More reliable requests, memory-informed decisions, and a more compact Home. Explore the fixes and refinements in 1.8.1 and the releases that came before it.
ABot 1.8.1 fixes approval-resume failures and interrupted Ollama streams, brings long-term memory into the first decision, and refines tool discovery and planning. A compact Home layout keeps everyday controls accessible on smaller screens.
Resume work without losing the update
Steering during a saved tool-approval resume no longer fails the request. The original approval or rejection stays in place, while the update reaches the existing continuation once.
Handle interrupted Ollama streams cleanly
An abrupt disconnect now fails the request explicitly instead of crashing the Gateway. Partial output is not accepted as a completed answer, and the same process remains available for later requests.
Bring memory into the first decision
Execution Agent and Supervisor prepare automatic long-term memory before their first decision and retain it through the request. Execution Agent can combine an eligible short answer and memory candidates in one model call.
Explicit memory recall remains available in both execution modes.
Find tools before opening every detail
Execution Agent, Supervisor, and Planner start with a compact catalog organized by tool groups. Detailed operations appear when the relevant group is opened. Supervisor chooses an available Planner for remaining work that spans multiple steps and at least two tools.
Keep Home compact and useful
At widths up to 1440px, Home shows compact previews for activity, conversations, approvals, runs, automations, and computer access. Open a widget to use its existing live panel in a scrollable dialog; wider screens keep the full layout.
Opening and closing widgets preserves their live state, drafts, actions, and focus.
More fixes in this release
Execution Agent keeps its direct-answer option when a conservative context estimate would previously hide it. Actual context admission, provider token counting, and compaction still enforce configured limits.
Initial memory preparation preserves supported conversation text and attachments while filtering request-local tool evidence and ignoring undeclared memory-reference fields.
The interface labels the existing Fast reasoning mode as Low. Stored values, API values, effort mapping, and defaults are unchanged.
Computer previews include accessible names and operating-system descriptions. Opening SPARK chat focuses the input.
v1.8.0
Fewer interruptions. More reliable everyday work.
ABot 1.8.0 focuses on reliability across model responses, tool results, and conversations. It also keeps Web UI preferences and Send next queues with your installation and fixes Marketplace installation links on Mac.
Handle incomplete responses explicitly
OpenAI decisions and file payloads use the final answer rather than preceding commentary, reject incomplete responses, and finish when completion is received instead of waiting for the connection to close. Ollama errors and missing completion markers now fail explicitly rather than accepting partial output.
FAST mode no longer adds a separate output limit that could cut off valid results. The model's configured limits still apply.
Keep tool results attached to the work
Execution Agent decisions and replies use the Runtime's recorded tool outcomes and exact result evidence. Saved approval waits retain that evidence after an upgrade, and steering reaches the request that owns the work when it resumes.
Keep drafts in the right conversation
Chat, Home, and SPARK keep text and files with their originating conversation during navigation or a rejected send. Unconfirmed sends remain separate for manual review, and ordinary sends no longer wait for unrelated preference saves.
Keep preferences with your installation
Web UI preferences and Send next queues are saved with your ABot installation, so they are available across browsers and local server addresses. This is local installation storage, not cloud sync.
Ordinary unsent drafts stay local to their window. Existing browser-only preferences and queues start fresh once; see the upgrade notes below.
Open Marketplace links on Mac
Marketplace installation links now handle the native macOS URL event and open the paired Runtime's plugin review screen. The link prepares a review; installing the plugin still requires your confirmation.
Existing Mac connections need their installation-link handler refreshed after upgrading. Update Companion and follow the macOS connection instructions.
More fixes in this release
Model and connection setup recover interrupted configuration changes. Context-window errors appear beside the field that needs attention.
Execution Agent context avoids repeated capability descriptions and work-plan schemas, with clearer capability selection and Planner handoffs.
File creation guidance distinguishes files from directories and ordinary edits from complete replacement. A file blocking a parent directory now produces a clearer error.
SYSTEM and Companion mask detected secret indicators in observed content and withhold associated desktop images before delivery. Detection is a safeguard, not a guarantee that every secret will be found.
Closing SPARK cancels pending openings, preventing a delayed opening from changing navigation or focus. Provider diagnostics retain bounded partial responses and stream timing to help investigate failures.
Memory review markers keep the latest reviewed revision per memory, with a limit of 200 records per environment, and retire confirmed deletions.
