# OpenWebUiCode - Open-WebUI+Code — .NET (Coding Agent Edition) A migration of [Open WebUI](https://github.com/open-webui/open-webui) to **.NET 10 · C# 14 · Blazor WebAssembly** (SvelteKit + FastAPI → ASP.NET Core + Blazor), faithful to the original layout and configuration — **extended into a full coding-agent platform** in the spirit of [opencode](https://github.com/sst/opencode): repo-bound chats running on dedicated git worktrees, sub-agents in parallel, mid-run steering, durable memory, scheduled routines, a VS Code-style IDE and a worktree review pipeline. [![CI Build & Test](https://github.com/afonsoft/OpenWebUiCode/actions/workflows/ci-build-test.yml/badge.svg)](https://github.com/afonsoft/OpenWebUiCode/actions/workflows/ci-build-test.yml) [![Code Quality](https://github.com/afonsoft/OpenWebUiCode/actions/workflows/code-quality.yml/badge.svg)](https://github.com/afonsoft/OpenWebUiCode/actions/workflows/code-quality.yml) [![Security Scan](https://github.com/afonsoft/OpenWebUiCode/actions/workflows/security-scan.yml/badge.svg)](https://github.com/afonsoft/OpenWebUiCode/actions/workflows/security-scan.yml) [![Accessibility Audit](https://github.com/afonsoft/OpenWebUiCode/actions/workflows/a11y-audit.yml/badge.svg)](https://github.com/afonsoft/OpenWebUiCode/actions/workflows/a11y-audit.yml) [![Quality Gate](https://sonarcloud.io/api/project_badges/measure?project=afonsoft_open-webui&metric=alert_status)](https://sonarcloud.io/project/overview?id=afonsoft_open-webui) [![Bugs](https://sonarcloud.io/api/project_badges/measure?project=afonsoft_open-webui&metric=bugs)](https://sonarcloud.io/project/overview?id=afonsoft_open-webui) [![Vulnerabilities](https://sonarcloud.io/api/project_badges/measure?project=afonsoft_open-webui&metric=vulnerabilities)](https://sonarcloud.io/project/overview?id=afonsoft_open-webui) [![Code Smells](https://sonarcloud.io/api/project_badges/measure?project=afonsoft_open-webui&metric=code_smells)](https://sonarcloud.io/project/overview?id=afonsoft_open-webui) [![Coverage](https://sonarcloud.io/api/project_badges/measure?project=afonsoft_open-webui&metric=coverage)](https://sonarcloud.io/project/overview?id=afonsoft_open-webui) [![Duplicated Lines](https://sonarcloud.io/api/project_badges/measure?project=afonsoft_open-webui&metric=duplicated_lines_density)](https://sonarcloud.io/project/overview?id=afonsoft_open-webui) [![Maintainability Rating](https://sonarcloud.io/api/project_badges/measure?project=afonsoft_open-webui&metric=sqale_rating)](https://sonarcloud.io/project/overview?id=afonsoft_open-webui) [![Reliability Rating](https://sonarcloud.io/api/project_badges/measure?project=afonsoft_open-webui&metric=reliability_rating)](https://sonarcloud.io/project/overview?id=afonsoft_open-webui) [![Security Rating](https://sonarcloud.io/api/project_badges/measure?project=afonsoft_open-webui&metric=security_rating)](https://sonarcloud.io/project/overview?id=afonsoft_open-webui) [![Technical Debt](https://sonarcloud.io/api/project_badges/measure?project=afonsoft_open-webui&metric=sqale_index)](https://sonarcloud.io/project/overview?id=afonsoft_open-webui) [![Lines of Code](https://sonarcloud.io/api/project_badges/measure?project=afonsoft_open-webui&metric=ncloc)](https://sonarcloud.io/project/overview?id=afonsoft_open-webui) > **Versão em português:** [README.pt-BR.md](README.pt-BR.md) ## Open WebUI (original) vs opencode vs oh-my-pi vs deepseek-harness vs this fork The agent roadmap was benchmarked against [opencode](https://github.com/sst/opencode) (incl. the [alltomatos/opencode](https://github.com/alltomatos/opencode) fork — see [the full analysis](docs/analise-fork-alltomatos-opencode.md)), [oh-my-pi](https://github.com/can1357/oh-my-pi) — whose session-tree, advisor and context-tooling design shaped several features below — and [deepseek-harness](https://github.com/deepseek-ai/deepseek-harness) (`dsh`), DeepSeek's "everything-is-a-plugin" agent harness built on [Cordis](https://github.com/cordiverse/cordis). | Capability | Open WebUI | opencode | oh-my-pi | deepseek-harness | **This fork (.NET/Blazor)** | |---|---|---|---|---|---| | Chat + streaming,
multi-provider | ✅ Ollama/OpenAI,
RAG/pipelines | ✅ any LLM
via gateway | ✅ multi-provider
(TUI) | ✅ `ctx.llm`
adapter plugins | ✅ parity +
OmniRoute-friendly | | Repo/workspace
bound to session | ❌ single
workspace | ✅ per-project | ✅ per-project | ✅ per-session
working dir + exp.
