Arc 0.51.85 - Review before replying

Arc 0.51.85 — Review before replying

Published October 25, 2026 · Windows x64 development release.

This release includes changes 0.51.84 and 0.51.85 since public 0.51.83.

Take another look before answering

Choose High or Ultra and open the thinking-level menu. Review before replying is enabled by default and can be switched off there.

• High: one review of the final draft against your request and the available evidence.
• Ultra: up to two reviews, including a second look after a correction when needed.
• One bounded correction: checks look for missing requirements, contradictions, calculation mistakes and unsupported claims. A good draft can remain unchanged.
• Visible progress: brief review/check/correction counters explain the result without presenting a model review as verified correctness.
• The same selected model: reviews run serially, with no extra reviewer model and no executable tools. They cannot repeat posts, send emails, modify files or expand permissions.
• Existing checks stay in charge: file, source, screenshot and completion checks still determine whether the requested task was completed. A short outcome checklist also guides eligible work calls.

Instant, Medium and voice retain their existing response paths. Extra review calls consume time and tokens; they do not increase the context window or guarantee a correct answer. Reviews have bounded deadlines and preserve the usable draft when a review fails or cannot fit. Large files and incomplete tasks still need their own real checks.

Music has moved out of Arc

As requested, 0.51.84 removes Arc's standalone YouTube Music player, saved-playlist panel and voice/text music playback commands, along with their music-only resolver and playback services. Existing playlist metadata is left untouched but is no longer read by Arc.

YouTube search, captions, transcripts, video research and screen/audio context remain available. TwinYou Music is a separate app and is unchanged. Earlier music release notes remain in the version history as historical records.

Validation and limits

For this publication, all 23 focused answer-review and signed-update tests passed, and packaged-runtime checks confirmed version 0.51.85, matching implementation files, a two-check correction and preference persistence. The installer matches the original successful October 21 build and packaging record.

The earlier Arc Focus run passed five neutral fixtures covering arithmetic, requested formatting, unsupported test-success claims, an already-correct answer and an actual agent conversation. This is not a broad accuracy or speed benchmark.

The earlier full suite recorded 3,851 passes, nine failures and ten skips. All nine original failing cases subsequently passed focused reruns, but the complete suite was not rerun to a clean result; a larger retry also hit the test runner's heap limit. These scopes overlap and must not be added together. No fresh GPU/model benchmark or installation over the user's personal profile was performed for this publication.

Download and verification

• Windows installer — 334,602,221 bytes (about 319 MiB).
• Website download · Version history · Review setup and limits.
• SHA-256: 097fae952d87f028ee5151174cad75dd22447d0c6f52f87af62062913aab45d5.

The installer is not Windows Authenticode-signed. Signed Arc updater metadata and SHA-512 download verification are separate integrity checks. No VirusTotal result is claimed for this exact installer. Model weights remain separate downloads; normal updates preserve models, conversations and settings.

Arc 0.51.83 - Music from Home, CPU-only search and playback

Arc 0.51.83 — Music from Home, with CPU-only search and playback

Published October 19, 2026 · Windows x64 development release.

Ask for a song or playlist, and keep working. Type a music request on Home or in chat, or say it after “Hey Arc”. Arc opens Music, fills in your search and plays the first public YouTube match inside the app. No pasted link or chat-model generation is needed.

Try it

• “Play Get Lucky by Daft Punk.”
• “Search for the song Take On Me on YouTube.”
• “Find an 80s hits playlist.”
• “Search YouTube for Get Lucky and play it.”
• “Pause music”, “resume music”, “skip this song”, “previous song” or “stop music”.

Use Settings → Music · YouTube, or Settings → Plugins → YouTube → Audio player, for the player and saved playlist library. Install Agent Reach from Plugins if prompted. Enable Hey Arc and microphone access for voice commands.

What CPU-only means

Direct music requests are recognized by a small local command parser before the chat model runs. Search and stream resolution use the managed YouTube runtime; the persistent audio player handles decoding and playback. Music does not load an LLM, acquire an inference/GPU allocation, or run Ultra reasoning. Typed music confirmations also skip text-to-speech.

Your selected AI model stays available for other work. This does not mean the entire Arc process always shows zero GPU usage: the interface can use GPU rendering, and Hey Arc still uses the configured speech engine. Internet access is required for YouTube search and streaming. General video research remains on Arc’s research path.

Music features included in this release

• Search from Home or chat: play, search and find requests for songs or playlists open Music automatically. The title and Song/Playlist choice are filled in. A missing title gets a quick local prompt instead of a model call or a guessed song.
• A player inside Arc: Play/Pause, Previous/Next, seeking, volume, Stop, a collapsible floating player and the current queue. Playback continues across chats, settings navigation and minimizing Arc. Voice searches do not bring a hidden window to the foreground.
• Playlist support: save, rename and filter up to 100 playlist shortcuts. An exact saved playlist name takes priority. Public playlist searches load the first match, up to 100 tracks; song searches queue up to five matches. These are search results, not personalized recommendations.
• Voice and text transport controls: Pause and Resume are explicit actions. A user-requested pause stays paused when a voice conversation finishes. Successful playback is confirmed by the audio player, not just a resolved URL.
• More resilient streams: complete and validate audio segments in small RAM buffers, retry temporary failures at the same byte position, and discard incomplete data before decoding. Stop and track changes cancel outstanding work. Clearer messages distinguish network, timeout, format and HTTP failures.
• Existing choices remain: pasted links, saved playlists and the official browser fallback are still available. Arc does not save music files or import browser cookies.

