Published release: v5.8.7 is officially published with Windows Setup/Portable, macOS (arm64/x64) and Linux (arm64/x64) installers, SHA256 checksums, a release manifest and five platform provenance records. Protected PR #210 merged as ; exact-main CI passed all 18 jobs, and Release workflow #38965349765 completed successfully.
lnwjud — local AI-agent runtime & MCP gateway
Étoiles
209
Outils
—
Disponibilité
—
1a0453ed0ef1de4cdf9ce1ba9edee8da0de8426b
Tunnel Reset All Settings / persistent alias: Factory reset reconciles only the verified stopped local lnwjud runtime alias through the official tunnel-client command, preserving unrelated tunnels and rejecting running, foreign or unverifiable ownership. Targeted desktop regressions passed; a separate real-notebook post-install smoke test is still recommended.
External MCP direct invocation: Content-only child providers no longer require an invented outer structured-result schema. Direct calls to Playwright, Context7 and codebase-memory-mcp were successfully verified on the v5.8.7 development runtime.
Issue #208 — guarded Git import: Dedicated git_clone_safe supports policy-approved, workspace-scoped HTTPS import pinned to an exact commit with native Windows create-only staging, NTFS identity checks, non-replace publication and fail-closed unknown-state recovery. Existing generic Git command restrictions and host authorization remain in force.
Native Git stage integrity: Verify trusted Git HEAD, reject tracked/untracked modifications and concealed index changes (assume-unchanged / skip-worktree) before publication. The Windows regression suite passed 5/5 after reproducing the original flaw.
Issue #209 — write safety: Strict create-only protection is enforced even under the Full permission profile. Workspace writer leases, global managed-writer admission, workspace_path_preflight, and process-lifetime fencing passed targeted and exact-main hosted validation. Explicit recovery of an abandoned same-client writer lease is allowed only when idle ownership is proven; unknown or live work remains blocked.
Security boundary: Safety claims cover untrusted caller input and writers mediated by the lnwjud Gateway. No guarantee is made against arbitrary same-Windows-user/administrator native processes operating outside the gateway; those processes are outside the tested contract.
Process startup safety: Failed pre-spawn process admission releases the reservation; a successfully spawned process retains it until verified zero-active native termination. The 35 process regressions and package typecheck passed.
Scheduled Task truthfulness: Disabled Native ChatGPT continuation was not silently recreated. No recurring native task has been claimed active for this release goal.
ภาษาไทย: เผยแพร่ v5.8.7 อย่างเป็นทางการแล้ว ตรวจ CI ที่ Merge SHA ตรงครบ 18 Jobs พร้อม Manifest, SHA256 และ Provenance ครบทุกแพลตฟอร์ม แก้ Git Import, Strict Create-Only, Writer Lease, Tunnel Reset และ Direct MCP; รับประกันความปลอดภัยในขอบเขต Writer ที่ผ่าน lnwjud Gateway เท่านั้น ไม่ครอบคลุมโปรเซสภายนอกที่ใช้สิทธิ์ Windows User เดียวกัน
What's new in v5.8.6
Engineering Gate checkpoint compatibility (Issue #198): Add checkpoint_goal_engineering with the same strict schema, Goal handler, lease/revision checks and evidence validation as checkpoint_goal. It follows the canonical tool's availability, so disabling checkpoint_goal disables both names.
Connector verification: After the v5.8.6 update, a fresh connector session advertised both checkpoint tools with the Engineering fields. Direct calls to both names and get_goal readback confirmed persistence. Connector display names are chosen by each user.
Engineering Harness audit routing (Issue #194): Keep Full Audit requests and read-only reviews of existing deployments or release history in the local review workflow. Explicit deploy, release, commit, push, PR and merge requests retain their matching delivery scope and gates.
Existing Engineering Goals: Reconcile workflow classification, gate contracts and source scope from the lease-acquired Goal snapshot. Preserve compatible findings, task tracking and proof; revalidate scope-dependent checks when the requested source scope changes.