`tool-worktree` | ✅ **per-chat git
worktree** — strict
scope, global
never leaks | | Sub-agents /
parallel sessions | ❌ | ✅ task tool
+ sessions | ✅ declarative
`.md` task agents | ✅ `tool-subagent` +
backend seam (in-proc,
SDK, ACP, Claude
Code, Codex) | ✅ `delegate_task`
(+ `*.agent.md`),
depth ≤3, worktree-
isolated | | Review before
merge | ❌ | ✅ orchestrator
reviews diffs | ❌ | partial — diff
view, no merge
tool | ✅ `worktree_diff` /
`worktree_merge` +
auto-merge opt-in | | Steer/queue
mid-run input | ❌ | ✅ durable
steer/queue | ❌ | ✅ durable inbox
+ `agent.inject()` | ✅ `…/runs/{id}/steer`
(steer + queue) | | Durable agent
memory | ❌ | ✅ memory tools
+ auto-sync | ✅ memory tools | ❌ no memory
subsystem | ✅ `memory_save` /
`memory_search` + 6h
auto-distill `auto:*` | | Routines /
reminders | ❌ | ✅ routine tool
+ scheduler | ❌ | ✅ `schedule` /
`tool-schedule` | ✅ `routine`/`reminder`
+ `once` schedules | | Task list tool | ❌ | ✅ todo write | ✅ | ✅ `tool-todo` | ✅ `todo_write`,
stable ids | | Context
compaction | ❌ | ✅ auto-summarize | ✅ auto + on-demand,
with pruning | ✅ compaction seam
+ result pruner +
image offload | ✅ LLM summary +
pruning + digest +
`history://` recovery | | Parent/child
run tree UI | ❌ | ✅ session tree | ✅ session tree | ✅ subagent
conversations | ✅ sidebar tree +
badge + parent
markers | | Parallel runs
console | ❌ | ✅ | ✅ | ✅ jobs +
subagent views | ✅ "Active runs"
sidebar | | Runs survive
browser close | partial | ✅ | ✅ | ✅ durable
session log | ✅ server-side +
`Last-Event-ID` | | Session branching
(fork any turn) | ❌ | ❌ | ✅ `parentId` +
`/fork` | ✅ turn-boundary
fork (`seedLength`) | ✅ `ParentId`/`Leaf`
tree + "branch
from here" + summary | | Model roles +
fallback chains | ❌ single model | partial | ✅ `modelRoles` +
`fallbackChains` | partial — per-cap.
providers +
`llm-retry` | ✅ `tasks`/`memory` /
`advisor` + vision
capability map | | Tool approval
tiers | binary
approve-all | ✅ permission
model | ✅ tier + tool policy
+ user override | ✅ presets +
approval +
`sandbox-policy` | ✅ `Read`/`Write`/`Exec`,
fail-closed, per-tool
overrides | | Advisor /
watchdog | ❌ | ❌ | ✅ read-only 2nd
opinion | partial — exp.
