Documentation
¶
Overview ¶
Package mood detects interaction *friction* — how a user is engaging with the agent — from their terminal input, using a deterministic, dependency-free lexical heuristic. No model, no network, no surprise deps (consistent with memcode's local-first doctrine and the BM25-over-embeddings choice for recall).
This is NOT a personality judgment about the user. It is a control signal: a rising friction reading (caps, expletives, "still broken", repeated corrections, interrupts, denials) usually means the agent is on the wrong track and should change strategy. And the *intensity* of a direction is part of its meaning — an instruction given with force ("do NOT add a paid vendor") is a stronger, more durable constraint than a calm aside, so memcode records the intensity alongside the direction and weighs it later.
Index ¶
Constants ¶
This section is empty.
Variables ¶
var W = struct { Caps, Expl, Neg, Punct, Rep, RepWord, Density, Combo, Directed, Pos float64 }{ Caps: 0.5, Expl: 0.65, Neg: 0.55, Punct: 0.4, Rep: 0.3, RepWord: 0.25, Density: 0.2, Combo: 0.25, Directed: 0.1, Pos: 0.7, }
W is the tunable feature-weight table for friction scoring — kept in one place (not scattered through conditionals) so the model can be calibrated easily.
Functions ¶
Types ¶
type Cadence ¶
type Cadence struct {
InterMessageMs int64 `json:"inter_message_ms"` // gap since the previous turn (0 = first)
Chars int `json:"chars"` // message length
Words int `json:"words"` //
Burst bool `json:"burst,omitempty"` // arrived right after the previous turn
RapidCorrection bool `json:"rapid_correction,omitempty"` // a quick corrective follow-up → repair
LongPause bool `json:"long_pause,omitempty"` // big gap → possibly a new topic
}
Cadence is the timing shape of a turn — a subtle but real interaction signal. Three short messages fired off in five seconds is a different state than one calm, considered prompt. We capture only DERIVED aggregates (never raw keystrokes — that gets creepy fast): the gap since the last turn, size, and a few booleans the runtime uses for routing + friction.
type CadenceTracker ¶
type CadenceTracker struct {
// contains filtered or unexported fields
}
CadenceTracker derives per-turn timing features from message arrival times.
func NewCadenceTracker ¶
func NewCadenceTracker() *CadenceTracker
NewCadenceTracker returns a tracker using the wall clock.
func (*CadenceTracker) Observe ¶
func (c *CadenceTracker) Observe(text string, corrective bool) Cadence
Observe records a turn's arrival and returns its cadence. corrective signals the turn is a correction/negation (from routing or friction), so a quick follow-up is flagged as a rapid correction (a strong "we're off track" cue).
type Reading ¶
type Reading struct {
Valence float64 `json:"valence"`
Intensity float64 `json:"intensity"`
Frustration float64 `json:"frustration"`
State State `json:"state"`
Signals []string `json:"signals,omitempty"`
}
Reading is the friction charge of one piece of text (or a smoothed aggregate). Valence is -1..+1; Intensity and Frustration are 0..1.
func (Reading) FrictionLevel ¶
Friction returns the gauge level for r's state, and Behavior the strategy.
type State ¶
type State string
State is the interaction state inferred for a turn (or the running aggregate).
const ( Calm State = "calm" // positive / satisfied Focused State = "focused" // neutral, working — the default Curious State = "curious" // exploratory: "what if", "I wonder", "how does X work" Confused State = "confused" // uncertainty, questions, "what / huh / unclear" Frustrated State = "frustrated" // negative, exasperated Angry State = "angry" // peak friction: caps + expletives + blame Urgent State = "urgent" // time pressure: now / asap / immediately Discouraged State = "discouraged" // low-energy defeat about the WORK ("stuck", "give up") )
type Tracker ¶
type Tracker struct {
// contains filtered or unexported fields
}
Tracker keeps a smoothed running friction across a session so the gauge reflects the general interaction rather than one stray message, and detects repeated corrections (the same complaint twice = the agent really isn't listening, which should raise friction).
Source Files