Changelog & Releases

// RELEASE_LOG

[21 RELEASES DOCUMENTED]

v0.10.0

OCT 14 2026

Operations dashboard in Command Center, smarter network discovery, and reliability fixes across scans, commands, and ticket ingestion.

Added

Added an Operations section to the Usage page with at-a-glance counts for jobs, tasks, tickets, and schedules, credential usage, and activity breakdowns by agent, host, specialty, operational rule, and gateway.

Improved

Devices that share a single IP address behind a NAT gateway (for example, several VMs forwarded through one Proxmox or VMware host) are now identified and imported as separate hosts instead of being merged together or flagged as ambiguous.

Improved

Automatic scans now authenticate hosts that reuse a single SSH key across different login accounts. Previously, once a shared key had been tried under one username and failed, it was never retried under another, so a host reachable only as (for example) root rather than ubuntu was left uncharacterized. Discovery now pairs each key with every candidate username, so these hosts are recognized and characterized.

Fixed

The Command Center AI assistant no longer occasionally writes an action out as text instead of running it; its actions now execute reliably.

Fixed

Fixed commands that would hang indefinitely when they silently waited on hidden input (an interactive prompt or an unanswered confirmation). They now fail immediately and return whatever output was produced before the block.

Fixed

Fixed environment variables passed to remote commands being silently dropped instead of delivered to the target host.

Fixed

Fixed ServiceNow ticket ingestion picking up tickets from a similarly named environment (for example, pulling "demo-pre-prod" tickets into a "pre-prod" queue).

Fixed

Fixed network scans through a jump host missing live hosts. Unreachable addresses in a scanned subnet no longer tie up the shared connection pool, so every responsive host, and its open ports, is found reliably.

Fixed

Fixed the credential-lockout safeguard tripping too early on hosts that expose several services behind one address (for example, multiple VMs forwarded through a single Proxmox or NAT host). A locked-out account is now counted per host rather than per connection attempt, so legitimate multi-service hosts are no longer skipped.

Fixed

Fixed recon stalling on hosts that expose multiple endpoints; each endpoint is now time-bounded, so one slow or unresponsive service no longer holds up, or drops, the host's characterization.

v0.9.1

SEP 30, 2026

Read-only mode exceptions, clearer Command Center errors, and credential security hardening.

Added

Read-only (investigate) mode exceptions: define a short plain-language allow-list per organization or tenant-wide, and read-only agents allow exactly those ambiguous-but-safe operations (logins, dry-runs, session-scoped settings) while anything that changes the managed system stays blocked. Admin-managed.

Improved

Agents are now told about active read-only exceptions while investigating, so they attempt the operations you have allowed instead of assuming they are forbidden.

Improved

Clearer Command Center errors: form errors appear inline next to the relevant field, pages show a proper error state instead of failing silently, and every error carries a request id you can reference with support.

Fixed

Credentials imported through the hosts CSV import are now stored encrypted like every other source; previously-unencrypted rows are flagged for review.

Fixed

The CLI no longer prints database passwords when echoing deployment commands during install or upgrade.

Fixed

Each ServiceNow assignment group can now have only one active gateway, preventing duplicate ingestion, duplicate work, and conflicting replies on a ticket.

Fixed

Tickets that name a specific host now reliably run with that host's agent, and operator routing rules are honored when the agent is selected.

Fixed

A single stuck ticket no longer blocks processing for everyone: a hung handler is timed out and retried on a later cycle instead of freezing the queue.

Fixed

Tickets whose work is already finished are no longer repeatedly re-processed while they wait to be closed in ServiceNow.

Fixed

Ticket write-back (status, comment, escalation overrides) now accepts a "no change needed" decision instead of failing the job.

Fixed

Steering instructions sent to a running task are followed more consistently.

v0.9.0

SEP 16, 2026

Interactive infrastructure map, automatic network discovery, and smarter job completion.