`auto-review` +
guard plugins | ✅ advisor inspects
runs, no approval
power | | Agent-directed
context | ❌ | partial | ✅ `checkpoint`/`rewind`,
`context_notes`,
`new_context`,
`context_save`/`recovery`/`
compact` | ✅ `agent.inject()`
+ context plugins
(file/session/time
refs) | ✅ same trio + pruning
+ `history://` | | Layered config +
provenance | env + admin UI | partial | ✅ env>runtime>overlay>
project>global | ✅ profile→bundles→
`cordis.patch.yml`→
`--patch` | ✅ global/admin/user
layers + source | | Declarative
agent defs | ❌ | ✅ AgentUI | ✅ `.md` + frontmatter | ✅ agent presets /
personas + skill
files | ✅ `*.agent.md` (tools,
model, spawns) via
`delegate_task` | | Plan mode →
structured output | ❌ | ✅ plan/build
modes | ❌ | ✅ `plan-mode`
plugin | ✅ `/plan` → **SPEC SDD**
markdown; `plan_exit`
validates + saves to
`.openwebui/plans/` | | Media attachments
in chat | ✅ uploads | partial | ❌ | ✅ attachments
+ image offload | ✅ composer thumbnails,
inline + click-enlarge,
vision analysis | ### deepseek-harness: architecture deltas & ideas to adopt `dsh` is a **harness kernel**, not a product: every subsystem — model adapter, tool registry, session log, agent loop, even each UI block — is a Cordis plugin composed at boot from ordered layers (profile → bundles → `cordis.patch.yml` → home → `--patch`), with reversible effects that unwind on unload, config HMR and a plugin market (`dsh-plugin` topic). This fork is a **product**: the same class of capabilities ships compiled into ASP.NET Core services behind a single multi-user web surface. | Axis | deepseek-harness | This fork | |---|---|---| | Extension model | everything-is-a-plugin — `ctx` services, typed events incl. `agent/*`/`tools/*` waterfall interceptors, HMR | compiled-in services + internal run-event bus with DI observers and ordered waterfall interceptors (`pre-step`, `tool.call`) — no dynamic loading; extensibility via MCP servers, `*.agent.md`, skills/personas | | Surfaces | `web` · `headless` · TS/Python SDK (JSON-RPC) · `acp` (editor) · Electron desktop | web (Blazor WASM) · `--headless` one-shot (JSON Lines) · `POST /api/v1/rpc` (JSON-RPC 2.0) · `open-webui-acp` stdio host | | Execution seam | swappable `ctx.fs`/`subprocess`/`sandbox` providers — SSH backend relocates Bash/PTY/LSP wholesale | local tools + approval tiers | | Session store | append-only JSONL event log, versioned format migrations (v0→v4), "model-visible means logged", incremental projections | EF Core/SQLite entities + persisted run events, checkpoint replay | | Multi-user | single-user (DeepSeek account/credentials plugins) | users/RBAC/admin UI, JWT + `sk-*` API keys | | Distribution | `npx @deepseek-ai/dsh` + plugin market | single multi-arch Docker image / `dotnet run`; ACP host ships as `src/OpenWebUI.Acp` (publish as single-file binary or run via `dotnet run`) | Ported ideas (SPEC-20261021-harness-seams, branch `feature/Devin-20261021-harness-seams`): internal `agent/*`/`tools/*`-style event seams — `IAgentRunObserver` + `RunInterceptorRunner` (ordered waterfall, 5s timeout, fail-closed on `tool.call`), shipping `TelemetryObserver` and `RepeatToolGuard` (vetoes a tool call repeated identically 3×) · composable prompt sections — `ISystemPromptContributor` with DI ordering; the former `systemParts` blocks are now contributors with byte-identical output · launch surfaces — `--headless` (JSON Lines stdout, exit code = run status), `POST /api/v1/rpc` (`ping`, `chats.list/create`, `runs.start/get`), ACP stdio (`initialize` gated by `OPENWEBUI_ACP_TOKEN`, `session/new`→chat, `session/prompt`→run streamed as `session/update`) · webhook-triggered runs — provider-aware `AutomationHook` (generic/github) with `X-Hub-Signature-256` HMAC, event filters, `{{path}}` prompt templates, repo/persona/mode binding and an `AutomationHookDelivery` audit/idempotency log · agent teams — `Chat.TeamId`/`TeamRole` roster over `delegate_task` children, `TeamTask` durable tasks, steer/queue as inter-agent inbox, `team_task`/`team_message`/ `team_status` built-in tools and `GET /api/v1/chats/{id}/team`. Still open: execution-provider seam (remote sandbox for `shell_exec`/terminal/`lsp_*`). ## Screenshots | Adaptive/Fusion picker | Leader + Sidekick pair | |---|---| | ![Fusion first in the model selector](docs/screenshots/fusion-selector.png) | ![Leader and Sidekick selects](docs/screenshots/fusion-pair.png) | | Leader delegating to a sub-agent | Reviewed fusion answer | |---|---| | ![builtin_delegate_task mid-run](docs/screenshots/fusion-delegation.png) | ![completed fusion run](docs/screenshots/fusion-completed.png) | | Delegated sub-agents in parallel | Session tree + completion markers | |---|---| | ![delegate_task spawning two child runs](docs/screenshots/delegate-parallel.png) | ![parent markers + subtask tree](docs/screenshots/parent-notify.png) | | Active-runs console | Sub-session with parent link + approval | |---|---| | ![sidebar active runs](docs/screenshots/runs-parallel.png) | ![child chat approval card](docs/screenshots/subsession-approval.png) | | Per-chat repo binding | Routines & reminders | |---|---| | ![workspace repo picker](docs/screenshots/repo-binding-picker.png) | ![automations page](docs/screenshots/automations.png) | | VS Code-style IDE | Model combos | Agents & personas | |---|---|---| | ![IDE workbench](docs/screenshots/ide-workbench.png) | ![Settings → Combos](docs/screenshots/settings-combos.png) | ![Settings → Agentes](docs/screenshots/settings-agents.png) | | Chat (light/dark) | Workspace panel | Terminal PTY | |---|---|---| | ![chat](docs/screenshots/chat-light.png) | ![workspace](docs/screenshots/workspace-panel.png) | ![terminal](docs/screenshots/terminal.png) | | Permissions v2 rules | Auth (pt-BR) | |---|---| | ![Settings → Permissões](docs/screenshots/settings-permissions.png) | ![auth](docs/screenshots/auth-light.png) | ## Agent features (new — additive layer over upstream parity) - **Adaptive/Fusion model pair** — first item of the chat model picker: a virtual `fusion:{leader}>{sidekick}` ref whose **Leader** (reasoning, plan, review) orchestrates the run while every `delegate_task` sub-agent defaults to the cheaper **Sidekick**. Both picks are restricted to the effective `/api/models` catalog — re-validated server-side against spoofed refs; the pair persists per chat (`ModelsJson`) and per browser (`webui.model`, `webui.fusion.{leader,sidekick}`), gates budget on both models, auto-injects the orchestration toolset (`delegate_task`, `team_task`, `team_message`, `team_status`, `agent_catalog`, `run_result`, `todo_write`) plus a coordination prompt, and publishes a `fusion` phase + note carrying the active pair. Sub-agents communicate through the shared team task list and member inboxes; the Leader keeps judgment calls and writes the final answer. - **`delegate_task`** — spawns a child chat + server-side run: `wait:false` returns `childRunId` immediately, `wait:true` (default) blocks up to 300s; `repo` and `persona` (a workspace Skill as system prompt) params; depth ≤ 3; each sub-agent inherits the parent repo but works in its **own chat-scoped worktree**. - **Per-chat repo worktrees** — binding a chat to `owner/repo` creates an exclusive worktree (`workspaces/{user}/worktrees/{owner}__{repo}/{chatId}`, branch `chat/{chatId}`) over a shared base checkout; resolution is strict per chat — a new conversation never inherits your global binding. Unbind/archive/delete releases the worktree; the base checkout leaves disk when unreferenced, and a diverged base self-heals (diverged HEAD preserved as `backup/*`). - **Three agent modes** — composer selector `chat` / `plan` / `build`: `chat` is repo-free (workspace tools denied), `build` requires a bound repo (`PATCH` 400 otherwise; unbind drops `build`→`chat`; `plan_exit` only promotes with binding). Workspace gate covers `delegate_task`, `lsp_*` and `skill`; the format hook honors the chat binding's `FormatCommand`. - **Autonomous decisions** — `smart`/`auto`/`always-allow` presets auto-answer `ask_user` prompts that carry options (recommended option, else first) without pausing the run; every auto choice emits an auditable `decision_auto` event. - **Mobile workspace** — touch targets and collapsing chips keep the workspace panel, repo picker and IDE reachable on small screens. - **`run_result` / `worktree_diff` / `worktree_merge`** — collect async results, review the child