Also included since public 0.51.75

• 0.51.76 — Research recovery: stalled read-only research gets one evidence-based final-answer attempt, preserving sources, uncertainty and explicit stops. Missing requested screenshots or files remain incomplete.
• 0.51.77 — Workstation task records: compact records retain created-file revisions, source reads, screenshots, vision analyses and test results. “Recorded work” appears in the existing interface, with eight detailed workstation workflows.
• 0.51.78 — Focus workflow repairs: better handling of negative constraints, file creation and chunk finalization, typed JavaScript assertions, test repair, multi-source research, and readable page screenshots with vision receipts.
• 0.51.79–0.51.83 — Music: playlist shortcuts, in-app audio, text/voice commands, stream resilience, and CPU-only Home search/playback. The current in-app player supersedes the initial browser-only music behavior.

Existing models, chat history, guides, research, coding, screen sharing and other tools remain available. Normal updates preserve the separately stored workspace and model files.

Validation and current limits

The 0.51.83 checks passed 43 focused music/wake tests, six main-process IPC/protocol tests, and a voice-controller test covering four initial music requests and subsequent playlist searches. Main-process tests confirmed zero model-runtime, browser and context-preparation calls for the covered direct music commands. The production TypeScript/Vite build passed.

Electron checks exercised the production player and navigation hook with decoded audio. Live song and playlist checks covered search, playback, pause/resume, seeking, hidden-window behavior and advancing a playlist. The packaged build also passed live playback checks. These runs muted speakers and verified decoded audio/time advancement; they are not an acoustic microphone/speaker test or a guarantee of access to every track. The previous stream-repair checks played complete tracks past the reported two-minute boundary; they did not reproduce or prove the exact original cause of every reported interruption.

Public YouTube stream playback is experimental, not an official YouTube integration. Private, sign-in-only, live, protected or otherwise inaccessible streams are not bypassed. Availability can change; use the official player when appropriate. Spotify and iHeartRadio playback are not included. No new model-speed or VRAM benchmark is claimed.

Download and verification

• Windows installer — 338,404,172 bytes (about 323 MiB).
• Website download · Full version history · Music setup and limits.
• SHA-256: 709b5726400e8fcfb6cba43debd8b2eeb45a51873da216ada00af0ce49bf563f.

The installer is not Windows Authenticode-signed. Signed Arc updater metadata and download hashes are separate integrity checks. No VirusTotal result is claimed for this exact installer. Model weights remain separate downloads.

Arc 0.51.75 — Email, Reddit, timed-work guides and chat recall

Arc 0.51.75 — email, Reddit and timed-work guides

This Windows alpha release includes 360 substantial new guides: 120 for email assistance, 120 for Reddit posting and 120 for timed work. It also brings the changes from 0.51.72–0.51.74 to the public download channel.

New guide library

• Find the guides in Settings → Plugins under Guides. Each includes preparation, ordered steps, a worked example, verification, recovery and completion criteria.
• Guides cover email categories, reply tone, sender filters, inbox progress and duplicate-send recovery; Reddit drafts, exact approval, browser state and one-time scheduling; and clock, timezone, reminders, missed times and task handoffs.
• Automatic selection sends a compact task-relevant core to the model. Full instructions are available through the existing paged reader instead of filling the context window with the entire library.
• Fixed guide selection so pause, cancel and status requests do not select Reddit post-creation guidance. Timed-work guidance requires a matching timed intent.
• The bundled catalog now contains 21,123 workflows, including 20,887 authored guides and 236 retained upstream workflows. These are instructions for existing capabilities, not 360 new executable tools.
• No new dependencies or model downloads.

Also new since public release 0.51.71

Saved-chat recall (0.51.72). Written-chat requests can retrieve a bounded selection of relevant earlier conversations and recognized corrections. Settings → Memories includes Use past conversations; a request can use this chat only. Retrieval uses local CPU indexing, not another GPU model or training. Current instructions take priority and changing online facts still need fresh sources.

One approved Reddit post, now or later (0.51.73). Review the exact account, community, title, body and resolved time in Settings → Reddit posts. Approve once in advance, then Arc can submit that one post through the connected browser while open and awake. This replaces interval batches. Changed drafts require new approval; missed times and uncertain submissions require review instead of duplicate retries.

Questionnaire-based email assistance (0.51.74). Settings → Email assistant lets the user choose the observed mailbox, email categories, reply instructions/examples, sender filters and pace. One review starts one inbox pass. Messages are handled sequentially with fresh local-model context. Recipient/body checks, delivery records, pause/stop and uncertain-send handling help preserve control.