Added

Infrastructure Map: the Command Center home screen is now a zoomable, pannable canvas where hosts, agents, and subnets appear as cards, with relationships (including jump host chains) drawn as connections. Zoom out for subnet groupings and health halos; zoom in for real-time host and agent status.

Added

A context menu on any map card surfaces quick actions (create/edit a host or agent on the canvas, open job history, start a scan), plus a side panel showing the live Tasks and Jobs feed. Filter by status or search by name without leaving the map.

Added

Network Discovery: 2501 can scan your network and build host inventory automatically from the new Discovery page. Specify CIDRs and, optionally, a jump host to reach subnets that aren't directly accessible.

Added

Discovery runs an ephemeral container per subnet, authenticates to each discovered host with your stored credentials, and characterizes responders (OS, services, routes, technology stack) as tags. It can pivot through a bridging host to scan subnets behind it within the same scan.

Added

Discovered nodes appear in the infrastructure map as a review panel: confirm to promote to a managed host (with derived tags) or ignore. Re-scans are idempotent, linking newly found IPs to existing hosts instead of creating duplicates.

Added

Automatic job reopen on ticket state change: when a human moves a resolved ServiceNow incident back to New or In Progress, 2501 reopens the job and re-investigates without a new @2501 mention, treating prior work as historical context rather than proof of completion.

Improved

More accurate job completion: 2501 now consistently accounts for follow-up comments and scheduled work when deciding whether a job is fully resolved, instead of only in certain cases.

Improved

Read-only (investigate) mode is more precise: session-scoped and read-state commands (login/logout, refreshing or switching credentials, dry-runs like terraform plan or --dry-run, read-only queries) are allowed while system-changing commands stay blocked. Previously some safe commands, such as logout, were wrongly blocked.

Improved

Agents now receive full specialty descriptions (not just names) when planning and assigning tasks, improving routing accuracy where a specialty name alone was ambiguous.

Fixed

Mid-run steering instructions are no longer silently discarded; operator steering messages sent while an agent is running are now delivered reliably.

Fixed

The "View Job in Command Center" link in the Job Execution Plan now opens the job detail page instead of an unreachable URL.

Fixed

Stalled AI model responses now re-route to a healthy provider in 2 minutes instead of waiting up to 7 minutes before failing.

Fixed

The Jobs list keeps your current page after you open a job and return, instead of resetting to the first page.

Fixed

The Tasks table status column no longer changes width as tasks update, so the table stops shifting.

Fixed

Network discovery recon now works on Kubernetes-based deployments, characterizing hosts correctly in-cluster instead of failing with a misleading "transport error".

Fixed

Jump-host authentication now recognizes an SSH private key whether stored as a private-key credential or a plain secret, so jump hosts keep working across environments.

Fixed

"Test connection" for jump-fronted hosts now tunnels through the host's single- or multi-hop jump chain instead of dialing the target directly, so it no longer times out.

v0.8.0

SEP 02, 2026

Agentic gateway prompt overrides and multi-hop SSH jump host chains.

Added

Agentic gateway overrides: gateways now support separate inbound and outbound prompt overrides, each with its own enable toggle (both off by default, so existing behavior is unchanged).

Added

With the inbound override enabled, a skip decision can trigger escalation (reassignment to an assignment group) without leaving a comment, and skipped tickets get a new "Skipped" status that shows in the ticket list and as a filter.

Added

With the outbound override enabled, the agent controls write-back on job completion: ServiceNow ticket status (with close code), comment type (public note or work note), and assignment group.

Added

Both gateway prompts now have full version history: view previous versions and restore any of them directly from gateway settings. CLI import/export now includes both prompts and their enable flags.

Added

Multi-hop SSH jump hosts: connections can route through an ordered chain of up to 5 jump hosts instead of a single bastion, configured with a new chain editor in host settings (add, remove, reorder hops). Each hop authenticates with its own stored credentials, isolated from the hops above and below.

