Newmark Agent dev-0.3.11

Newmark Agent dev-0.3.11

This prerelease closes the tester-facing blockers found in the dev-0.3.10 real-model pressure audit.

Fixes

• --help / -h and --version / -v now terminate before GUI, TUI, CLI, server, or first-run runtime initialization. The help text exposes the GUI, TUI, CLI, Flow, editor, and non-interactive command surface from one contract.
• Electron utility graceful shutdown now confirms Windows process-tree quiescence using the captured root creation identity when Electron delays or returns a false child-handle kill result. Unknown descendants still keep the runtime quarantined; a dead root handle is detached without clearing that quarantine.
• A stale renderer runtime state can no longer lose an ordinary user input as Guide · Rejected. Explicit lifecycle rejections caused by an ended run or mismatched run ID are recorded as deferred and replayed with the same payload as a fresh Build after an authoritative state refresh.
• Deferred Guide/Goal continuations are resumed after the finalization barrier settles, and a tracked Goal is audited and synchronized at every terminal Build Block.
• Provider empty responses receive bounded same-deployment retries instead of silently completing a Build. The renderer also clears a stale local send lock when an authoritative target snapshot is already idle, preventing the next prompt from disappearing during rapid user turns.
• BrowserUse is bound to the conversation's built-in browser and right sidebar. ComputerUse is an explicit conversation-bound toggle independent of Build lifetime, with one process-wide lease and computerUse occupied contention feedback. High-frequency public Build events are frame-coalesced while durable event history remains complete.
• Added source, real Electron, packaged CLI/TUI, and real-provider regression coverage for help termination, stale Guide recovery, graceful utility stop, identity-bound process-tree failure, quarantine cleanup, ComputerUse contention, BrowserUse target binding, long UI event streams, and rapid consecutive sends.

Validation

• The deterministic gate passes 1465/1465 assertions. The final packaged real-provider matrix passes 20 CLI rounds, 20 GUI rounds, three Goal continuations, queue drain, conversation isolation, and a 600-fragment long-context run; process cleanup reports remaining=0.
• Windows MSI, portable ZIP, CLI smoke, and SSH-TUI four-restart stress passed locally. Linux artifacts are built and smoke-tested by .github/workflows/release-linux.yml on Ubuntu; the remote hashes are filled after the workflow download audit.

Packages are unsigned prerelease builds. Verify the SHA-256 values below before installing:

Asset	SHA-256
Newmark-Agent-0.3.11-x64.msi	E9BFB514B583556BFB64E8B35815EE8CD916CAC4FACC43EE0AC1FFD87D305817
Newmark-Agent-0.3.11-win-unpacked-x64.zip	B8E287CDB393482EF1B3CEB9A91AB4BF86328C06AF8A13A80239DD4778ECFDFB
Newmark-Agent-0.3.11-x86_64.AppImage	CC3CB39D4F011E81330E4F097A5A1029E13CC00026E8799FCB7379AD8B008E0D
Newmark-Agent-0.3.11-amd64.deb	F4FE914FE8DA3BE65CF055C0F2E86DC44D4D078771D94009D4E6AB81496A2767
Newmark-Agent-0.3.11-linux-unpacked-x64.zip	3CDE1AAEFFBACA2E6B9F3E5AAB9FFE0C2899308716DDA83B6DD5C32868E3DAB2

The checksum manifest is attached to the GitHub release as SHA256SUMS; the tagged release URL is https://github.com/positer/Newmark-Agent/releases/tag/dev-0.3.11.

Newmark Agent dev-0.3.10

Newmark Agent dev-0.3.10

Development prerelease for the durable Goal/Flow lifecycle and Build history hardening shipped in the local-first desktop, TUI, and CLI runtime.

Highlights

• Every terminal Build Block writes one idempotent durable work_overview long-log entry at declaration end, including short replies and recovered terminal runs, and performs one Goal completion audit when a Goal is present.
• A user interruption pauses a running Goal/Flow; ordinary user input cannot accidentally resume or complete it. Explicit resume immediately claims the Goal/Flow continuation and starts a goal/flow-driven Build.
• Owner-scoped lifecycle markers distinguish a live backend from a frontend tracking timeout or a dead owner. Unexpected exits pause Goal/Flow on the first cold load and record recovery; backend timeout cold starts do not mutate live background state.
• Windows packaged Browser-Use native clicks now yield between focused native mouse events so webview guest transitions are delivered reliably.
• Context folding keeps evicted history in a conversation-isolated append-only archive while preserving the bounded hot cache and visible conversation history.