Modern MCP HTTP inventory and schema evidence: Notify subscribed clients when host tool availability changes, then serve the current tools/list on the next request. tool_describe reports a SHA-256 digest from the active ToolRegistry schema and identifies remote connector schema as unavailable.
Durable MCP Shell results (Issue #193): Preserve exact client/session/workspace ownership when a modern HTTP request or legacy MCP transport is replaced. Authorized owners can retrieve retained task status, stdout/stderr and terminal receipts; foreign owners remain denied.
Checked workflows and recovery: Execute reviewed recipe steps only when requested, evaluate each step under current permissions and Goal leases, and distinguish unavailable provider evidence from completed work. Workspace indexing and context reporting retain explicit error/completeness states.
Agent Swarm: Fix non-interactive Codex CLI waiting for extra stdin and keep monitoring spawned agents after a start request ends. Verified with real local child-process fixtures; hosted Codex results still depend on account quota and provider availability.
Tool and native provider contracts: Synchronize all 285 declared MCP tools (273 default), improve truthful Doctor readiness and native media lifecycle boundaries. macOS/Linux capabilities remain subject to platform permissions and target-native verification.
v5.7.4 primarily fixed oversized native approval dialogs. v5.8.0 adds six review-first workflows, richer Doctor evidence, CSV/XLSX data tasks, Git explorer upgrades, a consistent Desktop UI, more reliable Secure MCP Tunnel startup, and controlled dependency maintenance.
Six bilingual workflows: Project Check, Code Review, Release Readiness, Connection Check, Data Audit, and Template Report. Required inputs are validated before preparing a draft; a preview never silently launches a Goal, scheduler, or AI worker.
MCP Call History: retained audit-backed calls with bounded pagination, tool/Goal filters, outcomes, and server p50/p95 when samples exist. Transport/Tunnel timing is still unknown where no trusted sample exists; only Goals with attributable calls appear in the Goal filter.
Goal Results and checkpoints: inspect meaningful progress, completed steps, recorded checks, actual mutation receipts and verifiable file artifacts. Result filters exclude Goals without evidence; verification is not inferred from a Goal's completed status.
Process Resources: display measured Windows working-set RAM and lifetime-average CPU for verified Goal-owned child processes; separately show the Desktop main process. Refresh while viewing, suppress Goals without measured resources, and never substitute zero for unavailable measurements.
Context Economy observability: show real transport-scoped MCP HTTP ledger hits and context bytes, explicitly not per-Goal totals. Shared services are not cancelled with a Goal; owned-task cancellation stays permission- and identity-checked.
Safer restore and audit visibility: result and restore flows preserve workspace scope, exact checkpoint/Goal revision checks, verified current file hashes, and explicit user confirmation.
CSV/XLSX data workflows: typed audit and comparison of tabular data plus template-based report creation with readback. Preserve source files by default; native Office actions remain dependent on platform/provider support.
Git explorer and diffs: show individual untracked files in new folders, not merely folder counts. Expand/collapse the tree, search and filter files, retain code/diff navigation, and show safe text/image previews or bounded binary metadata.
Desktop controls and typography: shared inputs, filters, buttons and responsive searchable dropdowns across Settings/Tools/Doctor/Git; menus anchor below their fields only, reserve outer scrolling room and scroll internally for long lists. Options wrap when long, compact options remain compact. Restore Prompt/previous Git text sizing and improve validation/loading feedback.
Command Prism status visual: replace the plain golden orb with a faceted prism, using actual MCP states (gold = ready, blue = working, red = stopped/offline). A reserved neutral gray visual does not claim a new backend state. Workflow cards receive matching restrained line icons.
Modal and layout fixes: correct shared-button CSS precedence so close glyphs stay visible in What's New, Tool Detail, and the guided Tunnel screen. Increase close targets, preserve keyboard access, and separate search fields from actions.
Work Log and Doctor UX: default Work Log to all workspaces; improve workspace/session filters, loading states, long labels, goal descriptions and meaningful empty states. Doctor Goal selectors show only entries that have relevant results for each tab.
Secure MCP Tunnel reliability: do not reconnect over a matching managed runtime before a 45-second elapsed readiness grace period. Keep Tunnel identity, ownership protection, health polling and retry backoff; this avoids premature repeated connect calls but does not guarantee a fixed startup time.
