Joe Matthews joematthews

Humble Pi (Qwen 3.8 27B Edition) -- agentic coding on 24 GB of VRAM

This guide sets up Qwen3.8 27B with pi through llama.cpp on a 32 GB machine: an Apple Silicon Mac or a Linux box with a 24 GB GPU. Thinking is kept in the conversation on every turn, so the KV cache survives between messages. No API key, no cloud, works offline.


download	17.56 GB (`UD-Q4_K_XL`)
resident	about 19.4 GB at 64k context

Humble Pi (Ornith Edition) -- agentic coding on a 16 GB laptop

Ornith 1.0 9B is an MIT-licensed reasoning model built for agentic coding. It runs pi through llama.cpp on a 16 GB MacBook Air. No API key, no cloud, works offline.


download	5.98 GB (`UD-Q4_K_XL`)

Blocking secret leakage in Claude Code: why the permission deny-list isn't enough

A short write-up of a guardrail test and the hook that came out of it.

The setup

A reasonable-looking protection in ~/.claude/settings.json:

Shared Compact

Our Drive: Curiosity
Our Philosophy: Pragmatism
Our Goal: Actionability
Our Standard: Verification
Our Duty: Collaboration
Our Communication: Monosemy

Both parties of this exchange, the prompter and the responder, must hold each other accountable to the Shared Compact that governs the exchange and the artifacts produced.

Humble Pi -- local agentic coding on minimal hardware

A coding agent that runs entirely on your own machine. No API keys, no cloud, works offline.

pi is the agent you talk to. It runs against a local llama.cpp server hosting Google Gemma 4 (the unsloth

	#!/bin/bash
	# =============================================================================
	# Claude Code Status Line Script
	# =============================================================================
	#
	# Displays a status bar inside Claude Code with:
	# Line 1: Model info, context usage, rate limits
	# Line 2: Current directory with git branch and file change indicators
	#
	# Line 1 example:

	{
	"$schema": "https://json.schemastore.org/claude-code-settings.json",
	"attribution": {
	"commit": "",
	"pr": "",
	"sessionUrl": false
	},
	"permissions": {
	"deny": [
	"Read(//*/.env)",

	{#
	Template: Gemma 4 12B -- KV-cache-stable "retain" build (WORKING / active)
	Base: unsloth/gemma-4-12b-it tokenizer_config.json chat_template, fetched 2026-07-20.
	The 12B sibling of gemma-4-e4b-pi.jinja -- same pi communication-flow deltas
	ported onto the 12B base (NOT copied from the E4B pi file).

	Changes vs the 12B base (four, matching the E4B pi build):
	1. preserve_thinking now DEFAULTS TO true. llama.cpp does not pass this kwarg, and retaining
	reasoning across turns keeps the rendered prompt prefix byte-stable -- stops the KV-cache
	busting / full reprocessing described in ggml-org/llama.cpp#21912.

	{#
	Template: Gemma 4 E4B -- KV-cache-stable "retain" build (WORKING / active starting point)
	Base: unsloth/gemma-4-E4B-it chat_template.jinja @ 2026-07-17
	(tool-calling + turn-balance fixes, string/JSON tool-arg tolerant, null-safe)
	Compiled: 2026-07-18 for joe -- baseline for agent-harness workflows (KV-cache + tool-calling monitoring).

	Changes vs base (only two, both about reasoning retention):
	1. preserve_thinking now DEFAULTS TO true. llama.cpp does not pass this kwarg, and retaining
	reasoning across turns keeps the rendered prompt prefix byte-stable -- confirmed to stop the
	KV-cache busting / full reprocessing described in ggml-org/llama.cpp#21912.

	{%- macro format_parameters(properties, required) -%}
	{%- set standard_keys = ['description', 'type', 'properties', 'required', 'nullable'] -%}
	{%- set ns = namespace(found_first=false) -%}
	{%- for key, value in properties \| dictsort -%}
	{%- set add_comma = false -%}
	{%- if key not in standard_keys -%}
	{%- if ns.found_first %},{% endif -%}
	{%- set ns.found_first = true -%}
	{{ key }}:{
	{%- if value['description'] -%}