macOS alpha, 0.32.0-alpha.1

Implemented capabilities

What the current Moguri alpha implements, by protocol checkpoint.

Published unchanged from docs/features.md in Moguri's source repository at commit b686ddc (0.32.0-alpha.1). That repository isn't public yet, so references to files that only exist there appear as plain text with a dotted underline.

This page records what the current Moguri alpha actually implements, release by release. It moved here from the README so the README stays a short landing page. For the broader product vision, which is not a list of implemented features, see product vision. Remaining work is tracked in remaining milestone gaps.

The first executable checkpoint includes an OpenTUI terminal client and an Electron desktop app sharing one local service. It supports GitHub Copilot device sign-in, account-enabled model selection, streamed conversations, native file operations with script/unlisted-website approvals, persistent drafts and queues, distinct Queue/Steer/Interrupt delivery, stopping owned process groups, restart recovery, archival/restoration, and basic per-call accounting. Optional metadata-only OTLP export (not public yet) projects existing execution/accounting records into traces, canonical gauges and opt-in correlated diagnostic logs, with keychain-only credential headers. It is off by default, profile-isolated and explicitly reviewed; no prompts, tool content, terminal data or private names are exported. Ordinary active-turn input now uses shared live delivery, while first-topic chat names (not public yet) are generated by one separately budgeted call; explicit names and held/queued input keep their own contracts. After a round of requests, a wrap-up turn (not public yet) commits, publishes the draft PR and follows CI and reviews. Optional pull request automation (not public yet), off by default with per-repository overrides, reruns CI jobs that failed for transient reasons after checking GitHub Status, and resolves base conflicts while keeping both sides' features. Optional Claude Agent and Codex Agent (not public yet) sessions run your own installed Claude Code or Codex with your own sign-in, while Moguri approves their tool calls and keeps the transcript. Both are off by default; see setup (not public yet). Besides GitHub Copilot and custom OpenAI-compatible endpoints (with custom headers), Moguri connects OpenRouter, the Anthropic API, the OpenAI API and a ChatGPT account (not public yet).

The TUI uses Enter to send and Shift+Enter for newlines; Cmd/Ctrl+Enter holds active-turn prompts or opens scheduling when idle. Its compact chat-first client saves and leaves on exit, /exit or Ctrl+C while the shared service keeps working; /stop remains the turn-stop command. The layout has a growing input, formatted replies, expandable tool output, clickable navigation and a focus-safe sidebar. Terminal controls (not public yet) preserve drafts while reading history. Shared themes (not public yet) include ten presets in Dark and Light, without changing saved choices or factory Dark. Moguri includes its own verified source setup instructions (not public yet) in normal model context, including project-free chats. Desktop Settings keeps on-screen text short; the Settings reference holds each page's full behavior and safeguards. The Local website allowlist (not public yet) supports session, project and profile-wide grants through Guardrails and /firewall.

The importable Moguri SDK now powers the CLI, TUI and desktop. It exposes shared editable/restorable system instructions and permissions, opt-in local vector/graph and coding-style memory, standard skills, reviewed evolution proposals, curated/approved scripts, declarative plugins, optional tool-less subagents and one native MCP gateway. Desktop starts dark and uses Moguri's original face-and-bat-wing mascot. The mascot now also carries its small red antenna pom-pom.

The Plan sidebar (not public yet) shows only an explicit multi-step model checklist, with long-task estimates separately labeled as advice, never proof of completion. Session controls (not public yet) include ordered drag/drop, context-menu actions, serial held queues and explicit work/question/approval/completion indicators with verified draft/open/merged PR state. Unread names use subtle font weight rather than extra session-list dots. Center drops nest sessions, create named project folders or add a project to an existing group (groups never nest); folder chats capture all member repositories for explicit Local reads, edits and commands. Fuzzy session/settings search and compact settings controls share the session reliability contract (not public yet). The Dashboard (not public yet) follows active sessions and opens their exact conversations. Its Analytics (not public yet) page draws running and recent turns on a zoomable lane timeline, kept in memory for an hour only. Project settings select registered projects; project-group menus fold only their member folders. Settings / Repositories accepts reviewed URL/bullet lists for bulk import, and workspace folders (not public yet): trust a folder such as ~/workspace and Moguri adds its Git repositories as read-only projects, optionally kept in sync. Their sessions always work in Moguri-owned worktrees, never the original checkout. Project settings also reviews remote URL changes (not public yet) across the actual shared-worktree boundary, without fetching or rewriting old PR sources. Built-in work-item views remove hidden Markdown comments. Repository-script review (not public yet) delivers the full exact source to the agent before execution approval; unsafe findings need separate actual-user force and never lift hard policy.

