Instructions and skills

← README

Instructions (AGENTS.md)

Standing guidance for the model, loaded into the system prompt once at startup:

• Global: ~/.config/reika/AGENTS.md — your personal preferences, applied in every project.
• Project: <cwd>/AGENTS.md — the repo's conventions.

When both exist they are merged, global first, with the project file named as the winner where they disagree — so a personal "terse replies, no emoji" rides along without displacing the repo's build and test rules. CLAUDE.md is accepted as a fallback name in either location.

Each file is capped at 12KB (≈3k tokens). An oversized file is replaced by its heading outline plus a pointer to read the relevant section, so it can never crowd a small context window.

Skills (reusable prompt templates)

A *.md file dropped in a skills directory becomes a slash command — useful for saved workflows (/review, /deploy, /refactor). Project shadows global, which shadows bundled, on a name collision:

• Bundled: the skills that ship with Reika (below)
• Global: $REIKA_SKILLS_DIR if set, otherwise ~/.config/reika/skills/
• Project: <cwd>/.reika/skills/

Skills Reika ships with

Skill	What it does
/issue	Reads a GitHub issue with gh (title + comments, since bodies are often empty), then works on it
/review	Reads a PR and its linked issue with gh, pages the diff, and reviews it in the terminal

They only appear where they can work: gh must be on your PATH, and the current repo must have a github.com remote. Otherwise they are left out of the command list and never offered by plain-English routing. Reika only looks these up — it does not run gh at startup — so a gh that is installed but logged out still shows the skills, and the first call says what is wrong.

To change one, copy it into your global or project skills directory under the same name; your copy replaces the bundled one. A skill of your own can declare the same requirements with a requires: frontmatter line (requires: gh, github). A requirement Reika does not recognize hides the skill.

Layouts

• Flat file: ~/.config/reika/skills/review.md → /review
• Directory with SKILL.md: ~/.config/reika/skills/review/SKILL.md → /review. Lets a skill carry supporting files (scripts, reference docs) — only SKILL.md is used as the prompt body; other files are ignored. Matches the Claude Code skill packaging convention, so you can drop skill folders in verbatim.

File format

Plain markdown with optional YAML frontmatter:

---
description: Review the current branch end-to-end
triggers: review the branch, code review, look over my changes
---

Review the changes on this branch:

1. Run git diff main...HEAD
2. Check for missing tests on changed code
3. Check for inconsistencies with AGENTS.md
   Report a punch list.

Without frontmatter, the first non-empty line becomes the autocomplete description and the whole file is the prompt body.

Invocation:

• /review — sends the file body as your input
• /review focus on the API changes — appends extra args to the body, separated by a blank line
• /skills — list available skills

Plain-English routing (triggers)

You don't have to remember the slash command. triggers: lists phrases that route an ordinary prompt to the skill — matched deterministically in the harness, never by the model. Writing "review the branch for me" gets you a one-line hint that /review exists and that the prompt was sent unchanged; the turn runs normally either way.

The frontmatter accepts any YAML list shape (triggers: a, b, triggers: [a, b], or a - block list). The skill's own name is always an implicit trigger, so a skill called verify routes "verify my changes" with no triggers: at all.

Matching rules, in short:

• Whole-word only — a skill named test does not fire on "latest"
• Multi-word phrases score higher than single words; phrases under 3 characters are ignored
• A tie between two skills suggests neither (routing by array order would be arbitrary)
• Continuations (yes, ok, continue, do it, …) never route
• A skill is suggested at most once per session — decline it once and it stops asking

Confirm-to-apply

REIKA_SKILL_AUTO (default ask) controls it. On a strong match a two-row dialog opens over the input before anything is sent — 1. Send as typed (selected) and 2. Apply /review — with your prompt still in the box underneath. ↑↓, the digits and y/n move the cursor (y lands on Apply), Enter answers, and ctrl-c drops the submit and leaves the prompt for you to edit.

• Applying prepends the skill body to your prompt, with a scrollback receipt naming which phrases matched. Declining sends it as typed, with no hint line.
• Strong means the prompt has the shape of a command, not just its keywords: two distinct triggers, or one multi-word phrase opening the prompt (a please / can you lead is fine). So review pr 420 but first explain how the recap is built asks, while there is an issue with the pr review flow only suggests.
• It is refused when the body would take more than ~15% of the context window, and never offered in chat mode, which has no bash for a skill's first step. Plan mode asks like the others, and its own prompt keeps an applied body to planning.
• A prompt submitted while a turn is running is asked about right then, and the queued entry carries your answer to the replay.
• Headless has nobody to ask, so ask sends the prompt as typed there. REIKA_SKILL_AUTO=apply (or 1) lets -p apply a strong match without asking, and only when the prompt is 12 words or fewer — the cap that prices a wrong pick with nobody to catch it. The TUI still asks under apply: a human present is never a reason to inject silently. off leaves only the hint line.

It is on by default because a wrong pick now costs one keystroke rather than a turn.

Pasted URLs

Paste an http(s) link into a prompt and Reika fetches it before the turn starts, prepending the extracted content as a <url href="…"> block — the same mechanism as @file mentions, so the model reads the page instead of guessing at it. Up to 2 URLs per prompt, truncated to 8k characters each (fetch_url gets the rest). Every fetch leaves a scrollback line, and a dead link is reported as one rather than silently dropped. Set REIKA_PASTE_FETCH=0 to turn it off — worth doing on an airgapped machine.

Only URLs you type are fetched. Links inside an @-mentioned file are file content, and links the model writes are handled separately by URL grounding (REIKA_URL_GROUNDING).

Rules

• Built-in commands always win over skills with the same name — you can't shadow /help or /exit
• Skill names are lowercased filenames; only [a-z0-9_-] are accepted (skip files with weird names)
• Skills load at bootstrap and on /cd — edit a file mid-session, then /cd . to refresh
• Files under <cwd>/.reika/ (including .reika/skills/, .reika/handoff/, etc.) appear in @ autocomplete — handy for inlining a project-local handoff doc, an in-repo skill file, or any other Reika-scratch content. Other dot-dirs (.git/, .vscode/, etc.) stay hidden.