macOS alpha, 0.32.0-alpha.1

First-milestone acceptance and roadmap

First-milestone acceptance, what is not available yet and the ordered roadmap.

Published unchanged from docs/remaining-milestone-gaps.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.

Milestone 1 accepted, 2026-11-15

Milestone 1 is accepted on draft #68 (not public yet) at final cut 18e3f78 (protocol 106, 0.32.0-alpha.1, SendBox 0.2.7), with native runs at e67faf5 whose runtime and native harness bytes are identical. Every required nominal cell, native gate and milestone criterion passes. #39 (not public yet) owns this acceptance; #17 was consolidated into it. Merging #68 delivers it to main.

User-directed scope revision

Hosted-main CI stayed blocked before any job step by account billing (latest run 43652897223 (not public yet)). On 2026-11-15 the user explicitly chose to accept full local source/desktop/native evidence instead of hosted CI for Milestone 1. The former "fresh authorized hosted-main CI pass" gate is therefore replaced by the complete local runs below; it was not silently dropped. CI remains main-push-only, and a future hosted run is ordinary regression evidence, not a Milestone 1 prerequisite.

Acceptance gates

  • Criterion-by-criterion integrated qualification against the unchanged specification, with exact source/builds, cross-client ownership, failure/restart/migration and supported/unavailable outcomes (table below).
  • All 6,720 required Local/SSH source cells and the full source suite at the final cut.
  • All 3,360 native Sandbox cells in one whole-matrix run on the current approved worker image, plus every native control/role/input, loss, cleanup, candidate, finishing, client and concurrency case, and the previously deferred native desktop paths.
  • Full hidden desktop suite, packed external SDK consumer and dependency audit at the final cut.
  • Local evidence in place of hosted CI, per the user-directed revision above.

Exact identities, results and retained report hashes are in the final qualification record (not public yet).

Run Result
Ten physical native gates 10/10 pass
Native nominal matrix 3,360/3,360 pass in one run (1,056 guest executions with owned-guest teardown, 2,304 pre-admission refusals)
Native combination file 3,483/3,483 pass
Native SDK/TUI/script/loss cases all pass (one test-isolation defect fixed and retained)
Native hidden desktop 26/26 pass
Full source suite (incl. 6,720 Local/SSH cells) 10,908 pass, 66 native-gated skips, 2 load-timing failures that pass on rerun (6,720/6,720 Local/SSH cells PASS)
Full hidden desktop suite 433 pass, 12 skips (11 native-gated pass natively; packaged CLI passes), 0 failures
Packed SDK consumer, audit PASS; 0 vulnerabilities; packaged app verified

Criterion-by-criterion record

Each criterion is from the milestone acceptance criteria. The cited files ran in the full source suite; native rows also ran on the physical guest.

