agent

package
v0.11.7 Latest Latest
Warning

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

Go to latest
Published: Jul 20, 2026 License: MIT Imports: 20 Imported by: 0

Documentation

Index

Constants

View Source
const (
	DefaultMaxTurns         = 400
	DefaultMaxParallelTools = 8
	DefaultToolTimeout      = 10 * time.Minute

	DefaultContextWindow = 400_000

	DefaultReserveTokens = 32_000
)

Variables

View Source
var ErrEmptyInput = errors.New("agent input is empty")

ErrEmptyInput means Send was called without any content.

View Source
var ErrMaxTurnsExceeded = errors.New("agent: internal turn-loop safety bound exceeded — likely a runaway tool-call cycle")
View Source
var ErrTurnInProgress = errors.New("agent turn already in progress")

ErrTurnInProgress means Send was called while another turn was active.

Functions

func ContextWindowFor added in v0.9.3

func ContextWindowFor(model string, largeContext bool) int

func DefaultEffort added in v0.10.5

func DefaultEffort() string

DefaultEffort returns the reasoning effort requested via WINGMAN_EFFORT. Empty (or "auto") leaves the role-based default in place. Unrecognized values are ignored so a typo cannot silently pin an unexpected effort.

func DefaultModel added in v0.9.4

func DefaultModel() string

DefaultModel returns the model requested via environment; WINGMAN_MODEL takes priority over the OpenAI-standard OPENAI_DEFAULT_MODEL.

func DefaultPlanEffort added in v0.11.2

func DefaultPlanEffort() string

DefaultPlanEffort returns the reasoning effort for plan mode requested via WINGMAN_EFFORT_PLAN; empty uses the role-based default.

func DefaultPlanModel added in v0.11.2

func DefaultPlanModel() string

DefaultPlanModel returns the model for plan mode; empty selects the largest available model automatically.

func DefaultUtilityModel added in v0.11.2

func DefaultUtilityModel() string

DefaultUtilityModel returns the model for internal utility calls (recaps, compaction summaries); empty selects the smallest available automatically.

func SandboxDisabled added in v0.10.6

func SandboxDisabled() bool

Types

type Agent

type Agent struct {
	*Config

	Messages []Message
	Usage    Usage
	Revision uint64
	// contains filtered or unexported fields
}

func (*Agent) ContextStats added in v0.11.2

func (a *Agent) ContextStats() ContextStats

func (*Agent) MessagesSnapshot added in v0.10.3

func (a *Agent) MessagesSnapshot() []Message

func (*Agent) QueueInput added in v0.10.9

func (a *Agent) QueueInput(input []Content) bool

QueueInput adds guidance to the active run. The agent consumes queued input at the next safe model boundary. It returns false when no run is active so callers can preserve the input as a normal follow-up instead.

func (*Agent) Recap added in v0.11.2

func (a *Agent) Recap(ctx context.Context) (string, error)

Recap produces a short user-facing briefing of the conversation so far, for returning to a resumed session.

func (*Agent) Running added in v0.11.2

func (a *Agent) Running() bool

Running reports whether a turn is currently active.

func (*Agent) Send

func (a *Agent) Send(ctx context.Context, input []Content) (iter.Seq2[Message, error], error)

Send starts exactly one turn. It never queues implicitly: callers that want to guide the active turn must use QueueInput, while FIFO follow-ups belong in a caller-owned session orchestrator. Setup errors are returned immediately; failures after the turn starts are yielded by the returned stream.

func (*Agent) StateSnapshot added in v0.10.3

func (a *Agent) StateSnapshot() State

func (*Agent) UsageSnapshot added in v0.10.3

func (a *Agent) UsageSnapshot() Usage

type Config

type Config struct {
	Model        func() string
	Effort       func() string
	Tools        func() []tool.Tool
	Instructions func() string

	// UtilityModel, when set and non-empty, handles internal utility calls
	// (compaction summaries, recaps) instead of the main model.
	UtilityModel func() string

	// SubagentModel resolves a model role for per-subagent overrides: "plan"
	// and "utility" name the session's role models, "" the currently
	// inherited model (consulted for effort clamping). ok=false or an empty
	// ID keep the inherited model. Nil disables overrides and clamping.
	SubagentModel func(role string) (ModelOption, bool)

	// CacheKey routes provider-side prompt caching; keep it stable per
	// conversation (e.g. the session ID) to maximize prefix-cache hits.
	CacheKey string

	Hooks hook.Hooks

	// MaxTurns caps successful model invocations in one Send run. Stream
	// retries and tool calls do not consume turns. Zero uses the default;
	// negative disables the safety bound.
	MaxTurns int

	// MaxParallelTools bounds concurrently executing read-only tool calls.
	// Zero uses the default; negative allows the whole emitted batch.
	MaxParallelTools int

	// ToolTimeout is a hard ceiling on every tool call. When zero, tools may
	// extend the default via tool.Tool.Timeout; negative disables deadlines.
	ToolTimeout time.Duration

