Run any Skill in Manus with one click

professor-x-ops

Stars1

Forks0

UpdatedJune 12, 2026 at 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.

Installation

Install with Codex or Claude Copy this prompt, paste it into Codex, Claude, or another assistant, and let it review the skill page and install it for you.

Run Skill in Manus

Source

xbrxr03

xbrxr03/professor-x-AGI

View GitHub Repository View Creator Repositories

Download

Run Skill in Manus

Related occupationsSOC

Based on SOC occupation classification

Software DevelopersComputer and Mathematical Occupations·SOC 15-1252

SKILL.md

readonly

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.

More from this repository

same repository

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.