diff, merge explicitly (or `auto_merge` opt-in). - **Steer & queue** — `POST /api/v1/chats/{id}/runs/{runId}/steer` injects a user message into a live run (`steer` = next turn boundary, `queue` = when idle). - **Agent memory** — `memory_save`/`memory_search` (project + global scopes), auto-injected `` block; 6h scheduler distills durable facts into `auto:*` memories with per-user watermark + opt-out. - **Routines & reminders** — `routine` tool (list/create/update/delete/run_now, `once`/`interval`/`daily`/`weekly`) and `reminder` (fires an `automation.reminder` notification) over the built-in Automations engine. - **Context compaction** — when history grows, the pipeline summarizes old turns through the configured model and keeps a cutoff kv per chat. - **Plan mode → SPEC SDD** — `plan` mode (or the `/plan` slash builtin) analyses the requirement read-only, then emits a full SPEC SDD markdown: metadata, user story, scope, `RF-###` requirements, BDD acceptance criteria, task plan, DoD and open questions. `plan_exit` validates the structure (incomplete plans are refused with the missing sections), persists to `.openwebui/plans/` and only promotes to `build` on user approval. - **Model roles + fallback chains** — dedicated models for `tasks` (titles, follow-ups, tags, compaction), `memory` (distillation) and `advisor`; each role resolves through a capability map (e.g. vision) with an ordered fallback chain, configurable per layer. - **Fine-grained tool approval tiers** — every tool declares `Read`/`Write`/ `Exec` (fail-closed `Exec`); per-tool user overrides can relax but never weaken a tool-level deny. - **Run advisor (watchdog)** — a secondary, read-only model reviews the live transcript, can inspect the workspace with its own read tools and injects guidance mid-run — with no approval or mutation power. - **Agent-directed context** — `checkpoint`/`rewind` tool pair (mark → explore → collapse), `context_notes` (per-branch notebook that survives rollover) and `new_context` (on-request compaction); mechanical pruning elides superseded tool results and `history://` URIs recover full pre-compaction turns. `context_save`/`context_recovery`/`context_compact` persist versioned context snapshots per chat (table `ContextSnapshots`) and compaction auto-triggers at `compaction:threshold_tokens` estimated tokens (~256k-token window). - **Chat branching** — messages form a tree (`ParentId`/`LeafMessageId`); "branch from here" forks any turn into a sibling chat, with a branch summary for the abandoned path. - **Declarative agents** — `*.agent.md` files (frontmatter: tools, model, spawns) discovered in the repo define named sub-agents, spawnable through `delegate_task(agent: "…")`. - **Layered configuration** — effective settings resolve through global → admin → user layers with per-key provenance. - **Memory pipeline** — scheduled consolidation distills session outcomes into the durable memory layers that are re-injected into new runs. - **Chat media attachments** — image/video thumbnails inside the composer, acceptable inline size after send, click-to-enlarge; attached images are always analyzed together with the prompt by a vision-capable model. - **Model combos (failover)** — `priority` / `round-robin` / `priority-round-robin` strategies over ordered model lists, shared RPM/TPM limits and per-model cooldown; user-scoped or admin-global, managed under Settings → Combos. - **Agent personas** — named personas (system prompt, model, guardrail level, tool allowlist) invocable via channels/hooks with a per-persona audit log. - **External agent registry** — detects local agent CLIs (Claude Code, Codex, OpenCode, Devin, Gemini) and bridges them into delegated runs. - **Global `~/.agents` scope** — skills, commands and `mcp.json` are discovered from `~/.agents` plus project `.claude`/`.agents` dirs and merged with the workspace catalog. - **Multi-question `ask_user` forms** — one prompt can carry several questions with `options`/`multiple`/`reject`; autonomy presets auto-answer per question. - **Parallel tool calls** — tools marked `SafeInParallel` execute concurrently inside a single provider round. - **Provider resilience** — retry with