Imported from nickmccarty/ollama-pi-harness (AGENTS.md). Install upstream with npx skills add nickmccarty/ollama-pi-harness. Copyright stays with the author.

AGENTS.md — harness-engineering wiki schema

This file tells your LLM agent how to maintain the project wiki.

Environment

conda activate ollama-pi

Key dependencies: ollama, ddgs, markitdown, jinja2, chromadb, sentence-transformers, torch (cu118), graphviz, requests, Pillow, playwright (+ playwright install chromium for KG screenshots).

Directory layout

harness-engineering/
├── wiki/                    ← LLM-maintained knowledge base (agent writes this)
│   ├── index.md             ← content catalog; updated by wiki_tools.py
│   ├── log.md               ← append-only operations log
│   ├── architecture.md
│   ├── eval-framework.md
│   ├── experiments.md
│   ├── synthesis-instructions.md
│   ├── marginal-value-search.md   ← design spec: saturation-based search loop
│   └── chromadb-memory-migration.md  ← design spec: semantic memory retrieval
├── graphs/                  ← knowledge graph HTML outputs (gitignored)
├── experiment-*.md          ← raw experiment records (immutable — never edit)
├── runs.jsonl               ← agent run log (immutable — never edit)
├── journal.md               ← dev journal (narrative; append-only)
├── roadmap.md               ← planning and stage tracking
└── README.md                ← public-facing overview

Key modules

File	Purpose
`agent.py`	Main research + write + verify pipeline
`wiggum.py`	Verification loop (WIGGUM_MAX_ROUNDS env var caps rounds)
`memory.py`	Observation store — SQLite + ChromaDB semantic retrieval + `assess_novelty()`
`autoresearch.py`	Autonomous synthesis instruction improvement loop
`search_cache.py`	SQLite TTL cache for DDGS queries
`kg_gen.py`	Knowledge graph generator — Ollama → D3.js HTML via Jinja2; `--screenshot` flag captures PNG via Playwright
`kg_template.html.j2`	D3.js force-directed graph template
`eval_suite.py`	Regression harness — `--score --tasks T_D,T_E` for autoresearch
`wiki_tools.py`	Python-deterministic wiki maintenance (index / log / lint)

Python-for-determinism rule

Structural bookkeeping is done in Python, not by LLM:

Updating wiki/index.md → python wiki_tools.py index
Appending to wiki/log.md → python wiki_tools.py log "ingest | <source>"
Checking orphans / broken links → python wiki_tools.py lint

The agent writes wiki page content. Python handles structure.

Wiki operations

Ingest

When a new source arrives (experiment result, paper, article, eval output):

Read the source
Write or update relevant wiki page(s) in wiki/
python wiki_tools.py index — refresh index
python wiki_tools.py log "ingest | <source name>" — append to log

One source may touch multiple pages (e.g. a new experiment updates both experiments.md and synthesis-instructions.md).

Query

When answering a question that produces a reusable result:

Read wiki/index.md to find relevant pages
Synthesize answer
If the result is reusable, write it to wiki/<topic>.md
Run index + log tools

Lint

Check for: contradictions, orphan pages, stale claims, missing cross-references. Run python wiki_tools.py lint for a report.

Page frontmatter

Every wiki page opens with YAML frontmatter:

---
title: <page title>
updated: YYYY-MM-DD
sources: [experiment-04.md, runs.jsonl]
tags: [eval, synthesis, architecture]
---

Index format

wiki/index.md is machine-maintained — do not edit by hand.

Log format

Each entry in wiki/log.md:

## [YYYY-MM-DD] <operation> | <subject>

Operations: ingest, query, lint, update.

Grep recent entries: grep "^## \[" wiki/log.md | tail -10

What stays out of the wiki

Raw experiment outputs → experiment-*.md
Agent run data → runs.jsonl
Dev narrative → journal.md
Planning → roadmap.md

The wiki is synthesis, not storage.

ollama-pi-harness

About