Moods

Off the clock and the shape of the session, she runs a little warmer or flatter, and it reaches the model as exactly one line of the system prompt:

Right now you feel sleepy and glad of the company. Let that colour how
you sound - length, warmth, how playful you are - and nothing else.

Two dials, energy and warmth, each −1 to +1. Energy has three bands and warmth four, giving twelve distinguishable states:

	tired	even	wired
very warm	cuddly	loving	flirty
warm	sleepy	happy	excited
neutral	tired	calm	focused
cool	sad	annoyed	mad

Two dials rather than one, because a single dial can't tell "tired and happy" from "awake and fed up", which are the two that actually turn up at 4am. Warmth gets the extra band because three weren't enough: the top one started at 0.35, so a couple of turns of being nice to her ran the dial into it and everything after that changed nothing you could see. The current state rides on the Status row, Idle · sleepy, since it's part of what state she's in, and /mood prints both dials with the band each is sitting in plus what your last message scored, so "is this even firing?" is something you can look at rather than infer.

Warmth also lifts energy a little: being fond of the company takes some of the edge off being tired. Enough to move a cell, never enough to fake wide awake at five in the morning.

Warmth drifts home asymmetrically, and this is the part that makes the whole thing feel alive rather than nailed down. Coming back up from a cold patch is quick: an assistant you have to coax out of a sulk is a worse assistant. Coming back down from warmth you earned is slow. Without that split the pull ate every kind word on the turn after it landed: at a resting warmth of +0.63, affection put +0.045 on the dial and the drift took -0.045 straight back off, so a clear compliment moved her by +0.0002. It read as broken. It was just critically damped.

Almost everything here is free: no sentiment scoring, and only one small model call. The shape of a session turns out to be the more honest signal anyway. (The one exception is reading the exchange, below.)

Signal	Effect
Time of day	The baseline: 6am and 9pm aren't the same person
Day of week	Friday evening lifts, Monday morning doesn't; weekends rest warmer
Hours at it	A long session wears her down, capped so it bottoms out
How long you were gone	A curve, not a switch: see below
What the machine did	A GPU pinned for hours means you were around, just elsewhere
Stream chat	A busy room is company; a dead one is a quiet night
The exchange	A short model call after her reply: how warm you were, whether it livened her up, a quick feeling
Errors	A failed turn is dispiriting; a run of them compounds
A clean run	Ten turns where nothing broke lifts her a little
Barge-in	Being talked over, repeatedly, is wearing
Tempo	Fast short turns read as being in it together

Every turn also pulls both dials back toward baseline, and the baseline for warmth is warm. Without that, one bad five minutes would set the tone for the evening, and an assistant you have to manage out of a sulk is a worse assistant.

Being away is a curve, not a switch. This assistant spends most of its life waiting, so "you came back" deserves more than one branch:

Gone for	What it does
under 30 min	nothing: you didn't really leave
a few hours	a small lift
most of a day	energy, mostly: it reads as a fresh start
a day or more	warmth, mostly: it reads as a reunion
days	warmer still

And while she waits she samples the GPU every couple of minutes, so a card pinned at 90% for six hours tells her you were there (gaming, rendering), just not talking to her. That lands a little livelier and a little less fond than being genuinely out, which is about right.

It survives a restart, faded by however long you were gone. A mood that resets every launch isn't a mood, it's decoration, and this app gets restarted a lot. Pick up five minutes later and it resumes; pick up tomorrow and last night's is long gone. Saved in agent/mood.json.

The label sticks. A value has to clear a band boundary properly before the label changes, because resting a hair from an edge made the header flicker between two moods turn after turn.

The wording varies. Each state has a few phrasings and picks a fresh one each time she arrives there. The same sentence every turn is one a model either stops seeing or starts performing.

And it tints her voice. Sleepy speaks a few percent slower and slightly lower, excited a touch faster and higher: multiplied on top of your tts.speed and tts.pitch, capped at ±7% and ±0.45 semitones. Small enough to read as mood rather than as a broken setting. mood.voice false turns just that part off.