Before upgrading
Existing browser-only preferences and Send next queues start fresh once. Conversations, memories, configuration, and read status are retained. Tool permissions default to Ask until you explicitly change them.
Every installation needs a one-time SPARK collection acknowledgment, including existing installations. It is retained across updates and re-pairing; collection and learning remain separately controlled.
Update Companion before using SYSTEM operations that require the new host-side secret protection. Unsupported Companion versions are rejected before dispatch. On Mac, also refresh the installation-link handler using the macOS connection instructions; updating Runtime alone does not replace Companion.
v1.7.0
Connect your Windows PC. Make ABot your own.
Connect Windows through ABot AI Companion on Microsoft Store. Discover community plugins, review their tools, and configure them in your own Runtime. Pagination makes longer conversations easier to navigate, while complete command output stays available beyond the current step.
A clearer way to connect Windows
Install ABot AI Companion once through Microsoft Store, return to ABot, and open a temporary connection link. Companion shows the Runtime address and asks you to acknowledge computer access and data sharing before allowing the connection. Already installed? Go straight to connecting.
The Companion app and Runtime have separate releases. This guided connection flow requires Runtime 1.7.0. Runtime stays in Docker or WSL; computer-action permissions and optional SPARK controls remain separate.
Discover community plugins
Browse the ABot Marketplace and choose Open in ABot to open a plugin source in your paired Runtime, or paste a public GitHub URL into Plugins → Discover. Review its tools and configuration, confirm Install and enable, then Apply changes when the Runtime is idle.
Beta in 1.7.0: plugins only. Private repositories, plugin updates, and Studio packages are outside this first installation flow. A listing is not a security certification, and an install link cannot bypass local confirmation.
Configure a plugin where you install it
Fill in declared settings and required secrets during installation, then revisit them from Plugins → Configure. See whether a setting uses its default or a local override, manage individual tools, and explicitly apply saved changes to idle Runtime environments.
Settings and secrets stay in local configuration shared by the consumer root's environments. Removing a plugin retains these values; existing Runtime owners keep their applied configuration until Apply succeeds.
Find your way through longer conversations
Session lists, conversation history, and SPARK gain pagination, with 20 items per page by default. Open older conversations and run links without losing saved sidebar order, activity details, or the live response. Read tracking follows the messages visible to you.
Keep the full command result
Local SYSTEM commands retain complete stdout and stderr on disk. Updated Companion installations transfer large results in verified chunks while preserving the process receipt. Text reaches the model through the existing context-admission limits.
Retained output is not unlimited model context. Capture retention remains in force across Companion reconnects, and output-delivery failures preserve process receipts.
More fixes in this release
Repository review resolves a source to an exact commit and validates declared files without executing plugin code. Installation checks that same commit and reviewed digest. It does not run npm install, build commands, or lifecycle scripts; publishers supply a complete built package.
Remove is available for plugins installed through the GitHub installation flow, not built-in plugins or manually copied folders. It disables the selected plugin and successfully applies that change to idle Runtime owners before deleting its installed package. Reinstalling clears only that plugin's exact disable and refreshes its package-local module caches, preserving other plugins and tool restrictions.
Plugin installation, selection saving, removal, and rollback coordinate with Apply. A failed Apply or concurrent configuration change prevents deletion. After a partially completed removal, the UI shows active environments and distinguishes newer saved changes that still need Apply.
Plugin configuration preserves retained settings and inherited defaults. Failed installations leave settings editable; changing the source clears secret drafts. Secret values are not returned to the UI, and local plugin credential files require owner-only permissions on POSIX systems.
Windows connection links expire and stay in the current setup surface rather than browser storage. Open Companion can reopen a saved offline connection without issuing a new invitation. Set up again requires confirmation before revoking the previous pairing and requesting a fresh grant.
Compact Computer Access cards show connection status across Windows, macOS, and Linux. Computer Access and SPARK wait until onboarding completes; ordinary refreshes preserve an existing connection and the open SPARK conversation.
Studio draft authoring distinguishes substantive text from structured output and metadata, with compact operation declarations for selected steps rather than repeating Studio guidance in ordinary tool descriptions.
Before upgrading
Windows Companion installation and updates now go through Microsoft Store; the old Windows EXE download is no longer offered by Runtime. Companion remains separate from Runtime in Docker or WSL. The 1.7.0 connection invitation requires local HTTP loopback access to Runtime; publish or forward the Web UI port appropriately on the Windows host.
