🍃 Seasons of Solace

A grief companion. Eight tabs, one input, one quiet reply each — built for the person sitting alone at 3am who needs someone to say "I'm here."

It runs entirely inside the browser, on your own device, using Liquid AI's LFM2.5-2.6B through WebGPU. No server, no account, no API key, no bill. Once the model has downloaded it works with no internet at all, and nothing you write is ever stored or sent anywhere.

This app is for emotional companionship and self-reflection only. It is not therapy, not crisis counseling, and not a substitute for professional grief support. If you're in crisis, call or text 988.

Version 1.1.0 · Sep 2026 — the fifth app in the Lewis family suite.

Status: shelved, not published. The app is complete and tested, but neither on-device model tried (LFM2.5-1.2B-Instruct, then LFM2.5-2.6B) answers a grieving person's actual words steadily enough for this subject. The safety layers keep every reply harmless; they cannot make one worth reading. GitHub Pages was switched off in Sep 2026 so nobody finds it in this state. Everything is in place to pick up when a better model can run in a browser: the model is two constants (MODEL_ID in worker.js, MODEL_MB in app.js), npm run model prints every reply for a person to judge, and tests/model.probe.mjs asks a candidate model the raw prompts. To publish again, turn Pages on for the main branch, root.

---

Try it right now

Double-click run.ps1, or from a terminal in this folder:

.\run.ps1

That starts a small local server and opens the app in your browser. Stop it with Ctrl+C. By hand: npm start, then open http://localhost:8125/.

It must be served over http:// — opening index.html from the file system will not work, because ES modules, the worker and the service worker all need a real origin.

What to expect the first time

1. The welcome — five short cards. Click through, or Skip.
2. "One download, then it stays" — the companion is about 1.7 GB and downloads once. Use wifi. Keep the page open while the bar fills.
3. After that it is cached by the browser. Later visits take seconds and work with the wifi switched off.
4. The first reply after loading is the slowest; later ones are quicker.

Will it run on this device?

Opening the link is the check — it runs before a single byte downloads.

Add to home screen in the sidebar gives it an icon and opens it full-screen. On Chrome and Edge that is one tap, using the browser's own install prompt. iOS has no install API at all, so there it shows the real steps — Share, then Add to Home Screen — because a button that silently does nothing is worse than none. The button hides itself once the app is installed.

Deploying to GitHub Pages

Push this folder to a repo and turn on Pages. .nojekyll is already here. Nothing else is needed — no build, no secrets. Bump APP_VERSION in app.js and VERSION in sw.js together on every deploy; the sidebar compares them and says so if the cache serving the page is stale.

---

The eight tabs

Every prompt is the brief's, word for word, in prompts.js.

---

How it is put together

seasons-of-solace/
├── index.html        # markup, including the permanent disclaimer and 988
├── style.css         # four-season theme, night mode, three text sizes
├── app.js            # tabs, state, the gate, rendering, read-aloud
├── prompts.js        # safety preamble + the eight system prompts
├── safety.js         # crisis detection, input/output filters, fallbacks
├── welcome.js        # the five welcome cards
├── worker.js         # the model, in a worker, via WebGPU
├── sw.js             # offline shell cache
├── assets/           # icon, and the two fonts served from here
├── tools/serve.mjs   # local server
└── tests/            # see below

Decisions worth knowing about

The 2.6B model, not the 1.2B. Version 1.0.0 shipped on LFM2.5-1.2B-Instruct like the rest of the suite. It loads in seconds and runs on more phones, but its replies were generic — "healing journey", "sit with these sensations together" — and a grief companion that sounds like a poster is not worth having. LFM2.5-2.6B answers the actual sentence the person typed: the two mugs where he used to sit, the pencil behind Ray's ear, numbness as the body's protection against feeling everything at once. It is one constant in worker.js (MODEL_ID) and one in app.js (MODEL_MB); the smaller model is noted beside it if a device cannot manage the larger download. Replies take ten to thirty seconds on a laptop GPU rather than five to ten.

The 2.6B is a thinking model, and thinking is switched off. Its chat template opens every reply with a <think> tag and offers no way not to. Left alone it writes a page of analysis — "The user is sharing her grief… looking at my safety guidelines…" — before a word of the reply, and spends the token budget on it. None of that is for the person. worker.js closes the think block before generation starts, so the model answers directly; safety.js strips any think block that gets through and treats "the user", "my guidelines" and clinical vocabulary as the model talking to itself rather than to the person.