Scope and limits

• Email adapters initially support fixed English, paginated Gmail, Outlook and Proton layouts. Unknown layouts stop for review. Synthetic mailbox tests do not establish compatibility with every real account; live sending still needs a supervised provider trial.
• Reddit scheduling is one explicitly approved text post. Arc must remain running and awake. Login and CAPTCHA remain manual. There is no interval posting or bulk-posting capability in this release.
• Other timed-work guides support reminders, planning and handoff to available tools. They do not add a general background scheduler for research, file creation or recurring email polling.
• Guides cannot grant permissions, expand model context, promise task completion or override tool checks. No real email or Reddit post was sent during automated verification.
• Windows alpha; the installer is not Authenticode-signed. Signed updater metadata and download hashes are separate integrity checks. No new VirusTotal result is claimed.

Validation

The new library passed catalog integrity and reproducibility checks, routing tests for all 360 natural-language examples, full-reference hash/paging checks, and focused regression tests. The earlier email implementation passed 3,791 tests with 10 skips and zero failures. This release additionally runs the production UI build and packaged runtime checks, including the real reader over all bundled workflow references. The release regression run passed 44 tests with zero failures or skips. These checks do not constitute live-account testing or a new model speed/quality benchmark.

See email setup, Reddit setup, saved-chat recall, and the 360-guide index.

Download and integrity

Windows installer · Website download · Full version history

Installer: 335,240,500 bytes, approximately 320 MiB. Models are separate downloads. Normal updates preserve the separately stored workspace, conversations, settings and models.

SHA-256: 6b46b4833e93cc40ab6b1068fb1a327b16b994e76b15380810ed639bcb2ac3ad

The release assets include the installer, cumulative changelog and SHA256SUMS.txt. No matching VirusTotal report has been supplied for these exact installer bytes.

Arc 0.51.71 - Helper teams, complete extensions and personal memory

Arc 0.51.71 — helper teams, complete extension projects and personal memory

This Windows x64 development release brings together 0.51.66–0.51.71, following public 0.51.65. Substantial tasks can now use a coordinated local helper team, browser-extension requests produce complete source folders, and Arc can retain concise personal details you choose to share. Reviewed Reddit drafts and finite timed queues are also available as a beta feature.

Download Arc · GitHub release · Full version history · Feature overview

A team for harder tasks

With helper tasks enabled, substantial High or Ultra work can plan five distinct assignments: three for requirements, research and approach, followed by two reviewers examining the parent's actual work. Arc then uses the findings for a bounded correction and final response. Simple questions, fast voice and story writing skip this automatic team overhead. If the team planner cannot produce a valid plan, the main agent continues directly.

• Up to 24 helpers per parent turn, including failed or cancelled helpers. At most five are admitted at once; additional work waits in order. Time spent waiting for a slot does not consume the helper's execution deadline.
• One selected model, shared GPU queue. Helper model requests run in turns; Arc does not load five or 24 copies of the weights. Read-only tool work can overlap. More helpers can take longer and use additional bounded ordinary memory/storage; they do not provide free parallel inference or remove the context window.
• Preparation, work, check and repair. Helpers provide evidence, proposals and review. The main agent retains file changes, browser/SSH actions and supported execution checks through its existing tools and permissions. Reviewers do not acquire new authority or turn an opinion into a successful test receipt.
• Project builds participate. Completed source files and available check receipts can receive two result reviews. The parent may reopen one existing file for a bounded correction through the same revision checks and validation loop.
• Visible progress and controls. The helper panel distinguishes queued, working and finished tasks, includes individual Stop controls and supports all 24 entries on desktop and narrow screens. Reports remain available through bounded excerpts and paging. Older three-helper conversation records remain valid.

Use the existing General Settings → Helper tasks switch. The 24-slot maximum is a coordination limit, not 24-way GPU inference. Read the helper-team architecture and limits.

Complete browser-extension source folders

New Brave/Chrome extension projects now use the sequential project writer. manifest.json, popup HTML, scripts and styles keep their exact relative filenames in one exported directory. Every planned file must be saved before Arc can report completion. Small extension projects of up to 12 files can finish in ordinary text chat within a bounded 32-round run; larger work and longer corrections retain Ultra → Until done, cancellation and saved checkpoints.

Static checks cover supported Manifest V3 metadata, local resource references, missing scripts/popups and common extension-page script restrictions. A failure triggers a bounded repair or preserves an incomplete project. These checks are not a complete manifest/security audit and do not establish compatibility with every website.

Version 0.51.70 also corrects planning guidance after a real Strata test produced circular file prerequisites. File-writing dependencies now distinguish build order from runtime references: a popup and its script can refer to shared IDs without both requiring the other to be written first. Invalid dependencies receive explicit correction guidance instead of being silently removed.

Open Code → Local website workspace, select the project, then Export & open folder. Review the generated source and permissions before manually loading the extension. Creating source does not install it or perform its proposed account actions. Read the source and local-helper guide.

Clearer assessments and tool creation

An instruction such as “inspect this and tell me whether you can build it” now retains its findings as a feasibility report. It no longer requires a reusable helper to be published just to finish an assessment. Explicit implementation requests can create source files; a small isolated JSON helper remains a different deliverable with its own fixture checks.

