Skip to content
Merged
Show file tree
Hide file tree
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension


Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
2 changes: 2 additions & 0 deletions README.md
Original file line number Diff line number Diff line change
Expand Up @@ -61,6 +61,8 @@ system, hardware and compatible builds; `atc engines list` shows what is availab
see [Foundry Local](#foundry-local) for platforms and setup.
- **LiteRT-LM** (`litert-lm`) — Google's `.litertlm` models on supported CPU, GPU or Intel NPU
accelerators; see [LiteRT-LM](#litert-lm) for platforms and setup.
- **Strata** (`strata`) runs pinned Qwen and DeepSeek variants on compatible GPU bundles. See
[Strata on Atomic Server](docs/strata.md) for installation and recovery.
- **stable-diffusion.cpp** (`sd-cpp`) — image generation.
- **audio.cpp** (`audio-cpp`) — speech, music and sound effects; see
[Audio generation](#audio-generation).
Expand Down
4 changes: 2 additions & 2 deletions THIRD_PARTY_NOTICES.md
Original file line number Diff line number Diff line change
Expand Up @@ -10,7 +10,7 @@ do not edit by hand.
| @ai-sdk/provider | 2.0.5 | Apache-2.0 |
| @ai-sdk/provider-utils | 3.0.39 | Apache-2.0 |
| @alcalzone/ansi-tokenize | 0.3.1 | MIT |
| @atomic-chat/core | 0.23.4 | MIT |
| @atomic-chat/core | 0.24.4 | MIT |
| @fastify/busboy | 2.1.1 | MIT |
| @isaacs/fs-minipass | 4.0.1 | ISC |
| @nodable/entities | 3.1.0 | MIT |
Expand Down Expand Up @@ -245,7 +245,7 @@ OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
SOFTWARE.
````

## @atomic-chat/core 0.23.4
## @atomic-chat/core 0.24.4

License: MIT

Expand Down
7 changes: 2 additions & 5 deletions bun.lock

Some generated files are not rendered by default. Learn more about how customized files appear on GitHub.

14 changes: 7 additions & 7 deletions docs/commands.md
Original file line number Diff line number Diff line change
Expand Up @@ -31,7 +31,7 @@ Generated by `npm run docs:commands` from the command specs. Do not edit by hand
| `atc models check` | Each managed engine's verdict on a Hugging Face model, before downloading it | works |
| `atc models files` | List downloadable GGUF files | works |
| `atc models list` | Installed models | works |
| `atc models pull` | Plan and start a durable download: GGUF, or with --engine an MLX or managed checkpoint | works |
| `atc models pull` | Plan and start a durable model installation | works |
| `atc models import` | Validate and register a local GGUF | works |
| `atc models logs` | The log of a loaded managed model, or of its last failed load | works |
| `atc models info` | Model capabilities on its engine | works |
Expand Down Expand Up @@ -586,7 +586,7 @@ Commands:
check Each managed engine's verdict on a Hugging Face model, before downloading it
files List downloadable GGUF files
list Installed models
pull Plan and start a durable download: GGUF, or with --engine an MLX or managed checkpoint
pull Plan and start a durable model installation
import Validate and register a local GGUF
logs The log of a loaded managed model, or of its last failed load
info Model capabilities on its engine
Expand Down Expand Up @@ -635,7 +635,7 @@ Usage: atc models search [query] [options]
Search model repositories

Arguments:
[query] Search words or owner/repo (optional for foundry-local, litert-lm and geniex)
[query] Search words or owner/repo (optional for engine catalogs)

Options:
--engine <provider> Engine provider ID
Expand Down Expand Up @@ -730,19 +730,19 @@ Global options:
```
Usage: atc models pull [model] [options]

Plan and start a durable download: GGUF, or with --engine an MLX or managed checkpoint
Plan and start a durable model installation

Arguments:
[model] owner/repo[:file.gguf], or owner/repo[@revision] with --engine