Marketplace links require Companion release 12, the internal Companion release rather than the Microsoft Store package version. Existing pairings can update without unpairing. On Windows, choose Update in Microsoft Store; other platforms update through Computer Access. Upgrading Runtime alone does not replace Companion.
Finish active conversations and Jobs before installing, configuring, enabling, disabling, or removing plugins. Browsing and repository review remain available while work runs. Saving configuration alone does not change the tools or values loaded by existing Runtime owners.
Removing a plugin installed through the GitHub flow deletes its installed folder, including any data it stored inside that folder. Local plugin settings and secrets, files stored elsewhere, and tool restrictions are retained. Review those retained values before reinstalling.
v1.6.0
Build workflows. Keep the conversation going.
ABot Studio turns repeatable work into reusable automations. SPARK gets one ongoing conversation for your messages and proactive suggestions. Underneath, stronger context compaction and retained command output help the Runtime carry useful evidence through longer tasks.
Build workflows in ABot Studio
Combine tools and agent steps in a visual editor, connect their inputs, and run steps in sequence or in parallel. Start from a template, reuse saved agents, and choose text or validated JSON results with acceptance checks.
Conversations can help create and edit drafts. Authoring a draft does not grant permissions, publish it, activate it, or start a run.
One ongoing conversation with SPARK
Keep SPARK nearby in a movable popup, with attachments, its own permissions, and Fast, Reasoning, or Deep settings. Messages and proactive suggestions share a persistent conversation per environment. Take a saved suggestion into a regular work conversation with its retained context when you are ready.
SPARK uses the selected Proactive model and starts in Ask mode. Opening a work conversation does not start the task. Clearing SPARK messages preserves its identity, permissions, and long-term memory.
Run on demand, on a schedule, or as files arrive
Start Studio automations manually, schedule them, or poll a folder for matching files—individually or in batches. Follow step outputs, receipts, and run reports, including the difference between an attempted action, confirmed execution, and an uncertain result.
Publish and activate explicitly. Runs use saved Full or FULL+ permissions. Reruns need confirmation and current permission checks; interrupted Studio runs are not automatically replayed.
Make room without losing the thread
Context compaction can reduce both conversation history and active-task evidence. Summaries are rebalanced as results arrive, with instructions to preserve unfinished comparisons and exact record/value associations. Execution Agent stages reuse verified summaries for the same accepted objective, while Supervisor workflows retain tool and delegated-role evidence.
Changed objectives cannot reuse stale summaries. Shared checkpoint allowances and hard input limits still apply; context indicators distinguish measured usage from estimates.
Retain command output beyond the current step
EXEC keeps complete stdout and stderr on disk, including output waiting for approval or not yet accepted for presentation. Text enters the model through context budgets, separately from delivery metadata. Cancellation and expiry retain the process identity without rerunning the command.
Retained output is not unlimited model context. The default retention is seven days after capture writers close, and retention is an age limit—not a disk-space quota.
Prepare your Mac before the first task
Check Screen Recording and Accessibility from computer setup or Home, with separate checks for the Companion and native Runtime. The release also addresses screen-capture geometry, notification setup, and keyboard combinations including the Cmd alias.
Permissions for one route do not grant access to the other. Check permissions on the Mac that will perform the actions.
More fixes in this release
The Studio library separates Published automations from Drafts and sorts by last edit. Schema-aware editors preserve unsaved changes, support reset-to-published and guarded move undo, and retain the selected run and expanded evidence during refreshes. Restored automations stay paused.
Studio authoring validates complete draft batches before saving, preventing partial writes and stale updates. Isolated agent steps use their selected tools and result schema. Unchanged schedules retain their timing; folder intake reports capacity exhaustion instead of replaying forgotten file versions.
SPARK suggestions are recorded once, in order, in its conversation instead of separate suggestion cards. Recent human messages can guide their language and tone without changing permissions, learned facts, or eligibility. Proactive progress can survive restarts and daily quota deferrals for up to six days.
SPARK keeps unread state across reconnects and marks only visible messages as read. Its launcher supports keyboard movement and reduced motion. Notifications open retained SPARK messages or Studio run reports in the correct environment.
