Examples and compatibility status

Start with a prompt or skill and copy it into the matching project .octet/ directory. These resources use normal filesystem discovery and typed contribution boundaries:

• Prompts: Pi-compatible Markdown and compact octet TOML.
• Skills: explicit, inspectable skills with bounded text resources.
• Pi migration skill: low-token cleanup around the zero-token Pi inventory.

Current extension authoring#

Start from the Python API 0.4 tool recipe for a bounded local process. The canonical API 0.3 example and its generated conformance suite stay live, with their exact versions and host pins. This checkout is a local release candidate, not a new SDK or bundle publication.

Legacy extension examples#

These are legacy implementation references, not extension authoring quickstarts. New extensions use API 0.4. The Python Extension runtime handles the feature-negotiated API 0.4 wire and the retained API 0.1/0.2, and the generated API 0.3 types serve the separate canonical wire. Don't retag these examples.

• hello-world: API 0.1 initialization, model tool, command, hooks, context, status, renderer and notification.
• caffeinate: API 0.2 macOS sleep inhibition during owning and root turns, with bounded cleanup.
• git-tools: API 0.1 read-only checkpoint, bounded git-status tool, semantic status contribution and renderer. Generic extension status doesn't become persistent coding-TUI chrome.
• local-model-workflow: legacy deterministic prompt and context shaping, and semantic status, for small contexts.
• lsp-client: legacy read-only LSP definition, references, hover and diagnostics through one code_intelligence tool, with bounded, typed unavailable states.

For an existing legacy setup, install the dependency-free source SDK from a checkout before copying an extension:

python3 -m pip install ./sdk/python

Executable extensions need separate enablement and exact trust. They run under the default full-access policy and stay stopped under --safe-mode. Capability declarations aren't an OS sandbox, so use a trusted, separately isolated environment. See discovery and trust and the legacy Python runtime.