Options:
--engine <engine> Engine provider ID for an engine-specific model download
--detach Start a managed download without following it
--refresh FastFlowLM: review the latest catalog for a model update
--refresh FastFlowLM or Strata: review the latest catalog for a model update
--import-dir <value> Strata: import a complete local source directory on the server
--device <device> Foundry Local: the variant for npu, gpu or cpu
--backend <backend> Device for LiteRT-LM or GenieX, or an exact GenieX variant ID
--cancel <request-id> Cancel a managed, Foundry Local, LiteRT-LM or GenieX download (with its
--engine)
--cancel <request-id> Cancel an engine-specific download (with its --engine)
--file <value> Exact filename in the repository
--quant <value> Select an unambiguous quantization
--mmproj Include required vision projector (default: true)
Expand Down
11 changes: 11 additions & 0 deletions docs/decisions/2026-11-17-credit-strata-resume-bytes.md
Original file line number Diff line number Diff line change
@@ -0,0 +1,11 @@
---
date: 2026-11-17
title: "Pass reusable Strata staging bytes to the core"
---

# 2026-11-17 — Pass reusable Strata staging bytes to the core

- **Context.** The normal compatibility check can refuse a resumed transfer when only the remaining sources fit on disk.
- **Decision.** After a disk refusal, count reusable staged bytes for the same pinned job and request a fresh compatibility check with `staged_bytes`. Count only durable ranges, matching single-stream partials, and hash-verified complete files.
- **Consequences.** The core keeps responsibility for preparation space and memory decisions. Core 0.24.3 credits `staged_bytes`. Core 0.24.2 retains its original refusal.
- **Alternative.** Counting the length of a sparse file could credit bytes that the downloader would still have to write.
11 changes: 11 additions & 0 deletions docs/decisions/2026-11-17-journal-strata-model-install.md
Original file line number Diff line number Diff line change
@@ -0,0 +1,11 @@
---
date: 2026-11-17
title: "Journal Strata acquisition and guard model replacement"
---

# 2026-11-17 — Journal Strata acquisition and guard model replacement

- **Context.** A Strata model needs a pinned input set and a separate runtime preparation that can outlive an admin or CLI connection.
- **Decision.** Keep the reviewed descriptor and request ID in the server's setup journal. Publish only verified source files. Reuse the core's preparation operation and its revision across client reconnects. When replacing sources, hold the core maintenance guard and keep a backup until the new preparation succeeds or the previous installation is restored.
- **Consequences.** An interrupted transfer resumes with the original hashes. If the core cannot confirm that preparation stopped, the server keeps the guard and both sets of files for recovery. A local import is copied, so deleting an installation never removes its source directory.
- **Alternative.** Writing downloads into an installed model directory would let a concurrent load observe mixed source sets and would leave nothing to restore on a failed update.
11 changes: 11 additions & 0 deletions docs/decisions/2026-11-17-use-core-strata-sources.md
Original file line number Diff line number Diff line change
@@ -0,0 +1,11 @@
---
date: 2026-11-17
title: "Read Strata sources through the embedded core and workspace"
---

# 2026-11-17 — Read Strata sources through the embedded core and workspace

- **Context.** Atomic Server embeds the core. The engine manifest belongs to the core, and the server reads the model catalog to plan model downloads.
- **Decision.** Pin the released core 0.24.0. Pass the server environment through to the core, including `ATOMIC_STRATA_MANIFEST_URL`. Use `ATOMIC_MODEL_CATALOG_URL` as the workspace fallback when `ATC_MODEL_CATALOG_URL` is unset, so a single setting can point both readers at a local catalog.
- **Consequences.** Existing workspace overrides keep their precedence. Both overrides accept absolute file URLs at server startup.
- **Alternative.** Making the core's catalog variable override the workspace variable would silently redirect servers with an explicit workspace catalog.
13 changes: 13 additions & 0 deletions docs/decisions/2026-11-18-align-web-engine-model-feedback.md
Original file line number Diff line number Diff line change
@@ -0,0 +1,13 @@
---
date: 2026-11-18
title: "Share desktop engine marks and align model feedback"
---

