一键在 Manus 中运行任何 Skill

professor-x-ops

星标1

分支0

更新时间2026年6月12日 09:43

Operational interface for running and measuring Professor X. Use when running or interpreting --repo-fix-bench, --hiro-null, --hiro-smoke, --evolve-on-repofix, or --evolve-skill-on-repofix; when picking the model/data-dir; when a run needs Ollama; or when committing/PRing work on this project. Knows the K=2/MDE conventions, the trustworthy-vs-untrustworthy metrics, and the branch/PR policy.

安装

用 Codex 或 Claude 帮你安装复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。

在 Manus 中运行

来源

xbrxr03

xbrxr03/professor-x-AGI

打开 GitHub 仓库查看创作者相关仓库

下载

在 Manus 中运行

Professor X — operational cheatsheet

Single Rust binary at professor-x/. Build: cargo build --bins. Run from professor-x/ (paths like scripts/benchmarks/repo_fix/tasks.json and config/hardware.toml are cwd-relative).

Model & data dir

Primary model is qwen3:8b-q4_K_M (local, Ollama). Pass --model qwen3:8b-q4_K_M to use the exact installed tag. llama4:scout is deprecated. Thesis: small model + great harness.
Prefix runs with PROFESSOR_X_DATA_DIR=$HOME/.professor-x (state.db, artifacts).
Ollama must be up (curl -s localhost:11434/api/tags). It occasionally hiccups mid-run.

Benchmarks (the trustworthy scoreboard is deterministic)

--repo-fix-bench — the TRUSTWORTHY metric: copy fixture → confirm test red → agent edits in a /tmp workdir → re-run test → pass iff green. Ungameable. Fixtures: scripts/benchmarks/repo_fix/ (stdlib check.py, NOT pytest). Baseline ~0.85 pass@1 on qwen3:8b. Each run ~7 min (±0.1 variance).
--hiro-null N --hiro-limit M — HIRO benchmark; uses a stratified sample + a hybrid judge (deterministic expected + LLM-judge). The LLM-judge is UNRELIABLE — prefer repo-fix for headline numbers.
--hiro-smoke — validate the 60-task HIRO file deserializes.

Self-improvement (M4 — empirical fitness gate)

--evolve-on-repofix N — evolve the system prompt, gated on measured repo-fix pass@1 (K=2 reps, MDE=0.10). Accepts ONLY a candidate that beats baseline beyond noise. A failure-aware proposer is shown the actual failures and targets them.
--evolve-skill-on-repofix N — same gate, but evolves the skills/conductor/px-fix-bug.md skill and PERSISTS it only if it measurably helps.
A gate that accepts nothing when nothing beats noise is CORRECT (unlike the legacy loop / ARIS meta-optimize, which never measure).

Conventions

Long runs: launch in the background, capture to /tmp/*.log, grep for pass@1 / CURVE.
Branch harness-gaps; commit + push freely; PRs to main (direct main push is policy-blocked).
Record honest results in docs/research/eval-trust.md. Apply verify-the-ruler and adversarial-self-review before recording or committing.

同仓库更多 Skills

同仓库

diagnose-from-trajectory

xbrxr03/professor-x-AGI

Read the actual failing trajectory before fixing an agent/harness failure on Professor X — never guess. Use when a HIRO or repo-fix task fails, pass@1 stalls or won't rise, the agent loops/thrashes, an edit doesn't land, or you're about to 'improve' a prompt or tool. Guessed fixes failed; trajectory-diagnosed fixes lifted repo-fix 0.50→0.77. Ships a Tier-3 helper: scripts/pull_trajectory.py.

2026-06-121

add-repofix-fixture

xbrxr03/professor-x-AGI

Author a new repo-fix benchmark fixture for Professor X correctly. Use when adding/expanding coding tasks for --repo-fix-bench, growing the benchmark, or making it harder/more representative. Enforces the eval-trust rule: a fixture is only valid once you've verified it goes red→green (broken→fixed) with stdlib checks (NOT pytest — pytest is not installed).

2026-06-121

adversarial-self-review

xbrxr03/professor-x-AGI

Two-thread adversarial review of your OWN work before you commit it on Professor X: steelman it, then attack it. Use before any commit or PR, before recording a result in docs/memory, before declaring a milestone done, or when you're about to claim a 'win' or improvement. Adapted from ARIS's kill-argument; the same pattern lives in Prof X's agent.critic.

