docs: chat task models — OWUI task model pinned to greenpt/glm-5.3-flash, suggestions state

Documents the verified setup: config/env/MCP on the persistent volume,
brand assets via the pinned overlay image, /app/data/env precedence over
cloudron env, the manifest.dockerImage deploy mechanism, image-gen as an
agent tool, and the 2026-10-18 crash-loop root cause (hand-edited manifest).

getting-started.md

@@ -62,6 +62,20 @@ Your usage draws from your monthly allowance; you can see it on the
 If the chat ever says your balance is out, that's the allowance, not an error —
 see "Your allowance" below.
 
+### Add the chat to your phone
+
+The chat works like an app on your phone — no app store needed. Open
+[chat.inference.coop](https://chat.inference.coop) in your phone's browser and
+log in, then:
+
+- **iPhone (Safari):** tap the **Share** button (the square with an arrow),
+  then **Add to Home Screen**.
+- **Android (Chrome):** tap the **⋮** menu, then **Add to Home Screen**
+  (or **Install app**).
+
+You'll get an icon on your home screen that opens the chat full-screen, like
+any other app. Your login stays active — just tap and chat.
+
 ## 3. Get involved in governance
 
 You don't just use the co-op; you help run it. Decisions about models, pricing,

maintenance/chat-task-models.md

@@ -0,0 +1,27 @@
+# Chat task models (titles, tags, suggestions)
+
+Both chat front doors generate small helper outputs — conversation titles,
+tags, retrieval queries — with a model separate from the one answering the
+member. These tasks are pinned to a small, inexpensive model so they never
+bill against the member's chosen (larger) model.
+
+## Open WebUI (chat.inference.coop)
+
+- Title generation and tag generation are **enabled**; follow-up reply
+  suggestions and input autocomplete are **disabled** (October 2026).
+- The task model (`task.model.default` in the Postgres `config` table) is
+  pinned to `greenpt/glm-5.3-flash` — a small, inexpensive model — so titles
+  and tags always run on it regardless of which model the member is chatting
+  with. Before October 2026 it was empty, which meant helper tasks ran on the
+  member's selected model.
+- Config lives in the app's Postgres database (`config` table, key/value rows
+  with JSON values) and is read at startup: change a row, then restart the
+  app (`cloudron restart --app <id>`; note `restart` takes no `--timeout`).
+
+## LibreChat (librechat-test.inference.coop)
+
+- Title generation is configured in `/app/data/librechat.yaml`:
+  `titleConvo: true`, `titleModel: greenpt/glm-5.3-flash`,
+  `titleMethod: completion` — the same small GreenPT model.
+- LibreChat v0.8.8 has no reply-suggestion feature; there is nothing to
+  disable on that front.