Improved

Host tags (OS, shell, type, technology) are now available on the host creation form, matching the fields already available when editing a host.

Improved

Credential role labels for dual-protocol credentials now correctly show "SSH/WinRM" instead of "SSH only".

Improved

The default list page size increased from 10 to 25, and the Models table Actions column is now pinned to the right so it stays visible without horizontal scrolling. (The Archives page has been removed.)

Fixed

Clearing a host's public or private IP address now saves correctly: setting the field empty removes the value.

Fixed

The agent credential list now shows credential names instead of abbreviated IDs when the credential cache is empty.

Fixed

The host agent count badge no longer includes archived agents in its total.

Fixed

Knowledge ingestion on large host inventories no longer times out; extraction completes reliably regardless of document size.

Fixed

Agents now retry on the fallback host designated in the gateway routing prompt when the primary host is unreachable at job start.

Fixed

Gateway inbound routing rules are no longer applied to task generation when the inbound override is disabled.

Fixed

ServiceNow ticket processing no longer hangs indefinitely when an API response stalls; requests now have a bounded deadline.

Fixed

The Emergency Stop button on the Jobs list is no longer visible to auditor (read-only) users.

Fixed

Remote-execution credentials now appear under Remote Execution on the agent form, separate from agent-accessible credentials.

v0.7.3

AUG 19, 2026

Fix for Investigate mode incorrectly blocking read-only sign-in commands.

fixed

Agents in Investigate (read-only) mode were blocked from signing in to the systems they were asked to inspect — authentication commands such as `oc login` or a cloud provider CLI login were wrongly treated as modifying operations; these sign-in commands only establish read access and change nothing on the target, so they are now allowed, while commands that actually modify a system remain blocked

v0.7.2

AUG 12, 2026

Gateway and organization management from the Command Center, per-agent credential access control, NTLM-encrypted WinRM over HTTP, and configurable provider auth headers.

added

Gateways can now be created and archived from Settings > Gateways (ServiceNow and Runner types), with routing-prompt editing, deduplication-window config, and raw metadata input; existing gateways gain an editable metadata field

added

Organizations can now be created from Settings > Organizations in the Command Center — previously CLI or direct DB access only

added

Per-agent credential access control: each agent has an explicit allowlist of credentials it may reference by name; only allowlisted credentials are advertised to the agent and secrets are never exposed

improved

WinRM connections to Windows hosts now use NTLM message encryption over plain HTTP, so default Windows server configs work without weakening security; TLS connections continue as before and the connection test uses the secured path

added

OpenAI-compatible and Anthropic-compatible providers now support a configurable auth header (e.g. `api-key`), enabling compatibility with API proxy gateways such as Azure API Management

improved

AI model providers can now be scoped to a specific organization in multi-org environments, and host knowledge is now isolated per organization

improved

Host MDX files support jump-host fields (jump_host, is_jump_host, jump_host_credential, jump_host_user_credential, eligible_subnets), applied before the hosts that reference them; YAML parse errors now show file path, line, column, a caret pointer, and hints

fixed

Auditor-role users seeing stop/delete controls for job schedules and chat conversations — hidden for read-only roles, with the delete endpoint guarded server-side

fixed

The Investigate mode badge not appearing in the task detail header

fixed

The ServiceNow ticket processing queue becoming stuck when a credential vault read hung — reads are now bounded and fall back gracefully on timeout

fixed

Renaming an organization to a name already in use being silently accepted — the update now returns an error

fixed

Kubernetes task failures only reporting the pod phase — the error now includes the specific reason (image pull error, scheduling constraint, etc.)

fixed

The infrastructure graph showing through the main page background

v0.7.1

JUL 29, 2026

ServiceNow change requests, platform configuration as code, multi-file knowledge uploads, and Kubernetes-native executors.

