onboarding-context-ingestion

name	onboarding-context-ingestion
description	Methods for ingesting onboarding context from various sources. Covers repo analysis via Kilo CLI, document conversion, voice transcription, and prior session recall. Companion to the group-onboarding skill — fills the Phase 0 patterns that the main skill references.
version	1.1.0
platforms	["linux"]
metadata	{"hermes":{"tags":["onboarding","context","ingestion","kilo","repo-analysis"],"triggers":["/onboarding with context"]}}

Onboarding Context Ingestion

Purpose

When /onboarding is invoked with context (a repo, document, voice message, URL, or prior conversation), this skill captures the specific resolution methods for each source type. It complements the Phase 0 algorithm in group-onboarding with executable patterns.

Source-Specific Methods

1. GitHub Repo (Kilo CLI) — with stale-doc detection

Trigger: User says "el contexto es este repo", "lee mi repo X", or names a GitHub repository as onboarding context.

Rule: Do NOT use read_file, search_files, or terminal ls/cat to explore the repo yourself. Delegate to Kilo CLI — UNLESS the user explicitly says "lee mi repo" or "read the content" (direct instruction overrides the rule).

Steps:

1. Clone the repo if not already present:

   gh repo clone <org>/<repo-name>
   

   (Use gh CLI — SSH keys may not be configured. Detect org from URL: if user says "kirlts/x" use kirlts, if "jpgil/x" use jpgil, etc.)

2. Run Kilo CLI with an analysis prompt that includes stale-doc cross-checking:

   kilo run "Dame un analisis completo del proposito y contenido del repositorio /home/opc/<repo-name>. CRITICO: verifica si la documentacion declarativa (.cursorrules, CLAUDE.md, AGENTS.md, README.md) coincide con la realidad del codigo. Para eso, compara sus afirmaciones sobre versiones de Python, framework, y dependencias contra requirements.txt, Dockerfile, pyproject.toml, y git log. Si hay discrepancia (docs desactualizadas), reportala explicitamente. Que tipo de informacion contiene, cual es su estructura, cuales son sus objetivos, y que conclusiones puedo sacar sobre como deberia operar el grupo de WhatsApp '<group-name>' de Martin (kirlts). Incluye detalles como: estructura de directorios, proposito principal, principios operativos, estado actual. Responde en español." --auto --dir /home/opc/<repo-name>
   

3. Kilo returns a structured analysis. Extract from it:

   • Profile name suggestion (often matches group or repo name)
   • Description (one sentence)
   • Repository path
   • Skills the profile needs
   • Constraints derived from repo's rules/RULES.md
   • Tone inferred from the repo's voice
   • Worker profile type (default vs worker)

4. Feed these into the group-onboarding Phase 1-3 flow, skipping questions already answered.

Why Kilo CLI: The user requires Kilo for ANY repo interaction, even read-only analysis. This is particularly strict for repos with .agents/ (Kairós governance) where even reading a file directly is forbidden. The pattern kilo run "task" --auto --dir <path> is the universal entry point.

2. Attached Document

Trigger: PDF, DOCX, PPTX, XLSX, EPUB, TXT, CSV, JSON, XML, image, audio, ZIP file.

Resolution: markitdown <file> CLI converts to Markdown. Then read the output.

3. Voice Message

Trigger: Audio voice note from user.

Pipeline: STT (Groq Whisper) → transcribed text → analyze as plain text context.

4. URL

Trigger: User shares a URL as onboarding reference.

Resolution: Fetch and extract content via web tools, then analyze.

5. Prior Conversation

Trigger: User says "en base a lo que conversamos" or similar.

Resolution: recall(bank=<group-name>-profile) or session_search for the relevant session.

Pitfalls

Mistake	Correction
Using read_file on a repo with .agents/	Always use Kilo CLI for repo exploration
Trying to explore a repo before cloning it	Clone first with gh repo clone
Asking "should I read this repo" when user already said "lee mi repo"	Just do it — the user explicitly asked
Reading partial files from a repo to save time	Let Kilo handle the full analysis — it reads comprehensively
Kilo output is long but user is on WhatsApp	Summarize Kilo's output to 2-3 lines, offer the full analysis as context for Phase 1-3