Validation

• npm.cmd run typecheck: passed.
• npm.cmd run lint: passed with 0 errors; the repository retains 56 existing warnings.
• npm.cmd run dist:windows-release: passed, including the complete release gate and packaging.
• npm.cmd run dist:linux: passed through WSL with the cached Electron 43.1.1 archive and the complete release gate.
• Final package smokes passed: Windows portable ZIP, Windows MSI, Linux AppImage, Linux deb, and Linux unpacked ZIP.
• Deterministic package smokes reported real_api_called:false.
• Linux 20-run latency benchmark passed: hot first-event p95 13.548 ms, hot first-token p95 10 ms, cold local-before-provider 52.681 ms, two tool-provider requests, and zero hot persistence writes.

Assets

Asset	Bytes	SHA-256
Newmark-Agent-0.3.10-x64.msi	225847605	472E1445F32484861D7779E4054EB7C918234B62A2EE2FA968006221AFA4D869
Newmark-Agent-0.3.10-win-unpacked-x64.zip	291678940	DBD1745216220FFE11C035B4D43CE6258332C439B467FFF24ABE56035A657ACE
Newmark-Agent-0.3.10-x86_64.AppImage	175786564	89CA21E673FEBDC494674D155E377CDF32B3BF312E42D26975867FB7554ADB08
Newmark-Agent-0.3.10-amd64.deb	135458768	BC596453869FE5A542B7328A9261A780BD1C5A6569CE0B3EDA0621FD53A5320E
Newmark-Agent-0.3.10-linux-unpacked-x64.zip	171911069	5A92E4BC8EB428CCE335FE50D81B4E51C7308842BA53BD6C63E232E73D2A6160

These are unsigned prerelease packages. Verify the downloaded files against the SHA-256 values above.

dev-0.3.7

dev-0.3.7 release.

Features

• SubAgent name/id decoupling: the caller-supplied name is a stable human-readable label that never contains the id; identity lives in id/shortId/displayName/qualifiedName. The frontend renders only the display name and Agent tool interactions accept both name and exact id.
• Agent-invoked context tools (affect only the LLM context, never the displayed conversation history):
  • \context_compress: actively compress the LLM context on demand (force bypasses the token threshold).
  • \context_history_manage: list/remove/summarize context-history entries, persisted to conversation state.
• GUI-TUI-CLI Flow state isolation verified and locked: every surface keeps Flow full state bound to its owning conversation target.

Checksums

Artifact	SHA-256
Newmark-Agent-0.3.7-x64.msi	25B3CC286015638E9C13E9E0093EE4AB94ABB97936C22DCB1C48ED12263F8DB9
Newmark-Agent-0.3.7-win-unpacked-x64.zip	159FC6856FF6752D1DFFB466FA11910E8CAB5065C668D3DA04C0C644F748049F
Newmark-Agent-0.3.7-x86_64.AppImage	4B52A9B923F357E4E229CA8396441780BC5E685B1D29CA96D515D1A774679C09
Newmark-Agent-0.3.7-amd64.deb	47FE878DF7706521020890B75C6CC39FCBCACB4F7C530F8B52D189EEB0E6A68D
Newmark-Agent-0.3.7-linux-unpacked-x64.zip	2CA53330CD583EDCECDC7FFEA54330F28B6A085F80ED187FE37BBCDBCBBCC960

Newmark Agent dev-0.3.5

Newmark Agent dev-0.3.5 — prerelease with async archive loading, cross-platform (macOS/HarmonyOS) build environment, and performance optimizations.

Highlights

• Archive load no longer blocks the window. \�gent:listArchives/
  eadArchive\ now use async IPC with first-line-only scans and event-loop yields, so a large archive directory never makes the main process unresponsive.
• Cross-platform build environment. \check-cross-platform-env.cjs\ validates platform independence and the build matrix. macOS dmg is built on a macOS host (
  pm run dist:mac); HarmonyOS integration uses the headless core (\dist/server.js, \dist/launcher.js) consumed by ArkWeb/ArkTS over HTTP — \dist:harmonyos\ packs it with pure-JS deps and HTTP-smokes it. Docs: \HARMONYOS_INTEGRATION.md, \docs/macos-build.md.
• Performance. Work-run duration updates coalesce to one animation frame and skip identical text writes; computer-use performance verify gracefully skips sampling when native screenshot capture is unavailable while keeping transport/cleanup checks.