Protocol98 adds Local skill source admission and per-turn rewind (not public yet). Complete source/provenance review, expiring grants, least-privilege restrictions and actual-user revocation share the service, SDK and clients. Reviewed rewind supports Local, pinned SSH and private Sandbox workspaces while preserving external edits, current Git intent, later history and spent/unknown budgets. Sandbox reviewed-script execution remains unqualified; the wider skills-safety work is not complete.

Protocol99 adds a session Inbox and explicit skill-load indicators (not public yet). Inbox follows Dashboard, opens the owning conversation and supports independent notification deletion without answering or approving. Successful source loads show their actual skill name and book icon; discovery is not loading.

Model recovery and scheduled prompts (not public yet) share the same service and SDK. Transient model failures use current-profile 30/90/120/180-second recovery rounds, without replaying earlier tools or hooks. The primary button retries the original eligible failure without sending a new draft. Idle scheduling offers local-time controls, unsent ghost prompts and an Automations / Scheduled prompts group, with one-at-most queue admission. Customize and Artifacts tabs (not public yet) sit next to Automations: toggle installed skills, install skills from configured repositories, browse the agent's toolsets, write or ask the agent for script tools for every chat, and find each project's builds, files, images and links with their chat. Agents get typed GitHub reads (tags, releases, annotations, contents, settings, rules) and reviewed github_write (comments, threads, releases, settings; draft PRs only). An optional recurring check suggests script tools for requests you repeat. Local and SSH continuation are qualified. A rolled-back native workspace keeps its explicit inspection/preparation gate instead of automatic root substitution.

Closing or quitting the desktop app saves unsent drafts and open queued-prompt edits first. Chats a quit, restart or crash interrupted can be continued (not public yet) one by one, all at once, or automatically after every restart.

Automations (not public yet) run a prompt or a declarative workflow of prompts and repository-trusted scripts once, every N minutes, daily or on chosen weekdays. Each run starts a new chat in its project with captured controls; script steps run without a model call under normal approval. Missed times run once, overlapping times are skipped, and restarts never replay a step. Desktop, TUI (/automations, F9) and CLI (moguri automations, YAML/JSON workflow files) share the same state.

The owner-requested protocol-41 platform extension (not public yet) adds Sky/Terra/Glacier knowledge, reviewed lifecycle hooks and immutable repository-backed marketplace bundles. Imports/updates/rollbacks stay disabled; knowledge injection has a separate default-off consent. Nothing self-approves or starts a model merely because it was installed.

Local Git built-ins cover status/history/diff, branch creation/switch/rename, exact staging, selected staged commits, history-preserving rollback and new-branch pushes through the same SDK and agent tools. Usage and boundaries (not public yet). GitHub workflows expose exact-root reviewed main/force-with-lease pushes, rebases, tags, two-way issue/PR/review work, pinned CI watching, releases and exact-head PR merges. Linear workflows add team-scoped issue/project/comment operations and explicit conflict-checked ticket pull/push. Both retain original intent/results without replaying ambiguous writes. Dependency caching extends to explicit private npm registries, npm/pnpm workspaces, Python, Rust/Cargo and Go; mutable installs remain private and executable builds are not implied. Desktop exposes the same APIs under Settings; shared /git, /github, /linear, /knowledge, /hooks, /marketplace and /cache commands reach TUI/CLI.

Protocol100 moves repository-script assessments to Inbox (not public yet) with collapsible source and explicit user approval, separate from execution. Startup shows the original mascot on black with an optional Space-to-flap game. Inline code stays plain text; fenced blocks keep explicit copy controls. Missing files say Not Found, and provider failures retain their redacted diagnostic.

Protocol101 adds explicit repo trust, worktree/PR defaults and recoverable output (not public yet). Existing exact trusted script entrypoints skip model assessment, not approvals; new/package/generated scripts stay outside the baseline. New single-project Local Git tasks use managed worktrees and normally finish at a draft PR. Inbox has search and type/read filters. Qualified partial results can continue; uncertain effects retain enforced read-only recovery, never false success.

Protocol96 added original-owner startup reconnect (not public yet) and a default-on option to keep agents working when the last window closes. Explicit Quit still stops the service and known user terminals. Startup confirmation, private ownership and unsent drafts remain protected; the original unreproduced 90-second initialization delay is not claimed resolved.

Protocol95 added the unified My Work inbox (not public yet): explicitly configured repositories/accounts, deduplicated PR/review/failed-CI/ unresolved-thread attention, source freshness and retained read-only refresh. CLI, TUI and desktop share the same SDK, original CI watches and exact scoped follow-up review. Opening an item never starts polling or grants permission. The existing active-session and usage/cost/velocity dashboard remains available. See the ownership and qualification contract (not public yet).

