Claude CodeCodex
Generate a comprehensive retrospective for a (model, lane) pair that maps the explored search space, identifies coverage gaps from the topic taxonomy (model-optimization-index.md) and Pallas kernel catalog, surfaces premature-exhaustion signals (single-axis-probe vs kernel-port imbalance, frontier-stuck duration), and…
vlasenkoalexey/tpu_performance_autoresearch_wiki
Skill Claude CodeCodex
Coding-discipline guidelines for surgically modifying model code inside a per-experiment fork in the autoresearch loop. Adapts Andrej Karpathy's LLM-coding guidelines (think before coding, simplicity first, surgical changes, goal-driven execution) to this project's invariants — no semantics drift, one mechanism per…
vlasenkoalexey/tpu_performance_autoresearch_wiki
Skill Claude CodeCodex
Generate a structured optimization hypothesis for the autoresearch loop. Walks three layers — (1) the generic wiki/model-optimization-index.md for TPU envelope data + topic pointers + cross-model refuted-pattern principles, (2) the generic wiki/model-optimization-blueprint.md for ladder-keyed phase ordering +…
vlasenkoalexey/tpu_performance_autoresearch_wiki
Skill Claude CodeCodex
Generate ONE structured kernel-optimization hypothesis — step K2 of wiki/kernelexperiments/program.md. Self-contained for the kernel lane (any harness, bare-prompt solo or dispatched): reads the K1 bound diagnosis + the kernel index's signal→lever map + the family's refuted layers, and returns a proposal with the…
vlasenkoalexey/tpu_performance_autoresearch_wiki
Skill Claude CodeCodex
Ingest a paper, article, blog post, talk, or documentation page into the wiki. Encodes SCHEMA's INGEST-SOURCE operation as a callable skill. Reads the raw source (PDF/HTML under raw/sources/ or fetched from URL), confirms emphasis with the human before writing, files wiki/sources/ - .md, updates connections…
vlasenkoalexey/tpu_performance_autoresearch_wiki
Skill Claude CodeCodex
Run the LINT operation defined in SCHEMA.md across the wiki. Checks every invariant the schema declares — orphan pages, broken links, missing profile/HLO sections, missing hypothesis-firing audits, stuck status: inprogress stubs, missing stub labels (Hypothesis/Mechanism/Predicted signal/Falsification criterion), log…
vlasenkoalexey/tpu_performance_autoresearch_wiki
Skill Claude CodeCodex
Refresh the GKE TPU cluster inventory for one or more GCP projects. Runs .claude/scripts/scan-gke-clusters.sh to enumerate clusters with TPUs in each project (status, XPK presence, ready-node count, XPK type, topology, machine type), writes a per-project Markdown summary at .env/ -gke-tpu-cluster-scan.md, and merges…
vlasenkoalexey/tpu_performance_autoresearch_wiki
Skill Claude CodeCodex
Start the autoresearch optimization loop for a specific model + lane. Resolves the hierarchical program.md (root → model → lane), asks the user for hardware (local TPU VM or GKE cluster of a specified TPU type + topology), discovers available clusters from .env/, checks occupancy with USERPREFIX-aware attribution…
vlasenkoalexey/tpu_performance_autoresearch_wiki
Skill Claude CodeCodex
Stop the autoresearch optimization loop cleanly. Disarms the launch-armed process watcher, cancels pending wakeups, reaps orphan workloads on the project's cluster pool, files any missing experiment pages, runs lint, and appends a clean-shutdown marker to wiki/log.md. Invoke at the end of an autoresearch session.
vlasenkoalexey/tpu_performance_autoresearch_wiki
Skill Claude CodeCodex
Connect ingested code silos on the concept axis — cross-link the same concept (a Pallas kernel, an optimization technique) across repos, inline like a normal wiki: the host wiki's concept page links down to each repo's implementation, and each implementation links back up. Selective by design — you choose WHICH…
vlasenkoalexey/tpu_performance_autoresearch_wiki
Skill Claude CodeCodex
Ingest a code repo into a grounded, lint-clean markdown wiki an agent can answer internals questions from. Idempotent reconcile — first build, version bump (--ref), or an added concept are all the same operation. Trigger when the user asks to wikify/ingest a repo, build an internals wiki, or update one. Accepts a repo…
vlasenkoalexey/tpu_performance_autoresearch_wiki
Instructions file CodexOpenCode
Instructions for vlasenkoalexey/tpu_performance_autoresearch_wiki, covering codex instructions, compatibility, operating rules and codex translation notes.
vlasenkoalexey/tpu_performance_autoresearch_wiki
Instructions file
Instructions for vlasenkoalexey/tpu_performance_autoresearch_wiki, a project described as: Wiki built using Karpathy method containing information about TPU performance optimizations and hooking it up to autoresearch optimization engine.
vlasenkoalexey/tpu_performance_autoresearch_wiki
Instructions file Gemini CLI
Instructions for vlasenkoalexey/tpu_performance_autoresearch_wiki, covering gemini/antigravity operating rules, platform adaptation (claude code → gemini/antigravity), 1. skills — native, no emulation, 2. never-stop hook & retrospectives and 3. session and transcript resolution.