Plugin results can declare text separately from metadata for context admission. The ordinary SDK result allowance rises to 256 KiB; attachmentInput and MIME declarations let compatible tools receive Studio file sources. Model context limits still apply.
File operations reject foreign or ambiguous path namespaces, and relative project paths stay tied to the request’s working directory. Tool payloads preserve their selected objective and reject duplicate JSON keys. Agent guidance favors saved scripts for long programs and distinguishes parsing failures from zero matches.
Compaction and provider token counting obey cancellation and model-attempt deadlines. Verified summaries remain bound to their source and objective, and operation-supervision outcomes survive reduction. Reasoning-enabled OpenAI decisions are no longer restricted by the short non-reasoning output cap.
Successful embedding onboarding enables native memory and disables the separate Memory plugin without deleting its existing data. Skipped or failed setup leaves plugin selection unchanged.
Brave search keeps retrieved results when an optional source-page fetch times out and reports incomplete coverage. Document-reading tools explain supported formats and direct source-code reads to file readers.
Before upgrading
Back up scheduler data before upgrading if you may need to downgrade. The scheduler journal migrates to schema version 4, preserving Jobs and run history, but Runtime 1.5.0 cannot read the upgraded journal.
Update an installed Companion to release 10 through Home → Computer access for Studio/SPARK notification links and the Mac fixes. Check Companion and native Runtime permissions separately.
EXEC’s former outputMaxChars setting no longer clips output. outputRetentionMs defaults to seven days after capture writers close; reading or restoring saved output does not extend retention. Plan disk capacity separately from this age limit.
Older SPARK suggestions remain in their historical conversations; suggestions without a saved handoff brief cannot open a new work conversation. Compatible proactive review stages can resume, while changed authoring stages wait for quota. Learning and reassessment retain their existing 24-hour progress lifetime.
Automation step details are retained for 30 days and reports for 90 days. Source snapshots remain while reports reference them, and reruns require those snapshots. Unresolved effects delay cleanup; ordinary Job history retention is unchanged.
v1.5.0
An agent closer to your everyday work.
ABot Spark introduces optional learning from computer activity and proactive suggestions. Desktop tools, a redesigned workspace, and stronger memory and approval handling bring more of your workflow into the Runtime.
Meet ABot Spark
Learn from visible accessibility text on your paired computer and receive proactive conversation suggestions. Collection, learning, and proactive reviews have separate controls, activity hours, application rules, and shared daily review limits.
Collection and proactive messages are off by default. Learning and proactive reviews cannot invoke tools or act on suggestions themselves; activity coverage can be partial.
Observe your desktop. Act on what is there.
SYSTEM tools can discover windows, capture screens or regions, and click, drag, scroll, type, or focus an observed window. Actions bind to a fresh observation and reject detected focus, session, or geometry changes.
Available operations depend on the OS, helpers, permissions, and model support. Screenshots are temporary request media, excluded from persisted conversation history and ordinary provider traces—not a promise that images stay off the model provider.
A workspace that keeps you oriented
An ink-and-mint Web UI puts activity inline with conversations. Pin, reorder, archive, and restore conversations; use direct workspace URLs and browser navigation. A persistent inbox collects replies, failures, approvals, and Spark suggestions.
Native alerts use the paired Companion and OS permissions. They can arrive with the browser closed while Runtime and Companion remain running. Conversation organization preferences are browser- and environment-local.
Choose what your agent remembers
Manage saved records and embeddings in the Memory workspace, inspect Spark candidates, and keep or delete saved memories from Home. Automatic proposals share one admission policy, requiring at least two independent evidence opportunities and a configurable promotion score before recall.
Explicit requests to remember a fact can use protected core-memory saves. Manual records, edits, and deletions take precedence over automatic replacements.
Pause for approval. Stop when you need to.
Web UI approval waits survive Runtime restarts. An explicit decision can resume the exact prepared action when the saved continuation is compatible. The composer Stop control cancels an active request while preserving its stopped exchange for future context.
Stopping does not undo completed tool effects. Restart never automatically replays interrupted work. New messages and Jobs in a waiting conversation remain queued until it resumes or is cancelled.
More fixes in this release
Computer access shares one pairing across SYSTEM, Spark, and compatible plugins. Linux setup is available; Windows/WSL setup no longer requires interop changes or a distribution restart.
Jobs save Full or FULL+ permissions; new Jobs default to Full. Existing Jobs retain their mode unless explicitly changed.