Failed local-helper creation distinguishes disabled/unavailable authoring, an unattempted build, and failed validation or tests. Existing working revisions remain intact. Describing future browser actions does not authorize those actions during inspection or source creation.

Compact personal memories

Find Settings → Memories → Details you've shared. Arc can capture clear personal facts from new direct messages, deduplicate them, and recall a small relevant selection. The recognizer uses conservative English patterns, rather than an extra model; it does not scan past conversations or need another GPU allocation.

Facts are saved locally in encrypted storage, separate from manual notes. You can edit, delete or clear them. User edits are protected from automatic replacement. Stored facts have a 768-token budget; a reply receives at most a 256-token relevant pack when context space permits. Turning off automatic capture or the saved-memory master switch pauses both capture and recall.

Temporary voice, shared screen/audio, attachments, tool output and generated templates are excluded; saved direct microphone questions can contribute facts. Sensitive or ambiguous statements are skipped by the bounded recognizer, though reviewing the saved list is still useful. Memory controls and limits.

Reddit drafts and reviewed timed posting — beta

Open Settings → Reddit posts, also linked from Plugins. Use your connected browser account to prepare 1–10 editable text-post drafts with the selected local model. Review the exact title, body, account, subreddit and rules before approving a finite queue at 15, 30 or 60 minute intervals. The first post waits a full interval. The visible automation disclosure is part of the content reviewed before scheduling.

The model cannot approve its own drafts. Editing an approved draft revokes approval. Restart, Windows lock or sleep pauses the queue; resuming starts a fresh interval without a catch-up burst. Account mismatches, CAPTCHA, rate limits, required flair and unsupported layouts pause work. Unknown submission outcomes require reconciliation instead of an automatic duplicate retry.

Live Reddit posting remains unverified. Automated browser tests use controlled local pages and do not publish posts. A supervised check on the intended account/community is still needed. Setup, review and recovery instructions.

Verification and measured limits

For 0.51.71, 505 targeted Node regression tests passed across the tool loop, project/extension writer, helper engine, saved artifacts, context compaction, runtime queues and browser/SSH/PC tool boundaries. Four native Python protocol tests passed. The 24-helper UI passed desktop, tablet and 390px checks without renderer errors. TypeScript, the production UI build, workflow-catalog integrity/reproducibility and packaged-helper checks also passed.

These are scripted-model and isolated-fixture checks. Real-model helper-team answer quality, VRAM usage and task speed were not benchmarked. The test count is targeted coverage, not a claim that every feature or website was retested.

A separate 0.51.70 one-machine Strata test generated a neutral four-file word-counter extension through Arc's production project workflow. Generation/project work took 32.1 seconds after a 40.5-second model load. Exported hashes matched the saved project; Chromium 153 loaded it and passed five word-count checks with no page errors or external requests. This establishes that neutral popup task, not arbitrary extension or website reliability. Earlier release-specific test coverage is retained in the version history; counts overlap and should not be added together.

The existing 0.51.65 VRAM/workload report remains available with its original hardware, source and scope. Its measurements are historical and are not new benchmarks of 0.51.71 or helper teams.

Installation

Download Arc-Setup-0.51.71.exe from this release or the website. Models remain separate downloads, and these changes add no new model or runtime dependency. Normal application updates preserve models, conversations and settings.

The installer is 334,855,928 bytes (about 319 MiB). Its SHA-256 is:

1ef87546e639ce80a0c2914dfceee2ba0dd6fdbd0404a74ef27b89c7fc76bf6b

The release includes a cumulative changelog and SHA256SUMS.txt. The installer is not Windows Authenticode-signed. Signed Arc updater metadata is separate from executable signing. No matching VirusTotal result has been supplied for this exact installer; older scans do not cover these bytes.

Arc 0.51.65 — Measured memory tiers and task improvements

Arc 0.51.65 — Measured VRAM tiers and longer saved work

This Windows x64 development release brings together 0.51.61–0.51.65, following public 0.51.60. It adds optional lower-VRAM tiers, persistent output checkpoints and multi-file website builds. Models stay separate downloads; normal updates preserve models, settings and conversations.

Download page · Feature overview · Detailed workload measurements · Machine-readable evidence · Memory architecture

Choose a memory/speed trade-off

Open Settings → AI Performance → Lower VRAM usage. Tier 1 accepts at most 20% lower measured generation throughput, Tier 2 35%, Tier 3 50%, and Tier 4 75%. The same prompt-throughput limit applies when normal prompt timing is available. These are limits in the short selection test, not a forced delay or a guarantee for every longer task. Tier 4's speed tolerance does not promise 75% less VRAM.

Arc compares supported cache precision and processing-buffer settings on the selected model. Higher tiers can test a small number of CPU-resident blocks when the verified engine supports them. Tier 4 adds at most three directly measured combinations of Q4 cache and CPU placement. Five basic response checks, measured generation speed and memory admission must qualify. Prompt speed is compared when the normal rate is available, and first-token delay when both timings are available; missing timing is not a successful comparison. A higher tier may keep the same setting when more aggressive offload is too slow. Existing successful comparisons are cached. Crashes, out-of-memory results and unhealthy configurations cannot be revived; successful trials rejected only by a stricter speed threshold can be reconsidered at a more tolerant tier.

