60-pipeline

A set of rules for submitting and improving Terminal-Bench 2 or Harbor benchmark tasks. It covers the path from proposal and local checks through pull requests, automated review, and repeated scoring runs.

In plain words
What is it for?
Use it to submit benchmark tasks, create and update pull requests, run local oracle and no-op checks, read automated feedback, and judge pass@2 and pass@5 results.
Why use it?
It explains how to interpret failures, avoid invalid results caused by timeouts or broken tests, and keep checks focused on the final reviewed version of a task.

Cursor rule for Cursor

Install

Getting it into your agent

One page per mod, every tool's command on it. A separate URL per tool would split the same page into five that compete with each other.

agentmods
npx agentmods add rules/xclaw-bot/benchmark-task-authoring/60-pipeline
Clone the repo
git clone --depth 1 https://github.com/Xclaw-bot/benchmark-task-authoring

Made for: Cursor.

Per session 964 This file is loaded in full into every session.
When invoked 964 The same file — it is already loaded in full.
Security scan A 0 findings. Scan, not verified.
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 $0.00964 $0.00964
Opus 5 $0.00482 $0.00482
Sonnet 5 $0.00193 $0.00193
Haiku 4.5 $0.00096 $0.00096

Measured 2d ago against content hash 8682404c559b, method: parsed. Prices are Anthropic first-party input rates as of 2026-08-30, from the pricing page.

Security

Grade A, and why

60-pipeline 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 2d 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.

.cursor/rules/60-pipeline.mdc · 82 lines

How it starts

The opening of the file, as written. The whole thing — 82 lines — stays where its author put it; the contents beside it link to each section on GitHub.


description: Submission pipeline: fork/PR flow, the automated check stages, pass@2 and pass@5 economics, reading feedback, revision limits. alwaysApply: false

Pipeline and iteration

Flow

Proposal (gated before you build) → fork → build on submission → local oracle/nop → PR → validity check → pass@2 at your timeout → Automated Review (all Blocking Issues fixed) → pass@5.

gh pr create --repo <program-org>/<task-repo> --fill

Iterate by pushing to the same branch — checks re-run and sticky comments update in place. Use the house PR description shape: one-sentence problem, numbered success criteria mirroring instruction.md, calibration results (oracle 1.0 / nop <1.0), how to run, notes on anything you had to interpret.

The difficulty gate

Acceptance bar: pass@5 ≤ 2/5.

pass@5 Meaning Outcome
0/5 valid failures Fully stumped, oracle still solves ✅ strongest
0/5 invalid failures Timeout / agent / verifier error / unfair prompt ❌ fix the cause
1–2/5 Solvable and genuinely hard
3–5/5 Too easy

A valid failure is the model finishing and being wrong on a fair problem. Timeouts and agent/verifier errors never count. Gate formula: (good valid fails) + (soft-timeout fails) >= 3 and good valid >= 1 — soft timeouts count, infra/in-progress timeouts do not.

Never game the score. Lowering the timeout or padding busywork makes the task harder to finish, not harder to get right; reviewers send it back.

Economics — treat runs as scarce

  • Pass@2 is 6 runs per day, per fellow, per task. Never spend one before the oracle is green locally.
  • A push cancels queued runs for the same PR (per-PR concurrency), so superseded runs auto-cancel — no quota wasted.
  • Fork contributors cannot cancel upstream runs (actions:write → HTTP 404).
  • Pass@5 auto-starts only if pass@2 was valid and Automated Review passed.
  • Review pipeline: max 2 revisions, then Holding-Rejection. Official guidance is to revise sent-back tasks before claiming new ones — they are closest to the bonus.

Read the full file on GitHub · 82 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. 2d ago First seen · 82 lines · 964 tokens per session scan A 8682404c559b

Subscribe to this mod's changes

60-pipeline is a cursor rule published in the GitHub repository Xclaw-bot/benchmark-task-authoring (2 stars, last pushed 18d ago), licensed MIT. It adds 964 tokens to every session, about $0.0048 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 2026-08-31.