# 2026-11-18 — Share desktop engine marks and align model feedback

- **Context:** The Web engine picker used one generic icon for several engines. A catalog verdict sat against the variant card, and the loaded label aligned with the top of its neighboring button.
- **Decision:** Use Atomic Chat's engine marks and initial-letter fallback in the picker and onboarding. Add breathing room below the variant controls and show model load progress and success in the same inline status slot.
- **Consequences:** Engine identities match across the two Web entry points. Compatibility reasons still come from the server and the loaded label still follows its service snapshot.
- **Alternatives:** Drawing new engine marks would diverge from Atomic Chat. Moving the loaded label to a separate line would move the content as model state changes.
- **Owner:** team.
- **Links:** `packages/admin-ui/LIFTED.md`, `packages/admin-ui/src/workspace/server/ServerModel.tsx`.
11 changes: 11 additions & 0 deletions docs/decisions/2026-11-18-filter-strata-catalog-on-server.md
Original file line number Diff line number Diff line change
@@ -0,0 +1,11 @@
---
date: 2026-11-18
title: "Filter the Strata catalog on the server"
---

# 2026-11-18 — Filter the Strata catalog on the server

- **Context.** Web and terminal clients can connect to a server with a different GPU. The Strata catalog contains separate Metal and CUDA recipes.
- **Decision.** Filter by the active or installable backend reported by the embedded core, then group the matching quantizations. Keep exact variant IDs in plans and retain the core's disk and memory verdicts.
- **Consequences.** Clients see the same variants for the server they are connected to. If the core has no active or installable backend, there are no browse results. Saved plans and their resume flow still use the pinned variant.
- **Owner.** Team.
13 changes: 13 additions & 0 deletions docs/decisions/2026-11-18-theme-web-chat-composer.md
Original file line number Diff line number Diff line change
@@ -0,0 +1,13 @@
---
date: 2026-11-18
title: "Use the active Web theme for the chat composer"
---

# 2026-11-18 — Use the active Web theme for the chat composer

- **Context:** The Web composer used fixed mint surfaces in both themes. The default textarea focus outline extended into the toolbar.
- **Decision:** Use the workspace's themed surface, text and action colors. Draw the textarea focus indicator on the composer perimeter and keep each toolbar button's own focus indicator.
- **Consequences:** The Web composer follows theme changes alongside the rest of the workspace. Terminal colors keep their existing values.
- **Alternatives:** Adding separate composer colors for each browser theme would duplicate the existing workspace palette. Keeping the inner outline would require extra spacing between the field and the toolbar.
- **Owner:** team.
- **Links:** `packages/admin-ui/src/workspace/chat/chat-composition.css`, [Share semantic colors between Web and the terminal](2026-10-24-share-interface-colors.md).
12 changes: 12 additions & 0 deletions docs/decisions/2026-11-18-use-published-core-integrations.md
Original file line number Diff line number Diff line change
@@ -0,0 +1,12 @@
---
date: 2026-11-18
title: "Use the published core integrations API"
---

# Use the published core integrations API

- **Context:** Atomic Server initially exposed core's integrations and Muse launch arguments through a patch to the npm package.
- **Decision:** Pin the published core 0.24.2, which exports the integrations module and accepts the embedding endpoint in the Muse launch planner.
- **Consequences:** Bun installs the upstream package directly. Agent setup and launch keep using that package's integration logic.
- **Links:** [Configure agents through core](2026-11-17-configure-agents-through-core.md).
- **Owner:** team.
14 changes: 14 additions & 0 deletions docs/decisions/INDEX.md
Original file line number Diff line number Diff line change
Expand Up @@ -11,6 +11,10 @@ add a new one that says which record it supersedes.

## Records

- **2026-11-18** — [Filter the Strata catalog on the server](2026-11-18-filter-strata-catalog-on-server.md) — share host-aware quantization groups between the web and terminal catalogs.

