prompt

Documentation

Index

• func Build(opts BuildOptions) string
• type BuildOptions
• type BuildResult
  • func BuildResultFor(opts BuildOptions) BuildResult
• type FileImpact
  • func FileImpactFor(path string, content string) (FileImpact, bool)
• type PreviewMode
• type PriorContextPreview
  • func BuildPriorContextPreview(opts BuildOptions, budgetTokens int) PriorContextPreview
• type RelevantMemoryItem

Functions

func Build

func Build(opts BuildOptions) string

Build assembles a system prompt by reading workspace bootstrap files.

Types

type BuildOptions

type BuildOptions struct {
	WorkspaceDir string   // path to workspace root
	WorkDirs     []string // additional working directories the session can access
	CurrentDir   string   // session's current working directory (overrides WorkspaceDir for relative paths)
	SubAgent     bool     // if true, only inject AGENTS.md and TOOLS.md
	// PlanClarifyMode is one of "smart" / "auto" / "ask" — controls whether
	// the LLM asks clarifying questions before drafting a plan. Empty
	// defaults to "smart". Ignored for sub-agent prompts.
	PlanClarifyMode      string
	Query                string
	SessionID            string
	MemorySearcher       memory.Searcher
	ForceRelevantMemory  bool
	StaticBudgetTokens   int
	RelevantBudgetTokens int
	TotalBudgetTokens    int
}

BuildOptions configures system prompt generation.

type BuildResult

type BuildResult struct {
	Prompt               string
	StaticTokens         int
	RelevantTokens       int
	RelevantMemoryCount  int
	RelevantSection      string
	RelevantMemoryItems  []RelevantMemoryItem
	RelevantBudgetTokens int
	TotalTokens          int
}

BuildResult captures prompt assembly output and budget usage.

func BuildResultFor

func BuildResultFor(opts BuildOptions) BuildResult

BuildResultFor assembles a system prompt and returns budget usage details.

The identity line (\"You are TARS, a personal AI assistant.\") was removed from the hardcoded header in ID-002(a). It now lives in the workspace IDENTITY.md default content, which is loaded as the \"## Identity\" bootstrap section below — that lets users override their assistant’s identity without recompiling. Response Formatting rules and the dynamic \"Current time\" line stay in code: they describe runtime constraints, not user-tunable persona.

type FileImpact

type FileImpact struct {
	Section         string `json:"section"`
	Role            string `json:"role"`
	MaxChars        int    `json:"max_chars"`
	Chars           int    `json:"chars"`
	EstimatedTokens int    `json:"estimated_tokens"`
	WillTruncate    bool   `json:"will_truncate"`
	TruncatedChars  int    `json:"truncated_chars,omitempty"`
}

func FileImpactFor

func FileImpactFor(path string, content string) (FileImpact, bool)

type PreviewMode

type PreviewMode string

PreviewMode distinguishes how the prior-context preview was constructed.

• PreviewModeDefault: ranked recall against opts.Query (same logic as the live LLM prompt path)
• PreviewModeRecent: opts.Query was empty, so the preview shows recent experiences/memory entries as a discoverability fallback

const (
	PreviewModeDefault PreviewMode = "default"
	PreviewModeRecent  PreviewMode = "recent"
)

type PriorContextPreview

type PriorContextPreview struct {
	Mode                 PreviewMode          `json:"mode"`
	Section              string               `json:"section"`
	Items                []RelevantMemoryItem `json:"items"`
	BelowThreshold       []RelevantMemoryItem `json:"below_threshold_items"`
	RecentFallback       []RelevantMemoryItem `json:"recent_fallback_items"`
	RelevantTokens       int                  `json:"relevant_tokens"`
	RelevantBudgetTokens int                  `json:"relevant_budget_tokens"`
}

PriorContextPreview is the structured payload backing the console's "Prior" panel. Unlike BuildResult, it never feeds the LLM prompt — the extra fields (BelowThreshold, RecentFallback) exist to help users understand *why* the live recall produced no results.

func BuildPriorContextPreview

func BuildPriorContextPreview(opts BuildOptions, budgetTokens int) PriorContextPreview

BuildPriorContextPreview produces the preview for the console panel. When opts.Query is empty it falls back to the most recent experiences / MEMORY.md entries so the panel has something to display; the live LLM prompt path explicitly does not do this fallback.

type RelevantMemoryItem

type RelevantMemoryItem struct {
	Source    string `json:"source"`
	SourceTag string `json:"source_tag"`
	Snippet   string `json:"snippet"`
	Tokens    int    `json:"tokens"`
}

RelevantMemoryItem is the structured form of one line injected into the "## Prior Context" prompt section.