maintenance/librechat-sandbox.md

@@ -0,0 +1,156 @@
+# LibreChat sandbox maintenance
+
+The LibreChat evaluation instance at **librechat-test.inference.coop**
+(Cloudron app `4023ccbf-d57a-4500-97df-14de9cf3516e`, appstore package
+`ai.librechat.cloudronapp@1.4.1`, which bundles **LibreChat v0.8.8**).
+
+The single source of truth for the deployment is the repo
+**[co-op/librechat-custom](https://git.inference.coop/co-op/librechat-custom)**
+— its README is the full runbook (config deploy, image build/deploy, rollback,
+env vars, branding overlay, known state). This doc is the operator-facing
+summary.
+
+## Architecture: what lives where
+
+- **App config** — `/app/data/librechat.yaml` (persistent volume, mirrored
+  byte-identically in the repo). Holds the "Inference Co-op" custom endpoint
+  (gateway), the 8-model chat picker, speech, web search, and the two MCP
+  servers. Volume edits deploy via `cloudron exec` (base64), never via the
+  image — the image never touches `/app/data`.
+- **Env file** — `/app/data/env` (persistent volume). **`start.sh` sources it
+  AFTER the process env, so it wins over `cloudron env`** — that is where
+  `APP_TITLE="Inference Co-op"`, `ALLOW_EMAIL_LOGIN=false` (password login
+  removed; Cloudron SSO is the only login path), and
+  `SCHEDULES_SINGLE_PROCESS=true` must live.
+- **Other env vars** — via `cloudron env`: `TTS_API_KEY`/`STT_API_KEY`
+  (dedicated LiteLLM virtual key, alias `librechat-speech`, $5/30d, audio
+  models only — never the gateway master key), `IMAGE_GEN_OAI_BASEURL` +
+  `IMAGE_GEN_OAI_MODEL` (agent image tool → gateway), `SEARXNG_INSTANCE_URL`.
+- **MCP servers** — `/app/data/mcp/server.py` (docs search, `search_docs`,
+  starts at boot) and `/app/data/mcp/image_server.py` (image generation,
+  `generate_image`, `startup: false`); both mirrored in the repo under
+  `mcp/`, stdlib-only Python.
+- **Branding overlay** (logo, favicons, PWA icons, `<title>`, `custom.css`,
+  `custom2.js`) — baked into a pinned overlay image `FROM` the official
+  appstore image. Deployed tag (last verified 2026-10-19):
+  `registry.inference.coop/co-op/librechat-custom:20261019-fc9b11e`.
+- **Branding CSS/JS** — `custom.css` overrides LibreChat's Tailwind theme
+  variables on `html.light`/`html.dark` using the HOMEPAGE's canonical
+  palette (`--cream #faf8f5`, `--wash #eef4ea`, `--green-deep #5b8c5a`,
+  `--green-mid #6ba86b`, `--ink #2d3327`, `--muted #6b7a62`) and brands
+  the loading flash (`#loading-container`); the login logo is the green
+  pine-dots vector (white-plate variant was invisible on cream — the
+  'flash on load' glitch), cache-busted via `?v=3` in the client bundle.
+  and hides the register/sign-up prompt via
+  `p:has(> a[href*="register"])`; `custom2.js` is a MutationObserver that
+  hides the prompt when React renders it (the auth form renders after
+  `load`). The `interface.theme` block in the yaml remains
+  forward-compatible only (inert on v0.8.8).
+
+## Image generation (two member-facing paths)
+
+1. **MCP tool (default, no key ever)**: the `inference-image` MCP server
+   exposes `generate_image` in the composer's **+** palette — no agent setup,
+   no key prompt. Per-member billing via `customUserVars.IC_API_KEY` (seeded
+   from the portal) + `env: IC_API_KEY: "{{IC_API_KEY}}"`; a system-key
+   fallback (`system-image-fallback`, $5/30d) covers the rare empty-value
+   case. The `customUserVars` gate blocks a not-yet-seeded member until the
+   next sync (≤10 min) — accepted as a short delay over billing everything
+   to the system account.
+2. **Agent tool**: LibreChat's built-in "OpenAI Image Tools"
+   (`IMAGE_GEN_OAI_BASEURL`/`IMAGE_GEN_OAI_MODEL` → gateway; members paste
+   their own key — v0.8.8 limitation).
+
+## Update policy
+
+- The app is **image-sourced** (the `--force` update moved it off the
+  appstore source), so Cloudron will not auto-update it. This is deliberate —
+  same lockdown model as LiteLLM.
+- **Image deploys**: build `--no-cache`, write the new tag into
+  `CloudronManifest.json.dockerImage` (the only editable field), commit, then
+  `cloudron update --app <id> --force`. **`--image` alone is IGNORED** when
+  the manifest sits in the deploy directory. **The remote builder's cache can
+  serve stale COPY layers even when file content changed** (six consecutive
+  "Build successful" runs did on 2026-10-19) — so after every image deploy,
+  verify: the Cloudron API's `manifest.dockerImage`, container-file md5 vs
+  repo md5, the served HTML, and the rendered DOM. Full checklist in the
+  repo README.