	ContextWindow int

	// LargeContext compacts against the model's full hardware window instead
	// of stopping at the provider's long-context price threshold (e.g. 2x
	// input pricing on GPT-5.4/5.5 beyond 272k input tokens).
	LargeContext bool

	ReserveTokens int
	// contains filtered or unexported fields
}

func DefaultConfig

func DefaultConfig() (*Config, error)

func (*Config) Derive added in v0.6.2

func (c *Config) Derive() *Config

func (*Config) Models added in v0.6.9

func (c *Config) Models(ctx context.Context) ([]ModelInfo, error)

func (*Config) Utility added in v0.11.6

func (c *Config) Utility(ctx context.Context, instructions, input string) (string, error)

Utility runs a one-shot completion on the utility model (falling back to the main model) and credits its token usage to the session through the context's usage sink. It backs internal helpers such as fetch page extraction.

type Content

type Content struct {
	Text    string `json:"text,omitempty"`
	Refusal string `json:"refusal,omitempty"`

	// Hidden marks injected context (e.g. background-task notifications) that
	// the model must see but UIs must not render as user input. A user message
	// whose content is entirely hidden becomes a hidden message.
	Hidden bool `json:"hidden,omitempty"`

	File *File `json:"file,omitempty"`

	Reasoning *Reasoning `json:"reasoning,omitempty"`

	ToolCall   *ToolCall   `json:"tool_call,omitempty"`
	ToolResult *ToolResult `json:"tool_result,omitempty"`
}

func CloneContent added in v0.10.9

func CloneContent(in []Content) []Content

CloneContent returns an independent copy suitable for retaining after an API call. Content only contains value fields and one level of pointer fields.

type ContextStats added in v0.11.2

type ContextStats struct {
	Model  string
	Window int

	InstructionsTokens int
	ToolsTokens        int
	ToolStats          []ToolStat
	MessagesTokens     int
	MessageCount       int

	LastInputTokens int64
}

ContextStats estimates what occupies the model's context window, by category. Token counts are byte-based approximations (~4 bytes/token); LastInputTokens is the provider-reported figure for the latest request.

func (ContextStats) EstimatedTotal added in v0.11.2

func (s ContextStats) EstimatedTotal() int

type File

type File struct {
	Name string `json:"name,omitempty"`
	Data string `json:"data,omitempty"`
}

type Message

type Message struct {
	Role MessageRole `json:"role"`

	Content []Content `json:"content"`
	Hidden  bool      `json:"hidden,omitempty"`
}

type MessageRole

type MessageRole string
const (
	RoleUser      MessageRole = "user"
	RoleAssistant MessageRole = "assistant"
	RoleSystem    MessageRole = "system"
)

type ModelInfo added in v0.6.2

type ModelInfo struct {
	ID string
}

type ModelOption added in v0.11.6

type ModelOption struct {
	ID        string
	MinEffort string
	MaxEffort string
}

ModelOption is a resolved model for a per-subagent override. MinEffort and MaxEffort bound the reasoning efforts the model supports; empty means unbounded on that side.

type Reasoning added in v0.6.2

type Reasoning struct {
	ID string `json:"id,omitempty"`

	Summary string `json:"summary,omitempty"`

	// Part indexes the summary part a streamed delta belongs to; renderers
	// separate parts however suits their medium.
	Part int `json:"part,omitempty"`

	// Content is the provider's opaque (encrypted) reasoning payload, only
	// replayable to the model that produced it. Model tags the producer so the
	// agent loop can purge stale payloads when the session model changes.
	Content string `json:"content,omitempty"`
	Model   string `json:"model,omitempty"`
}

type State added in v0.6.2

type State struct {
	Usage    Usage     `json:"usage"`
	Messages []Message `json:"messages,omitempty"`
	Revision uint64    `json:"-"`
}

func (*State) Load added in v0.6.2

func (s *State) Load(path string) error

func (*State) Save added in v0.6.2

func (s *State) Save(path string) error

type ToolCall

type ToolCall struct {
	ID string `json:"id"`

	Name string `json:"name"`
	Args string `json:"args,omitempty"`
}

type ToolResult

type ToolResult struct {
	ID string `json:"id,omitempty"`

	Name string `json:"name"`
	Args string `json:"args,omitempty"`

	Content string `json:"content,omitempty"`
}

type ToolStat added in v0.11.2

type ToolStat struct {
	Name   string
	Tokens int
}

type Usage

type Usage struct {
	InputTokens  int64 `json:"input_tokens"`
	CachedTokens int64 `json:"cached_tokens"`
	OutputTokens int64 `json:"output_tokens"`

	// LastInputTokens is the input size of the most recent request — the
	// current context occupancy, unlike the cumulative counters above.
	LastInputTokens int64 `json:"last_input_tokens,omitempty"`
}

Directories

Path Synopsis
fs
lsp
mcp

Jump to

Keyboard shortcuts

? : This menu
/ : Search site
f or F : Jump to
y or Y : Canonical URL