Worker and Planner handoffs retain full result prose, subject to context admission and compaction. Invalid delegated output returns a failure to its parent instead of granting completion or review approval.
New model profiles use separate files. Removing a declaration preserves its file and credentials; existing inline profiles remain supported.
Spark retains eligible pending work across pauses, transient failures, budget exhaustion, and restarts while inputs remain valid. Companion observations redact recognized credential fields before transport; recognition is bounded and covers new observations only.
v1.4.0
Set up your agent. Connect it to your work.
Guided setup gets your models and tools ready, Projects organize related conversations, and computer access connects a Docker Runtime to applications on its host.
Start with guided setup
Choose OpenAI or Ollama in the Web UI, configure optional embeddings, and select plugins and individual tools. Add models using saved provider connections while preserving existing profiles and the default model.
Give each project a place to work
Projects group conversations around independently selected working folders. Preview and download files from Activity when a stable working-folder identity is available; supported documents can open in their default application on a locally connected Mac.
A project folder does not change the default workspace or create a permissions sandbox.
Choose FULL+ for system actions
FULL+ allows task-relevant application discovery, launching, and system commands without additional ABot approval prompts. Ask and Full require approval for each prepared system action, including shell execution and process controls.
Operating-system permissions still apply. Disabled tools remain unavailable in every mode. New conversations remember your last explicit mode selection; existing conversations retain their own mode.
Connect Docker to your computer
Guided computer access pairs a Docker Runtime with a native Windows or macOS companion, with connection status and disconnect controls. Chat guides you through setup when required access is missing.
Working native access needs no companion. The Windows setup script can repair WSL interop and restart the selected distribution; the setup screen explains that interruption before download.
Follow plans and review original evidence
Execution Agent can show an adopted plan and item progress without extra model decision turns for status updates. Its Auditor selects original evidence from a complete work inventory and can request another bundle when needed.
Review guidance keeps unperformed functional tests visible as gaps and avoids repeatedly offering unchanged completed reviews.
More fixes in this release
Activity and approval cards show recorded commands, arguments, operating-system targets, working directories, and elevation details. Prepared actions stay bound to their selected target through approval and execution.
Supervisor and delegated Planner receive canonical capability descriptions with context-aware admission and compact fallbacks. Structured-output repairs retain capacity for required inputs.
Setup retries preserve provider selections, credentials, and model profiles. Configuration saves coordinate concurrent changes and retain recovery for malformed files.
Long Worker activity descriptions are bounded without aborting valid actions or splitting Unicode characters. Computer-status checks avoid unnecessary configuration rendering, and Jobs keeps a consistent content width.
Shell permission decisions use declared sensitivity. Approved commands retain their original text and may use an existing working directory accessible to the OS user. Ambiguous Dev View locators return candidate line ranges and concrete numeric controls.
v1.3.1
A focused system compatibility patch.
Fresh environment setup now creates the paths it declares, while core local tools gain a macOS-compatible path without relaxing their existing safety boundaries.
Start with the directories you configured
Initialization now creates each environment's configured working and storage directories while preserving existing configuration and custom paths.
Bring guarded local tools to macOS
File creation and editing, directory inspection, local search, and shell commands now use macOS-compatible behavior while retaining path containment, atomic writes, and process cancellation safeguards.
Release validation was completed on Linux. The macOS compatibility path is included in 1.3.1, but it has not yet been verified on a MacBook.
v1.3.0
Work that can wait. Memory you can ask for.
Native scheduled Jobs let ABot return to a task later, while focused memory recall and a redesigned Home workspace make ongoing work easier to start, follow, and resume.
Schedule work inside the Runtime
Create one-time or recurring Jobs that retain their task, model, agent mode, conversation, and run history. Manage them from the direct local Web UI or through the model's built-in schedules capability.
Jobs run with their saved model and full tool permissions while the local Runtime host is active. Missed offline occurrences are recorded but not backfilled, and failed runs are not automatically retried.
Ask memory a focused question
When long-term memory is enabled, both Supervisor and Execution Agent can request a bounded recall, receive the result, and continue the same root decision without delegating to a plugin or child role.
Recall remains optional reference context—not a user instruction, action authority, or proof that work completed.
Start from a useful Home
The Web UI now opens on a dashboard with recent Job activity, conversations with unread state, and shortcuts into attachments, Jobs, conversations, and configuration. Its composer starts a new conversation without replacing an existing Chat draft.
