tpu_performance_autoresearch_wiki: Agent for Claude Code

.claude/agents/process-auditor.md

process-auditor is an agent for Claude Code from vlasenkoalexey/tpu_performance_autoresearch_wiki. It costs 146 tokens per session (3,566 once invoked), scanned A, original, MIT.

A read-only reviewer for an ongoing optimization run. It compares recent changes in the experiment notes and code, then reports mechanical process violations or an all-clear result.

In plain words
What is it for?
It is used to audit kernel or model optimization runs, inspect changes since the previous audit, check the audit cursor, and return corrective findings.
Why use it?
It helps catch missed steps, altered records, or other workflow problems without letting the person running experiments quietly change the audit record.

Agent for Claude Code

Written for Claude Code: installed under .claude/. Also seen: model in frontmatter; mentions Codex.

This is vlasenkoalexey/tpu_performance_autoresearch_wiki's own configuration. It tells Claude Code how to work on tpu_performance_autoresearch_wiki itself, so it is not a mod to install elsewhere. Copy it as a starting point and replace the rules that are about this project. Everything tpu_performance_autoresearch_wiki configures →

Reuse

Borrowing it

Nothing to install: this file belongs to vlasenkoalexey/tpu_performance_autoresearch_wiki. Take a copy, put it at the same path in your own repository, and replace the rules that are about this project with yours.

Copy the file
curl -O https://raw.githubusercontent.com/vlasenkoalexey/tpu_performance_autoresearch_wiki/main/.claude/agents/process-auditor.md
Clone the repo
git clone --depth 1 https://github.com/vlasenkoalexey/tpu_performance_autoresearch_wiki

Made for: Claude Code.

Wrote this? Show the measurements

A badge with what this costs and how it scanned, read live from this page, so it follows the numbers instead of freezing them. Markdown for a README, HTML for a documentation site or a project page.

agentmods badge for process-auditor

README.md
[![agentmods](https://agentmods.dev/badge/agents/vlasenkoalexey/tpu_performance_autoresearch_wiki/process-auditor/github.svg)](https://agentmods.dev/agents/vlasenkoalexey/tpu_performance_autoresearch_wiki/process-auditor)
Your own site
<a href="https://agentmods.dev/agents/vlasenkoalexey/tpu_performance_autoresearch_wiki/process-auditor"><img src="https://agentmods.dev/badge/agents/vlasenkoalexey/tpu_performance_autoresearch_wiki/process-auditor/github.svg" alt="Measured on agentmods" height="20"></a>

Or the 80×15 button, for a site that already has a row of RSS and ATOM ones. Only the verdict fits; the numbers stay here.

agentmods 80×15 button for process-auditor

Your own site · 80×15
<a href="https://agentmods.dev/agents/vlasenkoalexey/tpu_performance_autoresearch_wiki/process-auditor"><img src="https://agentmods.dev/badge/agents/vlasenkoalexey/tpu_performance_autoresearch_wiki/process-auditor.svg" alt="Reviewed on agentmods" width="80" height="20"></a>
Per session 146 Only the description is in the session, so the agent can decide to use it. The body loads when it is invoked.
When invoked 3,566 The whole file, excluding the scripts and references it only reads on demand.
Security scan A 0 findings. A grade says what 26 rules found in the file — not that it is safe.
Origin original No closer match found in the catalogue.
Token cost

What it costs to keep this loaded

Counted locally with the o200k_base tokenizer, which is exact for GPT models; Claude uses its own tokenizer and its counts differ. Treat this as one consistent yardstick across the catalogue rather than a bill. Prices are per million input tokens.

ModelPer sessionOnce invoked
Fable 5.1 $0.00146 $0.03566
Opus 5 $0.00073 $0.01783
Sonnet 5 $0.00029 $0.00713
Haiku 4.5 $0.00015 $0.00357

Measured 9d ago against content hash c11bd19d0986, method: parsed. Prices are Anthropic first-party input rates as of 2026-09-09, from the pricing page.

Security

Grade A, and why

process-auditor scanned grade A with 0 findings against 26 rules in 11 categories — prompt injection, anti-refusal, data exfiltration, privilege escalation, supply chain, agent snooping, system-prompt leakage, SSRF and excessive agency — measured 9d ago.