The model, installed weight quantization and selected context stay unchanged. Turning the option off restores the normal tested profile. No new idle-unload timer is introduced. Strata retains its own engine and is outside these llama.cpp tiers.

Real measurements, with their limits

The linked report compares Normal, Tier 1, Tier 2, Tier 3 and Tier 4 using real inference and workload tools. It distinguishes native token speed from complete task time and whole-device GPU memory from owned worker/server RAM. Source reads, screenshots, vision analysis and file validation count only when their evidence exists. Failures are retained in the report.

The earlier three-tier calibration is retained in the version history with its original scope. The new report separates short calibration from the five-mode workload comparison. These are measurements on one machine, not proof of fit or equivalent performance on 4/8/12/24 GB cards. Lower observed whole-device VRAM during text generation does not establish that a screenshot/vision handoff uses less memory or that every task finishes faster.

Measured on Arc Focus and RTX 5090

One Windows PC, 16,384-token context, three 512-token generation samples per mode. These are raw whole-device text-generation peaks, not exclusive model allocations; pre-load medians were not subtracted.

Mode	Median tok/s	Throughput change	Pre-load GPU median, MiB	Generation GPU peak, MiB	Raw peak MiB saved
Normal	187.28	Reference	1,711	8,278	Reference
Tier 1	185.38	1.01% loss	2,096	7,781	497
Tier 2	184.24	1.62% loss	2,008	7,927	351
Tier 3	182.09	2.77% loss	1,889	7,490	788
Tier 4	52.67	71.87% loss	2,210	7,432	846

All five final modes delivered the requested web answer, readable screenshot/vision result and exact 1,000-line file, and reused the model after work. Manual review still found factual and citation errors. Tier 4's file took 611.56 seconds versus Normal's 121.59 seconds, including recovered model-call failures and an unusually slow call. Its extra raw generation saving over Tier 3 was only 58 MiB in this run, so Extreme was a poor speed/memory trade-off here. Overall workload peaks, vision handoffs, CPU/RAM/disk counters, individual samples, earlier failures and measurement limits are in the full report.

Corrections found during workload testing

• A layout-only screenshot target returns a concrete instruction to choose visible content or request a page viewport. Current-page references beyond the first 100 elements are supported, with the same strict format and live-page checks.
• A fully ended tool response with correctable argument-schema errors keeps the healthy text worker ready for the correction. Malformed or unfinished streams still release the failed worker.
• Newly authored, line-counted JavaScript checks the actual saved line count. A missing section separator is repaired only at a safe, complete-statement boundary; exact-copy requests bypass that repair, and retries do not resave accepted bytes.
• A duplicate-section replay guard checks the current file ID/hash and the same raw section with an unchanged nonempty checkpoint, returning the existing receipt/tail without appending bytes or advancing chunk count. Intentional repetition with an advancing checkpoint and exact-copy requests is preserved. Its 63 focused tests include the diagnostic's 1,000-lines-but-900-unique failure shape.
• Direct-source HTTP 403/429 observations survive compaction so the agent can choose another publisher or an already authorized normal browser route. This does not bypass access restrictions or guarantee another source is available.
• Automatic source selection skips already denied hosts during the same turn. If the available lead pool is blocked, one bounded alternate-source search uses the original query with denied-site exclusions. Explicit user-required access remains distinct; this does not bypass restrictions or guarantee a readable alternative.
• Exact failed-URL guidance and no-tools synthesis for already fulfilled pure web-research contracts address a repeated-URL diagnostic failure. Focused recovery tests passed. Errors remain in the ledger, and a source-read count alone is not a completed answer.
• Combined research/screenshot requests retain a bounded follow-through phase for requested searches and page reads. Receipts distinguish actual retrieved text from search leads and indexed fallback excerpts; multiple URLs do not prove independent corroboration.
• Screenshot guidance favors readable viewport or element captures for text while honoring explicit full-page requests. Vision is instructed to flag unreadable text, with actual visual accuracy reviewed separately from tool completion.
• Capture coverage distinguishes a top-level document-root viewport from a full-page image. Oversized non-root and iframe regions receive guidance to choose a smaller visible target instead of silently widening the screenshot. Focused checks and a readable live capture passed; combined research completion is reviewed separately.
• Untargeted visual-research screenshots use the actual viewport unless the original user explicitly requested whole-page coverage. Receipts describe that scope and suggest scrolling for other sections. Explicit full-page requests, targeted captures and standalone saving keep their existing behavior; whole-page vision may still have legibility limits.
• Screenshot delivery uses the verified saved image and its file card, instead of a webpage URL or invented filename. Requested vision observations stay associated with that image and its recorded source, preserving unreadability caveats.
• Plan-only cleanup preserves substantive answers, citations, mixed code and quoted/fenced text after a Tier 2 diagnostic exposed overly broad removal. The focused suite passed 63 tests and the actual failed answer replay retained its substantive remainder. Successful tool calls do not by themselves establish answer quality; affected source revisions remain diagnostic.
• Final response instructions close the tool phase and retain combined research requirements. A confidently plan-only answer gets one bounded evidence-only composition retry without tools or repeated captures; persistent failure remains incomplete with actual images and receipts preserved. Real explanations, code, quoted plans and user-requested plans remain, and image evidence has a context-sensitive whole-payload bound. The focused repair suite passed 84 tests; this is completion handling, not factual grading.
• Research compaction preserves bounded, query-relevant exact passages and qualifications with omission notices, instead of a navigation-heavy page prefix. Page text remains untrusted evidence; final interpretation still needs accuracy checks.