Bundled native runtime: update the verified target-native tunnel-client to 0.0.16 and preserve native payload preparation for installation on supported platforms.
Dependency updates and delivery checks: use one managed dependency-update PR with conditional branch cleanup rather than duplicate updates. Keep functional tests and exact-commit target-native CI/build gates; Windows community artifacts can remain unsigned when no signing certificate exists, with SHA-256/provenance checks required.
Latest published release: v5.8.7. The download buttons above point directly to the verified v5.8.7 assets. The release was published after the exact tagged main commit passed the target-native release gates.
See RELEASE_NOTES.md for previous releases and the complete release history.
Download lnwjud-Setup-<version>.exe for the normal installer, or lnwjud-Portable-<version>.exe if you prefer a portable executable.
Launch lnwjud, add the project/workspace you want to use, then configure the MCP connection method you need.
Community Windows artifacts may be unsigned when production code-signing credentials are not configured. Release verification still checks the declared signing mode, SHA-256 manifests, runtime provenance, and packaged-app smoke evidence; verify the release assets if Windows shows an unknown-publisher warning.
Windows-only features such as WSL, Registry, Windows Sandbox, Windows OCR and Outlook/COM remain available only where the Windows provider is supported.
macOS 13+
Download the artifact that matches the Mac architecture (arm64 for Apple silicon or x64 for Intel). Do not swap architectures just because both filenames contain the word “Mac”. Computers remain unusually literal about this.
See Install lnwjud on macOS for DMG/ZIP installation, permissions, MCP launcher paths, Gatekeeper/signing expectations, and troubleshooting.
Linux
Choose the release artifact that matches the Linux architecture and package type. Ubuntu 24.04 LTS x64 is the primary Linux release target; arm64 remains architecture-specific and evidence-gated.
See Install lnwjud on Linux for AppImage/DEB installation, Wayland/X11 behavior, secure storage, MCP, and platform limits.
Local data vs workspace metadata
Installed lnwjud keeps per-user runtime data outside your source repository: %APPDATA%\lnwjud on Windows, ~/Library/Application Support/lnwjud on macOS, and $XDG_DATA_HOME/lnwjud (or ~/.local/share/lnwjud) on Linux unless LNWJUD_DATA_PATH is explicitly set. A workspace may contain .lnwjud/project-profile.json for project-scoped policy such as Ponytail or Engineering Harness mode/profile; .lnwjud/ is local metadata and is ignored by this repository.
Engineering Harness (opt-in)
Engineering Harness is Off by default. Enable it in Settings → Engineering Harness when you want lnwjud to turn a substantive coding request into a durable, risk-based workflow instead of treating every request the same way.
Presets: Standard, Senior (recommended), Strict, or Custom.
Scope: coding projects/tasks by default, with an optional per-workspace On/Off + preset override.
Policy precedence: explicit workspace/project opt-out wins; project policy can strengthen a preset but cannot silently weaken an explicit user opt-out.
Durability:engineering_start_task reuses the existing durable Goal authority. Requirement changes increment the Goal intent revision and stale only affected Engineering gates instead of resetting proven work.
Evidence: required checks remain Pending/Running/Failed/Stale until observed evidence satisfies them. Host-observed mechanical checks are not replaced by a model assertion.
Enforcement: first-party development mutations such as guarded file edits, Git, shell/process/project commands, verification, Codex/Agent Swarm and worktree operations require the current engineeringTask binding when Harness is active. Existing permission, Active Project, recovery, Full Bypass and rolling goalLease rules still apply independently.
Advisory boundary: external mcp_call, general Office/web/native-UI routes are not mechanically Engineering-guarded in v5.7.0; their existing security/permission controls remain authoritative.
No hidden writes: project assessment and the Settings project-profile preview do not write repository files, commit, push, deploy, or create a Scheduled Task by themselves.
lnwjud exposes 294 tool definitions through one local runtime and MCP gateway. The default advertised set is 282; all 294 are available when Codex delegation plus Agent Swarm is enabled.
