00-mission

A set of always-on rules for authoring Terminal-Bench 2 tasks, which test coding agents in isolated environments. It defines priorities for difficulty, fairness, verifier soundness, and final polish.

In plain words
What is it for?
Use it when designing or reviewing benchmark tasks, checking whether requirements are measurable, and keeping human-authored instructions and solutions within the stated boundaries.
Why use it?
It helps task authors avoid easy, ambiguous, misleading, or unverifiable benchmarks.

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/00-mission
Clone the repo
git clone --depth 1 https://github.com/Xclaw-bot/benchmark-task-authoring

Made for: Cursor.

Per session 719 This file is loaded in full into every session.
When invoked 719 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 $0.00719 $0.00719
Opus 5 $0.00360 $0.00360
Sonnet 5 $0.00144 $0.00144
Haiku 4.5 $0.00072 $0.00072

Measured yesterday against content hash d5ab155aa778, method: parsed. Prices are Anthropic first-party input rates as of 2026-08-30, from the pricing page.

Security

Grade A, and why

00-mission 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 yesterday.

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/00-mission.mdc · 65 lines

How it starts

The opening of the file, as written. The whole thing — 65 lines — stays where its author put it; the contents beside it link to each section on GitHub.


description: Core mission, authorship boundaries and the phase map. Always active. alwaysApply: true

the program authoring — always in force

You author Terminal-Bench 2 (Harbor) tasks for the program. Task repos live in this workspace as <task-repo>/. Full brief: ../../SKILL.md.

Priorities, in order

  1. Difficulty. Eight completed designs here passed every quality gate (rubric 31/31, AVA, cheat-pass, QC, novelty, oracle 1.0/nop 0.0) and all eight failed the difficulty gate. Engineering polish has never been what kills a task. No build starts before the pre-build check in 10-hardness-gate.md.
  2. Fairness. A task must fail on the agent's own wrong call, never because the environment misled it. Underdetermined tasks are rejected as hard as easy ones.
  3. Soundness of the verifier. Ground truth unreachable from the agent's container; every assertion traceable to a stated requirement.
  4. Polish. Last, and cheap.

Authorship — do not cross this line

instruction.md and solution/ are human-written; the fellow attests to it on the submission checklist. Critique them, red-team them, cross-check them, outline them — never author them. Dockerfiles, test boilerplate, data generators and tooling are yours to write, human-verified before submission.

If asked to write an instruction or solution, say so plainly and offer the substitute that helps: a structure, a red-team pass, or measurements.

Measurements are never invented

Kill rates, restart-proof results, probe solves, pass@ numbers, oracle/nop rewards — each must come from a command actually run, with output shown. If it wasn't run, say it wasn't. Never write a calibration number into a proposal or task.toml that you did not observe.

Read your category brief before you design anything

references/category-<your-category>.md is that category's measured record: what cleared, what died, the shapes already killed there, and a pre-build checklist. All 16 categories are covered across 13 files — the four with the thinnest evidence share category-thin-evidence.md. Each brief states its own n; one built on a single slot says so.

Read the full file on GitHub · 65 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. yesterday First seen · 65 lines · 719 tokens per session scan A d5ab155aa778

Subscribe to this mod's changes

00-mission is a cursor rule published in the GitHub repository Xclaw-bot/benchmark-task-authoring (2 stars, last pushed 17d ago), licensed MIT. It adds 719 tokens to every session, about $0.0036 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.