Documentation
¶
Overview ¶
Package pipeline provides types and configuration for stage-based pipeline execution. The legacy middleware-based pipeline has been removed in favor of the stage architecture. See runtime/pipeline/stage for the current implementation.
Index ¶
Constants ¶
This section is empty.
Variables ¶
This section is empty.
Functions ¶
This section is empty.
Types ¶
type Config ¶
type Config struct {
Stages []string `json:"stages"` // Pipeline stages in order
}
PipelineConfig represents the complete pipeline configuration for pack format
type ExecutionResult ¶
type ExecutionResult struct {
Messages []types.Message `json:"messages"`
Response *Response `json:"response"`
Trace ExecutionTrace `json:"trace"`
CostInfo types.CostInfo `json:"cost_info"`
Metadata map[string]interface{} `json:"metadata"`
PendingTools []tools.PendingToolExecution `json:"pending_tools,omitempty"`
}
ExecutionResult is the output of a pipeline execution.
type ExecutionTrace ¶
type ExecutionTrace struct {
LLMCalls []LLMCall `json:"llm_calls"` // All LLM API calls made during execution
Events []TraceEvent `json:"events,omitempty"` // Other trace events
StartedAt time.Time `json:"started_at"` // When pipeline execution started
CompletedAt *time.Time `json:"completed_at,omitempty"` // When pipeline execution completed
}
ExecutionTrace captures the complete execution history of a pipeline run.
type LLMCall ¶
type LLMCall struct {
Sequence int `json:"sequence"` // Call number in sequence
MessageIndex int `json:"message_index"` // Index into messages array
Request interface{} `json:"request,omitempty"` // Raw request (if debugging enabled)
Response interface{} `json:"response"` // Parsed response
RawResponse interface{} `json:"raw_response,omitempty"` // Raw provider response
StartedAt time.Time `json:"started_at"` // When call started
Duration time.Duration `json:"duration"` // How long the call took
Cost types.CostInfo `json:"cost"` // Cost information for this call
ToolCalls []types.MessageToolCall `json:"tool_calls,omitempty"` // If this call triggered tool execution
Error *string `json:"error,omitempty"` // Error message if the call failed
}
LLMCall represents a single LLM API call within a pipeline execution.
type Response ¶
type Response struct {
Role string `json:"role"`
Content string `json:"content"`
Parts []types.ContentPart `json:"parts,omitempty"`
ToolCalls []types.MessageToolCall `json:"tool_calls,omitempty"`
}
Response represents the output from a pipeline execution.
type RetryPolicy ¶
type RetryPolicy struct {
MaxRetries int `json:"max_retries"` // Maximum retry attempts
Backoff string `json:"backoff"` // Backoff strategy ("fixed", "exponential")
InitialDelayMs int `json:"initial_delay_ms,omitempty"` // Initial delay in milliseconds
}
RetryPolicy defines retry behavior for provider middleware
type StateStoreConfig ¶
type StateStoreConfig struct {
Store interface{} // State store implementation (statestore.Store)
ConversationID string // Unique conversation identifier
UserID string // User identifier (optional)
Metadata map[string]interface{} // Additional metadata to store (optional)
}
StateStoreConfig contains configuration for state store middleware
type ToolPolicy ¶
type ToolPolicy struct {
ToolChoice string `json:"tool_choice,omitempty"` // "auto", "required", "none", or specific tool name
MaxRounds int `json:"max_rounds,omitempty"`
MaxToolCallsPerTurn int `json:"max_tool_calls_per_turn,omitempty"`
MaxParallelToolCalls int `json:"max_parallel_tool_calls,omitempty"` // Max concurrent tool executions (default 10)
MaxCallsPerMinute int `json:"max_calls_per_minute,omitempty"` // Per-tool calls/minute (0=unlimited)
MaxCostUSD float64 `json:"max_cost_usd,omitempty"` // Cost budget in USD (0=unlimited)
Blocklist []string `json:"blocklist,omitempty"`
// MaxIdenticalToolCalls is the maximum number of times the same tool may be called
// with identical arguments before the loop is aborted as an infinite-loop guard.
// 0/unset means use the default (3). Never set to unlimited — use a large number instead.
MaxIdenticalToolCalls int `json:"max_identical_tool_calls,omitempty"`
// StopOnTool, when non-empty, stops the agent tool loop at the end of any round
// in which a tool with this name was called (RFC 0010 termination.tool_called).
StopOnTool string `json:"stop_on_tool,omitempty"`
}
ToolPolicy defines constraints on tool usage.
func MergeToolPolicy ¶ added in v2.10.0
func MergeToolPolicy(caller *ToolPolicy, prompt *packspec.ToolPolicy) *ToolPolicy
MergeToolPolicy combines the policy a caller configured for a provider stage with the tool_policy declared on the prompt that stage is running.
Where the limits come from. Each layer bounds something different; only the first two bound the same thing, which is why they are merged here:
- prompts.<task>.tool_policy (PromptPack spec) bounds ONE turn of that prompt: max_rounds is LLM → tool → LLM cycles, max_tool_calls_per_turn is tool calls, across all of the turn's rounds.
- compositions.<name>.steps[].termination.max_steps (RFC 0010) bounds the LLM-tool loop of ONE agent step — the same unit as max_rounds, not a count of composition steps. The spec requires a composition's step graph to be acyclic, so a composition never re-executes a step and has no "steps executed" limit; its total is the sum of its steps' loops. The composition executor passes max_steps in as the caller's MaxRounds.
- workflow.engine.budget (max_tool_calls, max_total_visits, max_wall_time_sec) and per-state max_visits bound the WHOLE workflow run, across states and turns. That is where looping lives (the spec says to encode loops there, not in a composition), and it is enforced by the workflow state machine on top of each turn's limits, never merged here.
The merge only ever narrows: every layer can tighten a limit, none can widen one another layer set. So:
- MaxRounds and MaxToolCallsPerTurn: the smaller of the values that are set (0 means unset). An agent step's max_steps therefore runs until the lower of max_steps and the prompt's max_rounds.
- Blocklist: the union.
- ToolChoice: the caller's when set, otherwise the prompt's. It is a mode, not a limit, and the caller (a composition step, an Arena scenario) is the more specific context.
- Every other field is runtime-only (not in the spec) and comes from the caller unchanged.
Returns caller unchanged when prompt is nil, so the result may be nil.
type TraceEvent ¶
type TraceEvent struct {
Type string `json:"type"` // Event type
Timestamp time.Time `json:"timestamp"` // When the event occurred
Data interface{} `json:"data"` // Event-specific data
Message string `json:"message,omitempty"` // Human-readable description
}
TraceEvent represents a significant event during pipeline execution.
type ValidationError ¶
type ValidationError struct {
Type string `json:"type"`
Details string `json:"details"`
Failures []types.ValidationResult `json:"failures"` // All failed validations
}
ValidationError represents a validation failure.
func (*ValidationError) Error ¶
func (e *ValidationError) Error() string
Error returns the error message for this validation error.
Source Files
Directories