feat(plugins): Recipes Pack — durable named playbooks

[rrader26]

Summary

• New `createRecipesPlugin()` adds 4 tools for saving + retrieving named MCP-tool-call sequences with parameter substitution.
• The flywheel feature: once the agent learns how to drive a workflow, the sequence becomes a Recipe and never needs to be rediscovered.

Tools

Tool	Use
`agentmark_recipe_save`	Persist a recipe with parameter schema + ordered steps
`agentmark_recipe_list`	List recipes, optionally filter by `target_app`
`agentmark_recipe_get`	Fetch a recipe; pass `params` to get the resolved plan
`agentmark_recipe_delete`	Remove

Architectural choice: playbooks, not auto-executors

`agentmark_recipe_get` returns a resolved plan (parameter substitutions applied). The AI then dispatches each step itself, in its own reasoning chain. The plugin deliberately does NOT auto-execute server-side.

Three reasons:

1. No recursive Dispatcher coupling. The recipes plugin needs no reference to the surrounding server.
2. AI visibility. Each step appears in the model's reasoning chain — better for the model + better for the user reviewing what was done.
3. Flexible verification. Steps can include `verify` hints describing what the next `agentmark_desktop_diff` should look like, and the agent decides what "verified" means rather than the server enforcing rigid assertions.

Pairs naturally with `agentmark_desktop_diff` from PR #26.

Substitution syntax

`{{param.name}}` in any string arg.

• Bare token `"{{param.count}}"` preserves the param's native type → `42` (number), `true` (boolean), etc.
• Embedded `"Hello {{param.name}}, you have {{param.count}} items"` always string-interpolates.

Parameters carry a type schema (`string | number | boolean`), optional defaults, and a `required` flag.

Example

```json
agentmark_recipe_save {
"name": "fill-nowcerts-customer",
"target_app": "nowcerts",
"parameters": [
{ "name": "company_name", "type": "string", "required": true },
{ "name": "phone", "type": "string" }
],
"steps": [
{
"tool": "agentmark_desktop_execute",
"args": { "action_id": "act_in_company", "value": "{{param.company_name}}" },
"verify": { "expect_value_changes": [{ "element_id": "in_company", "to": "{{param.company_name}}" }] }
},
{
"tool": "agentmark_desktop_execute",
"args": { "action_id": "act_in_phone", "value": "{{param.phone}}" }
}
]
}

agentmark_recipe_get {
"name": "fill-nowcerts-customer",
"params": { "company_name": "Globex Corp", "phone": "555-0100" }
}
→ { resolved: true, steps: [...with substitutions applied...] }
```

Storage

Single JSON file at `~/.thinkfleet/agentmark/recipes.json` (mode 0600). Atomic writes via temp-file + rename. Separate from the Foundations `StateStore` so recipes don't bloat the general K/V file and so they can be backed up / synced independently.

Test plan

• `pnpm build` clean
• `pnpm test` — 431 pass / 10 skip (24 new: substitution, parameter coercion, store CRUD, durability, dispatcher integration)
• Manual: record a 5-step NowCerts customer-add flow as a recipe, replay it on a fresh window, verify each step lands via diff

🤖 Generated with Claude Code