Newmark Agent

Newmark Agent is a local-first desktop workspace for AI-assisted software development, research, automation, and controlled computer operation. It combines a graphical desktop application, terminal interface, and CLI around the same local configuration, workspaces, conversations, tools, workflows, and archives.

The application connects to user-selected model providers through OpenAI-compatible, Anthropic-compatible, GitHub Models, and custom endpoints. Provider credentials and mutable workspace state remain under the user's Newmark data directory rather than inside the installation or repository.

Capabilities

• Multi-model work: fixed deployments and audited Auto routing with model capability, quality, cost, speed, privacy, and reliability signals.
• Persistent workspaces: conversations, drafts, Build history, plans, goals, queues, archives, media, skills, and workflows are scoped to the correct workspace and conversation.
• Three interfaces: Electron GUI, Newmark --TUI, and Newmark --cli share the same backend contracts and local state.
• Build, Plan, Goal, and Flow modes: direct execution, structured planning, persistent objectives, and reusable workflows with resumable execution.
• Controlled tools: terminal, files, browser use, GitHub, SSH, MCP, automation, image display, OCR, and Windows computer use are exposed through policy and approval boundaries.
• Conversation-bound BrowserUse: the built-in browser guest and right browser sidebar resolve to the same workspace/conversation target as BrowserUse calls; switching conversations never reuses another conversation's page.
• Conversation-bound ComputerUse: ComputerUse is an explicit per-conversation toggle independent of Build lifetime. Build blocks only switch it and request screenshots/actions; a process-wide lease permits one conversation at a time and reports computerUse occupied to competing conversations.
• Long-running work: context compression, durable work runs, public progress events, queue control, graceful interruption, and force-stop without replacing the runtime.
• Local memory and archives: Memory Lab, conversation archives, and workspace history provide inspectable continuity without requiring a hosted application backend. Conversation archive clicks remove the row optimistically, can interrupt any target runtime, and persist archive payloads without blocking the interface.
• Agent-driven context management: The Agent can actively compress its LLM context, inspect current entries, or search/read/restore folded history through context_compress and context_history_manage. The active Build block triggers at 70% of its model window, long-term history triggers at 20%, and the active Build keeps its own bounded working window so repeated compaction does not collapse every turn to the old 20% target. A bounded hot cache keeps normal turns lean while evicted folds remain in a conversation-isolated append-only local archive. These operations affect only model context; displayed conversation history is never altered.
• Cross-platform packaging: Windows MSI and portable packages, plus Linux AppImage, Debian, and unpacked distributions.

Interfaces

GUI

The Electron application provides the primary workspace surface: conversation navigation, model and reasoning controls, Build timeline, plan and goal panels, Flow execution, tool activity, file tree, Memory Lab, archives, and settings.

TUI

Launch the terminal interface from an installed package or source checkout:

Newmark --TUI

The TUI uses the same conversation and workspace model as the GUI, including drafts, Flow state, model selection, work runs, and theme preferences.

Its light theme hydrates readable foreground/background defaults from the shared user configuration and keeps a full ANSI canvas painted when legacy or low-contrast color pairs are present.

CLI

The CLI exposes non-interactive commands for provider configuration, model validation, skills, Memory Lab, archives, automation, and scripted agent use:

Newmark --cli --help

Installation

Download published packages from the GitHub releases page. The source now targets the dev-0.4.0 candidate; Windows MSI and win-unpacked ZIP artifacts will be rebuilt from the release gate and are not yet published to remote releases.

For first-time discovery, run Newmark Agent.exe --help for the desktop package or Newmark.exe --help for the console launcher. These commands terminate after printing the GUI, TUI, CLI, Flow, edit, and non-interactive command surface; --TUI --help and --cli --help use the same terminating help contract.

Windows

• Newmark-Agent-0.4.0-x64.msi: per-machine Windows installer for the next candidate build.
• Newmark-Agent-0.4.0-win-unpacked-x64.zip: portable Windows directory with GUI, TUI, CLI, and launcher files for the next candidate build.

Linux

• Newmark-Agent-0.4.0-x86_64.AppImage: portable Linux desktop package for the next candidate build (not yet rebuilt).
• Newmark-Agent-0.4.0-amd64.deb: Debian/Ubuntu package for the next candidate build (not yet rebuilt).
• Newmark-Agent-0.4.0-linux-unpacked-x64.zip: unpacked Linux distribution for the next candidate build (not yet rebuilt).

Availability depends on release publication. Packages are unsigned prerelease builds; verify downloaded artifacts with the SHA-256 values published in their release notes.

Configuration

Newmark stores user configuration, credentials, conversations, runtime state, caches, and archives under:

~/.Newmark/

The repository contains DESKTOP/config.example.json as a starting point. Providers can be configured from the GUI, TUI, or CLI. Never commit a populated config.json, API keys, tokens, or generated user data.

Provider selection is deployment-aware: two providers may expose the same model name without sharing credentials, validation evidence, routing state, or conversation selection.

Architecture

GUI / TUI / CLI
        |
        v
Target-bound desktop and terminal adapters
        |
        v
Conversation kernel and Agent runtime
        |
        +--> Electron utility runtime pool
        +--> WSL runtime pool
        +--> Context orchestrator and provider adapters
        +--> Native tools, MCP, browser, files, and automation
        |
        v
Local workspace, conversation, archive, and runtime stores

Conversation and workspace identity are composite target identities. Runtime events, queues, drafts, work runs, Flow suspensions, and snapshots are routed by target rather than by a mutable foreground selection.