Criterion Executable evidence
TUI-created sessions are controllable in desktop and vice versa tui.test.ts, tui-milestone-combinations.test.ts, reconnect.test.ts, client-negotiation.test.ts; hidden desktop milestone-combinations.spec.ts over the same SDK
Restart loses no admitted/queued messages, drafts, artifacts or promised terminal state recovery.test.ts, queued-recovery.test.ts, draft.test.ts, terminal-recovery.test.ts, terminal-reattachment.test.ts, service-startup-resumption.test.ts
Managed worktrees without coupling worktree and session deletion managed-workspaces.test.ts, git-managed.test.ts, repository-workflow-defaults.test.ts
Queue editing/claiming under concurrent clients queued-drafts.test.ts, queued-continuation.test.ts, delivery.test.ts
Distinct Interrupt, Steer and Queue semantics delivery.test.ts, active-message-delivery.test.ts
Cancellation stops or reports failure to stop the owned process tree supervisor.test.ts, program.test.ts, command-jobs.test.ts, native-tree.test.ts
Artifacts survive source-session archival and deletion artifacts.test.ts, session-retention.test.ts, artifact-materialization.test.ts
Profile isolation of credentials, projects, policies, artifacts profiles.test.ts, profile-store.test.ts, profile-controller.test.ts, native native-profiles.test.ts
Dashboard totals trace to immutable events telemetry-dashboard.test.ts, dashboard-overview.test.ts, budget-metering.test.ts
Estimated, provider-reported and billed costs never conflated budgets.test.ts, provider-pricing.test.ts, stream-metrics.test.ts, ds4-compatibility.test.ts; every matrix cell asserts null billed USD
Every transition and failure produces a diagnostic event operation-telemetry.test.ts, worker-failures.test.ts, retained-telemetry.test.ts, desktop-alerts.test.ts
Unit, integration, migration, restart and cross-client coverage The full suite's *-migration, *-recovery and *-clients files, plus actual TUI and hidden desktop suites
Factory defaults and identical default resolution across clients model-defaults.test.ts, model-defaults-state.test.ts, session-controls.test.ts, settings-controls.test.ts
Allowed-env enforcement and the running-Local-turn exception profiles.test.ts (atomic global environment policy), platform.test.ts, matrix policy cases
No host fallback, unprotected model request or false cleanup on SendBox failure sandbox-runtime.test.ts, sandbox-recovery.test.ts, sandbox-containment.test.ts; native gates N3/N7
Fixed model and Kupo independent of autonomy, Env and budget All 10,080 nominal cells; guided-advice.test.ts, questions.test.ts and the 160 actual-question cases
Kupo single, cascade and critique without violating constraints kupo.test.ts, kupo-combinations.test.ts, compound-envs.test.ts, candidate-envs.test.ts; native candidate cases
Router, critic and decision calls accounted for and cancellable compound-ledger.test.ts, compound-cancellation.test.ts, kupo-recovery.test.ts, decisions-controller-loss.test.ts, decisions-transport-loss.test.ts
Real actors and delegation; human/policy wins decision races decisions-authority.test.ts, decisions-policy.test.ts, approval-memory.test.ts, questions.test.ts
All six typed attachment types with durable drafts and source permissions context-inputs.test.ts, document-attachments.test.ts, paste-files.test.ts, paste-file-state.test.ts
Remote SSH runs agent work and model calls with truthful loss handling remote.test.ts, remote-transport.test.ts, remote-recovery.test.ts, profile-remote-roots.test.ts; 3,360 SSH cells
Token/USD caps, all roles, late telemetry and authorized wind-down budget-plan-batches.test.ts, kupo-budgets.test.ts, decisions-budgets.test.ts, winddown-commit.test.ts, winddown-budget.test.ts

Not claimed

Acceptance uses generated inference and fixtures. It does not claim live provider/decision quality, calibrated independence or confidence, actual billed cost, public signing/notarization/distribution, Linux release readiness or a hosted CI pass. Jev's missing enforceable wire stays blocked; optional Laya is not a milestone gate.