backoff, `finish_reason=length` continuation and context-overflow recovery inside the run pipeline. - **Durable runs** — persisted run events, checkpoint replay and heartbeat ownership: a live run survives an API restart and resumes on drain. - **Run steps + usage/cost** — per-agent step feed, last-step wrap-up and usage/cost metadata per run; the workspace **Changes** tab shows the run diff including changes the agent committed. - **Command expansion** — server-side expansion of repo/workspace slash commands with per-run overrides. - **Web search failover** — `auto` mode rotates across configured providers with a per-provider cooldown. - **Bounded tool output** — oversized tool results spill to a managed file (`builtin:big_output`) instead of flooding the context window. - **Markdown source view** — view/source toggle and copy on assistant messages. - **Agent secrets env** — key→value secrets (admin-global + per-user overrides) injected into `shell_exec` jobs and the terminal; the GitHub token from the integration is included by default. - **Browser visual loop** — Playwright browser tools (`browser_navigate`/`browser_observe`/`browser_act`/`browser_screenshot`) with local OCR for visual-reasoning loops. - **Settings redesign** — categorized sections (Combos, Agentes, Manutenção, …) with a unified save bar. - **Retention** — daily purge (automation runs >90d, notifications >60d, steers of terminal runs >30d) + worktree orphan prune + WAL checkpoint + weekly VACUUM; the **Manutenção** settings tab adds chat/memory retention with dry-run and "run now". - **Chat hierarchy** — `ParentChatId`/`ParentRunId`, `GET /chats/{id}/children`, nested sidebar tree, completion markers posted back to the parent chat. - **Persistent sidekick sub-agents** — `delegate_task` with `persistent` + `sidekick_id` reuses a named child chat (kv-backed) with a serialized `SidekickGate`; builtin `explore`/`general` profiles, `max_nesting` alias and foreground/background approval switch (`approve-mutations`/`allow-readonly`). - **Deferred tool loading** — core/deferred tiers (opt-in `BuiltinTools:Deferred`) keep the system prompt light: a `## Available on demand` index plus the `builtin_tool_search` loader reveals schemas on demand and merges them per iteration; MCP servers defer by volume. - **Plan mode v2** — builtin `write_plan` tool, 5-phase plan-mode prompt and `GET /workspace/repo/plans` listing; approved plans save to `.openwebui/plans/`. - **Lifecycle hooks** — command (`{matcher, command}`) and prompt hooks across `PreToolUse`/`PostToolUse`/`Stop`/`SessionStart`/`SessionEnd`/`SubagentStop`/ `Notification`/`UserPromptSubmit`/`PreCompact`, configured via `.openwebui/ hooks.json`; block decisions inject steering notes into the next round. - **Smart permission (LLM judge)** — `smart` preset auto-answers low-risk asks and escalates the rest through a per-run LLM classifier with nonce prompt and session dedup. - **Sandboxed execution** — `Sandbox:Enabled` confines `shell_exec`/background jobs inside `bubblewrap` (fake home, ro-binds, net off) behind a domain-allowlist network proxy; per-scope tool exclusions and a matching `autonomous` approval preset; `GET /api/v1/system/sandbox` reports status. - **Permissions v2** — ordered user rules `{action, resource glob, effect}` (last match wins) evaluated between admin policies and risk presets; "Sempre" answers persist as per-workdir `SavedApproval`s revocable individually; `external_directory` boundary covers file tools + `apply_patch`; all managed in Settings → Permissões (see screenshot below). - **Agent loop pack** — parallelism footer on tool results, `unseen_messages` steer readback, `instant_context` restricted-exec collector (`` XML), `extend_context` budget (skips 1 emergency compaction, cap 2), stall watchdog nudge (threshold 4) and prompt-cache keepalive during approvals. Full details per feature (with screenshots): [docs/en/AGENT-FEATURES.md](docs/en/AGENT-FEATURES.md) · [docs/pt/RECURSOS-AGENTE.md](docs/pt/RECURSOS-AGENTE.md) ## Classic features (upstream parity) - **Chat** with SSE streaming (Ollama / OpenAI-compat), attachments, 👍/👎, editing, regeneration, LLM-generated title/follow-ups/tags - **Decoupled