These describe the implemented fixes. Live success and remaining failures are recorded separately in the workload report.

Six substantial completion guides

Six workflows guide actual source verification, readable screenshots with vision, combined research and image delivery, checked file chunks, debug/verify/save work, and evidence-based recovery. The full references contain 8,100 words in total (1,310–1,390 each), with complete compact task/check/stop recipes. A serialized single-guide selection is 1,811–1,888 characters including the shared notice, identity and formatting; the actual recipe body is smaller. The existing context-dependent budget limits what is loaded; full references are read on demand.

Find them in Settings → Plugins using the guide index. They use existing tools and permissions, pres...

Arc 0.51.60 — File History, browser tools and guided work

Arc 0.51.60 — File History, browser tools and guided work

Arc 0.51.60 is a cumulative Windows x64 development release covering 0.51.53–0.51.60, following the previously published 0.51.52 installer. It adds recoverable file history, a browser task library, 60 guided work areas, Steam research, Twitch alerts, and more precise coding repairs.

Download Arc-Setup-0.51.60.exe from Assets below. Models remain separate downloads. Normal updates preserve separately stored models, conversations, settings and personal guides. Finish active work before updating.

Download page · Complete version history · Feature and model overview

Remember and recover your work

Created Files → File History now provides a searchable local backup library. Arc records saved files and revisions with their filename, date, file type, a text excerpt and the creation instructions when available. Ask “Find the PHP page we created” or “Restore my lost notes.” Arc can search earlier work, inspect verified source and return an exact restored file in chat.

• New Created Files saves and appends, verified local-PC file saves, local PHP project files and generated images are backed up automatically. Existing Created Files are imported after startup; a manual backup button is also available.
• Backups are independent of the working copy and conversation. Deleting either does not delete the backup. Restoration verifies SHA-256 and creates a numbered copy when the name is already taken.
• Identical content is stored once. The initial limit is 10 GiB of unique file bytes and 50,000 version records, with visible warnings and manual removal. Older versions are not automatically erased.
• Search is local and requires no extra model, embedding download or VRAM. It uses metadata and the first 8,000 characters of text; full source can be read in excerpts.

This is recovery on the same PC, not an off-device backup. It cannot recover files deleted before their first backup, survive loss of the disk or Arc application data, or automatically rebuild an entire multi-file project. Remote SSH backups retain their existing separate behavior. File History guide

Browser tasks you can find and use

Settings → Plugins → Browser tasks adds 120 detailed workflows across 12 groups, each with an editable example, a procedure, evidence checks and recovery instructions. They cover tabs, navigation, page elements, copying, screenshots, vision discussion, forms, research, structured data and reporting.

Four new controls extend the existing browser engine:

• Locate rendered text and its surrounding context.
• Inspect a selected region for text, headings, links, labeled controls or table rows.
• Highlight a region with a passive blue outline; Escape or ten seconds removes it.
• Scroll a page or container, or bring an observed target into view.

These are 120 task guides and four new controls, not 120 separate execution engines. Browser permissions, sign-ins and human CAPTCHA handling still apply. Small-context responses retain complete records and continuation positions. Canvas, frame and shadow-root content can still require the existing screenshot/vision paths.

Sixty work areas and seven practical utilities

Settings → Plugins → Work Areas adds 60 searchable areas for coding, research, data, writing, learning, planning, platforms, voice/vision and operations. Each provides a detailed guide, six stages with acceptance checks, prerequisites, failure handling, a concrete deliverable and an editable prompt.

Seven utilities help the model organize and verify work: a work-area planner, checkpoint audit, research-evidence audit, weighted comparison, measured-value summary, unit converter, and text-length/repetition review. Requested exports go through Created Files. These utilities calculate or check structured evidence; they do not establish factual truth, literary quality or successful remote actions by themselves. Existing personal guides and off switches are preserved.

More precise coding repairs

Arc can read exact line ranges, append to an existing code file and replace only the faulty lines while preserving surrounding content and originals. Current source hashes and observed ranges guard against stale or guessed edits. The local PHP workspace also gains append and inclusive line replacement.

Requests such as “test/debug this file and correct errors” route through read, check, repair and recheck. Supported syntax checks and existing JavaScript assertions run against corrected revisions when enabled. Check-only requests do not authorize edits; syntax-only checks do not execute JavaScript. Missing checks cannot be reported as passed. Existing sandbox and server-publishing boundaries remain in place.