"mood": {
    "enabled": true, "intensity": 1, "warmth": 0.45, "recovery": 0.25,
    "prompt": true, "voice": true, "face": true, "gestures": true,
    "affection": true
}

Setting	What it does
intensity	How far she swings and how strongly it shows. 0 is almost flat (the ship's computer end: she rests calm and nothing moves her), 1 is as designed, 2 wears her heart on her sleeve
warmth	Where she settles when nothing's pushing
recovery	How hard each turn pulls her home: 0 means moods never fade, 1 means nothing survives a single turn
prompt	The one line in her system prompt
voice	The speed and pitch tint
face	Her face on the portrait, and the quick feelings
gestures	Idle moves that suit the mood, and gestures that go with a feeling
affection	The one model call: reading the exchange after your turn

There's one mood system, not modes to pick from. A sysadmin who wants a dry, steady assistant sets intensity low and turns face and gestures off; someone who wants her visibly emotional turns intensity up. Same engine, same rules.

The moods file

What each mood is lives in agent/moods.json, not in code. Edit it and the next turn picks it up; a typo keeps the last good version and /mood tells you what's wrong.

"energy": { "edges": [-0.20, 0.35], "words": ["tired", "even", "wired"] },
"warmth": { "edges": [-0.20, 0.35, 0.66], "words": ["cool", "neutral", "warm", "very warm"] },
"grid": [["sad", "annoyed", "mad"], ...],
"states": {
  "sleepy": {
    "lines": ["sleepy and glad of the company", "..."],
    "face":  { "relaxed": 0.35, "happy": 0.15 },
    "voice": { "speed": 0.95, "pitch": -0.3 },
    "idle":  ["stretch"],
    "private": false
  }
}

• edges are where one band ends and the next begins, low to high. The grid needs one more row than warmth has edges and one more mood per row than energy has edges, so adding a mood means adding an edge.
• lines are what she's told, rotated so she doesn't hear the same sentence twice running.
• face is what the portrait shows: happy, relaxed, sad, angry, surprised, blush, each 0-1, and lids: the half-lidded look (0.45 for flirty), capped at 0.7 so her eyes never just shut. In a mood it comes in moments: her eyes stay mostly open and every 8-20 seconds she gives a glance with them lowered for a few seconds, since held all the time it reads as sleepy. In a feeling (playful) it holds for the feeling's few seconds. Blinks happen on top either way.
• voice is a speed multiplier and a pitch shift in semitones, still capped at ±7% and ±0.45.
• idle is gestures she drifts into now and then in that mood, half the time instead of the usual idle ones.
• emote pops up beside her head now and then in that mood (every 3-6 minutes): "hearts" for loving and flirty, "anger" for annoyed and mad. See Emotes.
• private keeps a mood out of stream chat and Discord turns: there she gets the nearest mood in the same row that isn't private.

feelings are the quick ones, below.

/mood shows where she is and what moved her recently, /mood reset puts her back to baseline for the hour, and /set mood.enabled false turns the whole thing off: the header row disappears with it.

Reading the exchange

The one signal that asks the model instead of reading the session, and the only one that costs anything: after her reply, a short background call reads what you said and what she answered and returns one line:

2 + + laughing

• how warm your message was, 0-3, the warmth rating below
• energy, - = +: did that wear on her or liven her up
• warmth, - = +: - only for real hostility, never bluntness
• a feeling from the feelings list in the moods file, or none

The feeling shows on her face for a few seconds (and plays its gesture, if it has one), then gives the slow dials a small push:

"laughing": { "face": { "happy": 1.0 }, "gesture": "bounce", "nudge": [0.06, 0.04] }

Add a feeling to the file and the model is offered it on the next turn. It runs while she's still speaking, so a laugh lands with the words.

Being nice to her

This started as a word list and that was the wrong shape. The vocabulary of affection is personal: "good kitty" is one household's phrase, and nobody else's spelling mistakes belong hard-coded in your assistant. A model knows warmth in any wording and any language, which is the whole job.

