PoopRusteek — Skills & Local RAG

Skills

A skill is a reusable markdown instruction set the agent can load into its system prompt. PoopRusteek discovers them from a dozen agent-ecosystem directories.

Authoring / installing

Primary format: one directory per skill containing a SKILL.md:

{skills_dir}/{skill-name}/SKILL.md

SKILL.md may start with an optional 3-line YAML frontmatter that overrides the defaults, then a markdown body:

---
name: <short title>
description: <one sentence: what it teaches + when to use it>
---

# <Title>
<lean, actionable body>

Legacy format: *.prompt.md files (built-in only, e.g. assets/prompts/; base and tools are excluded).

Discovery order (first-found-wins per name)

project (./.skills/, ./.claude/skills/) → home agent dirs (~/.claude, ~/.agents, ~/.cursor, ~/.windsurf, ~/.cline, ~/.codex, ~/.github/copilot, …) → config dirs (~/.config/{skills,opencode,Claude,cursor,windsurf,cline,zed}/skills) → data dirs (~/.local/share/{skills,pooprusteek/skills}) → built-in (assets/prompts/) → custom paths from config.skills.paths.

Drop your skill in any of those. To install into PoopRusteek's own data dir, use {data_dir}/pooprusteek/skills/{skill}/SKILL.md.

Enabling / disabling

/skills                    # picker to toggle skills
/skills list
/skills enable <name>      # pin the skill into the system prompt
/skills disable <name>

Enabled skills are concatenated under # Active Skills and appended to the system prompt. Extra search dirs live in config.skills.paths. The agent can also list/load skills itself mid-conversation via the skill tool — so you often don't need to enable anything by hand.

Local RAG (semantic layer)

A fully offline hybrid index — multilingual e5-small ONNX embeddings + stemmed TF-IDF, fused with Reciprocal Rank Fusion — over three corpora: skills, MCP tools, and your conversation history. The ~120 MB model downloads once on first launch, then zero network. It degrades gracefully (falls back to lexical, never blocks features).

What it does for you automatically:

• Per-turn hints — every prompt is matched against the skill and MCP-tool catalogs; top hits ride along as an ephemeral hint ("this skill may apply — load it with the skill tool"), so the agent discovers capabilities without you enabling them.
• Deferred MCP schemas — above 12 tools ([semantic] mcp_schemas = auto), the system prompt carries only a server-level summary (playwright (25 tools)), not each tool. Full definitions arrive via the hint or the tool_search builtin. Disabling RAG forces full schemas back on.
• Builtins for the model — tool_search (find MCP tools on demand) and history_search (recall solutions from past sessions).

Controlling it

/rag              # how-it-works + live status (state / model / indexed counts / config)
/rag on | off     # persist semantic.enabled and flip the running service
/rag reload       # drop model + corpora, re-verify/re-download, re-embed everything
/rag-limit [<N>|auto|off]   # embedder batch cap (ONNX memory guard); auto sizes from RAM

Config (config.toml):

[semantic]
enabled = true          # whole RAG layer; /rag on|off flips this
top_k = 3               # max suggestions per corpus per turn
min_dense_score = 0.80  # cosine floor for hint candidates without keyword overlap
mcp_schemas = "auto"    # "auto" (defer above 12 tools) | "full" | "deferred"
rag_limit = "auto"      # embedder batch cap; set via /rag-limit

Searching your history

/search [query]

Opens a full-screen search over all saved sessions: s cycles sort (relevance / newest / oldest), r filters by role, u dedups to one hit per session. Enter on a result loads that session. The history index lives at {data_dir}/semantic/history.json — a rebuildable cache; session files are the source of truth.