2026-06-121

rust-harness-change

xbrxr03/professor-x-AGI

The safe workflow for changing Professor X's Rust harness code (react.rs, executor.rs, hiro.rs, main.rs, proposer/loop_runner). Use when editing the harness/agent loop, tools, evaluation, or evolution code; before committing a Rust change; or when a change touches behavior a test or benchmark covers. Enforces: build → FULL test suite → measure on the trustworthy benchmark → commit on harness-gaps → PR.

2026-06-121

small-model-harness-design

xbrxr03/professor-x-AGI

Design principles for making a weak local model (qwen3:8b) into a capable agent on Professor X — the project's core thesis. Use when designing/changing the agent loop, tools, prompts, or evolution; when deciding how to fix a capability gap; or when tempted to blame the model. The recurring finding: the gap is almost always the HARNESS, not the 8B's reasoning.

2026-06-121

verify-the-ruler

xbrxr03/professor-x-AGI

Mechanism-check a metric before trusting it on Professor X. Use when about to report/record/act on a pass-rate, pass@1, pass@3, φ, or any score; when a number jumps or looks too good; before writing a result to docs/research or memory; before calling a milestone done; or when a metric reads a suspicious constant (0.000, always-same). This project has shipped fabricated 'confirmed' metrics — M0 exists to stop that.

2026-06-121

name	professor-x-ops
description	Operational interface for running and measuring Professor X. Use when running or interpreting --repo-fix-bench, --hiro-null, --hiro-smoke, --evolve-on-repofix, or --evolve-skill-on-repofix; when picking the model/data-dir; when a run needs Ollama; or when committing/PRing work on this project. Knows the K=2/MDE conventions, the trustworthy-vs-untrustworthy metrics, and the branch/PR policy.
allowed-tools	Bash(*), Read, Grep, Glob

Professor X — operational cheatsheet

Single Rust binary at professor-x/. Build: cargo build --bins. Run from professor-x/ (paths like scripts/benchmarks/repo_fix/tasks.json and config/hardware.toml are cwd-relative).

Model & data dir

Primary model is qwen3:8b-q4_K_M (local, Ollama). Pass --model qwen3:8b-q4_K_M to use the exact installed tag. llama4:scout is deprecated. Thesis: small model + great harness.
Prefix runs with PROFESSOR_X_DATA_DIR=$HOME/.professor-x (state.db, artifacts).
Ollama must be up (curl -s localhost:11434/api/tags). It occasionally hiccups mid-run.

Benchmarks (the trustworthy scoreboard is deterministic)

--repo-fix-bench — the TRUSTWORTHY metric: copy fixture → confirm test red → agent edits in a /tmp workdir → re-run test → pass iff green. Ungameable. Fixtures: scripts/benchmarks/repo_fix/ (stdlib check.py, NOT pytest). Baseline ~0.85 pass@1 on qwen3:8b. Each run ~7 min (±0.1 variance).
--hiro-null N --hiro-limit M — HIRO benchmark; uses a stratified sample + a hybrid judge (deterministic expected + LLM-judge). The LLM-judge is UNRELIABLE — prefer repo-fix for headline numbers.
--hiro-smoke — validate the 60-task HIRO file deserializes.

Self-improvement (M4 — empirical fitness gate)

--evolve-on-repofix N — evolve the system prompt, gated on measured repo-fix pass@1 (K=2 reps, MDE=0.10). Accepts ONLY a candidate that beats baseline beyond noise. A failure-aware proposer is shown the actual failures and targets them.
--evolve-skill-on-repofix N — same gate, but evolves the skills/conductor/px-fix-bug.md skill and PERSISTS it only if it measurably helps.
A gate that accepts nothing when nothing beats noise is CORRECT (unlike the legacy loop / ARIS meta-optimize, which never measure).

Conventions

Long runs: launch in the background, capture to /tmp/*.log, grep for pass@1 / CURVE.
Branch harness-gaps; commit + push freely; PRs to main (direct main push is policy-blocked).
Record honest results in docs/research/eval-trust.md. Apply verify-the-ruler and adversarial-self-review before recording or committing.