Protocol94 added authenticated GitHub search and composer fixes (not public yet): private-repository reads use the actual configured account, provider tool IDs are assigned before ownership is retained, and one Enter sends one ordinary message even during autosave. Autonomy settings are beside their choices; profile-wide command families and fenced-code copy controls share the existing policy and clipboard paths. It preserves protocol93 retained PR CI watches (not public yet): exact account/root/head ownership, cross-view deduplication, historical observations, startup-failure diagnostics and source-bound notifications. Reopening a pane never restarts polling; uncertain cleanup remains partial. This source update does not relabel the separately delivered protocol92 DMG. The required combination and native-gate map (not public yet) keeps complete Local/SSH generated coverage separate from missing current physical SendBox and account-blocked hosted-CI acceptance.

Whole-descendant containment can be required explicitly through Guardrails / Require Sandbox (qualified SendBox) or moguri containment require-sandbox --confirm-review. Ordinary Local/SSH still owns known groups only. Missing qualification blocks new work; selecting this policy does not silently change an existing session's Env, qualify an old workload or choose a model.

The macOS app bundles SendBox 0.2.7 and a compatible pinned Linux worker, Node runtime and native dependencies. Managed setup (not public yet) replaces developer commands with one explicit in-app review. SendBox 0.2.7 is the minimum supported release; a saved older selection is blocked until you explicitly review the bundled setup, never silently upgraded or replaced by Local. Private workspaces and selected-provider DNS aliases are automatic. Setup review discloses the broader GitHub Meta network delegation (the SendBox host caches Meta ranges and seeds each guest; guests never contact Meta). Public runtime updates stay separate from activation and tool permissions. The native combination record (not public yet) tracks current physical evidence separately from historical runs and the remaining milestone gates.

Explicitly configured local OpenAI-compatible models now share the same worker, permissions and accounting as Copilot. Endpoint discovery never scans hosts or downloads weights. DS4 with DeepSeek V4.1 Flash Q2 (not public yet) has a real installed-runtime qualification and an explicit, generated-context compatibility probe. New sessions resolve snapshotted global, current-profile and project controls, with explicit session overrides and visible provenance; existing sessions are not rewritten. Runtime and hardware inspection (not public yet) keeps actual execution-machine observations separate from endpoint availability and capacity estimates. Exact offline version execution needs a separate one-use user review; it never starts a model or installs weights.

Automatic Kupo exposes Balanced, Savings and Performance as routing modes, separate from exact shared role-effort and working-window controls. It selects only compatible actual roles within the selected boundary; Savings/Performance require reviewed evaluation and reference-price profiles, never guessed frontier classes. Bounded Cascade and Critique retain original role results, and text compaction (on by default, following the shared smart compaction setting) preserves readable history and source provenance. Finite automatic work uses complete all-role holds and selected-boundary finishing. The owned transport contract (not public yet) keeps SSH/native bounded decisions and metadata finishing on their original parent; neither approves code or Local Git. Managed native scheduling can prepare its source without a model turn. Snapshot retries and transport checkpoints (not public yet) add explicit opt-in original-snapshot restaging, separately accepted peer Git and disclosed native-model/Local-Git publication. Jev/System One still needs a real bounded output/accounting contract; generated fixtures are not live-model quality or billing evidence.

Reviewed candidate resources (not public yet) add pinned-SSH worktrees and native private-snapshot staging, exact project tests and separately accepted source application. Native implementation remains behind current released-runtime qualification, not a Local fallback. Private native acceptance does not publish to Local Git or replace that separate review.

Settings / Models / Providers can disable providers or unwanted models per profile. Matching upstream model IDs share one explicit provider choice, without duplicates or silent fallback. The same availability gates /model, routing and new calls. Disabling the selected model requires a fresh selection, even after re-enabling. Provider/model controls (not public yet).

Isolated profiles use immutable SDK clients over the same local authority. Desktop's profile selector, TUI F8 and CLI --profile ID preserve separate accounts, projects, drafts, memory and artifacts. Creating a profile neither copies data nor selects it. Global environment policy, shared startup preferences and optional global hard budgets cannot be escaped by switching. Old owned work stays in its original profile. Remote writers sharing a pinned peer and canonical workspace serialize across profile/target aliases and Local mappings; failed-root quarantine survives restart. Guided questions can include an explicitly labeled current-solver recommendation without another model call. You still choose the actual answer or tool approval; advice retains its model provenance and cannot grant permissions.

Context supports project/client text, locally normalized clipboard/file images, exact reviewed folders, typed GitHub issue/PR references and authorized session-text snapshots. Unsupported image models preserve the draft/queue rather than dropping context. Bare attachment prefixes remain literal until the user selects their source.

