daemon

Documentation

Overview

Package daemon wires up and runs the Kram daemon's HTTP server — extracted out of cmd/daemon so the unified cmd/kram binary can start a daemon in-process (a goroutine, not a subprocess) instead of shelling out to a separate binary. Subpackages (store, session, agent, tools, server, gatewayclient, compaction) hold the actual logic; this is just the wiring cmd/daemon and cmd/kram both need.

Index

• Constants
• func Run(ctx context.Context, cfg Config, logger *slog.Logger) error
• type Config

Constants

const TokenFileName = "daemon.token"

TokenFileName is the 0600 file, written next to the daemon's DB, that carries the per-process auth token. A standalone CLI reads it to attach to a running daemon; the single-binary cmd/kram threads the token in-process and only writes this for external attach/debugging.

Functions

func Run

func Run(ctx context.Context, cfg Config, logger *slog.Logger) error

Run opens the store, builds the agent loop, and serves the daemon's HTTP API until ctx is canceled, then shuts down gracefully. It blocks until the server has fully stopped.

Types

type Config

type Config struct {
	Host       string
	Port       int
	DBPath     string
	GatewayURL string
	Model      string // gateway combo used for new messages
	Workspace  string // project root the agent's tools operate within
	MaxTurns   int    // model calls per automatic continuation segment
	// PreferStreaming threads through to agent.Config.PreferStreaming —
	// see that field's own doc comment for the real tradeoff it accepts.
	// This struct's own zero value (false) is buffered, matching
	// agent.Config's own field; cmd/kram's -stream flag defaults to true
	// at that call site so the live indicator's reasoning excerpt works
	// out of the box, but can be turned off per deployment. The concrete
	// reason an opt-out turned out to be necessary, not just
	// theoretical: a large local model whose inference server sends
	// nothing at all (not even a reasoning fragment) during prompt
	// prefill trips router.BoundedPeek's fixed idle timeout well before
	// the first token streams, failing every turn outright — buffered
	// mode has no such window and works fine for the exact same setup.
	PreferStreaming bool
	// PromptProfile threads through to agent.Config.PromptProfile (#130)
	// — "frontier" or empty for compact. cmd/kram derives it from the
	// active combo's configured model names (see resolvePromptProfile);
	// the standalone daemon takes it as a flag.
	PromptProfile string
	// AuthToken is the bearer token the daemon's HTTP surface requires on
	// every route except /health (see server.New). Empty means Run
	// generates a fresh random one at boot. The single-binary cmd/kram
	// passes the same token it hands its in-process CLI, so the two agree
	// without a file round-trip; either way Run writes the resolved token
	// to a 0600 daemon.token file next to the DB so a standalone CLI can
	// attach.
	AuthToken string
	// GatewayClientTimeout caps the whole HTTP call the daemon makes to the
	// gateway. Zero means gatewayclient.DefaultTimeout. cmd/kram sets this
	// to a value derived from the gateway config's provider timeout and
	// longest fallback chain (config.Tunables.ResolvedGatewayClientTimeout),
	// so the client never cuts off a legitimate multi-provider round.
	GatewayClientTimeout time.Duration
	// MaxContextTokens is the model context-window budget that governs when
	// the agent loop compacts history. Zero falls back to
	// compaction.DefaultMaxTokens (via agent.Config). cmd/kram derives it
	// from the active combo's smallest provider context window
	// (config.Config.ComboContextWindow), or the --max-context-tokens flag
	// overrides both.
	MaxContextTokens int
}

Config configures one daemon instance.