Read the work, not just its status
Chat now renders tables, nested lists, quotes, code, links, and bounded cached link previews. Tool activity groups each evolving execution into a compact row with executor attribution and expandable recorded input and output.
More fixes in this release
Local applications share scheduler ownership, isolate generated state for named environments, and retain durable scheduling history.
Schedule management stays on the direct local Runtime backend and is unavailable through external bridge mode. Local run history is retained rather than automatically pruned.
Live activity describes the current phase below assistant content and stays static during approval waits or disconnections; animation respects reduced-motion preferences.
Direct Supervisor responses retain their accepted recommendation through response authoring and discard superseded preparation after steering.
Scheduled requests retain Job and run origin, memory continuations preserve their accepted lookup and chronology, and rejected tool lifecycles settle their correlated events.
v1.2.0
Clearer handoffs. More visible work.
Delegated work keeps its target, plans are easier to follow, and a new Light search option works without a search API key.
Keep delegated work on target
Each Planner-owned Worker receives one production plan item, with its exact assignment, working directory, and relevant dependency artifacts preserved through execution. Supervisor coordinates any subsequent independent review.
One item is one complete outcome—not necessarily one file or one tool call.
Search without a search API key
Light search discovers and ranks content within a bounded catalog of public Hebrew and English sources. It gives the web plugin a key-free search option alongside configured Brave Search.
Light is not whole-web search. A failed Brave search does not silently switch to Light.
Follow the work from the conversation
A current-turn Planning drawer, compact role activity, and role avatars make delegated progress visible. Model selection, reasoning depth, and context usage sit closer to the composer; history stays docked on wide screens.
See the evidence behind the handoff
Source receipts distinguish returned page content, snippets, references, omitted output, and failed fetches. Reviewer handoffs retain the actual completion target and delegated results, with checks against conflicting verdicts and missing evidence.
Receipts describe tool output and provenance. They do not prove that a model read a source or that its conclusion is correct.
More fixes in this release
Ollama structured output no longer gets an artificial token ceiling based on JSON schema size. Physical-context and core-decision limits still apply.
UTF-8 decoding preserves characters split across incoming chunks; tool counters reflect actual invocations.
Planning state restoration, live updates, replay, and steering receive fixes. Session history can copy the exact session filename for support.
Creation guidance preserves existing resources and asks for clarification when required read-only discovery cannot resolve an ambiguous mutation target or scope.
v1.1.0
Continuity without replacing your setup.
Optional memory across sessions, more focused configuration, and tighter links between an accepted operation and its execution evidence.
Optional cross-session memory
Passive long-term memory can use a consumer-selected OpenAI or Ollama embedding model, with bounded retrieval and background persistence. Memory is disabled by default.
Capability action, objective, payload, and returned evidence remain bound through execution. Explicit reasoningEffort: "none" stays explicit through provider adapters.
v1.0.0Initial public release
The public foundation.
The first public release established the runtime library, public configuration schema, sanitized configuration examples, and the runtime plugin contract with example integrations.
Using an earlier version?
Your local configuration stays yours.
Package upgrades do not automatically rewrite or migrate your local configuration. Legacy v1 request-runner files remain supported. Compare new guidance with your model calibrations and methodology files before changing them.
Install version 1.8.1
$ npm install @abot-ai/runtime@1.8.1
Before upgrading to 1.6.0: back up scheduler data if a downgrade may be needed. The journal upgrades to schema version 4; Runtime 1.5.0 cannot read it. Review the release-specific retention and Companion guidance above.
Before upgrading to 1.5.0: core memory automatically upgrades to schema version 5 and creates a private .pre-v5.backup. Version 1.4.0 cannot read the upgraded store; preserve that backup for a possible downgrade. The separate memory plugin store is unchanged.
After upgrading, update an installed Companion through Home → Computer access; updating Runtime alone does not replace a running Companion. Older approval events cannot become executable saved continuations. Inspect any interrupted action’s effects before retrying. Desktop control needs acceptance checks on your actual OS, permissions, applications, and provider; automated tests do not establish support on every desktop.
Using named environments with existing state in an inherited runtime directory? Assign that populated state to the intended environment before relying on the new isolation layout. ABot stops and asks for an explicit assignment rather than moving it or opening an empty environment silently.