Conversation archiving is a destructive target operation: the GUI removes the row immediately, the desktop backend force-stops a running or stopping target, and archive Markdown/manifest writes use unique names and asynchronous atomic files. Concurrent targets may write archive payloads in parallel; each final state deletion merges against the latest lock-protected workspace snapshot so rapid clicks cannot resurrect or overwrite sibling conversations. A failed backend receipt is reported without renderer-side rollback.

Flow takeover is strictly conversation-local on both the backend and the GUI. A running or paused Flow belongs to the conversation that started it: flow:run, flow:resume, flow:guide, and flow:stop all resolve the owning conversation target first, the persisted suspension is stored per conversation, and the takeover bubble only reflects the active conversation's own Flow state. Switching conversations never surfaces, mutates, or clears another conversation's Flow, and the whole takeover bubble is the single interactive affordance (click to pause while running, click to resume while paused).

Subagent identity is decoupled: the caller-supplied name is a stable human-readable label that never contains the id, while id/shortId/displayName/qualifiedName carry identity. The frontend renders only the display name, and Agent tool interactions accept both the name and the exact id (subagent_list returns both for precise targeting).

Development

The desktop source lives in DESKTOP/ and requires Node.js with npm. From the repository root:

cd DESKTOP
npm ci
npm run build

Run the development GUI or interfaces:

npm start
npm run start:cli
node dist/launcher.js --TUI

Verification

The dev-0.3.12 safe product gates remain the last verified package baseline for the temporary-root, rebuilt Windows-package, protected-install-shape, expanded GUI/TUI/CLI, and current Flow lifecycle scopes. The source version is now 0.4.0; the 0.3.14 Windows MSI and win-unpacked ZIP artifacts were rebuilt locally on 2026-08-14 (latest: MSI 226,040,216 bytes / SHA-256 1CAB2A1D558AC0C8DBF85D7DE60E539401EB828E361005F2F9320AEBAF4ABD7A, ZIP 291,866,935 bytes / 6E7298D42641B2393CC71665EB15E1E743A034D9D2D2589CDCC2653452CEBC3C) and passed packaging gates plus a real OpenRouter provider stress run with deepseek/deepseek-v4-flash that passed ALL 11 scenarios (Final Verdict release-usable). The earlier goal-continuation failure on meta-llama/llama-3.3-70b was re-adjudicated via deepseek-v4-flash as model exact-marker noncompliance, not an app regression — and goal_manage/conversation_rename tools plus cordis-core toolchain migration landed in this build. Remote publication, Linux packages, and UAC/MSI installation remain pending. The earlier independent no-context release-flow FAIL/HOLD remains historical evidence about a prior package boundary and was not silently reclassified. The last verified safe package boundary is app.asar 157,099,444 bytes / SHA-256 5A4EC8CF1E4BC3F1E16749A822E62A4BFEC43B2588A8DA26405BDD064078C8D4, MSI 225,884,568 bytes / 27CD340EBAA71CF088905007A171D4AE24A474F0617E1FF180F1A52FE0F02C45, and portable ZIP 291,752,665 bytes / D13376B03D541423C912049E82AD9FCF1FD9ADB3A4ED5641D0CA2928F0F9A688. No-provider and unavailable explicit-model requests still fail closed in the rebuilt source/package gate, while default validate-models loads fresh durable evidence read-only and never writes config unless --persist is explicit. Detailed evidence is archive/20260813-dev-0.3.12-flow-lifecycle-repair-and-release-gates.md; the new candidate boundary is DESKTOP/scripts/release-notes-dev-0.3.13.md.

The project includes source-contract, runtime-isolation, context, provider, tool exposure, GUI, TUI, CLI, SSH, WSL, and shared-backend stress suites.

2026-08-13 — dev-0.3.13 context compression, cache, and prompt policy

Source dev-0.3.13 now separates context pressure by lifetime: the active Build block uses a 70% trigger/working-window policy, while long-term history uses a 20% trigger/summary budget. Context status exposes both token counts, thresholds, retention budgets, and usage percentages. Repeated compression recognizes a prior summary anywhere in the retained history, and the Build tail is retained independently so an active block is not repeatedly collapsed toward the long-history budget.

The stable system prompt no longer re-ranks skills from the latest history message or injects the full linked-plan Markdown/revision on every turn. linked_plan remains available as an explicit tool and is disclosed in the base prompt. Multi-step work now has a mandatory bounded inline checklist contract with pending, in_progress, completed, and blocked states. Request-scoped task/bootstrap metadata is injected only for the first provider request of a Build or immediately after compression; tool sub-turns reuse the stable prompt prefix for better provider cache reuse.

Source verification passed node dist/tests/verify.js (1501/1501), performance/cache (17/17), normal-chat regression, Context System V2, Context V2 stress (1461/1461), and compression pressure (34/34). npm run build passed and npm run lint reported 0 errors with existing warnings. No dev-0.3.13 package rebuild, UAC installation, Git tag, or remote publication was performed; the package gate remains pending.

2026-08-13 — built-in editor Copilot prediction optimization

The native editor prediction path now uses a dedicated low-latency budget: 3,200 characters before the cursor, 800 after it, 96 output tokens, a 6.5-second deadline, and streaming deltas when the provider supports them. It falls back to bounded non-streaming chat for older gateways without changing the user's normal conversation API mode. Valid Copilot deployments are preferred only after credential/status checks; the active validated model remains a fallback.

Rapid typing and caret movement now cancel the owner-scoped request immediately, avoid full-file syntax re-rendering for caret-only changes, and reuse identical cursor anchors briefly. Empty, whitespace-only, aborted, superseded, and timed-out predictions are quiet instead of flashing No completion. The deterministic editor regression is included in node dist/tests/verify.js (1495/1495). The real packaged APInebula gpt-5.4-mini editor route returned an 11-character completion through the delta channel in 4,287 ms (firstDeltaMs=4,285); the provider/network latency is reported rather than hidden.