The desktop interaction update (not public yet) adds a compact five-source attachment menu, retained previews/removal, reviewed session drag-to-attach, persisted session order and confirmation preferences. Use the whole session row or Alt+Up/Down to reorder siblings, including children. Dragging a child out asks before detaching, with a saved opt-out in Settings / Confirmations. Unpin has a crossed-out pin. Global views live on the left; the right rail contains the session tools and per-session Code change and Work item links. Click an active right-rail icon again to collapse its pane back to the rail. Terminal belongs to a session: project-free consoles start in configured HOME, and project consoles keep their bound worktree. There is no global-terminal UI. Ordinary project-free Local chats also use their captured Home (not public yet) for approved file and command tools, without creating a project. Browser control keeps its explicit user toggle; GitHub commands need an explicit repository target. Opening Terminal or explicitly switching chats with it visible starts or reuses the exact selected console; + is independent of unrelated UI work. Enable local voice in Settings to download the sole pinned model; disabling voice deletes only that Moguri-managed model. Click the microphone to record and click again to confirm; hold Ctrl+Space for push-to-talk and release to finish. Shortcut toggle mode is an explicit setting, independent of the microphone button and consented focused-composer standby. Words stream into the draft without an Insert step or automatic prompt submission. Side chat opens immediately, read-only, with inherited controls and captured main context.

Session controls and interactive tools (not public yet) add visible stable IDs, reviewed ID context, success-only dependencies, same-project bulk actions and pinning, multiple verified PRs with an Actions/reviews pane, real JS/TS definitions/references, typed array helpers, isolated interactive HTML and owned Local Playwright/browser screenshots attached to chat. These are bounded alpha capabilities, not native OS control or a general-purpose automation engine.

Chat activity (not public yet) uses small operation icons, expandable original arguments/results, observed timing/output quantities and live timers. Responses have copy/info actions; the compact budget view puts estimated USD above tokens/context. Private reasoning and unknown costs are never invented.

The Agent terminal tab (not public yet) shows Moguri's approved commands and output. The agent can reread its own command history; the separate User and Global consoles remain private. Commands keep their existing approvals and per-command PTY lifecycle, not shared shell state.

Lean tooling (not public yet) adds lazy typed discovery, compact recoverable results, deterministic file/data/document operations and browser-free search. Root-bound main/child messaging and orchestration preserve read-only side chats, actual actors, approvals and aggregate budgets. Shared workflows include /review, /security, /performance, /debug and the new design, continuation, child-session and critique commands (not public yet). Rank equivalent integrations in Settings: Gateway MCP, direct MCP, REST, then CLI by default. Direct schemas keep the same exact MCP execution approvals.

The workspace experience (not public yet) adds chart-first My work, Usage, Cost and Velocity dashboard pages and a profile-private skill gateway for local folders and refreshed remote archives. Settings use compact information/copy controls and contain Artifacts, Candidates and unified Sandbox management. Suggestions offers global local usage-gap checks; /suggestions defaults to the current project's scope. My work also has the configured GitHub attention inbox (not public yet), with explicit refresh rather than a recurring scheduler. Model generation still requires exact source review. Alerts have a dismissible X without clearing the underlying work or draft. Read-only Kanban sync retains its exact source across presentation updates; see the alert and observation contract (not public yet). Automations holds recurring automations and one-time scheduled prompts. Work item links the project's Kanban cards or GitHub issues; session PRs open in Code change. Copilot connection status stays in Settings, not the left sidebar.

PDF, Excel (.xlsx) and Word (.docx) use locally extracted, encrypted document snapshots with bounded, approved page/section/cell inspection. Originals stay retained; formulas never execute and scanned PDF pages explicitly report unavailable OCR. Source/text files, including Java, and images remain supported. See document limits and usage (not public yet).

Implemented model controls offer the model maximum and half maximum, or actual provider-declared working windows, through the same service contract. Unsupported saved values block inference without clamping history. Pastes over 50 lines become exact scoped text files, not whole-paste prompt expansions. Desktop, TUI and the SDK preserve failed/conflicting unsent pastes; approved bounded file reads work through Local and pinned SSH workers without modifying a source checkout. See usage (not public yet) and the context boundary (not public yet). These context features are implemented. Final native and integrated acceptance remain separate in the current roadmap.