Area
Examples
Workspace & files
read/search/edit files, paging, full scans, project indexing, recovery trash
Local MCP: use the packaged stdio launcher or Desktop loopback MCP endpoint directly. The Local transport starts the local listener without ngrok, a public URL, or Remote MCP OAuth.
Remote MCP via ngrok + OAuth: remains the default/backward-compatible Remote MCP transport. lnwjud owns the protected loopback OAuth gateway and ngrok child runtime.
Remote MCP via Cloudflare: opt-in for a Cloudflare Tunnel/reverse proxy that you manage. lnwjud provides a stable local OAuth-gateway target and verifies the configured public HTTPS origin; it does not claim to create or manage the Cloudflare tunnel.
Remote MCP via Custom URL: opt-in for another externally managed HTTPS reverse proxy. No ngrok token/install is required; the public origin must route to the stable local protected gateway.
OpenAI Secure MCP Tunnel: separate outbound-only OpenAI tunnel path using the verified target-native bundled tunnel-client. One lnwjud tunnel endpoint is intended to serve multiple simultaneous ChatGPT chats/workspaces; do not create one tunnel/profile per chat. v4.62.0 explicitly gives the tunnel transport 32 active MCP-request slots so normal multi-chat tool fan-out does not hit tunnel-client's lower default ceiling. If several physical lnwjud hosts (for example Mac + Windows) must be independently selectable, give each host a distinct Tunnel ID/ChatGPT connection: HTTP replicas sharing one Tunnel ID are work-sharing replicas, so a request goes to whichever replica polls it first rather than to a chat-selected host.
External MCP servers: lnwjud can discover supported Cursor/Claude Desktop/custom MCP definitions and keep child MCP servers separate from the first-party catalog.
lnwjud can keep long-running work alive across chat turns with a durable goal plus the bundled lnwjud-scheduled-continuation skill. The intended model is simple: one stable goal, one recurring Native ChatGPT Scheduled Task, periodic checkpoints, and an explicit finish only when the work is genuinely complete.
The scheduled continuation should be a Native ChatGPT Scheduled Task running in cloud mode every 1 hour. It is a watchdog that wakes the workflow back up; the durable goal remains the source of truth for plan state, progress, blockers, evidence, and tracked background tasks. Do not create a new goal or timer on every wake.
@lnwjud
Workspace:
C:\path\to\my-project
Continue this task until it is genuinely complete. Do not stop only because the current chat turn ends.
- Use the lnwjud-scheduled-continuation skill.
- Create or resume the same durable goal using one stable goalKey. Treat that Durable Goal as the authoritative task state and never create duplicate goals for the same job. When plan or step progress changes, use update_goal_plan so the Watcher-visible plan stays current.
- Create exactly one recurring Native ChatGPT Scheduled Task, run it in cloud mode every 1 hour, and reuse that same scheduled continuation for this goal.
- Checkpoint immediately after every meaningful milestone, and refresh active progress at least every 10 minutes when sustained work has no natural milestone. Persist current phase, step status, next action, blockers, evidence, tracked background tasks, and reconstruction-grade resumeContext: changed files, exact commands/results, decisions, failed attempts, pending validation, resume prerequisites, state facts, and artifacts. A checkpoint records progress; it is not a reason to stop useful work.
- When the scheduled task wakes the workflow, claim the continuation for the same goal and resume from the latest checkpoint instead of starting over.
- If CI, build, test, deployment, or another process is still running, keep tracking the same task until its terminal result is known. If a failure is fixable, inspect the exact failure, diagnose it, repair it, and run the relevant validation in the same turn instead of stopping at a status report.
- Do not finish the goal while any planned step, acceptance criterion, required Engineering gate/review finding, blocker, or blocking task remains unresolved. Reconcile stale duplicate/superseded/abandoned goals instead of leaving finished work Active in Watcher.
- When the work is truly complete, cancel the scheduled continuation, make the exact Native ChatGPT Scheduled Task non-runnable, then call finish_goal with final evidence and read get_goal to confirm a terminal state before reporting completion.