The last verified local Windows candidate (dev-0.3.12) is app.asar 157,131,813 bytes / C197E6F992729C468B4CB4F6440CC0DBD1E1331F18F83EC067A49FA9A2F4B83F, MSI 225,896,856 bytes / 27447481E6FFF8D20CC4CC629E55269FF1DBD1C7102474DD14C1B0535CD4AE38, and portable ZIP 299,390,078 bytes / 30079E9CECB1B77AAF70486D6C7FDAC042894669CB2F0DAD08371A429DF3AB74. Packaged native editor, CLI, context-compression, console-wrapper, deterministic SSE, and real-model editor gates passed for that baseline. The dev-0.3.13 package must be rebuilt and re-gated; no additional global UAC install was performed in this version-bump slice. Detailed baseline evidence: archive/20260813-dev-0.3.12-editor-copilot-prediction-optimization.md.

cd DESKTOP
npm run test:desktop
npm run test:full-release

Package validation scripts cover the Windows MSI administrative image, Windows portable output, protected-install root redirection, Linux AppImage, Debian, Linux unpacked output, startup behavior, process cleanup, release CLI behavior, and the installed-local surface gate. The current rebuilt package boundary is release/win-unpacked/resources/app.asar (157,068,983 bytes, SHA-256 1A7C444A4DF068E2BBE62D72B6CE17EED7ADBAB40F51697A3619CC303E32A324), MSI (225,888,664 bytes, SHA-256 426F7A010C76AAA04C043C8E03644010C5A1317F5B254FD6F7B3C53E7386E774), and portable ZIP (291,743,396 bytes, SHA-256 6AAD7C293761939CF0C739F82EBB7D651B438994124A413322BD1437C3B118BA). The Browser guest keeps the five-second creation floor for background demand but bypasses it for an explicit visible-tab activation; startup recovery measured startupMs=1046 and browserOpenMs=45. The rebuilt package was not reinstalled machine-wide; the actual Program Files package remains stale. The _ref APInebula key was used only through an environment variable for one current real-model CLI probe; the provider path timed out in-app and was separately observed rate-limited, so real-model acceptance is not claimed. WSL TUI remains an environment skip because Ubuntu-24.04 is unavailable. The latest no-model GUI evidence is archive/2026-08-12-blackbox-gui-no-model-repair.md; broader evidence remains in archive/2026-08-12-dev-0.3.12-subagent-findings-repair-and-expanded-release-gates.md, archive/2026-08-12-dev-0.3.12-user-pressure-browser-interaction-and-release-gates.md, and archive/2026-08-12-real-provider-stress-debug.md.

Building Releases

Windows packages:

cd DESKTOP
npm run dist:windows-release

Linux packages are built natively or through WSL:

cd DESKTOP
npm run dist:linux

The resulting artifacts are written to the repository-level release/ directory. Release upload is intentionally separate from local packaging.

Security and Privacy

• Credentials are loaded from local configuration and are redacted from diagnostics and stress reports.
• Hidden provider chain-of-thought is not rendered as public work text or persisted as ordinary conversation prose.
• Tool access is schema-validated and policy-bound; high-risk capabilities require explicit authorization or approval.
• Browser actions use scoped capabilities, rendered-content boundaries, and stale-page guards.
• Workspace and conversation state is target-bound to prevent cross-conversation writes and UI leakage.

License

Newmark Agent is distributed under the license in LICENSE. Third-party notices are included in THIRD_PARTY_NOTICES.md.

Maintenance Log

• 2026-08-14 — dev-0.4.0 分支交流 + 思考活动 + 历史卸载延迟： 新增 Build 块「思考中/进行了思考」活动（对支持 reasoning 的模型展开可见，绝不进入聊天正文与最终回复）；context_history_manage 的 remove 改为「只卸载长期 log + 声明后当前 Build Block 结束后才对后续 Block 生效」并暴露 pendingRemovals；新增「允许分支交流」对话类型（GUI 新建对话框 / CLI --branch-communication / 对话列表特殊标记），移除运行分支唯一性（runningNodeIds），每活跃分支独立维护完整缓存，并新增 branch_list/branch_send/branch_read/branch_create 工具；SubAgent 缓存命中优化（delegatedPrompt 移除易变 flowPc、mailbox body 截断）。版本号升至 0.4.0，交叉压力测试编排器更名为 dev-0.4.0-cross-stress.cjs 并新增 branch-communication 维度。verify 1530/1530、branch/SubAgent/gui-tui-cli 相关测试全绿。证据：archive/20260814-dev-0.4.0-branch-communication-and-reasoning.md。

• 2026-08-14 — DeepSeek Harness GUI installed locally: Installed the official @deepseek-ai/dsh@0.1.0-rc.6 package globally, started dsh --profile web, verified http://127.0.0.1:3080/ with HTTP 200, and opened the GUI in the default browser. Evidence: archive/20260814-deepseek-harness-gui-install.md.

