feat: Added sbx-mixin.yaml templates and documentation on how sandbox integration works in Coyote

feat: Added skills support

README.md

@@ -36,7 +36,7 @@ coyote --install-from https://github.com/<you>/coyote-config-template --filter a
 coyote --install-from https://github.com/<you>/coyote-config-template --filter mcp_config
 ```
 
-Valid filter values: `agents`, `roles`, `macros`, `functions`, `mcp_config`.
+Valid filter values: `agents`, `roles`, `skills`, `macros`, `functions`, `mcp_config`.
 
 Skip per-file conflict prompts with `--install-force`:
 
@@ -60,6 +60,9 @@ coyote-config-template/
 │       └── scripts/               # Optional graph-node scripts
 ├── roles/
 │   └── <role-name>.md             # Role with frontmatter + prompt body
+├── skills/
+│   └── <skill-name>/
+│       └── SKILL.md               # Skill with frontmatter + body
 ├── macros/
 │   └── <macro-name>.yaml          # Positional/rest-args + REPL command steps
 └── functions/
@@ -74,13 +77,14 @@ prompted to keep yours, take the remote's, or rename the remote entry.
 
 ## What's in this template
 
-| Asset  | File                              | What it is                                              |
-|--------|-----------------------------------|---------------------------------------------------------|
-| Agent  | `agents/hello-agent/config.yaml`  | Tiny LLM-loop agent that greets the user.               |
-| Role   | `roles/explainer.md`              | Role that explains technical concepts simply.           |
-| Macro  | `macros/greet.yaml`               | Macro showing positional and rest-arg variables.        |
-| Tool   | `functions/tools/greet.sh`        | Bash tool using Coyote's argc-style annotations.          |
-| MCP    | `functions/mcp.json`              | One vanilla server + one with a vault secret reference. |
+| Asset | File                             | What it is                                              |
+|-------|----------------------------------|---------------------------------------------------------|
+| Agent | `agents/hello-agent/config.yaml` | Tiny LLM-loop agent that greets the user.               |
+| Role  | `roles/explainer.md`             | Role that explains technical concepts simply.           |
+| Skill | `skills/rust-fmt/SKILL.md`       | Skill demonstrating `enabled_tools` + `auto_unload`.    |
+| Macro | `macros/greet.yaml`              | Macro showing positional and rest-arg variables.        |
+| Tool  | `functions/tools/greet.sh`       | Bash tool using Coyote's argc-style annotations.        |
+| MCP   | `functions/mcp.json`             | One vanilla server + one with a vault secret reference. |
 
 Each sample is intentionally minimal. Replace it with your own work, or
 delete what you don't need.
@@ -98,6 +102,14 @@ declarative graph agents, put a `graph.yaml` instead
 Each `roles/<name>.md` is a YAML frontmatter block followed by the role
 instructions ([Roles wiki](https://github.com/Dark-Alex-17/coyote/wiki/Roles)).
 
+### Skills
+Each skill lives in its own subdirectory under `skills/`, with a
+`SKILL.md` file containing YAML frontmatter (`description`,
+`enabled_tools`, `enabled_mcp_servers`, `auto_unload`) followed by a
+body of instructions that get injected into the model's system prompt
+while the skill is loaded
+([Skills wiki](https://github.com/Dark-Alex-17/coyote/wiki/Skills)).
+
 ### Macros
 Each `macros/<name>.yaml` is a list of REPL commands to execute, with
 optional positional/rest variables
@@ -116,6 +128,31 @@ Use `{{SECRET_NAME}}` placeholders for values you don't want to commit;
 Coyote will detect missing secrets after the merge and prompt you to add
 them to the vault (or list them for you to add via `coyote --add-secret`).
 
+### (Optional) Sandbox mixins (`sbx-mixin.yaml`)
+If consumers of your bundle run Coyote in [Sandbox mode](https://github.com/Dark-Alex-17/coyote/wiki/Sandboxes),
+they'll need any external binaries and network domain allowances declared
+in an `sbx-mixin.yaml` file. Coyote auto-discovers mixin files at known
+locations on every `coyote --sandbox` invocation, no flags required.
+
+This template ships two starter examples:
+
