executor

package
v0.8.0 Latest Latest
Warning

This package is not in the latest version of its module.

 Go to latest Published: Aug 31, 2026 License: Apache-2.0 Imports: 15 Imported by: 0

Documentation

Overview

Package executor defines the result type returned by model CLI executors.

Index

Constants

View Source 
const DefaultExecutionMaxTokens = 8192

DefaultExecutionMaxTokens is the bounded output budget used when a caller does not provide one for a task execution. Admission probes use their own, much smaller fixed budget and never use this value.

Variables

This section is empty.

Functions

This section is empty.

Types

type AnthropicExecutor 

type AnthropicExecutor struct {
	// contains filtered or unexported fields
}

AnthropicExecutor calls the Anthropic Messages API.

func NewAnthropicExecutor 

func NewAnthropicExecutor(apiKey, model string) *AnthropicExecutor

NewAnthropicExecutor creates an executor for the given Anthropic model.

func (*AnthropicExecutor) EffectiveTools added in v0.3.0 

func (*AnthropicExecutor) EffectiveTools() []string

EffectiveTools returns nil because HTTP executors cannot invoke tools.

func (*AnthropicExecutor) Execute 

func (e *AnthropicExecutor) Execute(ctx context.Context, prompt string, options ExecutionOptions) Result

Execute sends a full task prompt with an explicit bounded output budget.

func (*AnthropicExecutor) Run 

func (e *AnthropicExecutor) Run(ctx context.Context, prompt string) Result

Run sends prompt to the Anthropic Messages API and returns the text response. It is the short admission-probe path and always uses the 512-token budget.

func (*AnthropicExecutor) RuntimeID added in v0.3.0 

func (*AnthropicExecutor) RuntimeID() string

RuntimeID returns the executor's runtime identifier.

type CLIExecutor 

type CLIExecutor struct {
	// contains filtered or unexported fields
}

CLIExecutor runs a subscription CLI (e.g. claude -p) to answer prompts. Cost is $0 marginal — the user pays a flat subscription, not per token.

func NewClaudeCLIExecutor 

func NewClaudeCLIExecutor(model string) *CLIExecutor

NewClaudeCLIExecutor creates an executor that shells out to the claude CLI. Requires claude (Claude Code) to be installed and already logged in.

func (*CLIExecutor) EffectiveTools 

func (e *CLIExecutor) EffectiveTools() []string

EffectiveTools returns the tool set available in claude -p during execution.

func (*CLIExecutor) Execute 

func (e *CLIExecutor) Execute(ctx context.Context, prompt string, _ ExecutionOptions) Result

Execute invokes the CLI for a full task. Claude Code owns its own context and output controls, so the shared token option is intentionally not added to the command line; the result still satisfies the same execution contract, with provider usage and truncation remaining unknown.

func (*CLIExecutor) Run 

func (e *CLIExecutor) Run(ctx context.Context, prompt string) Result

Run invokes the CLI and returns stdout as the model's response.

func (*CLIExecutor) RuntimeID added in v0.3.0 

func (*CLIExecutor) RuntimeID() string

RuntimeID returns the executor's runtime identifier.

func (*CLIExecutor) Stream 

func (e *CLIExecutor) Stream(ctx context.Context, prompt string, w io.Writer) error

Stream invokes the CLI and pipes tokens directly to w as they arrive. Returns an error if the process fails; output is already written to w on success.

type CodexCLIExecutor added in v0.6.0 

type CodexCLIExecutor struct {
	// contains filtered or unexported fields
}

CodexCLIExecutor runs Codex through the user's existing CLI login.

func NewCodexCLIExecutor added in v0.6.0 

func NewCodexCLIExecutor() *CodexCLIExecutor

func NewCodexCLIExecutorWithUnknownCost added in v0.6.0 

func NewCodexCLIExecutorWithUnknownCost() *CodexCLIExecutor

NewCodexCLIExecutorWithUnknownCost preserves API-key and unrecognized Codex authentication as billable-unknown rather than subscription-free.

func (*CodexCLIExecutor) EffectiveTools added in v0.6.0 

func (*CodexCLIExecutor) EffectiveTools() []string

func (*CodexCLIExecutor) Execute added in v0.6.0 

func (e *CodexCLIExecutor) Execute(ctx context.Context, prompt string, options ExecutionOptions) Result

func (*CodexCLIExecutor) ExecuteWithEvents added in v0.6.0 

func (e *CodexCLIExecutor) ExecuteWithEvents(
	ctx context.Context,
	prompt string,
	options ExecutionOptions,
	w io.Writer,
	emit func(RuntimeEvent),
) Result

ExecuteWithEvents consumes Codex's JSONL stream so long-running agent work remains observable without exposing command arguments or command output.

func (*CodexCLIExecutor) Run added in v0.6.0 

func (e *CodexCLIExecutor) Run(ctx context.Context, prompt string) Result

func (*CodexCLIExecutor) RuntimeID added in v0.6.0 

func (*CodexCLIExecutor) RuntimeID() string

func (*CodexCLIExecutor) Stream added in v0.6.0 

func (e *CodexCLIExecutor) Stream(ctx context.Context, prompt string, w io.Writer) error

type EventTaskExecutor added in v0.4.0 

type EventTaskExecutor interface {
	ExecuteWithEvents(context.Context, string, ExecutionOptions, io.Writer, func(RuntimeEvent)) Result
}