• 2026-08-13 — GUI 纯键盘覆盖与 DSH 启发的状态视觉： GUI 现由一份含 65 个命令的统一注册表驱动菜单、快捷键和稳定测试清单；Ctrl/Cmd+Shift+P 打开可搜索命令面板，F1 打开快捷键帮助，F6/Shift+F6 循环主区域，Ctrl/Cmd+K chord 覆盖新会话、工作区、Plugins、DSH、Memory Lab、Automation、Flow、设置与主输入，并保留编辑器、终端、浏览器和输入法的原生按键边界。对话框焦点圈定/返回、ARIA tabs/list、方向键、可键盘调整分隔条、终端 Delete 关闭、prefers-reduced-motion 和 Browser WebView 的受限命令转发均已接入。视觉层借鉴 DSH 的状态/所有权分层、语义 token、静态配置与运行状态分离和不换行终端信息密度，但不复制 Cordis 运行时、DSH 品牌皮肤或 preview phase。源码全套与真实 win-unpacked CDP 烟测通过；证据见 archive/20260813-gui-keyboard-dsh-inspired-optimization.md。

• 2026-08-13 — 官方 DSH Plugin 兼容与 MCP 管理/发现交互：source + packaged UI verified： 插件菜单新增独立 DSH Plugin 入口，按官方 developer preview 契约展示本地 @deepseek-ai/dsh CLI/package、profile、bundle、未知配置键与可审核的 MCP 候选；更新通道明确为 latest（非锁定），兼容发现采用只读、宽容解析，不执行 DSH 或插件代码、不安装/更新软件包、不改写 DSH 配置，并保留未知键供用户判断。MCP 页面统一候选审核、搜索/筛选、编辑与持久化反馈；候选只预填审核表单，保存后默认禁用，不会自动安装、启动或启用服务器。npm.cmd run build、DSH/MCP 专项、完整 test:desktop:built 和重建后的 release:ui-skills-smoke 均通过；完整 UI verify 为 1504/1504。设计、安全边界、产物哈希和截图见 archive/20260813-dsh-plugin-mcp-discovery-compatibility.md。这只验证本功能与本地 unpacked 候选，不把整体 dev-0.3.13 发布状态改判为可发布。

• 2026-08-13 — dev-0.3.13 --help/--root 黑盒边界修复与自然长压续测： 全新上下文黑盒员在真实候选包中报告命令级帮助偶发进入 GUI/runtime、以及 --root 单独时 Chromium profile 可能落到用户目录。已将 command help 前置到 Agent/Electron 初始化前，并让显式 root 同时绑定 Electron userData、sessionData、Chromium --user-data-dir，启动日志也跟随临时根。源级门禁 1501/1501、真实 Builder SSH/TUI/CLI/上下文压缩/MSI/ZIP、独立安全黑盒 28 cases、帮助顺序 8 variants、GUI/Console Wrapper 根隔离均通过。规则已改为黑盒自然完成，不再设置人工硬截止；新的 gpt-5.6-luna/max 黑盒批次仍在执行，尚未宣称三轮 Clear、UAC 或 release。证据：archive/20260813-dev-0.3.13-help-root-boundary-fix.md。

• 2026-08-13 — 全场景交叉压力测试设计与黑盒派发：INCOMPLETE/HOLD: 先按入口传输、工作状态、上下文/模型、持久化、生命周期/资源五层重建交叉模型，定义 X01–X12，并拆为 GUI、TUI/CLI/共享后端、上下文/cache/Copilot、耐久/重放四个独立黑盒批次。四个 gpt-5.6-luna/max 无上文黑盒员均在硬截止内未返回结构化报告，故不计通过；候选包安全门禁 28/28 与压缩 CLI 9 请求本地复核通过，所有候选包测试进程已清理。未执行 UAC/MSI 安装，未修改用户配置或 Program Files。设计与审计见 archive/20260813-dev-0.3.13-full-cross-scenario-design.md 和 archive/20260813-dev-0.3.13-full-cross-scenario-blackbox-dispatch.md。

• 2026-08-13 — dev-0.3.13 cross-scenario continuation: INCOMPLETE/HOLD: Rebuilt and verified the real Windows candidate (0.3.13), repaired two release-gate assertions that still expected 0.3.12, and passed the packaged 28-case safe black-box gate, CLI/context-compression/shared-root gates, Flow pause/resume/running-archive (133 ms immediate UI removal, three duplicate receipts), startup recovery, editor, fast-switch, queue/plan, multi-window shared-backend, and 30-window TUI viewer stress. Three fresh-context gpt-5.6-luna/max GUI/TUI/context black-box batches reached their hard deadlines without structured final reports and count as incomplete; GitHub Copilot latency was environment-blocked because no selectable Copilot model was available. No UAC, MSI installation, Program Files mutation, or user-config write was performed; all Newmark/Electron processes were cleaned. Evidence: archive/20260813-dev-0.3.13-cross-scenario-continuation.md.

• 2026-08-13 — dev-0.3.13 全场景交叉黑盒 canary：INCOMPLETE/HOLD: 设计了 GUI/TUI/CLI、共享后端、Flow/Build、归档打断、上下文压缩/cache、linked-plan/inline task、Copilot、Provider 失败和生命周期的正交/三元压力矩阵，并交给两个无上文 gpt-5.6-luna/max 黑盒员。真实 Program Files 入口实际仍为 0.3.12，不是当前 dev-0.3.13 候选；安全 Temp 原始证据覆盖 help/version/state/坏模型/未知 tool、GUI 输入/新会话和 TUI Console/CLI/script PTY 子路径，但 Flow、归档、压缩/cache、三入口并发、Copilot 与长压未完成。两个代理在汇总阶段未返回，已关闭；产品进程为 0，无 UAC、安装或源码变更。完整矩阵见 tasks/plan.md，审计记录见 archive/20260813-dev-0.3.13-blackbox-canary-incomplete.md。

• 2026-08-13 — version bump to dev-0.3.13: Updated DESKTOP/package.json, DESKTOP/package-lock.json, the console-wrapper gate output, and the provider timeout gate label to 0.3.13; added DESKTOP/scripts/release-notes-dev-0.3.13.md. The prior dev-0.3.12 artifacts remain the last verified package baseline. No dev-0.3.13 package rebuild, UAC installation, Git tag, or remote publication was performed in this slice.