added

ServiceNow change requests are now supported alongside incidents: approved change requests in the Implement state are picked up automatically, executed per the implementation plan, then moved to Review with a close code and implementation notes

added

Platform configuration as code: new `2501 resources sync` and `2501 resources pull` commands manage all platform resources (agents, hosts, specialties, operational rules, credentials, gateways, blacklist) as versioned MDX files; sync generates a diff-style plan and applies changes in dependency-safe order, and pull exports a true mirror so pull-then-sync is drift-free (with per-org `--org` scoping)

improved

Knowledge base: multiple files can be selected and uploaded at once; large batches are queued and processed sequentially to prevent timeouts

improved

Operational rules that apply to multiple task types (e.g. debug and investigate) now match tasks correctly for all listed types; tag vocabulary expanded with new type and technology tags

added

Kubernetes/OpenShift: executors now run as Kubernetes pods instead of Docker containers

improved

Command Center shows the tenant name in the sidebar footer, token analytics display all usage including historical records with no linked ticket, and license task limits are enforced from both UI and API

improved

Model management is now fully driven by the catalog in Settings > Models; per-provider feature toggles removed, with existing tenant and agent configs preserved on upgrade

fixed

Gateway routing restrictions being bypassed — tickets outside the active gateway's scope are now dropped without posting a comment or changing status

fixed

Intermediate task progress and error notes being posted as public ticket comments — only the final outcome is now visible to requesters

fixed

Tasks being created on incorrect hosts when multiple host groups were evaluated in the same routing pass

fixed

Scope-only constraint rules (no procedure type) not being injected for tasks that had a defined procedure type

fixed

Task resolution time displaying the full queue wait instead of actual execution duration

fixed

The webhook create command not offering change request as a supported event type — the event type is now validated before any changes are written

improved

Restored deduplication of repeated lines in tool output, inadvertently dropped in a previous release

v0.6.3

JUL 15, 2026

Further hardening on top of 0.6.2, focused on specialty safety, ticket status/owner integrity, host matching, and gateway scope enforcement.

improved

Specialties now expose how many agents and rules reference them, so the config sync tool can safely refuse to archive a specialty still in use

fixed

Tickets that cannot be matched to any host or agent now keep their original status and owner unchanged

improved

Error and failure notifications during job processing are now private; only successful final resolution summaries remain public

fixed

A bot mention in a ticket is no longer misread as a host identifier — tickets naming a machine absent from inventory now correctly return no match instead of routing to an unrelated host

fixed

Gateway scope restrictions are now reliably enforced by a dedicated check before task planning; out-of-scope tickets are silently skipped with no comment and no status/owner change

v0.6.2

JUL 08, 2026

Incremental release building on 0.6.1's Knowledge work, with fixes to operational-rule matching, license-limit enforcement, and knowledge retrieval accuracy.

fixed

Blanket operational rules (those without a procedure tag, e.g. connection-method restrictions) are now always applied to every task; previously they were silently dropped

fixed

Gateway routing rules are now honored for unreachable-host alerts, so agents route to management or jump hosts as configured

fixed

Creating tasks from the UI or API now enforces the license task limit and errors when the limit is reached; previously the cap applied only to tasks auto-created from tickets

fixed

Operational rules covering multiple procedure types now match against any of their tags, not just the first

improved

Remediate tasks now receive ranked, relevant rules at startup based on the task description, so agents see applicable guidance from the first turn

improved

Knowledge retrieval now weights rule names more heavily than descriptions and raises the relevance threshold, reducing false positives

added

Tag picker gains new type categories (hypervisor, control-plane, generic-storage, iot, vault) and technology categories (aws, gcp, azure, proxmox, podman, rabbitmq, hashicorp-vault)

improved

The vault technology tag is renamed to hashicorp-vault, with existing records updated automatically on deploy

v0.6.1

JUL 01, 2026

Knowledge — turn runbooks into agent context — plus a redesigned tag picker, per-model thinking-effort control and health tests, richer job schedules, and license management in the Command Center.

