rajatscode

MLX vector-steered prompt valence

This gist is a small experiment for Ashkelon request mediation. It does not call a second hosted model to rewrite text. It loads an open-weight MLX model, measures a contrastive hidden-state direction from a kind seed and a hostile seed, and adds that direction to a selected transformer block while the model generates a rewritten request.

python valence_vector_mlx.py \

rajatscode

name

rajatscode

local-code

An interactive Ollama launcher for people who do not want to learn model names, quantization suffixes, or provider configuration first.

Run local-code and get four choices:

1. Install new models — see detected memory headroom, then the five best benchmark-matched local variants at or below your comfortable threshold, ranked by the live Arena text-overall leaderboard. Each row shows

rajatscode

#!/usr/bin/env python3
"""cs — find and pick up coding-agent sessions (Claude Code + Codex), even after the transcript is gone.

    cs [query]                     fzf picker over every session, newest first
    cs pickup  <id> [claude|codex|opencode|gemini] pick up in that tool (default: its own)
    cs resume  <id>                same as `pickup <id>` in its own tool
    cs restore <id>                bring a purged transcript back from an Entire checkpoint
    cs attach  <id> [--branch B]   write an Entire checkpoint for the session straight to the
                                   checkpoint remote (never amends a commit, never touches main)
    cs --refresh                   (re)index Entire checkpoint remotes found in your projects

rajatscode

def convert_to_couples(row):
    return [n // 2 for n in row]

def trim_couples(couples):
    return [
        couples[i] for i in xrange(len(couples))
        if couples[i//2*2] != couples[i//2*2+1]
    ]

def reindex_by_appearance(sequence):

rajatscode

MIT License

Copyright (c) 2020 rajat mehndiratta

Permission is hereby granted, free of charge, to any person obtaining a copy
of this software and associated documentation files (the "Software"), to deal
in the Software without restriction, including without limitation the rights
to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
copies of the Software, and to permit persons to whom the Software is
furnished to do so, subject to the following conditions:

rajatscode

# To get this to work: install names (`pip3 install --user -U names`)
import names
import random

# copied from https://pe.usps.com/text/pub28/28apc_002.htm
# hardcoded because requests + BeautifulSoup got too ugly and it's faster to
# just extract and prettify the words from the page manually
STREET_SUFFIXES = {
    'Alley', 'Annex', 'Arcade', 'Avenue', 'Bayou', 'Beach', 'Bend', 'Bluff',
    'Bluffs', 'Bottom', 'Boulevard', 'Branch', 'Bridge', 'Brook', 'Brooks',

rajatscode

import argparse
import cProfile
import itertools
import random
import string

ALL_FUNCTION_NAMES=[
  'using_choice', 'using_sample', 'using_randint', 'using_randint_with_offset', 'using_randbits',
  'using_randbits_ascii_lowercase', 'using_one_long_randbits'
]