• 2026-08-13 — two-tier context compression/cache/prompt policy: Build-block history now triggers and retains its working window at 70%, long-term history triggers at 20%, and context_history_manage status exposes both budgets. Compression reuse finds an existing summary anywhere in retained history. Stable prompt assembly no longer injects linked-plan Markdown/revision or history-ranked skills; linked_plan is disclosed as an on-demand tool, inline task checklist management is mandatory, and request-scoped bootstrap/task metadata is omitted from repeated provider tool sub-turns. Source build, verify (1501/1501), performance/cache (17/17), Context V2 stress (1461/1461), normal-chat, and compression pressure (34/34) passed. No package rebuild or UAC install was performed. Evidence: archive/20260813-dev-0.3.13-context-compression-cache-prompt.md.

• 2026-08-13 — built-in editor Copilot prediction optimization: Fixed frequent No completion/slow prediction. Completion validates usable Copilot candidates, falls back to the active validated model, uses 3,200/800 context, 96 output tokens, 6.5s deadline, forced streaming with bounded non-stream fallback. Added owner-scoped cancellation, early delta IPC, 300ms idle debounce, anchor cache/fingerprints, quiet empty-result handling, and caret-only rendering. Source verification passes 1495/1495; packaged native editor, CLI, compression, console-wrapper, deterministic SSE, and real APInebula editor IPC gates pass. Evidence: archive/20260813-dev-0.3.12-editor-copilot-prediction-optimization.md.

• 2026-08-13 — Flow continuation/archive interruption and background lifecycle repair: A paused Flow no longer hits the stale Flow component #0 cannot start before the previous Build block has terminated guard on explicit resume: only the interrupted target's running WorkRun ledger is finalized before the resumed component starts, while genuine concurrent Build protection remains. Stop/Esc during a resumed Flow is resumable; archive synchronously marks and aborts the Flow owner, force-finalizes its target ledger, clears suspension state, starts runtime hard-stop in the background, and single-flights duplicate archive IPC. The GUI removes the target row and Flow takeover on the first click and never rolls the row back; late Flow completion cannot recreate archived state. Renderer/window lifecycle diagnostics now record unresponsive/responsive, close, hide-to-tray, and will-quit transitions. Source feature gate passed 1482/1482; Flow pressure, archive concurrency/runtime, compression, performance, TUI/SSH/WSL/CLI, GUI/TUI/CLI shared-backend, packaged Flow pause/resume/archive/switch, tray, direct-close, and startup-recovery gates passed. The rebuilt safe artifacts are app.asar 157,099,444 bytes / 5A4EC8CF1E4BC3F1E16749A822E62A4BFEC43B2588A8DA26405BDD064078C8D4, MSI 225,884,568 bytes / 27CD340EBAA71CF088905007A171D4AE24A474F0617E1FF180F1A52FE0F02C45, and portable ZIP 291,752,665 bytes / D13376B03D541423C912049E82AD9FCF1FD9ADB3A4ED5641D0CA2928F0F9A688. No global UAC installation or Program Files process termination was performed. Evidence: archive/20260813-dev-0.3.12-flow-lifecycle-repair-and-release-gates.md.

• 2026-08-12 — independent no-context release-flow audit: FAIL/HOLD: A fresh black-box tester passed the primary safe package probes but found a P1 release-boundary mismatch (release/win-unpacked has post-package conversations/state.json not present in the portable ZIP), a P2 --help side-effect risk in release-installed-readonly-validation-stress.cjs, a P2 PowerShell GUI exit-code gate hazard, and a P3 invalid-model validation result of [] with exit 0. TUI functional interaction was blocked by the host's lack of an interactive PTY. No source, release artifact, Program Files path, or MSI installation was modified. Evidence: archive/20260812-1630-dev-0.3.12-blackbox-release-flow-adjudication.md and delegated raw report archive/20260812-1610-dev-0.3.12-blackbox-release-report.md.

• 2026-08-12 — BLACKBOX_GUI_NO_MODEL GUI failure convergence repair: The renderer now latches IPC/rejected-send failures and settles a matching provisional conversation as error, preventing a no-model request from being painted or returned as a false completed run while the target-scoped error event is still in flight. Added the real packaged GUI/CDP gate DESKTOP/scripts/release-gui-no-model-smoke.cjs, wired it into release-safe-blackbox-gates, and added a source contract assertion. The rebuilt package passed the dedicated gate, safe black-box 28-case matrix, protected-install shape, and full source gate (1478/1478, TUI 55/55, launcher 25/25, SSH/PTY four restarts, CLI 46, shared GUI/TUI/CLI 21 requests). Latest fingerprints are app.asar 157,068,983 bytes / 1A7C444A4DF068E2BBE62D72B6CE17EED7ADBAB40F51697A3619CC303E32A324, MSI 225,888,664 bytes / 426F7A010C76AAA04C043C8E03644010C5A1317F5B254FD6F7B3C53E7386E774, and portable ZIP 291,743,396 bytes / 6AAD7C293761939CF0C739F82EBB7D651B438994124A413322BD1437C3B118BA. No UAC installation was performed. Evidence: archive/2026-08-12-blackbox-gui-no-model-repair.md.