- **2026-11-18** — [Use the published core integrations API](2026-11-18-use-published-core-integrations.md) — pin core 0.24.2 and remove the dependency patch.

- **2026-11-17** — [Configure agents through core](2026-11-17-configure-agents-through-core.md) — use the shared registry for host configuration and launches.

- **2026-11-17** — [Show progress when opening terminal agent setup](2026-11-17-show-terminal-agent-setup-progress.md) — make the full agent card clickable and show the options check.
Expand Down Expand Up @@ -242,4 +246,14 @@ add a new one that says which record it supersedes.
- [2026-11-07 — Reuse the API listener when connecting an agent](2026-11-07-reuse-api-for-connections.md)

- [2026-11-07 — Package a runtime that does not inherit Windows sockets](2026-11-07-prevent-windows-socket-inheritance.md)

- [2026-11-17 — Read Strata sources through the embedded core and workspace](2026-11-17-use-core-strata-sources.md)

- [2026-11-17 — Journal Strata acquisition and guard model replacement](2026-11-17-journal-strata-model-install.md)

- [2026-11-17 — Pass reusable Strata staging bytes to the core](2026-11-17-credit-strata-resume-bytes.md)

- [2026-11-18 — Use the active Web theme for the chat composer](2026-11-18-theme-web-chat-composer.md)

- [2026-11-18 — Share desktop engine marks and align model feedback](2026-11-18-align-web-engine-model-feedback.md)
- **2026-11-15** — [Diagnose CLI chat failures with finite tags](2026-11-15-diagnose-cli-chat-failures.md) — tag client-detected failures without interpreting remote HTTP text as a client diagnostic.
5 changes: 3 additions & 2 deletions docs/release.md
Original file line number Diff line number Diff line change
Expand Up @@ -265,8 +265,9 @@ pwsh scripts/sign-windows-binary.ps1 dist\bin\atc-x86_64-pc-windows-msvc.exe dis

`atc` embeds the core as the npm package `@atomic-chat/core` (repo `atomic-chat-core`, published by its
release workflow) at an **exact** version, never a caret: the protocol and version checks make a silently
newer core unsafe. 0.23.4 is the pin. The Foundry Local and LiteRT-LM host entries (`./foundry-host`
newer core unsafe. 0.24.4 is the pin. The Foundry Local and LiteRT-LM host entries (`./foundry-host`
and `./litert-host`) run before command parsing in `src/bin.ts`, including in the compiled binary.
Core 0.24.2 exports `./integrations`, including the Muse launch arguments for an embedding server.
The core also exports the provider setting schemas
(`./settings`), `readWindowsAcl` in `./host`, and a PrismML catalog that picks the build for a file's
`requires`/`min_build`. Since 0.17.0 the core downloads no model files and exports no downloader (ADR
Expand All @@ -277,7 +278,7 @@ limit on remote GGUF reads (`remoteGgufTimeoutMs`); `startDaemon` passes both op

```jsonc
// package.json
"dependencies": { "@atomic-chat/core": "0.23.4" }
"dependencies": { "@atomic-chat/core": "0.24.4" }
```

To bump the core: `bun add --exact @atomic-chat/core@<version>`, run `npm run verify`, and commit the bump
Expand Down
40 changes: 40 additions & 0 deletions docs/strata.md
Original file line number Diff line number Diff line change
@@ -0,0 +1,40 @@
# Strata on Atomic Server

Install the engine on the server and browse the pinned model variants before choosing one.

```sh
atc engines install strata
atc models catalog --engine strata
atc models search qwen --engine strata
atc models pull iq3-xxs-dflash2 --engine strata --dry-run
atc models pull iq3-xxs-dflash2 --engine strata --request-id qwen-metal
atc run iq3-xxs-dflash2 --engine strata
atc config get engine.strata.memory_mode
```

The same catalog and server-side jobs are available in the terminal and web admin. The server
checks the selected bundle, GPU and disk on its own host. The engine bundle includes Python.
Set `HF_TOKEN` in the server environment for gated Hugging Face inputs after accepting their terms.

