30-harbor-format

A file-format and packaging guide for Harbor/TB2 benchmark tasks. Harbor/TB2 is a task format with instructions, a Docker environment, a reference solution, and verification tests.

In plain words
What is it for?
Use it to create task folders, write task.toml, configure the Dockerfile, define solution and test files, and run preflight checks before pushing.
Why use it?
It helps task authors meet the required file layout and manifest rules before submitting 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/30-harbor-format
Clone the repo
git clone --depth 1 https://github.com/Xclaw-bot/benchmark-task-authoring

Made for: Cursor.

Per session 1,115 This file is loaded in full into every session.
When invoked 1,115 The same file — it is already loaded in full.
Security scan C 1 finding. 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 $0.01115 $0.01115
Opus 5 $0.00558 $0.00558
Sonnet 5 $0.00223 $0.00223
Haiku 4.5 $0.00112 $0.00112

Measured 2d ago against content hash d080ba7155ff, method: parsed. Prices are Anthropic first-party input rates as of 2026-08-30, from the pricing page.

Security

Grade C, and why

30-harbor-format scanned grade C with 1 finding 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.

Recursive force deletehighDestructive command

rm -rf with a variable or a broad path is one typo away from removing the wrong tree.

- apt: don't pin versions; `apt-get update` before install; `rm -rf
.cursor/rules/30-harbor-format.mdc · 101 lines

How it starts

The opening of the file, as written. The whole thing — 101 lines — stays where its author put it; the contents beside it link to each section on GitHub.


description: Harbor/TB2 mechanical format: file layout, task.toml schema, Dockerfile rules, base images, timeouts, preflight. Auto-attaches to task files. globs: /task/ alwaysApply: false

Harbor format — the mechanical gate

Per-repo references/ files (rubric, diversity-taxonomy.toml, check-base-image.sh) are authoritative and override anything here.

Layout

task/
  instruction.md          # the only thing the agent sees at runtime (HUMAN-WRITTEN)
  task.toml               # manifest
  environment/
    Dockerfile            # ONE image, for both the agent run and the verifier
    data/                 # inputs copied in; NEVER ground truth
  solution/
    solve.sh              # mounted at /solution, runs the reference (HUMAN-WRITTEN)
  tests/
    test.sh               # verifier entry point; installs NOTHING
    test_outputs.py       # pytest assertions
    expected/             # ground truth — overlaid at /tests only at verify time

Canonical TB2 uses a single image. There is no tests/Dockerfile and [verifier] environment_mode stays unset.

task.toml

  • artifacts = [...] at top level, above the first [section]. Every agent-produced path the tests read must be declared.
  • [task] name = "<program>/<kebab-name>" — inside the [task] table, org/name format, name part ≤3 words. A root-level task = "..." string makes Harbor resolve zero tasks and abort.
  • [metadata]: category/subcategory are pre-seeded, do not edit. task_objective[] and artifact_type[] are closed sets from diversity-taxonomy.toml. expert_time_estimate_hours non-zero and plausible. The three explanation fields must be congruent with the actual files, and verification_explanation must justify the calibration of every tolerance.
  • Timeouts: agent timeout_sec3600 (project-wide hard ceiling). Long enough that the model can finish — the challenge is correctness, not finishing in time. All five trials timing out means the timeout is too low, not the task too hard.
  • No invented fields; extras are silently ignored and create false impressions.

Read the full file on GitHub · 101 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 · 101 lines · 1,115 tokens per session scan C d080ba7155ff

Subscribe to this mod's changes

30-harbor-format is a cursor rule published in the GitHub repository Xclaw-bot/benchmark-task-authoring (2 stars, last pushed 18d ago), licensed MIT. It adds 1,115 tokens to every session, about $0.0056 per session on Opus 5. A static security scan graded it C with 1 finding (recursive force delete). No closer match exists in the catalogue, so it is treated as the original; first seen 2026-08-31.