+- **When a new appstore package ships**: pull its manifest copy into the repo
+  (`GET /api/v1/apps/<id>` → `.manifest`), repoint `dockerImage` to the
+  rebuilt overlay, bump the Dockerfile `FROM` to the new `manifest.dockerImage`
+  (pinned), rebuild, verify in the sandbox, then update.
+- **Rollback**: repoint `CloudronManifest.json.dockerImage` to the previous
+  tag and rerun `cloudron update --force` (or the original appstore image).
+  Cloudron also snapshots the app before every image update.
+
+## Hard-won pitfalls (do not re-learn these)
+
+- `cloudron update` deploys **`manifest.dockerImage`** from the manifest in
+  the deploy directory — the `--image` flag alone is ignored (and the update
+  output still prints the `--image` value, so it *looks* right). The repo
+  manifest is a byte-identical copy of the OFFICIAL appstore manifest except
+  `dockerImage`; editing anything else in it (or hand-writing one that omits
+  the addons) strips `postgresql`/`mongodb`/`oidc`/`sendmail`/`localstorage`
+  from the app and crash-loops it (`mkdir: Read-only file system`). That is
+  what broke the sandbox on 2026-10-18.
+- The remote builder's layer cache can pin COPY steps to old file content
+  even across "successful" rebuilds — use `--no-cache` after asset edits and
+  confirm by comparing container-file md5 against the repo file.
+- Deploy `librechat.yaml` only via base64 round-trip. A `sed` insert once
+  broke the YAML indentation and LibreChat **silently fell back to defaults**
+  (all custom branding/models/MCP gone) with a single log line.
+- Reading large files from the container over `cloudron exec` with a plain
+  `cat` resets the connection (`ECONNRESET`) — pipe through
+  `gzip | base64 | fold` and decode locally.
+- `cloudron env set` alone does NOT disable the password form —
+  `ALLOW_EMAIL_LOGIN` must be flipped in `/app/data/env` (precedence above).
+- Image generation in LibreChat is the **MCP tool** (composer **+** palette);
+  the agent "OpenAI Image Tools" path also exists. Image models in the chat
+  picker were the source of the old 500 errors and are deliberately excluded.
+- No `branding:` block exists in LibreChat's config schema, and the login
+  page has no `.login-page`/`.primary` classes to target — theming happens by
+  overriding the Tailwind CSS variables on `html.light`/`html.dark`
+  (`--surface-primary`, `--accent-primary`, …) from the overlay `custom.css`.
+
+
+## Per-member key auto-load (no API-key prompts)
+
+Members never enter an API key. The portal provisions each member's LiteLLM
+key at onboarding; the sandbox seeds it into LibreChat's user-key store:
+
+- `GET /admin/librechat/keys` on the portal returns active members' chat keys
+  (dedicated `X-Sync-Token` auth — `LIBRECHAT_SYNC_TOKEN` on both apps).
+- `/app/data/sync-keys.js` (in the LibreChat container, mirrored in the repo
+  as `sync-keys.js`) fetches that feed and seeds the `keys` collection
+  (chat endpoint, `apiKey: "user_provided"`) and `pluginauths`
+  (`mcp:inference-image` → `IC_API_KEY`, injected per-user via
+  `customUserVars` so image generation bills to the member's own team).
+- Members without a sandbox account are skipped until they first log in; run
+  the sync after onboarding waves and on a schedule (the command is in the
+  repo README). Idempotent.
+- Encryption matches LibreChat's own (`CREDS_KEY`/`CREDS_IV`, read from
+  `/proc/1/environ` — exec sessions don't inherit the app env).
+
+> Terminal output redacts values after `apiKey:` as `***` even for harmless
+> strings (`user_provided`, `${TTS_API_KEY}`). Probe lengths/counts instead
+> of trusting redacted output.
+
+## Verification (after any change)
+
+```bash
+cloudron logs --app 4023ccbf-d57a-4500-97df-14de9cf3516e --lines 60 | \
+  grep -iE 'custom config|error|invalid|MCP|schedules'
+curl -s https://librechat-test.inference.coop/api/config   # appTitle, emailLoginEnabled
+```
+
+Expected: `Custom config file loaded`, `[MCP][inference-docs]` initialized
+with tool `search_docs`, `inference-image` registered (its `Tools: undefined`
+boot line is expected — per-user startup), `[schedules] engine started`,
+`appTitle: Inference Co-op`, `emailLoginEnabled: false`. The login page shows
+the co-op logo, the co-op favicon, a single "Login with Inference
+Cooperative" (Cloudron SSO) path with no email/password form, **no
+"Don't have an account?" line**, and the cream/green co-op color scheme.
+
+After an IMAGE deploy, additionally run the four-step verification in the
+repo README (API `manifest.dockerImage`, container md5 vs repo md5, served
+HTML tags, rendered DOM colors).