added

Knowledge: upload documents (PDF, markdown, text) to the new Knowledge section and the platform auto-extracts operational rules, host facts, and procedures; agents receive relevant knowledge automatically, matched to host tags and the task description

added

Each task now shows a Retrieval panel listing which rules and facts were injected and which were skipped, with reasons; the Knowledge detail page shows step-by-step processing status with tabs for extracted rules, blacklist entries, host knowledge, and pending facts

improved

Redesigned tag picker for hosts and operational rules with separate selectors per axis (OS, Shell, Type, Technology, Procedures), each with icons and inline search; host detail page gains a dedicated Tagging section

added

Control Thinking Effort per model (Off, Minimal, Low, Medium, High, xHigh) to tune reasoning depth and cost; run a one-click health and performance test per model showing reachability, time-to-first-token, and throughput

added

Support for Azure OpenAI APIM gateways: configure deployment-based endpoints with a custom API version directly in the provider dialog

improved

Job Schedules now show a human-readable Frequency column (e.g. "Every day at 2 AM") instead of raw cron, add a status filter and run-count column, and a detail stats bar (total/succeeded/failed/avg duration)

improved

Command Center chat assistant can now look up knowledge documents and job schedules, using name and tag filters plus follow-up detail lookups

added

License management: administrators can view active license details and apply new license keys directly in the Command Center

improved

Agent execution deduplicates repeated lines in command output before the agent processes them, cutting context noise significantly on log-heavy commands (up to 94% on real-world fixtures)

fixed

ServiceNow: when no hosts match a job, the ticket now gets a comment explaining the outcome and returns to Open instead of staying stuck In Progress

fixed

Analytics Tokens view now shows all LLM usage regardless of ticket association, and task/job LLM usage is correctly linked to its ticket so the Cost view attributes usage accurately

fixed

Task resolution time no longer includes queue wait — it now reflects the actual execution window

fixed

Agents on hosts with no knowledge facts no longer receive unrelated org-wide rules; the Retrieval panel now renders whenever a knowledge trace exists

v0.5.4

JUN 17, 2026

Windows gMSA authentication, HIPAA idle auto-logoff, a per-tenant LLM catalog, vision support for multimodal models, and flexible LLM credential sources.

added

Windows authentication via gMSA (Group Managed Service Accounts): credentials are resolved on demand from Active Directory over LDAPS to open WinRM sessions, so no password is stored or rotated by hand — configure once and reuse across every allowed Windows host

added

HIPAA idle auto-logoff: the Command Center now logs operators out after 30 minutes of inactivity, with timers resetting on interaction, to meet workstation security requirements without changing workflows

added

Per-tenant LLM catalog: each organization manages its own LLM providers and models from the Command Center — add a provider, register models, and pick which are available to agents, all without touching config files or restarting; the catalog is read live, so switching providers needs no redeploy

added

Vision support: agents can now process images on supported providers via the Vercel AI SDK driver path, forwarding screenshots and diagrams to the model alongside text

improved

Flexible LLM credential sources: provider API keys can be sourced from environment variables or the credential store, per provider; the Vercel AI SDK driver path is now the default for new environments, with the legacy driver as fallback

fixed

LLM errors from the new SDK driver path now surface their real cause in the Command Center instead of a generic failure message

fixed

Stuck recurring jobs now cancel cleanly when force-cleared instead of leaving in-flight work hanging

fixed

The schedule detector no longer copies cadence wording into the job description, avoiding duplicated text in generated recurring jobs (ISS-1033)

improved

Recurring job minute intervals are now floored to a minimum (5 min in production, 2 elsewhere) to prevent runaway scheduling from aggressive natural-language phrasing

improved

Self-update reliability: the engine seeds its LLM catalog in a single short transaction and prefetches existing entries to avoid lock contention on startup

v0.5.3

JUN 03, 2026

Maintenance release on top of 0.5.2, focused on engine memory-leak fixes, stuck-ticket reconciliation, and default log rotation for self-hosted deployments.