English: Save meaningful progress: current phase, completed/pending steps, next action, blockers, evidence, and tracked background tasks. A checkpoint records progress; it is not an instruction to stop working.
English: Finish only after all planned steps are complete, blockers are cleared, blocking tasks are terminal, acceptance criteria are satisfied, and scheduled continuation cleanup is complete. Record final evidence instead of merely declaring success in chat.
Plan + Watcher projection:get_goal_plan projects the authoritative plan; update_goal_plan changes that same durable plan and keeps the plan/progress shown by Watcher aligned with real step state instead of creating a second planner.
Freshness + reconstruction: call checkpoint_goal immediately after meaningful step/task/blocker/commit/push/CI/package milestones, and at least every 10 minutes during sustained work without a natural milestone. Meaningful checkpoints carry reconstruction-grade resumeContext; a short summary alone is not enough for reliable resume.
Acceptance:update_goal_acceptance records explicit completion evidence. finish_goal(status=completed) is rejected while any criterion is still pending or blocked; Engineering Harness goals must also satisfy their required gates and blocking review findings.
Newest intent wins:revise_goal_intent increments userIntentRevision; stale checkpoints or delivery receipts from older accepted user intent are fenced/retired rather than replayed.
Context Capsule:create_context_capsule stores a bounded immutable objective/decision/result summary for compact/resume or handoff. It stores task state, not private chain-of-thought, and does not drive ChatGPT through browser/DOM automation.
Pressure + iteration:context_pressure reports a local estimate when exact provider usage is unavailable, while advance_goal_iteration is explicitly bounded by maxIterations and can stop when no new evidence appears.
Completion hygiene: after watchdog cleanup and finish_goal, read get_goal and confirm a terminal status before reporting completion. Reconcile stale duplicate, superseded, or abandoned Goal cards when authoritative evidence shows they no longer represent live work; completed work must not remain Active in Watcher.
ภาษาไทยแบบสั้น: v5 แยก plan / acceptance / user intent / context capsule / bounded iteration ออกจากกันชัดเจน โดย Durable Goal ยังเป็น source of truth เพียงชุดเดียว ถ้าผู้ใช้เปลี่ยนคำสั่งใหม่ งานเก่าต้องแพ้ revision ใหม่ และ Context Capsule ใช้เก็บสรุปสถานะเพื่อกลับมาทำต่อ ไม่ใช่เก็บ chain-of-thought หรือใช้ browser ไปสร้างแชทใหม่เอง
Short prompt — English
Use lnwjud-scheduled-continuation. Create or resume one durable goal for this job, keep exactly one Native ChatGPT Scheduled Task running in cloud mode every 1 hour, and resume from the latest checkpoint until the goal is genuinely complete. Checkpoint every meaningful milestone, never duplicate the goal or timer, and when all steps are complete with no blockers or blocking tasks left, cancel the scheduled continuation first and then finish_goal with final evidence.
This pattern is especially useful for long CI/release jobs, multi-stage refactors, deployments, packaging, migrations, or any task where “continue later” should be backed by durable state rather than wishful thinking.
lnwjud is intentionally powerful because it operates on the local machine. Tool availability does not bypass Active Project scope, permission policy, mutation approval, recovery, host OS permissions, or platform capability gates. Unknown/unsupported platforms and missing native providers should fail closed rather than silently substituting a different OS implementation.
Pas encore de liste d'outils. Les serveurs distants en obtiennent une après le premier contrôle d'état.
FAQ
Questions
Comment installer Lnwjud ?
Copiez sur cette page la configuration stdio pour Cursor, Claude Desktop, Claude Code ou VS Code. Elle utilise l'installation de l'éditeur, pas un proxy PlugMCP.
Lnwjud nécessite-t-il une clé API ?
Aucune variable d'environnement requise n'est indiquée sur cette fiche.
Quel transport Lnwjud utilise-t-il ?
Lnwjud s'exécute en local via stdio.
Quand Lnwjud a-t-il été vérifié pour la dernière fois ?
Dernière vérification : 2026-11-15. lnwjud — local AI-agent runtime & MCP gateway