models.md

@@ -16,20 +16,18 @@ provider today is best on both.
 |---|---|---|---|---|
 | **Tinfoil** | *private* | **Strongest — architectural** | Not a green claim | Runs in hardware enclaves (TEEs) with end-to-end encryption (EHBP), verifiable via remote attestation. Neither we nor Tinfoil can read your prompts or responses at inference time. Energy mix is not disclosed. |
 | **GreenPT** | *green* | **Policy-based** | **Strongest — 100% renewable** | Standard (non-enclave) hosting in the EU; privacy rests on GreenPT's commitments and EU data rules, not on cryptography. Compute runs on 100% renewable energy. |
-| **PublicAI** | *public* | Policy-based | Undisclosed | A gateway for **publicly developed, sovereign models** — open-weight models built by national/public programs rather than private labs. Privacy is policy-based (no enclave); energy mix is not disclosed. |
+| **Regolo** | *green* | Policy-based — **zero data retention** | **100% renewable** | EU-hosted (Italy, Seeweb) with a stated zero-data-retention policy: prompts are not stored or reused. Compute runs on 100% renewable energy. |
 
 **Reading the gradations:**
 
-- **Privacy:** *architectural* (Tinfoil) beats *policy-based* (GreenPT, PublicAI). An
+- **Privacy:** *architectural* (Tinfoil) beats *policy-based* (GreenPT,
-  enclave protects you even if the provider is compromised; a policy protects
+  Regolo). An enclave protects you even if the provider is compromised; a
-  you only as far as the provider keeps its word. Both are better than a
+  policy protects you only as far as the provider keeps its word. Both are
-  provider with neither.
+  better than a provider with neither — and Regolo adds a written
-- **Green energy:** *100% renewable* (GreenPT) beats *undisclosed* (Tinfoil,
+  zero-data-retention commitment on top of its policy.
-  PublicAI). GreenPT reports its energy mix; the others do not, so we can't
+- **Green energy:** *100% renewable* (GreenPT, Regolo) beats *undisclosed*
-  claim greenness for them.
+  (Tinfoil). GreenPT and Regolo both run on 100% renewable energy; Tinfoil's
-- **Sovereignty:** *publicly developed* (PublicAI) is its own value — the models
+  mix is not disclosed.
-  come from public/national AI programs, built as public goods rather than
-  proprietary products. It's a third axis, independent of privacy and energy.
 
 Trade-offs are real: the most private option (Tinfoil) isn't the greenest, and
 the greenest option (GreenPT) isn't the most private. The point of the co-op is
@@ -61,30 +59,39 @@ per-token billing). The cached-input rate applies when a model reuses a
 shared prompt prefix (multi-turn conversations, agent loops); most providers
 charge a steep discount for it. Models with `—` don't offer a distinct
 cache-read price. **Vision** means the model accepts images as input
-(screenshots, photos, documents to reason about) — all models output text.
+(screenshots, photos, documents to reason about). **Image generation** is
-None of our current models generate images.
+available via the chat's image feature (the 🎨 button) using
+`regolo/qwen-image`.
 
 The emoji in the model name marks its **value tier** (the same icons used in the chat picker):
 
 - 🔒 **Private** — Tinfoil: architecturally private (TEE enclaves)
-- 🌱 **Green** — GreenPT: 100% renewable energy
+- 🌱 **Green** — GreenPT and Regolo: renewable-powered EU hosts (Regolo adds
-- 📖 **Public** — PublicAI: publicly developed, sovereign open models
+  a zero-data-retention commitment). The 🌱 emoji marks renewable-powered
+  providers, not a specific company.
 