runs** — responses execute server-side; closing/reloading the tab does not interrupt; on reopen the client reattaches with `Last-Event-ID` replay - **Notifications** — in-app toasts, Notification API and Web Push (VAPID) - **Tool streaming + approval gate** — `tool_call`/`tool_result` on SSE, approval presets per chat (readonly/approve/always/**auto by risk** LOW-MED-HIGH), deny with instructions for the model to correct course - **Builtin tools** — `shell_exec` + background jobs, `file_*` (list/read/grep/ glob/write/edit with diff), `code_interpreter`, `fetch_url`, `web_search`, `generate_image`/`generate_video`, `browser_screenshot`, `ask_user`, `n8n_list_workflows`/`n8n_trigger`, `todo_write`, `skill`, `lsp_*`, `plan_exit` - **Workspace panel** — Tasks, Changes (git diff), Jobs, MCPs, Terminal (PTY) and run Info; draggable resizer; `/ide` full-screen view - **Chat/Plan/Build agent modes** + permission modes; **checkpoints & revert**; **pause/resume** runs; slash commands from repo skills - **RAG, Knowledge, Arena, Channels/Notes/Calendar, Admin (users, connections, MCP servers, integrations), i18n in 8 locales, PWA offline shell** ## Quickstart ```bash dotnet restore OpenWebUI.slnx dotnet build OpenWebUI.slnx --configuration Release dotnet run --project src/OpenWebUI.Api # http://localhost:8080 ``` Docker (`:3032 → :8080`) — published **multi-arch image (amd64 + arm64)** on [Docker Hub](https://hub.docker.com/r/afonsoft/open-webui) and GHCR: ```bash docker run -d -p 3032:8080 -v openwebui-data:/data afonsoft/open-webui # or compose against the published image (no clone needed) curl -LO https://raw.githubusercontent.com/afonsoft/OpenWebUiCode/main/docker-compose.hub.yaml docker compose -f docker-compose.hub.yaml up -d # or build from source docker compose up -d --build ``` `docker-compose.prod.yaml` ships a production layout with watchtower auto-updates (5min poll). Full Docker guide: [docs/en/DEPLOY-DOCKER.md](docs/en/DEPLOY-DOCKER.md) · [docs/pt/DEPLOY-DOCKER.md](docs/pt/DEPLOY-DOCKER.md) ## Tests ```bash dotnet test OpenWebUI.slnx # 1900+ API tests + client tests (NUnit) ``` Coverage gate via Coverlet + ratchet (`.ci/coverage-baseline.txt` only goes up). ## Quality & CI | Workflow | Covers | |---|---| | `ci-build-test.yml` | Release build, all tests + coverage gate (Coverlet + baseline ratchet in `.ci/`), WASM drift guards, Tailwind freshness, Docker build, admission gate | | `a11y-audit.yml` | axe-core + usability/visual smoke (Playwright, mobile+desktop) | | `chat-e2e.yml` | Real-provider chat E2E (secrets-gated, skips green without `LLM_*`) | | `code-quality.yml` | Qodana + [SonarCloud](https://sonarcloud.io/project/overview?id=afonsoft_open-webui) | | `security-scan.yml` | CodeQL (C#+JS), gitleaks, actionlint+zizmor, OSV-Scanner, hadolint, Snyk, Trivy | | `scorecard.yml` | OpenSSF Scorecard (SARIF, weekly) | | `perf.yml` | Weekly load test — aiohttp harness in `tools/perf/` + aggregate-rps gate | | `release.yml` | Tag `vX.Y.Z` → artifacts + multi-arch (amd64+arm64) GHCR + Docker Hub image | | `dockerhub-overview.yml` | Syncs `docs/dockerhub/overview.md` to Docker Hub (manual) | | `cleanup-artifacts.yml` | Nightly cleanup of artifacts older than 30 days | Full pipeline & tooling map: [docs/en/CI-CD.md](docs/en/CI-CD.md) · [docs/pt/CI-CD.md](docs/pt/CI-CD.md) ## Documentation - [docs/en/AGENT-FEATURES.md](docs/en/AGENT-FEATURES.md) — agent platform details (EN) - [docs/pt/RECURSOS-AGENTE.md](docs/pt/RECURSOS-AGENTE.md) — detalhes do modo agente (PT-BR) - [docs/MIGRACAO-DOTNET.md](docs/MIGRACAO-DOTNET.md) — upstream → .NET parity map - [docs/architecture/architecture.md](docs/architecture/architecture.md) — architecture (Mermaid) - [docs/analise-fork-alltomatos-opencode.md](docs/analise-fork-alltomatos-opencode.md) — opencode-fork analysis that drove the agent roadmap - `.specs/` — in-flight SPECs; `docs/specs/` — delivered SPECs - [CHANGELOG.md](CHANGELOG.md) ## Star History [![Star History Chart](https://api.star-history.com/chart?repos=afonsoft/OpenWebUiCode&type=date&legend=bottom-right)](https://www.star-history.com/?repos=afonsoft%2FOpenWebUiCode&type=date&legend=bottom-right) ## License BSD-3-Clause — see [LICENSE](LICENSE).