Core Remote SSH execution uses an explicitly pinned host and approved remote workspace. The existing worker runs inference and file/command tools there; the local service still owns approvals and history. Disconnect/restart reconciliation retains late accounting without replaying effects or turning partial work into success. Files, Git changes and read-only plan views now use the actual SSH workspace through the same SDK and slash commands. The local service never substitutes an attached Local checkout for a Remote workspace or index. Explicit Remote plan saves now bind the reviewed peer/root/path/digest and run one owned filesystem operation without a model, preserving independent Remote-boundary snapshots. Stops, lost receipts, changed scopes and recovery keep partial outcomes honest and never replay writes. Qualified 0.18.0-alpha.2 adds reviewed peer/resource Remote MCP grants, actual model/external SSH calls and no-replay retained-result recovery, with shared CLI/TUI and desktop review/inspection. Forwarding access is separately opt-in; refresh credentials never leave their backend. This is still an initial macOS alpha, not the completed first milestone. Qualified 0.19.0-alpha.1 / protocol 25 also adds explicit Remote user consoles with captured SSH host/root, original-owner restart recovery and ephemeral Ghostty/CLI/TUI input. The exact SDK and macOS app are independently qualified; setup (not public yet) records their immutable identities. No terminal is silently relabeled Local or Sandbox. Qualified 0.20.0-alpha.1 / protocol 26 also adds reviewed private repository metadata through the selected SSH host's GitHub CLI account, with retained peer/root/ build provenance, no provider request and no laptop-account fallback. Its exact installed SDK and actual packaged app are independently qualified; setup (not public yet) records the immutable identities. At that checkpoint, Remote repository memory remained unavailable. Qualified 0.21.0-alpha.1 / protocol 27 now adds session-bound Local/SSH lexical indexing, actual-host freshness checks, optional model context/direct search and reviewed builtin memory MCP tools. SDK, CLI, native TUI and desktop share exact peer/root identity and no-replay index jobs. The exact installed SDK and actual packaged app are independently qualified; setup (not public yet) records their immutable identities. Old two-tool builtin MCP grants require fresh review, never silent widening. Earlier archives remain unchanged. Qualified 0.22.0-alpha.1 / protocol 29 adds a shared Local candidate workflow: explicitly stage text proposals in a separate clean owned worktree, review again, then apply only exact base-bound bytes. SDK, CLI, native TUI and desktop preserve independent history, real user acceptance, raw index/HEAD and partial/restart outcomes without a model or automatic commit. The exact installed SDK and actual packaged app are independently qualified; setup (not public yet) records their immutable identities. Compound model routing, finite kupo wind-down and Max still remain unavailable, not implied by the private Cascade/Critique transport proofs. Qualified 0.23.0-alpha.1 / protocol 31 adds independently retained planned budget capacity and one explicitly cost-reviewed, tool-less Local model proposal through the same SDK/CLI/TUI/desktop. Generated source/provenance and actual charges remain inspectable; staging and application still need their two separate reviews. Its exact installed SDK and actual packaged app pass independent qualification; setup (not public yet) records immutable identities. This is not completed compound routing or the first milestone; see the generation contract (not public yet). The 0.24 source foundation adds internal complete planned-capacity batch calculation, not another public model workflow or new native delivery. Recorded and prospective roles share the same cumulative budget; a failed whole bound leaves no prefix. The qualified 0.23 archives stay frozen. Its private acknowledged-role probe reuses the actual worker and budget transactions, not another model loop. Public compound admission and real-model quality remain gated. Qualified 0.25.0-alpha.1 / protocol 32 adds independently retained exact role observations and shared read-only inspection through source deletion/ restart. Model data is not a user answer, code approval or normal solver context. Its exact installed SDK and actual packaged app are independently qualified; setup (not public yet) records immutable identities. This includes the 0.24 prerequisites, not public compound scheduling, real-family quality or finite kupo. Earlier archives stay unchanged. The 0.26 source connects those complete retained role envelopes to the actual control-plane worker, with conservative future contexts and no-replay ownership. This remains an internal Node primitive, not public compound scheduling or a new qualified native delivery. Qualified 0.27.0-alpha.1 / protocol 33 links a final retained proposer to its canonical candidate origin through shared SDK/CLI/TUI/desktop inspection and separate staging/application. Its independent copy retains raw model evidence, not quality or user acceptance. Setup (not public yet) records complete source, independently installed SDK and exact native delivery qualification and immutable identities. Earlier archives remain frozen. Qualified 0.28.0-alpha.1 / protocol 34 adds explicit provider-declared Copilot metadata inspection through the SDK, CLI, native TUI and existing desktop Model menu. Normal login/discovery/inference is unchanged; unknown family stays unknown, and declarations never attest independence, quality or approval. The contract (not public yet) and setup (not public yet) separate its source, exact installed SDK and actual native qualification from the frozen 0.27 delivery. Earlier archives remain unchanged. Qualified 0.29.0-alpha.1 / protocol 35 adds explicit output/context contracts to the existing internal owned executor and shared read-only contract inspection. Router explanations cannot enter a contracted solver's selected context; rejected or malformed advice stays charged and cannot start the next role. This is not public compound scheduling, independent-family proof or acceptance. The contract (not public yet) records complete source, independently installed SDK and actual native qualification. The qualified 0.28 artifacts remain frozen.

