Documentation
¶
Overview ¶
Package harness assembles middleware, context providers, and history defaults for provider-backed agents.
Pass AgentConfig's result to the normal provider constructor. The harness reuses the agent lifecycle and the provider's tool-invocation middleware. Tracing covers agent runs, function tools, and model calls made by supported model adapters.
Index ¶
Constants ¶
const DefaultInstructions = `` /* 561-byte string literal not displayed */
DefaultInstructions guides tool use and progress updates.
Variables ¶
This section is empty.
Functions ¶
func AgentConfig ¶
AgentConfig adds harness defaults to cfg without modifying its input slices. Call it once for each agent configuration, before the provider constructor.
Caller middleware remains outermost; built-in context providers precede caller providers. Harness instructions precede caller run options. Existing history, message injection, tools and explicit history-conflict settings are retained. Per-service-call history persistence is enabled.
File memory uses a unique persisted working folder per session below WorkingDirectory/agent-file-memory. That subtree is excluded from default skill discovery and resource access. Its handles are cleaned up automatically when the provider is no longer reachable; filememory.Provider.Close can release them earlier.
Provider constructors remain responsible for tool-invocation configuration. Configuration errors from component constructors retain those constructors' panic contracts; filesystem setup failures return an error.
Types ¶
type Config ¶
type Config struct {
// WorkingDirectory supplies default skills and the agent-file-memory
// directory. Empty uses the current directory. Relative paths are resolved
// when AgentConfig is called.
WorkingDirectory string
// Instructions replaces DefaultInstructions. Nil uses the defaults.
// An empty or whitespace-only value omits harness instructions.
Instructions *string
DisableTodo bool
DisableAgentMode bool
DisableFileMemory bool
DisableSkills bool
DisableWebSearch bool
// AgentMode configures the mode provider unless DisableAgentMode is true.
AgentMode agentmode.Config
// Skills configures the skills provider unless DisableSkills is true.
// When both Skills.Skills and Skills.Sources are nil, skills are discovered
// from WorkingDirectory. Non-nil collections replace that default source.
Skills skills.ContextProviderOptions
// DisableToolApproval omits standing-approval and auto-approval middleware.
// It does not remove the provider's approval requirements from tools.
DisableToolApproval bool
ToolApproval toolapproval.Config
// DisableOpenTelemetry disables agent, tool, and model-call tracing, not
// feature-usage tracking. When tracing is enabled, OpenTelemetry can
// independently disable model-call instrumentation.
DisableOpenTelemetry bool
OpenTelemetry otelprovider.MiddlewareConfig
}
Config selects harness components. Its zero value enables the defaults. Agent identity, history, additional tools and providers belong in the agent.Config passed to AgentConfig. Tool-invocation settings remain in the provider constructor's existing configuration.
Source Files
Directories
¶
| Path | Synopsis |
|---|---|
|
Package agentmode provides a context provider that tracks the agent's operating mode (e.g.
|
Package agentmode provides a context provider that tracks the agent's operating mode (e.g. |
|
Package filememory provides a context provider with tools for persistent file-based agent memory.
|
Package filememory provides a context provider with tools for persistent file-based agent memory. |
|
Package loop provides middleware that re-invokes an agent until one of its evaluators decides no more work is needed.
|
Package loop provides middleware that re-invokes an agent until one of its evaluators decides no more work is needed. |
|
Package todo provides a context provider that gives agents todo list management tools for tracking work items during long-running complex tasks.
|
Package todo provides a context provider that gives agents todo list management tools for tracking work items during long-running complex tasks. |
|
Package toolapproval provides middleware that manages human-in-the-loop tool approval with support for "don't ask again" standing rules.
|
Package toolapproval provides middleware that manages human-in-the-loop tool approval with support for "don't ask again" standing rules. |
|
Package toolautocall provides the middleware that automatically invokes the function tools a model calls and feeds their results back for the next turn.
|
Package toolautocall provides the middleware that automatically invokes the function tools a model calls and feeds their results back for the next turn. |