 | Model | Provider | Privacy | Green | Vision | Context | Pricing (in / out / cached) | Notes |
 |---|---|---|---|---|---|---|---|
-| `🌱 greenpt/deepseek-v4.1-flash` | GreenPT | policy | 100% renewable | ✅ | 1M | $0.22 / $1.10 / $0.011 | **Default.** DeepSeek's multimodal reasoning model, renewable energy; TEE twin below. |
+| `🌱 greenpt/deepseek-v4.1-flash` | GreenPT | policy | ✅ 100% renewable | ✅ | 1M | $0.22 / $1.10 / $0.011 | **Default.** DeepSeek's multimodal reasoning model, renewable energy; TEE twin below. |
 | `🔒 tinfoil/deepseek-v4-1-flash` | Tinfoil | architectural | undisclosed | ✅ | 1M | $0.65 / $1.45 / $0.13 | TEE-protected version of the default; image input, tool calling, long docs/agents. |
 | `🔒 tinfoil/glm-5.3` | Tinfoil | architectural | undisclosed | — | 1M | $1.80 / $5.75 / $0.45 | GLM flagship, TEE-protected; strongest private-tier model for agentic coding and long-horizon tasks. |
 | `🔒 tinfoil/gpt-oss-120b` | Tinfoil | architectural | undisclosed | — | 131K | $0.15 / $0.60 / — | Lightweight open-weight reasoning; cheapest private tier. |
-| `🌱 greenpt/glm-5.3-flash` | GreenPT | policy | 100% renewable | — | 1M | $0.11 / $0.44 / $0.022 | Fast GLM model, renewable energy. |
+| `🌱 greenpt/glm-5.3-flash` | GreenPT | policy | ✅ 100% renewable | — | 1M | $0.11 / $0.44 / $0.022 | Fast GLM model, renewable energy. |
-| `🌱 greenpt/glm-5.3` | GreenPT | policy | 100% renewable | — | 1M | $1.10 / $4.40 / $0.275 | GLM flagship — strongest agentic model in the green tier. |
+| `🌱 greenpt/glm-5.3` | GreenPT | policy | ✅ 100% renewable | — | 1M | $1.10 / $4.40 / $0.275 | GLM flagship — strongest agentic model in the green tier. |
-| `📖 publicai/apertus-v1.5-8b` | PublicAI | policy | undisclosed | ✅ | 262K | $0.10 / $0.20 / — | Fully open sovereign model, small/fast, image understanding. |
+| `🌱 regolo/apertus-70b` | Regolo | policy + ZDR | ✅ 100% renewable | ✅ | 65K | $0.46 / $2.44 / — | Fully open sovereign model (Apache-2.0), large/capable, image understanding. |
-| `📖 publicai/apertus-v1.5-70b` | PublicAI | policy | undisclosed | ✅ | 262K | $0.82 / $2.92 / — | Fully open sovereign model, large/capable, image understanding. |
+| `🌱 regolo/qwen3.5-122b` | Regolo | policy + ZDR | ✅ 100% renewable | ✅ | 262K | $1.16 / $4.87 / — | Large vision+reasoning MoE (Qwen 3.5, 122B-A10B); 262K context. |
 
 Vision is verified working through the co-op gateway (base64 image input via
 `/v1/chat/completions`, standard OpenAI format). In the chat UI, just attach
 an image to your message while using a ✅ model.
 
+**Reading the Privacy column:** *architectural* means the protection is built
+into the infrastructure (TEE enclaves) — even the provider can't read your
+prompts. *Policy* means the provider commits, by contract and policy, not to
+store or reuse your data. *ZDR* (zero data retention) is a stronger policy:
+the provider keeps no copy of prompts or outputs at all after the response is
+delivered.
+
 **Same model, two providers:** `glm-5.3` (Z.ai's GLM flagship) is offered both
 on GreenPT (renewable, $1.10/$4.40) and on Tinfoil (TEE-protected, $1.80/$5.75)
 — pick the axis that matters to you. The fast `glm-5.3-flash` variant is
@@ -131,6 +138,22 @@ curl https://gateway.inference.coop/v1/chat/completions \
       ]}]}'
 ```
 
+### Image generation
+
+`🌱 regolo/qwen-image` (Regolo, EU-hosted, zero data retention) generates
+images from a text prompt. In the chat, press the **Integrations** button
+below the message box, choose **Image**, and describe what you want — the
+generated image appears in the conversation. Via the API:
+
+```bash
+curl https://gateway.inference.coop/v1/images/generations \
+  -H "Authorization: Bearer ***" \
+  -H "Content-Type: application/json" \
+  -d '{"model": "regolo/qwen-image", "prompt": "a watercolor painting of a mountain lake", "n": 1, "size": "512x512"}'
+```
+
+The response returns base64-encoded JPEG(s). Billed per pixel generated.
+
 ### About the Apertus models
 
 Apertus is the **Swiss AI Initiative's** open foundation model, built by a