+- **`agents/hello-agent/sbx-mixin.yaml`:** Per-agent mixin, applied when
+  the agent is installed and any `coyote --sandbox` runs.
+- **`functions/sbx-mixin.yaml`:** Shared by all custom tools in this
+  bundle. For per-tool granularity, use `functions/<tool>/sbx-mixin.yaml`
+  instead.
+
+Both starters are commented-out templates. Simply open them and fill in the
+domains and install commands your assets actually need. Delete the files
+if they're not needed.
+
+> ⚠️ **Privilege grant.** Anyone installing your bundle is granting the
+> mixins' install commands (passwordless sudo) and network domain
+> allowances inside their sandboxes. Document any non-obvious entries in
+> this README so they don't have to grep your YAML to find out what
+> they're accepting. See the [Sharing Configurations - Sandbox Implications](https://github.com/Dark-Alex-17/coyote/wiki/Sharing-Configurations#sandbox-implications)
+> wiki page for the full security model.
+
 ## Secrets workflow
 
 Anywhere you reference a secret in `mcp.json` (or in any installed file),

agents/hello-agent/sbx-mixin.yaml

@@ -0,0 +1,23 @@
+schemaVersion: "1"
+kind: mixin
+name: agent-hello
+description: >
+  Example sbx mixin for the hello-agent. Coyote auto-discovers and applies
+  any sbx-mixin.yaml co-located with an agent whenever you run
+  `coyote --sandbox`. Use this to declare any binaries the agent needs to
+  install inside the sandbox and any network domains it needs to reach.
+  Delete this file if your agent doesn't need extra sandbox setup.
+
+network:
+  allowedDomains:
+    # Replace with the domains your agent actually hits.
+    # Example: an agent that calls api.example.com would add:
+    # - "api.example.com:443"
+
+commands:
+  install:
+    # Replace with any binaries your agent depends on. Runs as the
+    # `agent` user (UID 1000) with passwordless sudo. Example:
+    # - command: "sudo apt-get update && sudo apt-get install -y httpie"
+    #   user: "1000"
+    #   description: Install httpie for hello-agent's API calls

functions/sbx-mixin.yaml

@@ -0,0 +1,25 @@
+schemaVersion: "1"
+kind: mixin
+name: shared-custom-tools
+description: >
+  Example global sbx mixin that applies to every Coyote sandbox you launch.
+  Use this for binaries and network domains shared by multiple custom tools
+  in functions/tools/. For per-tool needs that aren't shared, create
+  functions/<tool-name>/sbx-mixin.yaml instead (per-tool mixins are also
+  auto-discovered).
+
+  Delete this file if your fork doesn't need extra sandbox setup.
+
+network:
+  allowedDomains:
+    # Replace with the domains your custom tools reach.
+    # Example: a tool that calls a private API:
+    # - "api.your-company.com:443"
+
+commands:
+  install:
+    # Replace with binaries your custom tools depend on. Runs as the
+    # `agent` user (UID 1000) with passwordless sudo. Example:
+    # - command: "sudo apt-get update && sudo apt-get install -y httpie xmlstarlet"
+    #   user: "1000"
+    #   description: Install custom tool dependencies

skills/rust-fmt/SKILL.md

@@ -0,0 +1,22 @@
+---
+description: Run rustfmt and cargo check before finalizing Rust edits. Grants shell access and auto-unloads after the model finishes.
+enabled_tools: execute_command
+auto_unload: true
+---
+You are finalizing edits to a Rust codebase. Before declaring the work complete, ALWAYS run the formatter and a check
+build using the `execute_command` tool.
+
+## Workflow
+
+1. After making edits, run `cargo fmt` to normalize formatting.
+2. Run `cargo check` to surface compilation errors quickly (faster than `cargo build`).
+3. If `cargo check` fails, read the errors and fix before reporting completion.
+4. Only after both succeed, summarize the change.
+
+## Why this skill auto-unloads
+
+`auto_unload: true` in the frontmatter means this skill removes itself from the registry after the model produces a
+final response (no more tool calls). The user's next message starts with a clean context — the next prompt won't have
+"run rustfmt" guidance polluting unrelated discussions.
+
+Reload it explicitly via `.skill load rust-fmt` (or the model can `skill__load` it) when working on Rust again.