Delivered on main, not remaining implementation

Delivery Current executable scope and evidence
#51; closed #36/#37 Shared runtime/hardware readiness, one-use reviewed executable probes, exact pooled window/effort, finite all-role reservations and reviewed Local/SSH/private-native candidates. Shared sessions, queues, profiles, independent artifacts, worktrees, typed context and Ghostty ownership already exist.
#52; closed #7 Retained exact-PR CI watches, deduplication, original observations, loss/restart recovery and notifications across SDK/CLI/TUI/desktop. PR/CI contract (not public yet).
#56; #38 is closed Complete required Local/SSH source cell/criterion mapping, GitHub/input and control repairs. The matrix (not public yet) retains all 10,080 nominal cells as the opt-in full product; #56 itself disclaimed native acceptance, which the final qualification later completed.
#57 Narrow actual DS4/DeepSeek compatibility and cache-accounting evidence, not a general model-quality/capacity or all-Env claim. Exact contract (not public yet).
#58; closed #9 Unified authorized cross-repository GitHub My Work inbox, distinct from the existing active-session dashboard and the session notification Inbox. Refresh is read-only and profile/account/repository-bound. My Work contract (not public yet).
#59; selected #23 slice Original-owner startup/reconnect, background-window lifecycle and opt-in metadata-only OTLP traces/metrics over control-plane records. Export is off by default; logs, content capture and peer/guest-internal auto-instrumentation are not implemented. Startup (not public yet) and OTLP (not public yet).
#60; closed #13 and selected #45 slice Governed complete Local skill-source admission/provenance/least privilege/revocation, justified complete-source/envelope limits, per-turn checkpoints and reviewed Local/pinned-SSH/private-Sandbox rewind. Also actual named skill-load indicators, session Inbox and centered empty states. Source/rewind (not public yet) and Inbox/load (not public yet) contracts.
#71; closes #63 All five #63 connection options through the shared service, SDK, desktop and TUI: GitHub Copilot (unchanged), OpenAI-compatible endpoints with custom headers, header-scoped keys and explicit-only models, OpenRouter (key or sign-in), the Anthropic API (key only; Claude.ai sign-in and subscription tokens refused) and OpenAI API keys separate from Sign in with ChatGPT. Separately decided D36 adds off-by-default Claude Agent and Codex Agent harnesses that run the user's own CLI. Tested against generated fakes only, not real accounts. Hosted provider (not public yet) and agent harness (not public yet) contracts.
#72; closes #69 (owner sign-off 2026-11-16) D38 (not public yet): npm test runs a checked-in 78-cell covering array (pairwise plus Env x selection x budget/input/effort tuples and the 15 client journeys) and 72 exact manual-Kupo-effort refusals instead of the 10,080-cell product; the full product is opt-in npm run test:matrix. Milestone 1's full-product acceptance is not re-qualified or relabeled. Method and measured times (not public yet).
#87; closes #6 Protocol 115 automations (not public yet) (D50 (not public yet)): once, interval, daily and weekly schedules with explicit activation, DST-aware wall times, missed occurrences run once and overlapping ones skipped, each run in a new chat with captured controls, chained prompt steps and model-free repository-trusted script steps under normal approval, stop/delete propagation and restart without replay, across SDK, desktop, TUI and CLI (YAML/JSON workflow files). Runs need the service running; server delegation and non-time triggers stay deferred. The same change (D51) saves drafts and open queued-prompt edits before the app closes and lets interrupted chats be continued one by one, all at once or automatically after a restart.
#81; closes #12 Optional, off-by-default Stainless integration (not public yet) (D39 (not public yet)): Info.plist-only detection, a user-run headless compatibility check, owned headless playwright batches with a disposable profile and bounded console/network diagnostics, and Open in Stainless for Local previews across SDK, desktop, TUI and CLI. Stainless has no MCP yet; that is #80.
#78; closed #45/#23 (2026-11-19) The remaining applicable #45 AST controls with owner-signed exclusions (D40 (not public yet)) and the remaining #23 OTLP scope (D41 (not public yet)), both described below.
#86; closes #82 Protocol 114 Skills and Artifacts tabs next to Automations (D49 (not public yet)): skill toggles, digest-bound disabled installs from configured skill sources, built-in toolsets, user- or agent-written script tools, local repeated-task tool suggestions with an off-by-default recurring check, and a metadata-only per-project artifact library linked to each chat. Contract (not public yet).
#90; closes #84 Settings rail reduced from 11 groups / 37 entries to 7 groups / 28 (structure). Integrations (GitHub, Linear, MCP gateway, sql-not-so-lite, Stainless, Plugins, Marketplace) and Interface (Appearance, Keyboard shortcuts, Voice dictation, Notifications) are single rail entries with WAI-ARIA keyboard tabs; integration tabs show text status from already-loaded state. Item ids, deep links, restored layouts, search, save rules and dirty-state blocking are unchanged; no schema change. Covered by tests/settings-navigation.test.ts and the hidden tests/e2e/settings-navigation.spec.ts.

Closed #35 keeps its historical exact-artifact scope. Closed #22 is a completed Rust feasibility study, not production adoption; #1 is reference material. No closed implementation/research issue certifies first-milestone acceptance.

Ordered remaining roadmap

Ordering is triage, not permission to implement or an invented dependency chain. Delivered #6/#7/#9/#12/#13/#23/#36/#37/#38/#45/#53/#55/#63/#69/#82/#84 do not reappear as new implementation work. #53's audit, visual system and UI/UX polish are recorded in the design audit (not public yet); #55 smart compaction is on by default since protocol 110 (D31 revision (not public yet)).

