---
name: researching-with-rumilo
description: Dispatches AI research subagents via rumilo to search the web or explore git repositories. Use when the user asks to research a topic, look up library usage, explore a codebase, read documentation from the web, or when you need information from an external repository or website to complete a task.
license: GPL-3.0-or-later
metadata:
  author: Amolith <amolith@secluded.site>
---

`rumilo` spawns disposable AI subagents for two kinds of research: **web search** and **repo exploration**. Each mode runs in isolation and returns a synthesised answer on stdout.

## Important: timeouts

Rúmilo is an LLM-based subagent that may take multiple turns to complete its research. _Never_ impose a timeout.

## Modes

### Web mode

Search the internet or ask questions about a specific page.

```bash
# Open-ended search (no URL)
rumilo web "what is the current stable version of Zig"

# Question about a specific page (URL pre-fetched into context)
rumilo web "how does the slog package handle log levels" -u "https://pkg.go.dev/log/slog"
```

- Without `-u`, the agent searches the web on its own.
- With `-u`, the URL is fetched first and injected into the agent's context alongside the query, so it can answer questions _about_ that page.

### Repo mode

Clone a git repository into a temp directory, spawn an agent inside it, and let it explore the source to answer questions.

```bash
# Public repo
rumilo repo -u "https://github.com/owner/repo" "How does the routing system work?"

# Private repo (works with any clonable URI — SSH, HTTPS with auth, etc.)
rumilo repo -u "git@git.sr.ht:~user/repo" "What does the Config struct look like?"
```

The `-u` flag is **required** in repo mode — it needs something to clone.

## Flags

| Flag                       | Modes | Effect                                                            |
| -------------------------- | ----- | ----------------------------------------------------------------- |
| `-u <url>`                 | both  | Seed URL (web) or clone URI (repo)                                |
| `--model <provider:model>` | both  | Override the configured default model                             |
| `--ref <ref>`              | repo  | Checkout a branch, tag, or commit after cloning (requires --full) |
| `--full`                   | repo  | Full clone instead of shallow (depth 1)                           |
| `--no-cleanup`             | both  | Keep the temp workspace after exit                                |
| `--verbose`                | both  | Print tool calls and results to stderr                            |

## Crafting queries

Rumilo works best with focused queries. Think about **scope** and **coupling** when deciding what to ask and how many questions to include.

- **Tightly coupled, narrow scope** — multiple simple questions in one call is fine when they're about the same small surface area (a single file, one struct, one API endpoint).
- **Broad scope or loose coupling** — one question per invocation. If answering requires exploring many files or pages, keep it to a single clear question so the agent can focus.
- **Write natural questions, not keyword lists.** Rumilo is an LLM, not a search index.
- **Sequence calls when later questions depend on earlier answers.** Get the overview first, then drill in. Rumilo is amnesic — each invocation starts fresh. If a later query depends on an earlier answer (repo URL, version number, file paths), you must include that context explicitly in the prompt.
- **Run independent queries in parallel** (e.g. via separate tmux windows or backgrounded commands), but keep dependent queries sequential.

### Bad

Overloaded — five loosely-coupled questions spanning an entire diff, each requiring different context:

```bash
rumilo web "In this diff, find: 1. What files are modified and created \
2. What config options each channel exposes 3. How messages are formatted \
4. The architecture for adding a new channel 5. What UI components exist" \
  -u "https://example.com/commit/abc123.diff"
```

### Good

Sequenced calls, each with a clear scope. Later calls include non-obvious context from earlier results — file paths, symbol names, patterns — so rumilo can go straight to the answer instead of searching.

```bash
# 1. Lay of the land — broad overview first
rumilo web "What files were modified or created, and what's the broad shape of the change?" \
  -u "https://example.com/commit/abc123.diff"

# 2a and 2b can run in parallel — neither depends on the other,
#    but both benefit from context learned in step 1.

# 2a. Architecture deep-dive, pointing to the key file discovered in step 1
rumilo web "How is the notification system in pkg/notify/ architected? \
  Each channel seems to implement a Sender interface — what's the process \
  for adding a new one?" \
  -u "https://example.com/commit/abc123.diff"

# 2b. Config drill-down, referencing the settings struct found in step 1
rumilo web "What configuration options does each channel expose in \
  ChannelSettings? Can message content be customised?" \
  -u "https://example.com/commit/abc123.diff"

# 3. UI specifics, narrowed by what steps 1–2 revealed about the component tree
rumilo web "The channel config UI lives in web/src/components/settings/channels/. \
  Do the per-channel forms follow a shared pattern or component?" \
  -u "https://example.com/commit/abc123.diff"
```

Don't include context that's obvious or redundant — telling rumilo to "search for notif" when asking about notifications wastes tokens. Focus on non-obvious anchors: exact file paths, struct names, interface signatures.

Tightly coupled questions in a single call are fine when the scope is small:

```bash
rumilo repo -u "https://github.com/owner/repo" \
  "What fields does the Config struct have and which are optional?"
```

## Choosing a mode

- Need to find or learn about a library? **Web** to locate it, then **repo** to explore its source.
- Need to understand how a dependency works in detail? **Repo** with the dependency's clone URL.
- Need current docs, release notes, or general knowledge? **Web**, optionally with the URL of a relevant page.