• 2026-08-12 — dev-0.3.12 TUI light-theme repair: Shared-config hydration now chooses theme-specific readable defaults when light mode has missing legacy color fields, and the renderer normalizes low-contrast persisted pairs before painting the complete ANSI canvas. Added regressions for missing colors and the #B7E4FF/#F0F2F8 low-contrast case. TUI passes 55/55, built GUI/TUI launcher checks pass 25/25, SSH/PTY light-theme plus four-restart stress passes, the complete source release gate passes, and the rebuilt safe package passes protected-install-shape plus 28/28 black-box gates. Latest package fingerprints are app.asar 157,067,747 bytes / 82C2CA9A59FC5EBCE733F5FCEB985C9931EDE25539742182C1671B49422D6DB7, MSI 225,884,568 bytes / 0BCF5A7B05C778B213A433ED577C4FAD2F7A55D6C46377AA1F4CE27DAE82FEE6, and portable ZIP 291,743,065 bytes / 4C1FB4FAA470077B748AE8ADFF1961BE17038A1D49D51B160CF4EFCB4FBB4846. No UAC installation was performed; the existing Program Files copy remains outside this latest package gate. Evidence: archive/2026-08-12-tui-light-theme-repair.md.

• 2026-08-12 — dev-0.3.12 subAgent finding repair and expanded release gates: Repaired D1 protected-install shared GUI/TUI restart routing, D2 strict top-level/CLI argument validation, D3 cross-entry workspace live refresh, D4 no-model terminal error versus false completion, D5 embedded PowerShell CR/echo corruption, and D6 archive-list refresh lag. Archive performance remains parallel/atomic and target-force-stopping; the renderer removes archive rows optimistically and coalesces refreshes, while the backend single-flights duplicate target archives. Added the protected-install shared-root/restart stress gate and repaired the single-instance shared-backend release contract to verify forwarded launches preserve the original renderer and conversation. Final source test:full-release passed (1477/1477, Context System V2 1461/1461, TUI 53/53, SSH TUI with four restarts, CLI 46, shared GUI/TUI/CLI 21 requests); expanded packaged Flow/Subagent, Memory Lab, model settings, option feedback, Skills/MCP, ZIP/MSI, startup, compression, archive, terminal, queue, conversation-isolation, GUI↔CLI, and shared-backend gates passed. The current _ref APInebula real-model probe was not a pass: the app reported a provider timeout and the endpoint returned a rate-limit response; no secret was recorded, no user config was changed, no UAC installation was performed, and the stale Program Files package remains a separate HOLD/NO-GO. Evidence: archive/2026-08-12-dev-0.3.12-subagent-findings-repair-and-expanded-release-gates.md.

• 2026-08-12 — dev-0.3.12 Browser interaction repair and expanded release gates: A user-visible Browser tab activation no longer waits behind the five-second background guest-creation floor; background/prewarm requests retain the floor, while the target-bound Browser partition remains isolated per workspace/conversation. The stale startup-recovery assertion was updated to accept the intended persist:newmark-browser-<target> partition shape. Typecheck, build, focused startup/performance/compression regressions, the full source release gate (1477/1477, Context System V2 1461/1461, Browser-Use 81, SSH TUI with four restarts, shared GUI/TUI/CLI backend 21 requests), current-package safe black-box 28/28, protected-install-shape, startup recovery (1046 ms startup / 45 ms Browser open), ZIP/MSI/tray/GUI↔CLI sync, console boundary, context compression, and two-round real-model read-only validation passed. ~/.Newmark/config.json remained hash/size/mtime identical; no UAC installation was performed and the stale Program Files package remains outside the latest package gate. Evidence: archive/2026-08-12-dev-0.3.12-user-pressure-browser-interaction-and-release-gates.md.

• 2026-08-12 — independent no-context installed-directory gate: NO-GO: Without reading project documentation or source, the tester exercised the actual C:\Program Files\Newmark Agent GUI/TUI/CLI and shared-state restart paths. It reproduced a P1 shared GUI/TUI restart failure that attempted C:\Program Files\Newmark Agent\conversations, found non-strict unknown/missing CLI arguments, GUI/TUI workspace refresh lag, a no-model GUI completion/error mismatch, malformed PowerShell terminal echo, and a manual-refresh archive-list gap. Temporary roots and installed files were cleaned; no UAC or real credentials were used. Real-model pressure was not passable because the compliant temporary root had no configured model. Evidence: archive/2026-08-12-no-context-installed-directory-blackbox-release-gate.md.

• 2026-08-11 — dev-0.3.12 user-pressure cache/root repair and final safe gates: Read-only validate-models now loads fresh durable evidence without writing config, explicit CLI model selections fail before provider transport, and unambiguous provider/model selections are normalized. The new protected-install-shape gate passed CLI/TUI/GUI root redirection with Program Files mutable paths unchanged. The final source gate, rebuilt Windows package, safe black-box 28/28, ZIP/MSI administrative-image smoke, tray lifecycle, context compression, archive pressure, and final real-model validation/cache-reuse gate passed. Real-model TUI/CLI was blocked only by provider HTTP 403 bad response 808; no UAC install was performed. Evidence: archive/2026-08-11-dev-0.3.12-user-pressure-cache-root-repair-and-final-gates.md.

• 2026-08-11 — fresh no-context real-install audit: NO-GO: A long-running tester used only the actual C:\Program Files\Newmark Agent installation, executable probes, and observed GUI/TUI/CLI behavior. It found a P0 TUI EACCES crash while creating Program Files\Newmark Agent\conversations, a P1 validate-models process-leak/no-output case, a P1 GUI close/child-cleanup delay, and a P2 invalid-model request reaching a real Provider. GUI chat/Plan/Goal/Flow/archive, CLI discovery/exit contracts, and TUI startup passed; the rest of the three-entry pressure matrix was blocked by the P0 and real-provider quota. No UAC operation was performed. Evidence: archive/2026-08-11-no-context-installed-directory-full-pressure-audit.md.