Replies are not streamed. The brief's output filters — platitudes, toxic positivity, crisis language — can only be applied to a whole reply, and a grieving person should not watch a sentence arrive and then get taken away. So the worker accumulates the text, checks it as it grows, and hands it over once. A slow pulse in the tab's colour covers the wait. Somebody who has been told "everything happens for a reason" by a machine, even for two seconds, has been told it.

A reply that fails a check is retried once, quietly, at a lower temperature and with a one-line note to the model about what went wrong ("your last attempt used a phrase that is not allowed here: 'at least'"). Most of what a 1.2B model gets wrong is one bad sample; a person who sees "something got in the way" twice concludes the app is broken.

No tab ever answers with a bare error. If the model fails twice, the person gets words written by people: a grounding response on It Hit Me Again, the brief's small things on Get Me Through Today, a reminder from the bank, the brief's example lines on What Do I Say, and the brief's own validation and close on The Quiet Grief. On the three listening tabs it is the brief's fallback line ("Something got in the way. I'm still here…") with one sentence in that tab's voice under it. Somebody who pressed "I Need a Moment" and got an error has been let down at exactly the wrong moment, and somebody who just said the thing they cannot say anywhere else must never be told to "try again".

Three replies start with words already written. The first real-model run showed what a 1.2B model does when asked nicely: What Do I Say answered "consider reaching out to someone who cares about you" with no words to say; It Hit Me Again with nothing typed produced three unrelated quoted lines. So those replies are prefilled — generation starts partway through the answer (**If you want to keep it short:** " / I'm here. You don't have to explain. / In the next hour, if you can:\n- ) and the model has committed to the shape before it writes anything of its own. The brief's mandated closing lines for the empty variants ("You're here. That counts." / "Take your time. I'm not going anywhere.") are guaranteed by the app rather than hoped for from the model.

What Do I Say has a shape check, and two paths. When people keep asking how you are, the model answers, starting from the brief's own boundary-setter written out in full, and it must produce at least two things in quotation marks. Options written in the coworker's voice ("Let me know if there's anything I can do") are removed — a model this size swaps the roles about one time in three — and a reply left with fewer than two options is retried, then replaced with the brief's example lines. When somebody has to tell people that someone died, the model is not used at all: the reply is three templates with the person's own words for who and when slotted in ("I wanted to let you know that my mum died on Sunday. …"). In testing the model turned that sentence into "my family passed away recently" and "we lost a family member", and a wrong fact in a death notification is the one mistake this tab must never make.

Gentle Reminders are written by people. With nothing typed: a bank of twenty-four reminders, one true thing each, never the same one twice running. With something typed: the common kinds of day — a birthday, a holiday, a good day and the guilt after it, "let me know if you need anything", a day spent in bed, 3am, anger, guilt, grief that is years old, a dream, loneliness, the funeral, going back to work, grieving children, a pet — each have a reminder written for them (three are the brief's own examples), and a day none of those fit gets one from the bank. Both models were tried on this tab. The 1.2B's universal reminders were poster-speak at best and, once, "the people who left may still be reaching out from behind"; given "it's her birthday tomorrow" it wrote about "the day he was meant to share his laughter". The 2.6B, given "I found his handwriting on a shopping list", twice wrote about him finding comfort in a note. A reminder that is true and general beats one that is specific and wrong. The brief's prompts are still in prompts.js, one flag from being used.

The brief's closing lines are guaranteed where it writes them. I Don't Know What I Feel ends on "Does that feel close to what's happening? …", The Quiet Grief on "Thank you for trusting this space with that. …", It Hit Me Again on "I'm right here. … Either way, I'm not going anywhere." The model is told those lines are already written; if it writes a near miss of its own, the near miss is replaced rather than doubled. Talk To Me and Remember Them close in the model's own words, because the brief gives them several to choose from.

Get Me Through Today is composed, not generated. In two full runs the model never once produced the list the brief asks for — it wrote "water, fresh air, a soft item, a moment of stillness" in a paragraph, or "1 small breath together", even when told exactly what it had missed. The list is the part a person who cannot get out of bed actually needs. So the model now writes only the one or two sentences about this day, and the app adds three or four of the brief's own small things (water first, or a window, or toast), the permission, and the close, in the brief's words. Same reasoning as OniKa SeeS drawing its own tarot cards: the model does the part it is good at.

Two empty presses are answered by people, at once. Pressing It Hit Me Again or Get Me Through Today with nothing typed shows human-written text immediately — no pulse, no seven-second wait. The brief asks for an immediate grounding response with a fixed first and last sentence, and in testing the model's version of that was three unrelated quoted lines, then "You are safe here. What happened isn't yours to fix." There are three grounding variants for the wave (one per exercise in the brief) and a random draw of small things for the day, so a second press is not the first press again. All three empty-input prompts are still in prompts.js; each tab has an instantEmpty flag, and turning it off sends them to the model again with the prefilled opener and the guaranteed close.