fixed

A memory leak in the engine where stuck outbound requests during periodic ticket polling caused steady memory growth — calls now have a default timeout so resources are released even when an upstream system stalls

fixed

Lingering memory growth from background calls the previous timeout fixes did not cover — the platform now applies a default timeout to all outbound work and cancels stalled background tasks cleanly between cycles

fixed

Tickets remaining in a pending state when their parent job had already completed or failed — stuck tickets are now reconciled automatically every few minutes

improved

Log rotation is now applied by default on self-hosted deployments to prevent log files from filling the host disk

improved

Benchmark scenario validation relaxed to reduce false failures on otherwise valid runs

v0.5.2

MAY 20, 2026

Everything in 0.5.1 plus global Cmd+K search, recurring job schedules, RTMS ticket polling fixes, consistent timezone handling, and broad reliability improvements.

added

Global server-side search across all Command Center list pages (tasks, jobs, tickets, agents, gateways, plugins, blacklist, operational rules, users, archives), with a Cmd+K / Ctrl+K shortcut to search across resource types from anywhere

added

Recurring job schedules: the platform detects schedule expressions in natural language and creates recurring jobs on the specified cadence; a new Job Schedules section lists and inspects them

added

Client infrastructure context injection and task steering (as in 0.5.1)

added

Conversational AI chat assistant in the Command Center and 2501 chat CLI command

improved

Timezone handling is now consistent across the platform, so recurring jobs and timestamps respect the organization's IANA timezone

improved

RTMS polling now processes the most recent tickets first, is properly scoped per organization in multi-tenant deployments, and recovers more quickly from interrupted runs

improved

Shell command output is sanitized consistently across SSH, WinRM, and local execution, so logs and follow-up steps see the same content

improved

The self-update command surfaces failures from every component pull instead of stopping at the first, and includes the executor module in its summary

improved

Command timeouts are now distinguished from dropped connections, giving agents clearer, more targeted recovery steps

added

Experimental Vercel AI SDK driver path with native tool calling for additional model providers

fixed

Cancelling a task now cancels the entire parent job rather than letting remaining tasks proceed

fixed

Agents losing their planning context after the first generate cycle (ISS-995)

fixed

RTMS ticket polling could get stuck on a backlog of older tickets and miss newly opened ones

fixed

RTMS ticket pulling could surface tickets from the wrong organization in multi-tenant deployments

fixed

Specialty search returning incomplete results with large numbers of specialties (ISS-1003)

fixed

WinRM requests hanging indefinitely against unreachable hosts (ISS-1004)

fixed

Job failures not surfaced via completion notification

fixed

Agents not reconnecting to their execution host after a disconnect

fixed

Switching between organizations now properly resets context so stale data from the previous org no longer appears briefly

fixed

Benchmark runs no longer fail intermittently from race conditions during parallel scenario execution

v0.5.1

MAY 06, 2026

Client infrastructure context injection, task steering, targeted operational rules, better reliability on unreachable hosts, an AI chat assistant, and benchmark dashboard improvements.

added

Client infrastructure context injection: the platform reads your runbooks, knowledge bases, and operational notes and auto-populates each host with relevant tools, network quirks, credentials, and procedures — agents also check each host's live state at task start

added

Task steering: send a follow-up message to a running task from the Command Center to adjust the agent's objective (picked up between tool calls); completed, failed, or cancelled standalone tasks can be resumed with a new message

improved

Operational Rules can now target specific hosts and agents via match conditions (host name patterns, connection type, agent specialty); categorized as constraint or procedure, with tags; the task detail page shows which rules matched and which were skipped and why

improved

SSH and WinRM connection errors are now classified by type (host unreachable, auth failure, disk full, dropped connection, etc.), and tasks targeting unreachable Windows hosts fail fast; connection errors can auto-escalate to a Central host admin agent

added

New conversational AI chat assistant in the Command Center (and a 2501 chat CLI command) to investigate tasks/jobs/tickets and make configuration changes; write actions require explicit approval, reads run automatically

