Quick Task Definition

A Quick Task Definition describes a Quick Task: a standalone prompt, together with the session presets, used to start an AI conversation with a single click. Unlike the prompts embedded in an agent, a Quick Task is not bound to any agent, and it can be featured to the whole org and pinned by users.

A Quick Task is either:

The name and description of a Quick Task belong to the Quick Task resource itself, and are therefore not part of the definition below.

Managing Quick Tasks from CLI

Quick Tasks are managed with cs llm quick-task (alias qt), using a YAML file holding the definition described on this page:

cs llm quick-task create fix-test quick-task.yaml --description 'Fix a failed Go test'
cs llm quick-task create fix-test quick-task.yaml --shared   # an org-level shared Quick Task (admin only)
cs llm quick-task update fix-test quick-task.yaml
cs llm quick-task list                 # shared and personal Quick Tasks
cs llm quick-task show fix-test
cs llm quick-task remove fix-test

A Quick Task is referenced by name in the same way as an agent:

The tiles on the Home page are managed with:

cs llm quick-task pin fix-test         # pin to my Quick Tasks, listed after the featured ones
cs llm quick-task unpin fix-test
cs llm quick-task feature fix-test     # feature a shared Quick Task to all users (admin only)
cs llm quick-task unfeature fix-test

To start a conversation from a Quick Task:

cs llm session run --quick-task fix-test --input TEST=TestCreateSandbox

The session presets of the Quick Task are applied, and flags specified explicitly override them. Add --dry-run to print the expanded prompt and the effective presets without sending them.

Structure Overview

prompt:                   # The prompt, optionally with inputs
  inputs:
  - name: ...
  body:
    language: mustache
    script: ...

context:                  # Session presets for a new conversation
  agent: ...
  task: {}
  authorized_templates:
    enabled: true
    allowed:
    - pattern: ...
  model_mapping:
    mappings:
      generic: fast

execution_control:        # Execution presets for a new conversation
  approval_mode: ALL_APPROVED

icon: mui:BugReport       # The icon shown on the tile
disabled: false

Top-Level Fields

Field Type Description
prompt Snippet The prompt inserted into the composer. It is expanded by the client before sending, so the agent only sees plain text. When it defines inputs, a form is shown to collect them.
context Context The session context to preset when the Quick Task begins a new conversation, saving the user from picking options.
execution_control ExecutionControl The execution control to preset when the Quick Task begins a new conversation.
disabled boolean Takes the Quick Task out of use without deleting it — for example, while it is being revised. A disabled Quick Task is skipped in the featured and pinned lists and can't be used, but it is still listed to those who can update it.
icon string Distinguishes the Quick Task at a glance where it is offered, such as on the Home page. Uses the same format as the template icon customization (crafting.dev/sandbox/icon): a built-in name like docker, or mui:NAME for a Material icon like mui:BugReport. If unspecified, the default Quick Task icon is shown.

The presets in context and execution_control only apply when the Quick Task begins a new conversation. They are ignored when the prompt is inserted into an existing message or conversation, as that must not change the session.

Context

The session context to preset. Only the fields below are allowed; the other fields of a session context describe a running session (such as the target sandbox) and are rejected.

Field Type Description
agent string The AI Agent to start the conversation with. It is only a default, which the user can change. If empty, the default agent is used. The name is resolved in the same way as --agent of cs llm session run: a personal agent if it exists, otherwise an org-level one, and my: or org: restricts the scope. Only the owner's personal agents can be referenced from a personal Quick Task.
task Context.Task If specified, the conversation is started as an autonomous task, which runs without human interaction and is archived right after completion.
authorized_templates Context.AuthorizedTemplates The policy limiting the templates which can be used by the agents.
model_mapping ModelMapping Remaps the models used by the conversation. Every entry is validated against the org's LLM configuration. If unspecified, the mapping of the agent applies, if any.

Context.Task

Field Type Description
labels map[string]string Task specific labels.
interactive boolean If true, the session is started interactively, but owns the resources created from it. It is not archived upon completion, so further conversations are possible, and it is subject to the auto archiving policy of the org. The resources are not claimed until the session is archived or deleted.
initiator object The user who initiated the task, with id (the user ID) and name (the name or email for display). It is usually recorded by the system, rather than preset.

Example — an autonomous task:

context:
  task:
    labels:
      kind: triage

Context.AuthorizedTemplates

The policy defining which templates can be used by the agents, matched by template name. Once the policy is enabled: when allowed is non-empty, templates are denied by default and only matching ones are allowed; when allowed is empty, all templates are allowed by default. Templates matching excluded are then denied.

Field Type Description
enabled boolean Activates the policy. When false, all templates are allowed, regardless of allowed and excluded.
allowed SimpleNameMatcher[] Templates allowed to be used.
excluded SimpleNameMatcher[] Templates excluded from use.

Example:

context:
  authorized_templates:
    enabled: true
    allowed:
    - pattern: 'backend-*'
    excluded:
    - regexp: '.*-legacy$'

ExecutionControl

Field Type Description
approval_mode enum How tool calls are approved: DEFAULT — request the user's approval when needed; ALL_APPROVED — approve everything automatically.
streaming object The streaming mode when calling LLM APIs, with a single boolean field enabled.

Example:

execution_control:
  approval_mode: ALL_APPROVED
  streaming:
    enabled: true

Full Example

A shared Quick Task that asks for a failing test and a sandbox, and starts the coding agent with everything approved automatically:

prompt:
  title: Fix a failed test
  inputs:
  - name: TEST
    title: Test name
    placeholder: e.g. TestCreateSandbox
    required: true
    text: {}
  - name: sandbox
    title: Sandbox
    required: true
    reference:
      kind: SANDBOX
  body:
    language: mustache
    script: |
      In sandbox {{sandbox}}, run the Go test {{TEST}},
      find the root cause of the failure and fix it.
context:
  agent: coding
  authorized_templates:
    enabled: true
    allowed:
    - pattern: 'backend-*'
execution_control:
  approval_mode: ALL_APPROVED
icon: mui:BugReport

Create it, and feature it to everyone in the org:

cs llm quick-task create fix-test quick-task.yaml --shared --description 'Fix a failed Go test'
cs llm quick-task feature fix-test

See Also