EventTaskExecutor streams task text while reporting structured runtime lifecycle events and returning complete usage/cost telemetry.

type ExecutionOptions 

type ExecutionOptions struct {
	MaxOutputTokens int
}

ExecutionOptions controls a full task execution. MaxOutputTokens <= 0 uses DefaultExecutionMaxTokens so callers cannot accidentally request an unbounded provider response.

type OpenAIExecutor 

type OpenAIExecutor struct {
	// contains filtered or unexported fields
}

OpenAIExecutor uses OpenAI Responses for GPT-5.6 and Chat Completions for older OpenAI models and compatible providers.

func NewOpenAICompatibleExecutor 

func NewOpenAICompatibleExecutor(apiKey, model, endpoint string) *OpenAIExecutor

NewOpenAICompatibleExecutor targets any OpenAI-compatible chat-completions endpoint (Ollama, LM Studio, vLLM, llama.cpp server). apiKey may be empty.

func NewOpenAIExecutor 

func NewOpenAIExecutor(apiKey, model string) *OpenAIExecutor

NewOpenAIExecutor creates an executor for the given OpenAI model.

func NewOpenRouterExecutor 

func NewOpenRouterExecutor(apiKey, model string) *OpenAIExecutor

NewOpenRouterExecutor creates an executor for OpenRouter (OpenAI-compatible API).

func NewXAIExecutor 

func NewXAIExecutor(apiKey, model string) *OpenAIExecutor

NewXAIExecutor creates an executor for xAI Grok models via the OpenAI-compatible endpoint at api.x.ai. Use XAI_API_KEY.

func (*OpenAIExecutor) EffectiveTools added in v0.3.0 

func (*OpenAIExecutor) EffectiveTools() []string

EffectiveTools returns nil because HTTP executors cannot invoke tools.

func (*OpenAIExecutor) Execute 

func (e *OpenAIExecutor) Execute(ctx context.Context, prompt string, options ExecutionOptions) Result

Execute sends a full task prompt with an explicit bounded output budget.

func (*OpenAIExecutor) Run 

func (e *OpenAIExecutor) Run(ctx context.Context, prompt string) Result

Run sends prompt to the OpenAI-compatible API and returns the text response. It is the short admission-probe path and always uses the 512-token budget.

func (*OpenAIExecutor) RuntimeID added in v0.3.0 

func (e *OpenAIExecutor) RuntimeID() string

RuntimeID returns the executor's runtime identifier.

type Result 

type Result struct {
	Output       string
	Error        error
	Usage        Usage
	CostUSD      float64
	CostKnown    bool
	FinishReason string
	Truncated    bool
}

Result holds the output of a single model invocation.

type RuntimeAdapter added in v0.3.0 

type RuntimeAdapter interface {
	Run(ctx context.Context, prompt string) Result
	TaskExecutor
	ToolProvider
	RuntimeID() string
}

RuntimeAdapter is the complete transport contract used by veto. Admission and execution remain separate methods with independent output budgets.

type RuntimeEvent added in v0.4.0 

type RuntimeEvent struct {
	Kind   RuntimeEventKind
	Name   string
	Status string
	Count  int
}

RuntimeEvent is the allowlisted event bridge from an agent runtime into Veto's ledger. Name is a validated tool or artifact kind, never arguments, output, paths, or file contents.

type RuntimeEventKind added in v0.4.0 

type RuntimeEventKind string

RuntimeEventKind identifies safe, structured lifecycle events emitted by an agent runtime. Runtime adapters must not include prompts, tool arguments, tool output, credentials, or file contents in these events. 

const (
	RuntimeToolStarted       RuntimeEventKind = "tool.started"
	RuntimeToolCompleted     RuntimeEventKind = "tool.completed"
	RuntimeToolError         RuntimeEventKind = "tool.error"
	RuntimeApprovalRequested RuntimeEventKind = "approval.requested"
	RuntimeApprovalGranted   RuntimeEventKind = "approval.granted"
	RuntimeApprovalDenied    RuntimeEventKind = "approval.denied"
	RuntimeArtifactCreated   RuntimeEventKind = "artifact.created"
)

type TaskExecutor 

type TaskExecutor interface {
	Execute(ctx context.Context, prompt string, options ExecutionOptions) Result
}

TaskExecutor is implemented by transports that can perform a full task execution. Run remains the short admission-probe method for compatibility.

type ToolCapabilityStatus added in v0.4.0 

type ToolCapabilityStatus interface {
	EffectiveToolsKnown() bool
}

ToolCapabilityStatus distinguishes a known empty tool list from a runtime whose project/session-specific tools have not been discovered yet.

type ToolProvider 

type ToolProvider interface {
	EffectiveTools() []string
}

ToolProvider reports tools the transport can actually invoke. Text-only runtimes return nil so model-declared capabilities never grant tool access.

type Usage 

type Usage struct {
	InputTokens  int
	OutputTokens int
	TotalTokens  int
	Known        bool
}

Usage contains provider-reported token counts. Known distinguishes an omitted usage object from a provider that explicitly reported zero values.

Source Files

View all Source files

Jump to

Keyboard shortcuts

? : This menu
/ : Search site
f or F : Jump to
y or Y : Canonical URL
go.dev uses cookies from Google to deliver and enhance the quality of its services and to analyze traffic. Learn more.