• 2026-08-11 — dev-0.3.12 user-pressure fail-closed repair and post-build gates: Empty-provider and unavailable fixed-model requests now reject as real terminal errors, persist an error WorkRun, omit synthetic done/final_response, and return CLI exit 1; archive remains immediate, parallel, atomic, and force-interrupting. The final source gate, safe packaged black-box 28/28, ZIP/MSI administrative-image smoke, active compression, queue/Plan/Flow, rapid conversation switching, real-model GUI/TUI/CLI, and two read-only validation rounds passed. WSL is an environment skip; the optional multi-window probe was not counted because the product's single-instance focus behavior differs from that stale test contract. This slice did not re-run global UAC. Evidence: archive/2026-08-11-dev-0.3.12-user-pressure-fail-closed-and-release-gates.md.

• 2026-08-11 — dev-0.3.12 user-level pressure repair, package gate, and authorized global install: Fixed GUI Electron --root profile isolation, terminating -version/unknown-command discovery, invalid-model fail-closed semantics, CLI error exits, empty-Flow feedback, Stop/Esc and large-stream behavior, and the Windows console wrapper's Electron -- argument boundary plus ConPTY handoff. The source full-release gate, packaged GUI/TUI/CLI/shared-backend stress, ZIP smoke, MSI administrative-image smoke, active model-dispatched context compression, and console colon/spaced-root probe passed. After explicit authorization, the MSI was elevated-installed and repaired with REINSTALL=ALL so the native wrapper component was present; the installed-directory GUI/TUI/CLI/shared-backend/compression gate, PATH registration, artifact hashes, config immutability, and process cleanup passed. Real-provider validation was not claimed without supplied credentials. Evidence: archive/2026-08-11-dev-0.3.12-safe-temp-gate-wrapper-and-optimization.md.

• 2026-08-11 — latest no-context installed-directory audit blocks release: The independent tester used only the real installation and executable/probe feedback, without README or source context. It confirmed installed binary integrity, terminating version/unknown-command behavior, and real endpoint validation for APInebula/gpt-5.4-mini (44.724 s), but Newmark.exe help hung for 8.2 s, real GUI/TUI/CLI conversation pressure and shared-backend/compression scenarios were not completed, and C:\Users\12252\.Newmark\config.json changed from SHA-256 428F... to 9A0B... during validation. Verdict: P1 release blocker; no automatic config restoration was performed. Evidence: archive/2026-08-11-installed-no-context-blackbox-pressure-audit.md.

• 2026-08-11 — latest independent no-context release audit: A long-running first-time-user tester used README/--help discovery and real installed/portable GUI-TUI-CLI entrypoints. It passed 72/72 public help/version probes, 26/26 safe black-box cases, 20 real CLI rounds, six GUI rounds, installed TUI/CLI recovery, compression 34/34, SSH PTY, and shared-backend 21 fixtures. It blocked MSI/UAC because GUI --root wrote Chromium user data under the real %APPDATA%, the installed and portable app.asar hashes differed, literal -version and an unknown command hung, GUI Esc was not proven to stop, and invalid-model CLI returned exit 0. Real config hash stayed unchanged and all processes were cleaned. Evidence: archive/2026-08-11-no-context-subagent-blackbox-final.md.

• 2026-08-11 — dev-0.3.12 safe portable release gate: Repaired CLI error exit semantics, empty Flow feedback, the GUI provisional-Esc stop race, and large-stream rendering cost. Safe temporary-root black-box testing passed 26/26 help probes and empty-provider failure behavior; the deterministic full-release gate passed, compression passed 34/34, and final portable GUI/TUI/CLI real-model lanes passed with the real user config hash unchanged and no leftover Newmark processes. The MSI was generated and verified but deliberately not installed machine-wide; global UAC installation is waiting for explicit authorization. Evidence: archive/2026-08-11-dev-0.3.12-safe-portable-release-gates.md.

• 2026-08-10 — independent no-context black-box gate: A long-running tester using only README/help completed 20/20 real-model CLI Build rounds and basic GUI/help/security checks, but blocked release on incomplete TUI and three-entry shared-backend coverage, Plan/Goal provider/runtime failures, CLI empty-config exit-code behavior, GUI Esc/Flow risks, and a hard-link-based isolation test that changed the real config inode. The hard-link finding is classified as invalid test isolation plus an unresolved configuration-boundary risk, not silently accepted. Evidence: archive/2026-08-10-no-context-subagent-blackbox-final.md.

• 2026-08-10 — dev-0.3.12 full-feature black-box retest: A first-time-user pass used only README/help discovery, then exercised the real installed GUI/TUI/CLI with APInebula gpt-5.4-mini. Help and local CLI management passed; GUI queue/isolation/long-context/Goal, TUI Stop/recovery/restart, and CLI Plan/Goal/Flow passed. Sustained GUI round 5, CLI 20-round, and model-validation timeout risks were reproduced and retained, so an unqualified release claim is not accepted. Evidence: archive/2026-08-10-blackbox-dev-0.3.12-full-feature-pressure.md.

• 2026-08-10 — dev-0.3.12 source repair and package gate: Repaired help termination, provider timeout cascades, stalled stream cancellation, TUI recovery serialization, context-preparation repeated serialization, and compatibility metadata redaction. The full deterministic source gate exits 0, performance optimization passes 12/12, compression passes 34/34, and the direct Windows MSI/portable packaging body exits 0 with the package boundary recorded in OVERVIEW.md and the repair archive. WSL TUI is recorded as an environment skip; publication/upload remains a separate action.