Milestone 1 is accepted at protocol 106 with SendBox 0.2.7: all 10,080 required Local/SSH/Sandbox cells, ten physical native gates and the full source, hidden desktop, SDK and package runs pass. Regression runs now use a 78-cell covering array; the full product is opt-in npm run test:matrix (D38 (not public yet)). The acceptance record and roadmap maps every milestone criterion to its evidence, records the user-directed replacement of account-blocked hosted CI with local evidence, and lists the remaining roadmap owned by #39 (not public yet). It is still a development alpha: not signed, notarized or publicly distributed.

The earlier issue-17 checkpoint used protocol 37 for shared model windows, scoped long-paste files and reviewed SendBox execution. SDK, CLI/TUI and hidden desktop run the same qualified private-stage worker, Manual decisions, budget ledger, exact-owner stop/recovery and separate actual-user source publication. Actual guest/control-loss checks retain partial work and unknown usage without replay. Setup (not public yet) records the generated provider evidence and remaining gates; this is not a replacement for frozen archives, live-model quality or full first-milestone acceptance.

Chats can have no project, without inheriting any workspace root or coding tools. One desktop Start chooser offers a new chat, an existing project, local-folder import or explicit repository clone. General conversations live under Chats. /suggestions previews scoped prior prompts before model analysis; ask Moguri and /meta open a settings chat whose changes require exact human review. MCP HTTP servers support explicit OAuth with PKCE, resource-bound keychain credentials and refresh. Shared question, approval and done indicators appear in both clients. Desktop OS banners are off by default and use generic content unless the user opts into previews.

The desktop composer is an icon toolbar: attach, environment, model name, autonomy, unlimited-budget/usage, microphone and paper-plane Send. Holding Cmd/Ctrl turns idle Send into a queue clock; active turns show a red Stop icon in that same slot. A blocked session uses a Resume/play control there instead, preserving the unsent draft; inspect effects before continuing. Enter uses the current action; Shift+Enter adds a line. Both sidebars resize, and message copy appears on hover or keyboard focus. Toggle/dropdown preferences, model context and supported thinking effort save automatically through shared SDK controls; text settings retain explicit saves. Model switches preserve the original effort preference, with a supported fallback and restoration on capable models, not a reset to Model default. /thinking and /effort set supported effort; /context-window sets an exact token ceiling or min, half, max without running a model. See desktop controls and settings (not public yet).

New chats can use a fixed global default model/effort or the current profile's last explicit selection. Existing chats and higher-priority overrides stay unchanged. See model defaults (not public yet).

The session workspace adds queue-only Cmd/Ctrl+Enter, shared chat slash commands, enforced read-only sessions and parallel question-only side chats. Opening Side chat immediately gives the same composer, model/thinking controls and voice input with captured main-chat context, permanently read-only. Hold Ctrl+Space to dictate locally: words stream into the textbox and release flushes them without sending. The microphone button instead uses click-to-record, click-again-to-confirm, never press-and-hold. Session names and settings support fuzzy search. Desktop embeds Ghostty through ghostty-web, alongside local diffs, plans, isolated browser previews, linked PRs and Insights. Double-click a session to rename it. Sessions support pin/group/ attach/unread/archive/delete actions; project settings and optional archived retention share the same SDK. Permanent session deletion also deletes its originating candidates and generation proposals; archiving alone preserves them. Fresh verification of an exact linked PR's merge also cleans up its candidates. Ordinary artifacts, applied code, worktrees and external PRs survive. See candidate retention (not public yet). A terminal is an explicit Local or pinned SSH user console, not an agent permission or a claimed sandbox.

The Local Browser pane also supports approved agent navigation, inspection, clicks and text input. Its top Agent control switch starts off and revokes access immediately when disabled. It never adopts a personal browser. See browser control (not public yet).

Attachments stay compact in the composer/chat. Clicking an image opens the large, uncropped viewer without changing captured context.

Project-free sessions can open private HOME consoles without granting model workspace access. Project-free Autopilot cannot invent workspace access, and side chats expose an explicit concrete-model selector when needed. User consoles retain authenticated per-profile/peer owners across explicit service restart or controller loss. Ghostty and shared SDK/CLI/TUI inspection reattach the same process, never replay input or automatically replace an old console. Ordinary service stop closes known groups; private cleanup failures retain explicit retryable intent. Output stays bounded and volatile, not conversation/model context. See the Local ownership (not public yet) and Remote ownership (not public yet) boundaries.

