llm

Documentation

Index

• Constants
• func BuildCommitMsgPrompt(transcript []TranscriptEntry, diff string) string
• func BuildSuggestPRPrompt(diff, commitSubjects, prTemplate string) string
• func BuildUserPrompt(commitHash, subject string, stats ExplainContext, transcript []TranscriptEntry, ...) string
• func FormatCondensedTranscript(entries []TranscriptEntry) string
• func StripHTMLComments(s string) string
• type ExplainContext
• type GenerateResult
• type GenerateTextResult
• type NarrativeResult
• type SuggestPROutput
  • func ParseSuggestPROutput(raw string) (*SuggestPROutput, error)
• type TranscriptEntry
• type Writer
  • func Claude() Writer
  • func Codex() Writer
  • func Copilot() Writer
  • func Cursor() Writer
  • func Gemini() Writer
  • func KiroCLI() Writer
• type WriterRegistry
  • func NewWriterRegistry(writers ...Writer) *WriterRegistry
  • func (r *WriterRegistry) Generate(ctx context.Context, prompt string) (*GenerateResult, error)
  • func (r *WriterRegistry) GenerateText(ctx context.Context, prompt string) (*GenerateTextResult, error)
  • func (r *WriterRegistry) List() []Writer

Constants

const CommitMsgPrompt = `` /* 990-byte string literal not displayed */

CommitMsgPrompt instructs the model to generate a commit message.

const SuggestPRPrompt = `` /* 832-byte string literal not displayed */

SuggestPRPrompt instructs the model to author a pull request.

const SystemPrompt = `` /* 1247-byte string literal not displayed */

SystemPrompt instructs the model to analyze a development session.

Functions

func BuildCommitMsgPrompt

func BuildCommitMsgPrompt(transcript []TranscriptEntry, diff string) string

BuildCommitMsgPrompt assembles transcript context plus a bounded change summary and diff excerpt for commit message generation.

func BuildSuggestPRPrompt

func BuildSuggestPRPrompt(diff, commitSubjects, prTemplate string) string

BuildSuggestPRPrompt assembles the prompt for PR generation. It reuses the summary+excerpt strategy from commit-msg prompts.

func BuildUserPrompt

func BuildUserPrompt(commitHash, subject string, stats ExplainContext, transcript []TranscriptEntry, diff string) string

BuildUserPrompt assembles transcript, diff, and stats into the prompt.

func FormatCondensedTranscript

func FormatCondensedTranscript(entries []TranscriptEntry) string

FormatCondensedTranscript formats transcript entries into a human-readable format for LLM consumption.

func StripHTMLComments

func StripHTMLComments(s string) string

Types

type ExplainContext

type ExplainContext struct {
	FilesChanged int     `json:"files_changed"`
	LinesAdded   int     `json:"lines_added"`
	LinesDeleted int     `json:"lines_deleted"`
	AIPercentage float64 `json:"ai_percentage"`
	AILines      int     `json:"ai_lines"`
	HumanLines   int     `json:"human_lines"`
	SessionCount int     `json:"session_count"`
	RootSessions int     `json:"root_sessions"`
	Subagents    int     `json:"subagents"`
	TopFiles     []struct {
		Path       string  `json:"path"`
		Added      int     `json:"added"`
		Deleted    int     `json:"deleted"`
		TotalLines int     `json:"total_lines"`
		AILines    int     `json:"ai_lines"`
		HumanLines int     `json:"human_lines"`
		AIPercent  float64 `json:"ai_percentage"`
	} `json:"top_files"`
}

ExplainContext is the subset of explain stats needed for the prompt.

type GenerateResult

type GenerateResult struct {
	Narrative *NarrativeResult
	Provider  string
	Model     string
}

GenerateResult holds the narrative plus metadata about how it was generated.

type GenerateTextResult

type GenerateTextResult struct {
	Text     string
	Provider string
	Model    string
}

GenerateTextResult holds a raw text response plus metadata.

type NarrativeResult

type NarrativeResult struct {
	Title     string   `json:"title"`
	Intent    string   `json:"intent"`
	Outcome   string   `json:"outcome"`
	Learnings []string `json:"learnings"`
	Friction  []string `json:"friction"`
	OpenItems []string `json:"open_items"`
	Keywords  []string `json:"keywords"`
}

NarrativeResult holds the structured LLM response.