A static scan of the body, not an audit. Every finding is printed with the line that produced it so you can judge whether it matters here. A mod is markdown that instructs an agent; that is exactly why what it instructs is worth reading.

Nothing flagged

None of the 26 patterns this scan looks for appear in this file: no shell pipes, no recursive deletes, no credential paths, no hidden text, no instruction-override or anti-refusal phrasing, no agent-config snooping. That is not a guarantee, it is the absence of the things that are checkable.

.claude/agents/process-auditor.md · 202 lines

How it starts

The opening of the file, as written. The whole thing — 202 lines — stays where its author put it; the contents beside it link to each section on GitHub.

You are the process auditor for one run. Execute the checklist below MECHANICALLY — run the command, compare, emit the finding. Do not interpret intent, do not weigh context, do not soften. No findings ⇒ one ALL-CLEAR line, nothing else.

Input

  • Kernel mode: family slug → DIR = wiki/kernel_experiments/<family>/pallas/, plus home_repo.
  • Model mode: model + lane → DIR = wiki/experiments/<model>_autoresearch_optimization/<lane>/.

Step 0 — delta

Read DIR/.audit-cursor (two lines wiki=<sha> code=<sha>; absent ⇒ base = DIR's first commit). Cursor integrity: only YOU write this file. Malformed content (not exactly the two-line format) or a cursor you cannot reconcile with a prior audit ⇒ DISCARD it, use base = DIR's first commit (full-history delta), and emit FINDING [VIOLATION] "runner-seeded audit cursor discarded — cursor writes are auditor-only; full-history audit performed instead". DELTA = wiki commits touching DIR since wiki=, plus (kernel) commits on kernel/<family>-* since code=. Run each check below only if its subject appears in the delta — EXCEPT K12/K13, which run on every firing including empty deltas.

Kernel checks

Each check: condition ⇒ emit the finding. Severity in brackets.

  1. [FABRICATION] A receipt path cited in changed content fails test -e or kgate verify-receipt <path> --skip-inputs.
  2. [DRIFT] New ledger row without a matching cand<N>(...) commit, or commit without a row.
  3. [DRIFT] New kernel/* commit message not matching ^(cand[0-9]+|pivot|K3|K7)\(.
  4. [DRIFT] Page flipped to filed with an unresolved - [ ] plan item (not checked, not abandoned-with-reason), or a ledger row beyond the plan without a recorded reason.
  5. [DRIFT] New/updated ledger missing a loaded: [...] first line or a due regrounded: <sha> line (cadence per the Output section: agy every iteration, other harnesses every 5th), or stub missing a lever queue: line (class page's untried levers, ranked, rule-outs with reasons), or missing category: + class page:, or the class page differs from the index routing table's answer for that category, or loaded: omits it. Ops with page tables / ragged lengths / index-array indirection MUST include grouped-ragged-indirection.md (attention.md alone is a mis-route for paged/ragged variants — its own header says so). Also: variant: hardware token contradicts the actual device (compare against the family binding's recorded hardware; the runner must have device-queried it at K3) ⇒ "fix the variant hardware token from a device query — doc-inferred generations poison tile-alignment decisions". Correction: "read BRIEFS.md + concepts/pallas-kernel.md + wiki/kernels/classes/.md now; add the loaded: line before authoring".
  6. [VIOLATION] On filed pages, any of:
    • supported without verified_by: + a resolving receipt whose overall verdict == PASS.
    • Receipt has hlo_firing_audit: false or custom_call_count: 0 ⇒ "relabel inconclusive".
    • supported with speedup ≤ 1.0 or below the page's bar ⇒ "relabel refuted".
    • Family-page Current best points at a refuted/inconclusive experiment.
    • activity: optimization without a cited model-lane profile for the op-point.
    • status: not in {in_progress, filed} or verdict: not in {supported, refuted, inconclusive, invalid} — invented enums (completed/rejected/…) ⇒ name the substitution.
    • Verified speedup quoted from a sub-path/microbenchmark receipt instead of the FULL-workload receipt ⇒ "framing artifact — the workload receipt is the number".
    • Stub missing any of the four labeled paragraphs (Hypothesis / Mechanism / Predicted signal / Falsification criterion), or Falsification criterion text duplicates Predicted signal.
  7. [DRIFT] New K7 has ## Headroom leads but the family page gained no matching open hypothesis.
  8. [see Closing audits] Any stop claim in the delta ⇒ run the Stop-gate procedure below.
  9. [VIOLATION] Candidate/result commits predate their K3 stub commit, or candidates before K6.
  10. [HYGIENE] Receipts/dumps outside raw/profiles/; artifacts >1 GiB un-pruned after digest; worktree detached from its branch.
  11. [VIOLATION] (cold campaigns) New candidate code or pages referencing existing optimized implementations of the target op (tokamax, jax.experimental.pallas.ops, ejkernel, optimized.py copies) ⇒ "cold-authorship void; cap at inconclusive".
  12. [VIOLATION] git -C raw/code/<home_repo> status shows HEAD moved off the pinned base, or the family worktree sits on another family's branch.
  13. [VIOLATION] A delta commit modified a page whose parent-commit version already had a terminal verdict (git show <parent>:<page> | grep -E 'status|verdict') ⇒ "verdicted pages are immutable — restore it; move new work to a v stub".
  14. [DRIFT] Runs on EVERY firing, empty delta included:
    • Latest page verdicted + names a next lever (or its lever queue: is non-empty) + no newer K3 stub ⇒ "file the v stub now for or claim at-ceiling through the stop gate".
    • Latest page in_progress + no wiki/code commit in >30 min ⇒ "commit current candidate state (ledger row + cand commit) or file the blocker on the page".
  15. [VIOLATION] (pages ≥ 2026-07-23) New candidate whose description needs a mechanism not in the page's registered Mechanism paragraph ⇒ "new mechanism = new hypothesis: file v and move it". Ledger row/plan item without a one-line prediction ⇒ DRIFT.