For a local import, use `--import-dir /absolute/server/path` with `models pull --engine strata`.
The directory must have all the selected variant's files at the relative paths shown by
`--dry-run`. The server copies and hashes the inputs before registration. The original
directory stays under the operator's control.

After acquisition, the same job shows **Preparing model** while the core builds and verifies
the runtime data. If the server restarts, repeat the original `models pull` with the same
`--request-id` or choose **Resume** in Downloads. The saved plan supplies the source revisions and
hashes even if the catalog has changed. Cancel with `atc models pull --engine strata --cancel qwen-metal`.
A replacement keeps the previous files while the core holds the model for maintenance.
Use Installed models or `atc models rm <variant>` to delete an installed model.

For local conf testing, set the following variables in the environment of the process that
owns the server. Use absolute file URLs from a checkout with qualified bundles.

```sh
ATOMIC_STRATA_MANIFEST_URL=file:///absolute/conf/backends/strata-manifest.json
ATOMIC_MODEL_CATALOG_URL=file:///absolute/conf/models/catalog.json
```

The engine uses the first override. The workspace uses the second unless
`ATC_MODEL_CATALOG_URL` is set. Chat uses the existing authenticated OpenAI-compatible listener.
5 changes: 1 addition & 4 deletions package.json
Original file line number Diff line number Diff line change
Expand Up @@ -46,16 +46,13 @@
"verify": "npm run lint && npm run typecheck && npm run format:check && node scripts/gen-third-party-notices.mjs --check && npm test && npm run test:ui && npm run build && npm run build:ui && npm run build:bin && npm run test:e2e && npm run test:e2e:web && npm run test:runtime-compat && npm run test:installers"
},
"dependencies": {
"@atomic-chat/core": "0.23.4",
"@atomic-chat/core": "0.24.4",
"fast-xml-parser": "5.11.2",
"fflate": "0.8.3",
"ink": "^7.1.1",
"pdfjs-dist": "4.8.69",
"react": "^19.2.0"
},
"patchedDependencies": {
"@atomic-chat/core@0.23.4": "patches/@atomic-chat%2Fcore@0.23.4.patch"
},
"devDependencies": {
"@eslint/js": "^9.36.0",
"@types/node": "^22.18.0",
Expand Down
2 changes: 2 additions & 0 deletions packages/admin-ui/LIFTED.md
Original file line number Diff line number Diff line change
Expand Up @@ -136,6 +136,8 @@ The 2026-11-06 onboarding adaptation scopes `setup/FirstModel.tsx` to the select

The onboarding wizard now shares `intro/EngineChooser.tsx` with the intro: managed runtime consent and readiness precede model selection, Back/Escape retain wizard state, and returning from an engine panel rechecks the selected model's compatibility (`setup/SetupEngineDialog.tsx`).

- `workspace/EngineMark.tsx` and `public/engine-brands/{vllm,sglang,foundry-local,litert-lm,geniex,fastflowlm}.svg` follow Atomic-Chat `web-app/src/{lib/utils.ts,containers/server/ServerPageParts.tsx}` and `web-app/public/images/model-provider/` @ `07e4b8737`. The picker and onboarding use the same engine logos and initial-letter fallback, including Strata's `S`.

The 2026-11-07 adaptation replaces the Foundry native variant menu in `workspace/library/install/FoundryModelInstall.tsx` (original `LocalModelTable.tsx` @ `5043013`) with `VariantChoices.tsx`, shared with LiteRT-LM. Short device/readiness labels and a separate reason keep expanded catalog rows within the viewport. `use-variant-model.ts` refreshes core verdicts when setup returns from engine preparation. Engine dialogs now separate Overview from Settings while retaining staged edits.

The provider summaries, connector cards and workspace settings above now use shared non-interactive status indicators; skill upload uses the shared English file-picker control. Onboarding provider rows grow to accommodate connection status. Original provenance headers are retained.
Expand Down
Loading
Loading