Managed Local worktrees isolate branch checkouts without moving the attached base folder or index. Create, discover, inspect, open and safely remove them through the shared SDK, CLI/TUI or project settings. Worker tools, Files, Changes, Plan, terminals, memory and dependency installs resolve the same bound root. Writers serialize within a root; different worktrees can run concurrently. Removal requires a clean, idle, unbound owned checkout and preserves its branch. Discovery does not adopt or grant access to unmanaged worktrees.

Independent artifacts retain plans, decisions, answers, research notes and reports with revisions and actual source provenance. Create them without a conversation, or explicitly retain a completed user/model/tool result. Each has its own encrypted retention grant. Project removal previews the exact set and requires retain, export-then-delete or delete; ordinary documents survive session retention. Candidate/proposal reports are the explicit session-owned exception. Real structured clarifications record actual user answers and retain them by default, with an explicit opt-out. Answers never grant tool/Git permissions. Desktop answer controls (not public yet) use a scrollable choice list and a blank Other field that sends on Enter. Documents can associate with profile folders and project work items. Explicit reviewed repository copies preserve complete history under artifacts/, with an optional new local Git branch. No checkout, push or unrelated staging changes; project deletion re-verifies the actual copy and any requested commit.

Dictation uses local NVIDIA Nemotron 3.5 through a pinned native runtime. The user explicitly accepts and downloads the model once. NVIDIA documents 40 locales: 19 ready, 13 broad coverage, 8 adaptation-only, not 22 verified ready languages. No cloud ASR fallback or automatic model-weight download occurs. You can instead trust an existing streaming model file, such as a Nemotron streaming GGUF or Handy's Parakeet/Moonshine streaming GGUF; Moguri locks its SHA-256, asks again if it changes and never deletes it (setup (not public yet)).

Projects also have a local Kanban for feature, bug and misc tickets, with their own board chat, shared Show/Hide preference and same-panel GitHub/Linear sources, per-column creation and drag/drop. /continue-work or the prompt continue work asks each time whether to use Kanban or send the prompt normally. Kanban takes one eligible item in an isolated Local checkout through its draft PR; it does not merge or drain the backlog. Project secondary click opens actions; dragging or Alt+Up/Down reorders. Work item can select a Kanban card. Setup and boundaries (not public yet). Sync with GitHub updates linked status into Kanban only and can optionally import open repository issues. Draft PRs mean in progress, ready PRs mean review, and completed/won't-do closed issues mean done. Optional Linear imports start off, use a user-supplied keychain credential and never upload local tickets.

The shared dependency cache reuses SHA-512-verified archives across projects and branches, with npm and pnpm adapters. Exact lock/manifest plans and fresh OSV advisory checks gate downloads and reuse. Managed installs block lifecycle scripts and hooks, use private staging and keep mutable trees worktree-local. This is not a malware-free guarantee; script execution stays unavailable until a real SendBox boundary is qualified.

This is a macOS development alpha that completes Milestone 1; it is not a signed, notarized public release. Explicit estimate-based token/USD caps and separate profile/global hard policies now share cumulative role-aware reservations. Unknown completion holds capacity; deletion, defaults and restart do not refund it. Named Copilot/reseller models can use originating-provider API-equivalent estimates, not subscription bills. Finite USD needs a fresh user rate or explicit reference-rate consent; missing/stale prices remain unknown. These are admission limits, not exact provider-billing guarantees. CLI/TUI, desktop composer, settings and the profile usage dashboard share the SDK. Dashboard views now add UTC presets/year/custom ranges, calendar buckets, captured price history and retained per-call/owning-turn records. Observed stream delivery metrics disclose missing coverage and are not GPU-throughput or productivity claims; cache subsets never inflate reported totals. Tool/MCP outcomes now have a sanitized, session-independent ledger with an explicit known-result failure rate and separate denied/stopped/partial/unknown counts. Dashboard, CLI/TUI and SDK also expose current-profile key-backed ciphertext sizes, not physical disk usage or reclaimed space. Earlier deleted tool history remains unavailable. Repository/organization/visibility filters now use explicitly refreshed, root-bound GitHub metadata and immutable per-admission snapshots. Every tagged step rechecks its actual Local/SSH Git origin. Unknown history is not backfilled, and the open-source category is explicitly public plus detected OSI-approved license metadata, not a legal review or a synonym for public.

Reviewed text-context compaction now uses one explicitly authorized, tool-less Local/SSH model call through the same SDK. Previewing does not infer; generation, actual history/content review, apply/discard and original-context restore are separate actions. Complete recent turns, original encrypted history and drafts remain retained; rejected/unknown calls still count. Desktop exposes review under Model, and CLI/TUI share /compact. Images and implicit kupo compactor selection are not supported.