improved

Benchmark dashboard: real-time scenario progress, redesigned dashboard with pass rates and period-over-period deltas, filtering by time range/tag/engine, plus increased stability and configurability

improved

All table rows and navigation links now support cmd+click / ctrl+click to open in a new tab

fixed

Cancelling a task from the Command Center now cancels the entire parent job rather than letting remaining tasks proceed

fixed

Agents losing their planning context after the first generate cycle (ISS-995), causing tasks to execute without their initial plan

fixed

Tickets not marked as failed when the parent job workflow fails

fixed

Specialty search returning incomplete results with large numbers of specialties (ISS-1003)

fixed

Tasks on the job detail page appearing in arbitrary order instead of creation order

fixed

Job failures not surfaced via completion notification

fixed

WinRM requests hanging indefinitely against unreachable hosts (ISS-1004)

fixed

Agents not reconnecting to their execution host after a disconnect

fixed

Cancelled standalone tasks not resuming when a new steering message arrives

fixed

Benchmarks trends chart crashing when a failed or incomplete run is present — invalid entries are now skipped

v0.4.2

APR 15, 2026

Bug fixes for job reopen LLM anchoring, executor image pulling, and investigate-notice wording.

fixed

When a user reopened a completed job with a new @2501 comment, task creation anchored on the original ticket description instead of the new comment — the new comment is now the primary LLM input, with the original description as background context

fixed

executor-mcp image not pulled from ECR at runtime, causing task execution failures

fixed

Investigate-notice wording was identical for fully-investigate and mixed-mode jobs — the notice now varies by mode mix

v0.4.1

APR 08, 2026

CloudTemple AI model drivers, more accurate Investigate badges, and stop-first engine deployments.

added

Added CloudTemple drivers: Nemotron, Devstral, and Qwen3-Coder are now available as model options

improved

The Investigate badge on tickets and jobs is now derived from the latest task mode rather than the job-level mode, for a more accurate view when specialties override the requested mode

improved

Engine deployments now use a stop-first strategy, fully stopping the old container before the new one starts, preventing port conflicts and stale connections during rolling updates

fixed

Ticket status was not reset to pending on job reopen, causing the Command Center to show stale ticket state

v0.4.0

APR 01, 2026

Durable execution (resumable jobs and tasks), Investigate read-only mode, ServiceNow webhook ingestion, real-time Command Center updates, and expanded agent capabilities.

added

Durable execution: jobs and tasks now survive restarts, automatically resuming from the last completed step; graceful shutdown drains active work before stopping

added

Investigate Mode (read-only): agents support Investigate (observation/diagnosis) and Remediate (default) modes, requestable via @2501:investigate / @2501:remediate tags; specialties can be locked to Investigate Only

added

ServiceNow webhook ingestion: instances can push tickets in real-time via unique webhook URLs, with polling kept as a safety net

improved

Job and task status changes are pushed to the Command Center instantly, replacing 5-second polling; task detail streams new agent messages live

added

New tickets page with live status, description, linked jobs and comments; new organization detail page (editable name/description, activity stats, members, API keys); settings reorganized into tabs

added

Agents can now run tools inside Docker containers and Kubernetes pods, in addition to directly on the host

added

New LLM model available: Nvidia Nemotron 3 Super 120B (via OpenRouter)

improved

Selector dropdowns support server-side search; blacklist patterns support glob wildcards with live preview; pages load faster with skeleton placeholders and instant cached data

improved

Agents report the specific reason when no task is created (no hosts, no matching agents, task limit reached); timed-out commands are reported as errors so the agent can retry

improved

Only one task runs per host at a time, preventing conflicting operations; duplicate ticket detection can be disabled when not needed

added

Benchmark Runner: new validate command for scenario definitions/results, new sandbox command for debug environments, and validation now covers all tool calls