Steam, platform research and Twitch

• Steam game research: search for a game and retrieve current player counts, region-specific prices, review totals, metadata, global achievement percentages and news. Results include source links and retrieval times. No account or API key is needed for these public tools. SteamDB links are provided as references; historical SteamDB statistics are not scraped or represented as available.
• 150 platform guides: 30 each for eBay, Steam, YouTube, X/Twitter and Twitch, with examples, collection instructions, source checks and failure handling. Find them under Settings → Plugins → Platform research guides. Account and platform restrictions remain in effect.
• Twitch alerts: a dedicated settings page supports channel selection, live-stream alerts while minimized, optional stream-ended/category alerts, game filters, quiet hours, device-voice announcements and recent activity. It runs without model inference. Setup requires a Twitch Public developer application Client ID and user sign-in; instructions are included. Automated checks passed, but real account-connected notification delivery still awaits that setup. Twitch setup guide

Existing Hey Arc / Arc Mini, Show Arc screen and audio conversation, Watch with Arc for YouTube, Agent Reach, up to 20 requested screenshots, Ultra continuation and model residency remain available. No model weights or inference backends are changed by these releases.

Validation and practical limits

• For 0.51.60, 267 regression tests, 19 focused recovery tests, and the plugin-upgrade check passed. These overlapping runs are not a combined unique-test total.
• The real File History UI passed search, preview, deleted-file recovery, collision preservation, Explorer handoff, version filtering, keyboard closing and desktop/narrow-layout checks using isolated real file storage.
• TypeScript, production build and Windows packaging passed. The packaged app recovered deleted files byte-for-byte, searched after restart, preserved originals and loaded its tool registry and guide.
• Earlier component runs recorded 470 browser regression tests for 0.51.59 and 231 Work Areas regression tests for 0.51.58. Counts describe those runs, not a rerun of every app test against 0.51.60.
• No new live-model speed or accuracy benchmark is claimed. The latest work did not rerun the entire application suite or establish that the two previously reported memory-budget failures are resolved. External platform access, authenticated eBay sold-item results and real Twitch account delivery retain their documented limits.
• Local models can still make mistakes. Guides and continuation do not remove context limits or guarantee task completion. This is a Windows x64 release; no macOS installer or Android APK is attached.

Installer verification

• File: Arc-Setup-0.51.60.exe
• Size: 334,724,151 bytes (about 319 MiB)
• Windows Authenticode: not signed
• VirusTotal: no matching report supplied for this exact installer. Older hashes and timed-out scanners do not establish a result for this file.

SHA-256:

7822b5d0405cc0e4f351948915f8cf0a464c598abef5ffe499c8c807c48c9cb6

SHA256SUMS.txt and the cumulative changelog are attached. Arc's signed update metadata is separate from Windows Authenticode signing.

Arc 0.51.52 — Voice, YouTube, browser research and reliability

Arc 0.51.52 — voice, YouTube, browser research and reliability

Arc 0.51.52 is a cumulative Windows x64 development release. It includes the changes made since the previous public 0.50.4 installer, including Hey Arc/Arc Mini, Watch with Arc for YouTube, broader Agent Reach connections, screenshot-and-vision research, long-form file work, local PHP development, and the latest browser/eBay and feedback repairs.

Download Arc-Setup-0.51.52.exe from Assets below. Models are separate downloads. Normal updates preserve your separately stored conversations, settings, guides and installed models.

Full version-by-version changelog · Feature and model overview

Talk about what you see and hear

Show Arc combines optional shared-screen samples, local vision, local shared-audio transcription and voice conversation. Ask about a page, a coding problem, a chart or something said in a video. A blue boundary indicates desktop sharing, and Escape stops sharing/listening. Visual memory is sampled still frames, not continuous video comprehension.

Hey Arc now has a visible Arc Mini desktop helper, configurable automatic startup, prewarmed Focus/speech and minimized-window operation. Say “Hey Arc,” pause for the greeting, then ask. A local speech fallback improves wake recognition; microphone access and device-specific limitations remain. Turning the feature off is remembered. Explicit requests to go deeper can select a longer reasoning path while ordinary voice questions stay on the fast path.

Watch with Arc · YouTube is available in Plugins, Research and Voice. It prepares available captions or permitted audio, follows your playback position, optionally inspects selected frames, bookmarks moments, compares up to three prepared videos, and saves timestamped notebooks and actual analyzed images. Simple questions can reuse prepared evidence directly; background observation yields to foreground speech and questions. Enabling the plugin alone does not start observation.

The desktop also includes Android Companion preview pairing and private remote connectivity support. The separate APK is not included in this release. Phone hardware, capture/echo behavior and end-to-end latency still need physical-device validation.