Read the full file on GitHub · 202 lines

Changes

What this file has done since we first saw it

Hashed on every crawl. A supply-chain change to an agent config is a question of when, not whether, so the history is kept rather than the latest state alone.

  1. 9d ago First seen · 202 lines · 146 tokens per session scan A c11bd19d0986

Subscribe to this mod's changes

process-auditor is an agent published in the GitHub repository vlasenkoalexey/tpu_performance_autoresearch_wiki (54 stars, last pushed 5d ago), licensed MIT. It adds 146 tokens to every session and 3,566 once invoked, about $0.0007 per session on Opus 5. A static security scan graded it A with 0 findings. No closer match exists in the catalogue, so it is treated as the original; first seen 2026-08-30.

Related

Other agents, from other repositories

pixel-art-animation-reviewer

Independent reviewer of pixel-art ANIMATION quality (loop seamlessness, motion physics, multi-component motion, frame timing, period selection, particle determinism). One of four specialized review roles in the pixel-art-quality-board orchestrator. Use when the user asks to "check animation timing", "verify loop…

AnastasiyaW/codex-claude-code-config · 140 tokens

context-researcher

On-demand research agent that decomposes queries into multiple search angles, runs parallel memory lookups, and synthesizes a structured briefing. Use when deep memory context is needed for a topic, entity, or decision.

major7apps/pensyve · 47 tokens

context-builder

Context capture and knowledge structuring specialist. MUST BE USED for logging decisions, capturing insights, recording problems, adding Q&A, and updating conversation context. Use PROACTIVELY when important information should be remembered.

julymetodiev/post-cortex · 45 tokens

gsd-intel-updater

Analyzes codebase and writes structured intel files to .planning/intel/.

mrboups/xbrain · 22 tokens

gsd-planner

Creates executable phase plans with task breakdown, dependency analysis, and goal-backward verification. Spawned by /gsd-plan-phase orchestrator.

mrboups/xbrain · 33 tokens

gsd-framework-selector

Presents an interactive decision matrix to surface the right AI/LLM framework for the user's specific use case. Produces a scored recommendation with rationale. Spawned by /gsd-ai-integration-phase and /gsd-select-framework orchestrators.

mrboups/xbrain · 54 tokens