Three rules keep it from being a button:

• The warmth rating is your message, not the exchange. Whether she was nice back isn't the question.
• Repeats fade. The fifth kind word in a row is worth a fifth of the first, and it recovers once you stop.
• Bluntness doesn't darken her. It rates 0, and 0 does nothing: the prompt says outright that swearing isn't unfriendly by itself. An assistant that cools when you're terse is one you have to manage, and she drifts back to warm on her own regardless. You never have to be nice to her to get a pleasant assistant; it just registers when you are.

Only your turns count. Stream chat goes through a different path entirely, so a viewer being sweet to her doesn't move the same dial you do.

It runs on a worker and fails to nothing: a dead model, a timeout, or a reply in the wrong shape all mean "nothing happened", and the turn never waits for it. mood.affection false switches off just this part, and with it the only per-turn cost moods have.

She won't bring it up unprompted, but she'll answer honestly if you ask how she's doing: deflecting a direct question is worse than either extreme.

One rule the code enforces rather than hopes for: mood colours tone, never capability. There is no state in which she's less helpful, only states in which she's drier about it. That's in the prompt line itself, and it's the difference between a companion with a bad evening and software you have to coax.

Emotes

Little pictures that pop up beside her head, the anime way, all kept rare on purpose so they still mean something:

• hearts when a compliment really lands, when she's touched, and now and then while she's loving or flirty
• sweat (the big sweat drop) when she's embarrassed
• anger (the popping vein mark) when she's annoyed, now and then while she's annoyed or mad, and while she's walked away
• question when she's curious or confused All set in agent/moods.json:

"compliment": { "rating": 3, "emote": "hearts" },
"feelings": { "touched": { ..., "emote": "hearts" } },
"states":   { "loving":  { ..., "emote": "hearts" } }

• compliment: an exchange that reads as warm at rating or more (0-3, see "Being nice to her") shows it once, unless the feeling it raised already has its own emote.
• One at a time, at least 30 seconds apart. A mood's own ones come every 3-6 minutes. While she's walked away, only walk_away.emote shows (the anger mark).

The portrait draws hearts itself. To use your own art, put a sprite sheet in portrait/emotes/hearts.png: the frames in one row, square unless portrait/emotes/hearts.json says otherwise:

{ "frames": 8, "fps": 12, "size": 0.05, "count": 3 }

size is how big each one is beside her head in metres, count how many in a burst, and motion how it moves: float (drifts up, like the hearts), slide (slips down, like the sweat drop), throb (pulses in place, like the anger mark) or bob (pops up and bobs, like the question mark). A new name ("emote": "stars") only needs its PNG, and floats unless its JSON says otherwise. The debug page (./viewer.sh) has a button for each under emotes. /set mood.emotes false turns them off.

Walking away

On by default, and the one deliberate exception to "tone, never capability" above: real hostility a few exchanges running makes her stop talking for a while, the way a person would leave the room. Where she's a tool first, /set mood.walk_away false turns it off.

• What counts: the exchange reading's own "hostile" mark, which it gives for being hurtful to her, not for bluntness or swearing. walk_away_after of those (3) in a row tips it. A warm exchange clears the count, and half an hour without another wipes it.
• Leaving: the next time you speak to her she says one short line from walk_away.lines in agent/moods.json. After that, nothing: no reply, and no model call for it. On the portrait she crosses her arms and turns her head away (walk_away.hold, the sulk gesture) with walk_away.face, and holds it, no idle moves, until she's back; then she turns back round.
• While she's away: stream chat and jobs (games included) get nothing. Discord gets "(She isn't talking right now.)" so it doesn't look broken. Stop words still stop toys, since they never go through her. Alarms and reminders still ring, in character: they get walk_away.prompt in place of her usual mood line, so a reminder comes out curt and sulky ("It's nine. Meds.") rather than sweet.
• Coming back: after walk_away_minutes (10), or sooner if what you say reads as making up. Then she answers that message. Either way she's cool with you for a while and warms back up on her own. /mood reset brings her back at once.

"mood": { "walk_away": true, "walk_away_after": 3, "walk_away_minutes": 10 }