Smart compaction is on by default: when a chat nears its model's working window (75%, or 200k tokens), Moguri summarizes older turns with that chat's model in one metered call and keeps going, with original history restorable. Type /compact to compact now. Built-in slash commands always win over a skill with the same name, and Moguri asks you to rename such skills.

Kupo (not public yet) is always in the model picker. Its automatic policy uses configured models: none blocks inference, one runs directly in Single, and multiple models support task-based Single, gated Cascade or eligible read-only Critique with one revision. Declared capability categories guide routing, including pools of three through twenty models. Only the selected response reaches normal approvals; every actual role is metered. Independent critique still needs reviewed lineage evidence. Finite automatic budgets, images, synthetic tuning and Sandbox routing remain unavailable without fallback. Existing explicit Single policies (not public yet) remain supported separately.

Reviewed Local Cascade/Critique (not public yet) now has a shared development service/SDK, CLI, actual TUI and desktop path. An actual classifier binds a declared concrete route; tool-less proposals retain the task plan, architecture and implementation. Actual-user review separately authorizes staging and exact project tests in distinct owned worktrees. Failed drafts remain isolated, and a recommended candidate still needs its separate source-application review. Reviewed lineage sources are limited user policy claims, not proof of independent training or calibrated quality.

Finite Local routing now reserves one actual metadata-only finishing role and has separately reviewed actual-user accepted-work checkpoint branches that preserve the source index/HEAD. A summary never accepts code or approves Git. Local Autopilot Max (not public yet) now uses an explicitly configured bounded-model System One-compatible alternative with fixed or Single-routed solvers. Exact existing tool delegation, user-declared technical choices and bounded safe-boundary steering share the real model ledger and cancellation owner. The actual actor remains decision-provider, not the user; human intervention and hard policy win. Jev itself remains blocked where its documented API cannot enforce the required output bound. No generative substitution or calibrated reliability is claimed.

The current generated source, packed SDK and exact native image pass the full acceptance sweep (not public yet). Live-provider/calibrated-quality gates, complete remote/native side-panel/extension transport, Remote/native compound/Max and remote/native wind-down remain incomplete. Unknown usage and Copilot billing are not presented as zero cost. Manual remains the default. Guided still needs human confirmation; Autopilot can use only explicit bounded user grants. SendBox has a visible fail-closed configuration/readiness surface. The development SDK/client lifecycle now passes actual native generated-provider checks, but readiness still needs explicit reviewed release/image/network pins and private staging. Endpoint-only cold networking remains blocked; no alias or Local fallback is added automatically. Installing SendBox alone does not enable native execution.

npm ci
npm run build
npm start -- --project /absolute/path/to/project

Run npm run desktop in another terminal to open the same sessions in desktop. In the TUI, Ctrl+L starts Copilot sign-in, Ctrl+M selects a model, and Ctrl+K shows commands. Desktop exposes the same setup in GitHub Copilot. Complete GitHub's device authorization yourself; Moguri does not read IDE login files or require another provider's API key.

See setup and usage (not public yet), the SDK API and CLI, the initial architecture, and Phase 0 evidence and remaining release blockers (not public yet). The session workspace contract (not public yet) records Ghostty qualification, deletion/recovery and project-policy boundaries. The engagement contract (not public yet) records OAuth, project-free isolation, suggestions, reviewed configuration and notifications. The runtime foundation (not public yet) records local models/defaults and the genuine SendBox qualification blocker; typed inputs (not public yet) records snapshot consent and image/source boundaries. The managed workspace contract (not public yet) records Local ownership, root consistency and partial-operation reconciliation. npm run package:mac produces an ad-hoc-signed development app, DMG and ZIP under release/. It is not a notarized distribution. The app includes the CLI, TUI and their shared Node runtime. Copying it does not start the TUI or add a shell command. Use Settings > General > Command line > Install moguri command, then follow the PATH setup instructions (not public yet). That guide also documents removal, moved-app recovery and the manual workaround for older builds. CI runs only on pushes to main, not pull requests or feature branches. The explicit 30-minute cap is removed; GitHub-hosted jobs still have the platform's six-hour limit.

The in-app updater follows the upstream mbianchidev/moguri GitHub Releases. Public releases are checked, verified and staged without a token, so once the repository is public, packaged apps update automatically, and installs still wait for your confirmation. An optional read-only token covers the private period and GitHub's rate limit; see GitHub updates.

Settings has a shorter rail: 7 groups and 28 entries, down from 11 groups and 37. Integrations and Interface are each one rail entry with keyboard-navigable tabs, and each integration tab shows its status. See Settings rail structure.

Back to the user guide