Artifacts (unsigned prerelease)

File	SHA-256
\Newmark-Agent-0.3.5-x64.msi\	\EE9E523930AFEDAE4147423BCFB9CE286B519FDE15C9C42F68550F9ABA474708\
\Newmark-Agent-0.3.5-win-unpacked-x64.zip\	\2EB0CD180B7E41927AC177B35D2FF3EFB2C9347F812CD80D12381CDFF28F9F75\
\Newmark-Agent-0.3.5-x86_64.AppImage\	\9096D5DBD5B80988937DBDFE3283D6C7664D93C4AE2088DCB2199194A90440C3\
\Newmark-Agent-0.3.5-amd64.deb\	\D4B08EDD8F368825966E80EC9658463DA6A5832BFC13D3292F77A47D86811B6A\
\Newmark-Agent-0.3.5-linux-unpacked-x64.zip\	\6665E6673AE96D5EB65C2CE0690C5808B181112FAF40A54F18E05507DA63A16D\

HarmonyOS headless bundle is produced by
pm run dist:harmonyos\ (dist-harmonyos/).

Newmark Agent dev-0.3.4

Newmark Agent dev-0.3.4 — prerelease build fixing Flow conversation binding, Flow completion running-flag cleanup, long-conversation memory bounds, and adding a comprehensive mode-conversation stress gate.

Highlights

• Flow strictly bound to its owning conversation. \low:run\ now executes on an isolated conversation agent pinned to the conversation that started the Flow (same as \low:resume). Switching conversations mid-run can no longer redirect the next build component into another conversation or strand the running flag. The takeover bubble, build output, and completion all stay on the owning conversation.
• Flow completion clears its running flag. \stopFlowRunInternal/queue settlement are target-scoped, so a Flow that finishes in the background (while the user views another conversation) clears its own running flag.
• Long-conversation memory bounds. Work-run events are windowed to the last 250 per run and the run list to the last 120 per conversation; the live UI ledger is capped too. Combined with snapshot message windowing and load-earlier pagination from 0.3.3, very long same-conversation runs no longer grow the ledger without limit.
• Comprehensive user-level stress gate. New \mode-conversation-state-stress\ runs in \ est:full-release: Build/Plan/Goal/Flow modes, mid-Flow conversation switching, background Flow completion, per-conversation takeover isolation across six concurrent conversations, 5-minute foreground-tracking expiry, message windowing, work-run/event bounds, and click-pause→resume.

Build and verification

• \�erify\ suite: 1421 / 1421 passed
• Full release gate: green (desktop gate, mode-conversation stress, TUI, SSH/WSL TUI, CLI, GUI/TUI/CLI shared-backend)
• flow-click-pause-resume real-backend smoke (incl. mid-run conversation switch): all checks passed
• Fast conversation switch smoke + flow-pause-stop stress: all checks passed

Artifacts (unsigned prerelease)

File	SHA-256
\Newmark-Agent-0.3.4-x64.msi\	\89E69DE983AA41F7F92C4831E5F184370B3F1B803C8B3B08A412E053EC8339DB\
\Newmark-Agent-0.3.4-win-unpacked-x64.zip\	\25C7364732B0AA5E9A82981C0C5BB6425C82DA388082CCB2CACD7AE7CF41C391\
\Newmark-Agent-0.3.4-x86_64.AppImage\	\8C87FAF6AB954A77B626693B39FC9D906B52E254C4D58D8054A8D9C226A2BC1C\
\Newmark-Agent-0.3.4-amd64.deb\	\922FD5D22234CFF18314DDF7AA42983F371FADD288BBACB8B4198B9FE92E9B08\
\Newmark-Agent-0.3.4-linux-unpacked-x64.zip\	\24CF3BE27354FB95183BA374C56087853D98A1F6C51A1C2EC963E7CE97F72140\

Newmark Agent dev-0.3.3

Newmark Agent dev-0.3.3 — prerelease build fixing Flow takeover pause/resume, a startup hang on large conversations, and context compression triggering.

Highlights

• Flow takeover pause/resume fixed. Clicking a running takeover bubble pauses the Flow; clicking the paused bubble resumes it. A stale resume click on a still-running Flow returns \�lreadyRunning\ instead of an error, so the running takeover is kept (no teardown, no spurious error), and the paused bubble flips to running optimistically on resume.
• Startup hang on large conversations fixed. Token estimation no longer uses a per-char JS loop (replaced with a native regex count that stays accurate for CJK/JSON yet runs in ~5ms on 2000 messages). Snapshot IPC now windows \chatMessages\ to the most recent 200 messages, so a long-running conversation is never transferred wholesale across IPC.
• Load-earlier pagination. Conversations longer than the window show a \u201cload earlier messages\u201d button; clicking fetches the preceding window and prepends it. Full history stays persisted.
• SubAgent multi-spawn display. A Build Block that spawns several SubAgents collapses to a single \u201cOperated N SubAgents\u201d chip with names in the tooltip instead of one chip per SubAgent.
• Context compression trigger hardened. \�stimateContextTokens\ prices CJK (~1 token/char) and JSON/tool-call structure, and \ ransformContext\ holds a 0.9x-window safety net so a request can never reach the provider oversized.

