ico-your-internals

name	ico-your-internals
description	Dog-food intentional-cognition-os against the operator's own system. Discovers a target — a folder of markdown docs, a project, or an ecosystem — creates an isolated ICO workspace outside the target tree, ingests and compiles its docs, runs a hand-authored question bank against it, verifies each citation by deterministic source-grep, and produces a structured receipts trail with receipts, verifications, friction logs, and a sanitized summary. Sanitized per-run summaries are written into the target repo's dogfood directory; raw answer content and workspace stay in the user cache. Use when validating ICO works on real corpora, generating dog-food proof artifacts, or surfacing ICO bugs to file as beads. Trigger with /ico-your-internals, dog-food ICO, analyze my system with ICO.
version	0.1.0
author	Jeremy Longshore <[e-mail]>
license	Apache-2.0
model	inherit
effort	medium
argument-hint	--target <path> --bank <bank.yaml> [--dry | --probe <run-id>]
allowed-tools	Read, Grep, Bash(ico:*), Bash(node:*), Bash(jq:*), Bash(find:*), Bash(ls:*), Bash(cat:*), Bash(mkdir:*), Bash(cp:*), Bash(bash:*), Bash(python3:*), Bash(git:*), AskUserQuestion
tags	["dogfood","rag","citation-verification","knowledge-compilation","audit-trail"]
compatibility	Requires Claude Code, intentional-cognition-os installed globally via npm install -g intentional-cognition-os, an ANTHROPIC_API_KEY environment variable, plus python3 and jq on PATH. The target repo should have a dogfood directory with question-banks subfolder.

ico-your-internals — dog-food ICO against your own system

Invocation: "/ico-your-internals" · "dog-food ICO" · "analyze my system with ICO" · "run a dog-food session against ".

The operator-facing skill that exercises intentional-cognition-os end-to-end against a real corpus and produces a structured receipts trail. v0.1 covers single-target runs against a hand-authored question bank with deterministic citation verification.

Overview

This skill answers one question: does ICO actually work on real corpora? It exercises the full ICO pipeline (init → mount → ingest → compile → ask) against a doc-heavy target the operator points it at, then verifies every citation ICO produces against the source it claims via deterministic substring grep. The output is a structured receipts trail — committed sanitized rollups for proof-of-use, plus local-only raw answers for forensic inspection.

The skill is the kernel of a build-in-public dog-food loop. Every run produces one number (citation-verify rate) that goes into dogfood/progress.md and trends over time. Friction in ICO itself surfaces as friction.jsonl entries that become bead candidates on the ICO repo.

What this skill does NOT do:

• Build new corpora (it only runs ICO against existing markdown documentation)
• Auto-commit results (operator decides what lands in the repo)
• Make any Claude API call without an explicit budget estimate and operator confirmation above the $0.50 threshold

Prerequisites

Before invoking:

1. ICO installed: npm install -g intentional-cognition-os. Verify with ico --version returning a real semver string (≥ 1.1.2 recommended).
2. Anthropic credentials: ANTHROPIC_API_KEY exported in the shell environment.
3. POSIX tooling: python3 (≥ 3.10) and jq on PATH. Standard bash, find, awk are assumed.
4. A target: an absolute path to a project or doc directory with at least one .md file. The target is mounted read-only — the skill never writes inside it.
5. A question bank: a YAML file matching the schema in references/question-bank-spec.md. For session-0 scaffolds the bank must have non-empty questions — empty banks halt the skill.
6. Optional: the target repo has a dogfood/ directory (created when running against the ICO repo itself). When absent, sanitized rollups land alongside the cache and the operator manually moves them.