Browser, Agent Reach and screenshots

• Choose Brave, Chrome or Edge and connect the usual browser with its own required permission. Arc can find the tab matching a request and start a closed browser when needed. A separate research profile has its own sign-in; no saved-password/cookie import is performed.
• Agent Reach adds supported public readers and searches, account-connection help, platform permission switches and optional diagnostics in batches of three. Signed-in features remain subject to the platform's availability, account state and limits.
• Ask for up to 20 screenshots, saved and previewed in chat/Created Files, with optional local-vision summaries. A blue outline marks the screenshot area. Distinct captures, retained page evidence, partial file receipts and retry behavior reduce repeated navigation and plan-only endings.
• A request-understanding panel shows the outcome and steps Arc inferred. The plan guides tool routing but is not treated as completed work or additional permission.
• Research → Security adds bounded domain-research workspaces with module progress, cancellation and report export. These are native adaptations, not full installations of every upstream security CLI.

Long-form work, guides and local coding

• Ultra Until Done supports bounded continuation with checkpoints, saved evidence and preserved partial results. Long-story mode writes a TXT file with a 20–90 paragraph target, continuing verified sections and keeping the exact saved ending.
• Stronger repetition detection catches identical and rotating paragraph/sentence loops. Failed story sections get bounded rewrites without feeding the rejected draft back as the next section. Model reasoning budgets and context accounting were repaired for Strata/tool work.
• Import personal JSON, TXT or Markdown guides from Plugins, using clear field descriptions and downloadable examples. New collections include 300 story-writing guides, 300 advanced Hey Arc guides, and 120 Ultra Until Done guides. Guides are instructions, not executable extensions.
• Code → Local PHP workspace offers persistent local source, isolated lint/preview, optional copying from an SSH project, and separately reviewed publication back to the server with revision/conflict checks. Requires the optional Docker/PHP runtime; this is not a full XAMPP/MySQL stack.
• Generated Python runs in a fresh Pyodide/WebAssembly environment with bounded virtual files and no host shell or UAC access. Third-party integrations and authorized remote operations have their own permissions.
• The selected text model stays resident between tasks where possible. Temporary vision can restore the original model/context; shared GPU scheduling and failures can still require handoffs.
• Strata's optional Q2_0/IQ2_XS choices now support the pinned alesha-pro abliteration projection variant while keeping the Strata engine. Eligible existing weights are reused with a small projection download. Short paired RTX 5090 Q2_0 measurements found similar speed; they do not establish broad accuracy or performance on every PC. No Hy3 backend is included.

Latest fixes in 0.51.51–0.51.52

Discord feedback: the Feedback button sends a reviewed UTF-8 text attachment through Arc's server relay instead of email. No Discord sign-in is required. The webhook credential stays on the server; it is absent from the desktop bundle. Only the submitted fields/version/reference are included, not automatic conversation or log uploads.

Browser launch and eBay: screenshot requests open a visible research browser when necessary. A missing tool registration that could reject unrelated research before inference is repaired. Market requests use a smaller relevant tool catalog and preserve exact page excerpts through compaction. “Take 3 of them and screenshot them” keeps the requested count, and the final research explanation remains alongside image attachments.

Sold-price evidence: distinct sold-item pages and captures are tracked separately. Active listings, search results and repeated screenshots cannot fulfill a request for three distinct sold listings. Undisclosed accepted offers are excluded from averages; currency, shipping, grade and lots must be distinguished. Sign-in and human-verification blockers stop the loop with a clear explanation.

Find the new instructions at Settings → Plugins → eBay research · Sold prices & screenshots. Sign in yourself when eBay requests it, then return to the same chat and say “continue.”

Verification and remaining limits

• 216 focused Node tests, 91 Python worker tests, and 96 voice-boundary tests passed.
• TypeScript/production UI build, Windows packaging checks, wide/narrow eBay help UI, installer execution and installed-file verification passed. Forty-two packaged/installed files matched the verified source/build output.
• The complete existing run recorded 2,978 tests: 2,970 passed, 3 failed, 5 skipped. One screenshot/file-card regression was fixed and passed its focused rerun. Two existing memory-budget failures remain; the complete suite is not represented as clean.
• Real Arc Focus Ultra work in visible Brave reached eBay's Sold Items login screen in six rounds, recovering one typing error. Its strict no-error-events smoke assertion therefore did not pass. Three authenticated sold listings, their screenshots and a live average remain unverified until sign-in. Controller/fixture tests are not substituted for that live result.
• Platforms can require sign-in, impose rate limits or refuse access. Arc does not solve CAPTCHAs. Vision, speech and local models can make mistakes. Continuation and SSD-backed retrieval do not remove active context limits or guarantee completion.
• Windows CPU/CUDA/Vulkan paths are implemented where verified; there is no universal ROCm/SYCL/Metal/TensorRT-LLM execution support or macOS installer in this release. The default Arc system-RAM ceiling remains 32 GiB.

Installer and scan information

• File: Arc-Setup-0.51.52.exe
• Size: 334,292,623 bytes (about 319 MiB)
• Platform: Windows x64
• Authenticode: not signed
• Matching VirusTotal report: not yet supplied for this exact file; older-file reports are not claimed for this release.

SHA-256:

7aaf881bcdf8507272c2e2a72d1f6a6907257f38b08aa7ac97a79617148de380

SHA256SUMS.txt and the downloadable cumulative changelog are also attached. This is an active-development release, not a claim that all workflows, hardware or external services have been validated.

Arc Master

Master Copy (beta)