Design goal (large long-running conversations)

Startup loads only the initially displayed conversation; other conversations are loaded on foreground switch; a conversation\u2019s foreground tracking expires 5 minutes after leaving it (the backend run is unaffected).

Build and verification

• \�erify\ suite: 1419 / 1419 passed
• Full desktop gate: green (context V2 stress 1461, startup prewarm 74)
• Flow pause/stop UI stress (incl. click-to-pause → resume): all checks passed
• flow-click-pause-resume real-backend smoke: all checks passed
• Fast conversation switch release UI smoke (incl. workspace-switch Flow isolation): all checks passed
• App starts cleanly with real user data

Artifacts (unsigned prerelease)

File	SHA-256
\Newmark-Agent-0.3.3-x64.msi\	\FE5F9B354E1171405C82E44B510870DFCFDD9E1BAF29AC747E27944ACEE45AA3\
\Newmark-Agent-0.3.3-win-unpacked-x64.zip\	\95407021E3B641E6673128660757D21F50DAFF5B06CD1DD0533FB488D1F0F6DC\
\Newmark-Agent-0.3.3-x86_64.AppImage\	\933E14A324EA782974700CF361168EA9C81403C6F6DE3E31E73E07ABA3753E07\
\Newmark-Agent-0.3.3-amd64.deb\	\21E62C406FEF503765FF7028DE08CEC2299E344E4F1AE39E6F3A39470ED295AC\
\Newmark-Agent-0.3.3-linux-unpacked-x64.zip\	\BE5B84486648AE7014006796B98C70EBA391E17D7CC163A3E6A6A281E563ACE3\

Newmark Agent dev-0.3.2

Newmark Agent dev-0.3.2 — prerelease build with the Flow takeover fully bound to its owning conversation.

Highlights

• Flow takeover is conversation-local, not global. A running or paused Flow belongs to the conversation that started it on both the backend and the GUI:
  • low:run / low:resume / low:guide / low:stop resolve the owning conversation target first and only act on that conversation's Flow.
  • Sending a new Build/Plan/Goal instruction or enqueuing Guide in one conversation no longer clears another conversation's paused Flow.
  • Flow suspensions persist per conversation and survive restarts.
• Workspace switches reconcile the takeover. Switching workspaces changes the active conversation; the takeover bubble now reconciles on that switch too, so a Flow paused in another workspace's conversation never leaks into (or gets stuck in) the current view.
• Takeover bubble UX: the whole bubble is the single interactive affordance — click it to pause a running Flow, click it to resume a paused Flow (the old local Resume button is gone).
• The renderer binds the takeover display to per-conversation backend state and reconciles it on every conversation/workspace switch.

Build and verification

• �erify suite: 1417 / 1417 passed
• Full desktop gate: green (context V2 stress 1461, startup prewarm 74, and all suites)
• Flow pause/stop UI stress: all checks passed
• Fast conversation switch release UI smoke: all checks passed (draft isolation, Flow target isolation, workspace-switch Flow isolation, response speed)
• dev018 mode UI smoke: all checks passed
• Full release test suite passed on Windows and inside WSL before packaging

Artifacts (unsigned prerelease)

File	SHA-256
\Newmark-Agent-0.3.2-x64.msi\	\67186F442CC738F199887C4E6E05E92914E74CE2E1F08BEB092127C625C1FD5C\
\Newmark-Agent-0.3.2-win-unpacked-x64.zip\	\

Newmark Agent dev-0.3.0

dev-0.3.0 — context-system/Provider rework (Layer A-F migration complete).

• Every model request assembles its system prompt through the Context Orchestrator (18-section order, layered hashes, byte-identical output); no inline prompt concatenation remains.
• OpenAI-protocol streaming routes exclusively through V2 provider adapters; legacy inlined serialization removed (explicit dispatch with Anthropic / GitHub Models / adapters-V2 / hard error).
• Legacy tool-surface router and buildSystemPrompt inline join removed; migration.md §2F cleanup landed.
• Full release gate green: verify 1410, contextSystemV2Stress 1461, providerBridgeV2Verify 25 (pure V2 contract), toolProvisioningVerify 64/64, guideInsertionStressVerify 40, TUI 53, SSH/WSL PTY stress, CLI contract 32+30, GUI/TUI/CLI shared-backend stress.

