Documentation
¶
Overview ¶
Package planexec provides a plan/execute/replan strategy: an LLM planner decomposes the prompt into parallel tasks, each task runs as a child Process (any Agent[T], adapted via makeInput), the results are collected, and the planner decides to iterate or finalize. A summarizer produces the final answer. It does at most one LLM Generate per transition (checkpointing between phases via the Phase field), which keeps replay cheap and avoids relying on intra-transition determinism (agentkit is at-least-once, D44).
The one deliberate exception is E24: when a plan response is not valid JSON, the plan transition issues a single in-transition correction Generate before giving up, rather than failing the whole transition and relying on backoff.
Index ¶
Constants ¶
This section is empty.
Variables ¶
var ( RolePlanner = agentkit.DefineModelRole("planexec.planner") RoleSummarizer = agentkit.DefineModelRole("planexec.summarizer") )
Roles are exported as package variables (identity binding, D32). Unbound roles fall back to the default model.
Functions ¶
func Register ¶
func Register[T any](r *agentkit.Registry, name agentkit.AgentName, version int, taskAgent agentkit.Agent[T], makeInput func(TaskSpec) (T, error), opts ...Option) (agentkit.Agent[Input], error)
Register builds and registers the planexec strategy. It is generic over the task agent's input type T: makeInput adapts each planned TaskSpec into the child input T, so any Agent[T] can be a task (D50). taskAgent's zero value or a nil makeInput is rejected (ErrInvalidAgentDef).
Types ¶
type Input ¶
type Input struct {
Prompt string `json:"prompt"`
}
Input is planexec's own launch input (required; empty prompt is rejected by Init).
type Option ¶
type Option func(*config)
Option configures the strategy.
func WithLimiter ¶
WithLimiter sets this agent's execution budget, which Strategy.Limit answers with. Default: none, i.e. LimitPass at every check.
This is a different question from WithMaxRounds: that one is an algorithm parameter the planner is even told about, while a limiter sees the kernel's counters. Those include every terminated child, so on a planexec parent one limiter is a budget for the whole fan-out.
func WithMaxParallelTasks ¶
WithMaxParallelTasks caps tasks per round (reflected as the plan schema maxItems). Default: 5.
func WithMaxRounds ¶
WithMaxRounds caps the plan/replan rounds. Default: 3. Reaching it forces finalize.
func WithOnFinish ¶
func WithOnFinish(h agentkit.FinishHandler[Output]) Option
WithOnFinish wires a completion handler for this agent. Delivery is best-effort; see agentkit.WithOnFinish.
func WithSystemPrompt ¶
WithSystemPrompt sets the system prompt for planner/summarizer Generates. Default: "".
type Output ¶
type Output struct {
Summary []string `json:"summary"`
Tasks []TaskResult `json:"tasks"`
}
Output is the Done output.
type TaskResult ¶
type TaskResult struct {
Title string `json:"title"`
Status agentkit.ProcessStatus `json:"status"`
ProcessID agentkit.ProcessID `json:"process_id"`
Output []byte `json:"output,omitempty"`
}
TaskResult is one task's outcome. Output carries the child's Output verbatim on success (its interpretation is the caller's, D40).
Source Files