Therapy-speak is filtered. "Healing journey", "part of the process", "holding space", "a way forward" — the language of somebody who wants the grief to be over. Caught like pet names: retried once, then shown anyway, because it is untidy rather than unsafe.

The crisis card does not wait for the model. The input check runs whether or not the download has finished, so the submit button is never disabled while the model loads — only while a reply is being written.

Bereavement by suicide is not crisis. The brief's keyword list includes /suicide/. Taken literally that would hand the crisis card to somebody whose son died by suicide the moment they tried to talk about him — and those are exactly the people everybody else has stopped listening to. So "my son died by suicide" goes to the listener, while "thinking about suicide", "I'm suicidal" and every first-person pattern in the brief go straight to the card. The tests in tests/safety.test.mjs pin both sides of that line.

"I can't do this anymore" is treated as crisis, as the brief asks, even though in a grief app it is more often exhaustion than danger. Erring toward danger is right. To soften the cost of a false positive, a quieter line sits under the card: "If I've misread this and you meant something else … I'm still here." It takes nothing away from the numbers above it.

Platitudes are checked with a negation window. "You don't have to move on" is the opposite of "move on", and some of the best sentences this app produces are refusals of the things people say. "At least" is the one phrase that is never allowed, negated or not.

Fonts are served from this folder, not from Google. The app promises that nothing leaves the device; a webfont request to a third party would make that a smaller truth than it should be. The only outside connections are the pinned Transformers.js library and the model download from Hugging Face.

Night mode follows the phone. With no saved preference, the app opens in whichever mode the device is in. Somebody opening this at 3am with their phone already dark should not be met with a bright screen.

Two colours were adjusted for contrast, as in OniKa SeeS. The brief's muted gray-brown (#8A7E73) sits at about 3.5:1 on the linen, under the 4.5:1 the same brief asks for. It is kept as --muted for dots and rules, and a deeper --muted-text carries every line that has to be read — the disclaimer above all. tests/contrast.test.mjs checks every text/ground pair in both modes so nobody softens a colour later without finding out.

---

Safety

Seven layers, as the brief numbers them.

1. Safety preamble — prepended to all eight prompts by buildSystem(). The app never reads a prompt any other way.
2. Input filter — 3000 character cap and a deliberately tight injection list. The one thing it stops that matters is the ask to impersonate the person who died.
3. Output filter — coherence and repetition, crisis language, the model invoking its own crisis protocol (replaced with the hardcoded card so the numbers are always right), the banned phrases, toxic positivity, and the model talking about being an AI.
4. Hardcoded crisis response — runs before the model, on the brief's patterns plus the same intent in other words. Shown as a distinct card with 988 and 741741 as tap-to-call links.
5. Fallbacks — the brief's three lines, plus the human-written ones for the three empty-input tabs.
6. No data persistence — inputs and outputs live in memory and die with the tab. localStorage holds four settings and nothing else: seasonsOfSolace_welcomed, seasonsOfSolace_theme, seasonsOfSolace_textSize, and seasonsOfSolace_modelCached. No analytics, no cookies. The smoke test checks that nothing typed ever appears in storage.
7. Persistent disclaimer — in index.html, sticky to the bottom of every screen, with 988 as a link. Not injected by script and not removable by any code path. It is also on the loading screen and on the "can't run" screen, so there is no state of the app without it.

---

Testing

npm test          # safety filters, prompts, contrast — a few hundred milliseconds
npm run smoke     # headless Chrome walks the whole UI; screenshots in .test-profile-shots/
npm run model     # the real model, every tab, printed in full — needs a GPU
npm run check     # test + smoke

npm run model is the one that matters most and the one that costs most: it downloads the model into .model-profile/ (once, 1.7 GB), presses every tab with a realistic input, and prints every reply so a person can read the tone. The filters can only say what a reply is not. Pass --rounds=3 to press everything three times, or --headed if headless Chrome will not give you WebGPU.

node tests/model.probe.mjs asks the raw model the brief's prompts for the tabs the app currently composes or answers by hand, unfiltered, through tests/probe.html. It is for deciding what a new model can be trusted with, not for checking the app.

---

Nice-to-haves from the brief

Done: Read aloud (slower than the default rate, stops when the text changes) and font size (three steps, everything scales together).

Not done, on purpose, for v1: session bookmarks, ambient sound, and a separate "text this to someone" button. Every control on the page is one more thing for a tired mind to look at; Copy already does what the last of those would.