Newmark Agent dev-0.2.6

dev-0.2.6 — pause-aware Flow Stop/Esc and per-conversation drafts

Flow Stop/Esc is now pause-aware. While a Flow is running, the first Stop/Esc cooperatively aborts into the persistent paused takeover — the suspension is saved to the conversation store and the takeover bubble shows the failure reason with a Resume button, nothing is torn down. A second Stop/Esc force-stops the run, exits the takeover, and returns to Build. Sending a new Build/Plan/Goal/Flow instruction while paused exits the pause into that new process without restoring the previous mode.

Draft persistence. Unsent input is stored per conversation and survives conversation switches, workspace switches, and app restarts, restored when you return. Clearing a draft is durable too.

Verification

• verify.js 1410/1410 (previously 1404), including new runtime draft persistence assertions
• New stress suites: flow-pause-stop-draft-stress.cjs (backend: cooperative abort → persisted suspension, force-stop, new-instruction, 40 draft round-trips, restart survival, 8 agents × 200 drafts, churn), flow-pause-stop-ui-stress.cjs (jsdom renderer state machine)
• The first stress run caught and fixed a real draft-clear persistence bug, now regression-locked
• Windows release gate + packaged MSI/zip smokes (22 assertions each) pass
• Linux release gate + packaged AppImage/deb/linux-unpacked-zip GUI smokes pass (WSL Ubuntu-24.04)

Artifacts (SHA-256)

Package	SHA-256
Newmark-Agent-0.2.6-x64.msi	8EE2385127326AFC4663B2D255D3A78C241F770D4AE56ABEC3E7BA1095C2159B
Newmark-Agent-0.2.6-win-unpacked-x64.zip	CAEBDFDD572C49BBEF05E7F99D09A2E59A863D7EC132C8C75841E2895CF33638
Newmark-Agent-0.2.6-x86_64.AppImage	ABAC8A64D441C3E39195435B8EA9D9378A8C59A55360A81067FA2760EC66F7B9
Newmark-Agent-0.2.6-amd64.deb	8E15AF7E4EEED18741E4D0C7357F6179E1AE8CA31DCC5E01881079B5575F9CA2
Newmark-Agent-0.2.6-linux-unpacked-x64.zip	1CE717403586E2303043F2EF21AABD3E77B337868ACE171EE0AADD97EB15E529

Newmark Agent dev-0.2.5

dev-0.2.5 — SubAgent work streams live into GUI and TUI

GUI: new SubAgent work overlay (draggable panel) renders task lifecycle, streamed text, tool calls/results, and completion per peer; caches up to 500 events per actor; reopening a finished SubAgent replays its work from cache; Esc / chat-tab close.

TUI: Agents view expands each SubAgent row (Enter) into its conversation history and final result.

Core: peer work events (emitPeerWorkEvent/runSubagentJob) strip runtime/workspace keys and attach actor ids; conversationKernel broadcasts peer events without persisting them; runtime.unsubscribePeer() stops delivery.

Tools: subagent orchestration core (task + subagent_*) always preloaded by routeToolSurface/ToolProvisionSession.reconcile, appended after the 8-schema intent slice.

Verification

• verify 1402/1402 · toolProvisioningVerify 64/64 · TUI 53/53 · packaged MSI/zip smokes 22 assertions each
• New stress suites: subagent-streaming-relay-stress.cjs, subagent-gui-overlay-smoke.cjs

Artifacts (SHA-256)

Package	SHA-256
Newmark-Agent-0.2.5-x64.msi	E974BBBF463EECCC2A23D59B3269CD747A41BCE0E5C4A6990523D6EBFAA1F03F
Newmark-Agent-0.2.5-win-unpacked-x64.zip	DD56C86D3EF987071C4092733526ED4E76AA3311139956D8674D3129E0DDAC0A
Newmark-Agent-0.2.5-x86_64.AppImage	C54767C6D6D13290DA60B5BB4113901788766EC7C04411D88430ACC468318E8D
Newmark-Agent-0.2.5-amd64.deb	E35852FCDB1D57FE609CA3D783D3B4FD77AF7163437EC7213F5C8E12F8328F4D
Newmark-Agent-0.2.5-linux-unpacked-x64.zip	BA87310DD995DA8799BEEDFC207D42027FBB172CBC69707F10CE63F09F01623D