type SuggestPROutput

type SuggestPROutput struct {
	Title string `json:"title"`
	Body  string `json:"body"`
}

SuggestPROutput is the structured JSON the LLM must return.

func ParseSuggestPROutput

func ParseSuggestPROutput(raw string) (*SuggestPROutput, error)

ParseSuggestPROutput extracts the structured title/body from the LLM response.

type TranscriptEntry

type TranscriptEntry struct {
	Role     string
	Summary  string
	ToolName string
	FilePath string
}

TranscriptEntry is a lightweight event for the condensed transcript.

type Writer

type Writer interface {
	// Name is the stable identifier used in logs, attribution
	// records, and error messages (e.g. "claude_code", "codex").
	Name() string
	// Model reports the model the writer's CLI will invoke. "unknown"
	// is acceptable for CLIs that don't expose a stable model alias.
	Model() string
	// Find returns an executable path for this writer's CLI, or
	// empty when the binary is not installed on the host. Must be
	// cross-platform (PATH lookup honors .exe on Windows etc.).
	Find() string
	// Generate sends prompt to the writer's CLI subprocess and
	// returns the model's response as text. binPath is the resolved
	// binary location from a prior Find() call. Returns an error
	// the registry uses to fall through to the next writer.
	Generate(ctx context.Context, binPath, prompt string) (string, error)
}

Writer is the per-LLM CLI integration that the WriterRegistry walks in fallback order. Each writer locates its binary, generates a response, and reports its model and display name independently. The registry owns redaction, ordering, and the fallback chain; writers own the subprocess-level details of how their CLI is invoked.

func Claude

func Claude() Writer

Claude returns a Writer for the Claude Code CLI. Used by the composition root to build the production WriterRegistry.

func Codex

func Codex() Writer

Codex returns a Writer for the OpenAI Codex CLI. The composition root places it right after Claude Code in the production fallback order - Claude stays the primary daily driver; Codex is the obvious secondary now that we capture its sessions first-class.

func Copilot

func Copilot() Writer

Copilot returns a Writer for the GitHub Copilot CLI.

func Cursor

func Cursor() Writer

Cursor returns a Writer for the Cursor agent CLI.

func Gemini

func Gemini() Writer

Gemini returns a Writer for the Gemini CLI.

func KiroCLI

func KiroCLI() Writer

KiroCLI returns a Writer for the Kiro CLI.

type WriterRegistry

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

WriterRegistry holds the ordered list of writers tried by Generate and GenerateText. The fallback contract is: redact once, try each writer with a non-empty Find() in order, return the first success; when every writer either skips (no binary) or errors, return a chained error naming each failure in fallback order.

Redaction is performed once before any writer sees the prompt. The redactor field is unexported and seamable from tests in the same package; callers outside the package always get the production redactPrompt.

func NewWriterRegistry

func NewWriterRegistry(writers ...Writer) *WriterRegistry

NewWriterRegistry constructs a registry over the given writers in fallback order. Order matters: the first writer with an available CLI that returns a successful response wins. Production wiring lives in internal/providers/composition.go.

func (*WriterRegistry) Generate

func (r *WriterRegistry) Generate(ctx context.Context, prompt string) (*GenerateResult, error)

Generate is the narrative variant: same fallback contract as GenerateText, but each writer's text response is parsed as a NarrativeResult JSON blob before being returned. A successful CLI response that fails to parse as a narrative is treated as a writer failure (falls through to the next writer); the parse error is chained into the final message.

func (*WriterRegistry) GenerateText

func (r *WriterRegistry) GenerateText(ctx context.Context, prompt string) (*GenerateTextResult, error)

GenerateText sends a redacted, UTF-8-safe prompt to the first available writer and returns its raw text response. Prompt prep runs exactly once regardless of how many writers are tried. Writers whose Find() returns empty are skipped (no subprocess attempt). Writers whose Generate returns an error fall through to the next writer; their error is appended to a chained message returned only when every writer fails. When no writer is even installed, returns a single-line "no AI CLI found" message that enumerates every registered writer so the install hint stays honest as the registry grows.

func (*WriterRegistry) List

func (r *WriterRegistry) List() []Writer

List returns the registered writers in fallback order. Used by the "no AI CLI found" error message to enumerate install candidates, and by health checks that want to inspect the registry shape.