pi-mood

Why

LLM agents routinely ignore simple rules, even when they are clearly stated in AGENTS.md. The rules are there — in the system prompt — but after a few dozens of tool calls they get buried under conversation history.

Yet the same LLM follows plan mode instructions flawlessly. Plan mode works because it injects a reminder directly into the user message, not from the distant system prompt. So the model can follow rules — it just needs them close to the generation point.

Periodic injection of your own rules throughout a session should therefore make the agent substantially more reliable. The same principle powers steering mode in Roo-Code.

pi-mood is a minimal implementation of this idea. You annotate existing AGENTS.md headings with desired relative frequency.

How it works

Reads ~/.pi/agent/AGENTS.md and <cwd>/AGENTS.md, parses headings annotated with @N, and injects the rule text as a persistent system message between LLM calls. Rules are selected with frequency-weighted probability and a cooldown to avoid repetition.

Headings without @N are ignored. Any heading level works (##, ###, etc.). @N must appear at the end of the heading line (regex \s*@(\d+)\s*$). Parent sections include their subsection headings and bodies.

## Design review @5

Before any change, write a design brief and send to reviewer. If trivial, show
brief to user and request skip.

### What should be in the brief @3

...

• @5 — relative frequency. Higher = selected more often.
• After selection, the rule's effective frequency is halved for the next pick (2^(-times_shown)).
• A rule is injected every N LLM calls. N defaults to 5, configurable in .pi/mood.json (see Config below).

Status bar

mood: — · — · —                          (fresh start)
mood: 1.2kt · now · «Design review»      (just injected)
mood: 3.5kt · 4 ago · «Say it straight»  (between injections)

Install

pi install npm:pi-mood

Or clone and install from local path:

git clone git@github.com:NikolayXHD/pi-mood.git
pi install ./pi-mood

Config

Create .pi/mood.json in the project root:

{ "injectionFrequency": 5 }

• injectionFrequency (number, default 5) — inject a rule every N LLM calls.