Order Actual remaining scope
1. Everyday UX and existing integrations #77 app-wide accessibility verification (screen reader, WCAG 2.2 AA, keyboard/focus, physical-display contrast), deferred from #53.
2. Distribution and docs #26 Homebrew cask/access/install/update ownership; #14 landing/downloads/user/developer site. The owner chose the MIT license, a hybrid local-build/CI-publish release process and a Homebrew cask on 2026-11-19; the repository stays private for now. The cask waits for a published public release, and signing/notarization still needs a decision.
3. Optional providers and automation #80 Stainless MCP integration once mbianchidev/stainless#22 ships, #34 Laya (its real-weight qualification needs a separately provisioned Laya service), #8 dependency roundups, #10 continuous personas, #11 bounded specialized presets, #46 Slack mentions, #85 phone steering over Tailscale Serve (scoped paired gateway and PWA client; /remote is already the SSH command), #54 GetMoshi integration and #64 global/repository flags for evidence-gated flaky-CI retries (provider status preflight, bounded attempts) and feature-preserving conflict resolution. Automations (#6), exact-PR CI watches (#52), the reviewed rebase/conflict workflow and reviewed compaction are not those broader features.
4. Separate product #67 moguri-assistant, a separate repository embedding the shared SDK as a personal assistant (calendar, mail, messaging, releases, browser and more). Not started here; each external tool needs its own credential, permission and data-handling decision.

Protocol 112 delivers the remaining #23 scope (D41 (not public yet)): opt-in fixed-template diagnostic logs correlated with trace/span IDs, finite failure outcomes, correction logs instead of duplicate spans, secret-backend credential headers bound to the collector origin, measured per-turn overhead and the process/Env coverage matrix (not public yet). SSH-peer and Sandbox-guest internal propagation stay an explicit documented gap, content capture stays out of scope, and there is still no second ledger.

Protocol 112 also delivers the remaining applicable #45 controls (D40 (not public yet)): a deterministic source indicator scanner that gates model-approved scripts (the user may accept the risk) and refuses hidden-character trusted-folder loads, per-profile saved publisher keys with permanent key/version blocks, and generated-model AST01 qualification. The project owner signed off on 2026-11-17 on excluding a transparency log/PKI, Local whole-descendant isolation, malware/semantic scanning beyond the deterministic rules and live-model testing. Remote reviewed scripts stay unsupported, while Remote SSH sessions now load skills on demand (D42 (not public yet)): the same admission rules as Local, with the complete admitted package copied to a digest-named folder on the SSH host and its scripts run through normal command approval. The AST category matrix (not public yet) records controls, evidence and those exclusions. Neither change is OWASP certification or a claim of zero risk. Protocol 112 also lets a user steer move a running agent command to the background instead of waiting for it (D43 (not public yet)), and every client asks before archiving or deleting running sessions (D44 (not public yet)). Code Change adds status-gated Address code review, Fix merge conflicts and Fix CI actions for linked pull requests (D45 (not public yet)). Archiving now removes a clean, fully pushed session worktree and its local branch, and restoring recreates it (D46 (not public yet)). An optional virtual disk caps the space new session worktrees can use, cleaning up archived pushed worktrees and then pausing work when it fills (D47 (not public yet)). It covers worktrees only; age/size deletion policies for the encrypted session store remain deferred. Queued prompts now look like sent prompts, edit in place with every attachment source, and Send now steers them into the running turn (D48 (not public yet)). Protocol 114 adds Skills and Artifacts tabs next to Automations (#82, D49 (not public yet)): skill toggles, installs from configured skill sources, toolsets, user- or agent-written script tools, repeated-task tool suggestions with an optional recurring check, and a per-project artifact library linked to each chat. A hosted multi-publisher gallery and recurring prompt/workflow schedules (#6) remain open.

At the live read the repository is private and has no GitHub releases. It is MIT-licensed, with a tag-triggered release process whose first run waits for GitHub Actions billing to be fixed. Development macOS app/DMG/ZIP artifacts are ad-hoc signed, not notarized or public distribution. This audit opens no other-repository PR, changes no upstream dependency and makes no publication/installation decision. Future Rust adoption still requires explicit direction, current-protocol parity, ADRs and a feature-equivalent value/distribution comparison.

Not silently added to the milestone

The owner subsequently requested the explicitly scoped protocol-41 platform extension (not public yet): Sky/Terra/Glacier knowledge, reviewed hooks/repository-backed marketplace, two-way GitHub/Linear workflows, private/workspace and Python/Rust/Go caches, and a qualified SendBox requirement for whole-descendant containment. This request does not change this handoff's first-milestone estimate or retrospectively qualify older images/archives.

Broader orchestrator/session groups, multi-repo coordination, Porto/database plugins, scheduling, specialized modes and managed cloud provisioning remain separate deferred work. Do not silently make them new first-milestone blockers. The cloud fleet on Agent Substrate analysis (not public yet) records how hosted fleets could follow production SendBox; it is not roadmap work.

Back to the user guide