Agent Skill

A packaged Agent Skill that teaches a coding assistant to write svir correctly: the same material as this site, reorganized for a model rather than a reader.

Download svir-skill.zip

Source: skill/svir.

Why you might want it

svir's shape is unlike the LLM SDKs most code was written against. The conversation is the request, the system prompt is a field, the answer is one stream whose last item is the whole answer, and an error is a kind to match on. An assistant writing from habit reaches for a system role message, choices[0].delta.content, or a client that remembers the chat. That code does not compile here, or worse, compiles and loses the tool calls.

The skill front-loads exactly those traps, fourteen non-negotiables, and then routes to detail on demand.

What is in it

File	Covers
SKILL.md	Establishing the version and the features, a call and a tool loop that work, the non-negotiables, routing
references/requests.md	The system prompt, messages and parts, attachments, reasoning effort, structured output, conversations, storing and restoring
references/streaming.md	Events, the completion and its finish reasons, reasoning, usage and speed, cancelling, strict and lenient, limits, failures
references/tools.md	Tool, the Tools registry, schemas from types, the loop, streamed, requiring or forbidding a call, a Toolbox of one's own
references/client.md	Base URL, API keys, extra headers, timeouts, listing models, layers, TLS providers, a custom HTTP backend, tests without a server
references/codec.md	A proxy that relays the stream, Encoder and Decoder alone, a transport of one's own
references/errors.md	Error kinds, and symptom to cause at runtime and at compile time

SKILL.md stays short on purpose: an entrypoint the agent always reads, and six references it loads only when the task needs one.

Install

The SKILL.md format is a shared standard, so installation is the same everywhere: unzip and copy the svir/ directory into the tool's skills folder, keeping the folder name. It has to match the name in the frontmatter.

Tool	Personal	Per project
Claude Code	~/.claude/skills/svir/	.claude/skills/svir/
opencode	~/.config/opencode/skills/svir/	.opencode/skills/svir/
Codex CLI	~/.codex/skills/svir/	.codex/skills/svir/

unzip svir-skill.zip
mkdir -p ~/.claude/skills && cp -r svir ~/.claude/skills/

Restart the assistant afterwards: skills are discovered at startup.

opencode also reads .claude/skills/ and .agents/skills/, so a single copy inside a project can serve more than one tool.

Anything else

Any assistant that can read a file will do. Point it at SKILL.md and let it follow the links, or add a line to the project's AGENTS.md:

For Rust code that talks to LLMs with the `svir` crate, read
`.agents/skills/svir/SKILL.md` and the reference file it routes you to.

The code in it compiles

Every Rust snippet in the skill is a complete set of items, and every one is compiled against the published svir crate in this repository's CI, with the schemars and tracing features on. What an assistant copies out of it builds. That is the point of shipping a skill rather than a prose summary: an assistant that pastes a plausible-looking API is worse than one that pastes a verified one.

The code on this site is held to the same check. To run it yourself after editing:

python3 ci/check-snippets.py --docs-dir skill

Version

The skill tracks svir 0.1.5: OpenAI-compatible Chat Completions, streamed. The frontmatter records it, so an assistant can tell whether the skill matches the crate in front of it:

metadata:
  svir-version: "0.1.5"
  msrv: "1.85"
  wire-api: "OpenAI-compatible Chat Completions, streaming"

It also tells the assistant what changed between 0.1 releases, for a project locked on an older one.