• 2026-08-10 — dev-0.3.12 final installed acceptance: The independent installed leak is closed by exact same-version product removal and elevated MSI installation: final package and C:\Program Files\Newmark Agent app.asar are byte-identical, the nested compatibility key is [REDACTED], and the four-executable help matrix passes 72/72. Final installed real-model GUI passes six UI rounds, two Goals, queue drain, conversation isolation, and long context; final installed TUI/CLI passes Stop/Esc/recovery/restart plus direct CLI; an additional installed CLI lane passes 8/8 rounds in 61.0 seconds; all exact target processes and temporary roots are cleaned. The previous 20-round provider timeout remains an explicit sustained-risk boundary. Evidence: archive/2026-08-10-dev-0.3.12-user-pressure-repair-and-gates.md.

• 2026-08-10 — dev-0.3.11 installed black-box release pressure audit: From the real C:\Program Files\Newmark Agent installation, top-level help/version, APInebula gpt-5.4-mini CLI/GUI matrix, independent Newmark.exe CLI 20-round control, TUI 10-round/restart control, package hash, and npm.cmd test all passed. The full-feature release remains not ready because TUI long-task Stop/Esc recovery stalled in a real PTY and subcommand --help was inconsistent, with validate-models --help not terminating within 30 seconds. Scenario design and evidence: archive/2026-08-10-blackbox-release-pressure-test-dev-0.3.11.md.

• 2026-08-09 — dev-0.3.10 Windows/Linux prerelease: Bumped package metadata to 0.3.10, rebuilt Windows MSI/portable ZIP and Linux AppImage/deb/unpacked ZIP, and passed the complete release gate plus all five final package smokes. GitHub release dev-0.3.10 was then re-downloaded and hash-audited successfully; artifact names and SHA-256 values are recorded in OVERVIEW.md, the release notes, and archive/2026-08-09-dev-0.3.10-win-linux-release.md.

• 2026-08-09 — dev-0.3.10 black-box real-model stress: A tester-facing run using only packaged help discovery and isolated real-model sessions passed sustained CLI, queue, conversation-isolation, and long-context lanes, but found a fresh-start Goal activation failure (Electron utility runtime stop could not confirm child exit), an intermittent 20-round UI Guide · Rejected timeout, and non-terminating/incomplete --help behavior. Verdict: not ready for real-model acceptance. Full matrix: archive/2026-08-09-blackbox-real-model-release-test.md.

• 2026-08-09 — dev-0.3.9 local Windows package/install: Ran npm.cmd run dist:windows-release from the dev-0.3.9 source; the full release gate plus final verify.js passed 1461/1461. The generated MSI and portable ZIP were installed/validated locally: installed 0.3.9.0 at C:\Program Files\Newmark Agent, installed resources/app.asar matched the packaged SHA-256, installed CLI smoke passed, and ~/.Newmark stayed byte-identical. Artifacts and MSI logs: archive/2026-08-09-dev039-local-package-install.md.

• 2026-08-09 — dev-0.3.10 Goal/Build lifecycle hardening: Every terminal Build Block now writes one idempotent work_overview entry with the startup input, final summary, event/Guide counts, and one terminal Goal audit. User Stop pauses an active Goal, ordinary input cannot resume or complete a paused Goal, and explicit resume immediately starts a Goal-driven Build. Runtime lifecycle markers distinguish a live backend from an unexpected process exit; owner-PID checks preserve backend WorkRuns after a frontend tracking timeout, while the first load after a real crash pauses Goal/Flow and records interrupted recovery. Focused lifecycle assertions, typecheck, build, and the complete 1461-assertion desktop gate passed. Evidence: archive/2026-08-09-dev-0.3.10-goal-build-lifecycle.md.

• 2026-08-08 — dev-0.3.9 installed real-model stress: Executed the machine-wide 0.3.9.0 installation against APInebula gpt-5.4-mini over the OpenAI protocol with an isolated temporary root and secret redaction. The full pass covered 3 CLI rounds, 3 UI rounds, Goal continuation, queue drain, conversation isolation, long-context compression, and exact installed-process cleanup. Two first-pass harness false negatives were corrected: Goal completion is asynchronous, and 40,600 ASCII characters estimate below the 12,800-token trigger of a 16k test window. A targeted 8k-window retest passed Goal in 22.173 s with 2 assistant calls and compressed 52,419 original characters to 463 estimated tokens in 36.441 s; no installed process remained and the real user config hash was unchanged. See archive/2026-08-08-dev039-real-model-stress.md.

• 2026-08-08 — dev-0.3.9 local package/install stress: Built and file-smoked the Windows MSI/portable ZIP, upgraded the machine-wide installation to 0.3.9.0, and proved the installed app.asar matches the package while ~/.Newmark/config.json stayed byte-identical. Stress testing fixed a real utility/WSL second-stop termination defect and refreshed two stale packaged-smoke contracts. Source, package, installed dev-0.0.8/dev-0.1.0, startup recovery, tray lifecycle, SSH/WSL, and shared-backend gates pass. Installed dev-0.0.9 remains unstable in its Browser/UI tail, and the formal performance preflight fails at 584.53 MiB startup private memory and 534.1 ms input latency under 4× CPU throttling; the 20-run acceptance set was therefore not claimed. See archive/2026-08-08-dev039-local-package-install-stress.md.

• 2026-08-08 — dev-0.3.9: Added Prime Agent-informed durable context folding: bounded hot cache, append-only per-conversation cold archive, explicit bounded read, hot/cold search, cold restore tombstones, and archive status. Typecheck and 1461 desktop assertions pass. See archive/2026-08-08-dev039-prime-agent-context-archive.md.