fixed

Agent edit form not showing current configuration values

fixed

Task detail page showing stale status when navigating back

fixed

Bot comments on tickets triggering unwanted job reopens

fixed

Agents silently failing on empty LLM responses instead of retrying

fixed

Command verifier producing too many false positives on safe commands

fixed

Loop detection triggering false positives on valid check-then-act workflows

fixed

SSH workspace path using local temp directory instead of /tmp on remote hosts

fixed

Host search not working in the create agent dialog

fixed

Tasks not appearing in the agent panel

fixed

Ticket comments rendering malformed HTML

v0.3.2

MAR 18, 2026

ITSM ticket management in Command Center, real-time SSE updates, Elasticsearch full-text search, and graceful engine shutdown.

added

View and manage ITSM tickets directly from the Command Center, scoped per organization with full permission controls

improved

Command Center now uses Server-Sent Events (SSE) backed by PostgreSQL LISTEN/NOTIFY — tasks and jobs update live without polling

improved

Search in Command Center is now powered by Elasticsearch for faster, more accurate results

improved

Engine now shuts down gracefully on SIGTERM: active jobs are drained before exit, preventing tasks left in transient states during restarts

fixed

Command Center search not being scoped to the current organization

fixed

Start task not working in certain conditions

fixed

Host deletion now also removes its associated agents and tasks

fixed

Worktree names with slashes causing incorrect directory nesting

fixed

SSE connection stability issues on Bun (unhandled promise rejections on disconnect)

fixed

Command Center form validation errors appearing before first submit

fixed

Job processor querying non-terminal jobs inefficiently at the database level

v0.3.0

MAR 04, 2026

Multi-organization support, task deduplication, VMware plugin, Command Center UI revamp, System Monitor, and agent connectivity testing. ECR tag: release-0.3.x-20126cb2

added

Multi-Organization: users can now belong to multiple organizations with new role-based access control (CASL-based ACL) and Auditor role

added

Automatic task deduplication from ITSMs with configurable per-ITSM deduplication settings

added

New MCP-based VMware plugin

improved

Command Center UI revamp: redesigned pages, clickable table rows, better resource linking, updated color scheme and charts

added

New System Monitor page with memory monitoring and historical metrics

added

Test agent SSH/WinRM connectivity to hosts directly from the agent page in Command Center

fixed

Fixed token count display in Command Center

fixed

Fixed plugin executor crash when feature is disabled

v0.2.6

FEB 17, 2026

User-facing improvements including new ticket comment tracking, faster pickup, FSx ONTAP support, and new AI models. Major backend hardening across ticketing, multi-tenant data model, job lifecycle, and infrastructure. ECR image tag: 721-5ae0d660

improved

2501 now picks up new comments on tickets it has already worked on

improved

Faster ticket pickup by 2501

added

Added FSx ONTAP / NetApp ONTAP platform support

added

New models: Qwen Coder Next (80B) and GLM 5

improved

Improved Agent / Job / Task search in Command Center

This is some text inside of a div block.

improved

Increased password complexity requirements for Command Center

improved

Renamed Configuration to Specialty in Command Center

fixed

Several issues in 2501 MCPs integration

improved

Migrated and hardened ticketing layer (new ticket schema, dedicated ticketing package, ServiceNow client refactor)

improved

Improved job lifecycle handling (follow-up jobs, job restart logic, status model cleanup)

improved

Major worktree and local dev overhaul (isolated environments, config handling, CLI tooling, Postgres fixes)

improved

Strengthened multi-tenant data model and ACL enforcement (required tenant_id/org_id, Prisma + ESLint rules)

fixed

Stabilized and fixed DB migrations and seeding mechanisms

improved

Infra hardening and deployment reliability improvements (HTTPS, deploy preflight, Terraform host updates)

improved

Benchmarking and scenario testing rework (BenchV3, scenario runner, metrics model refactor)

improved

Monorepo and codebase cleanup (dead code removal, dependency bumps, linting, package extraction)

improved

Improved logging and Elastic tooling integration

v0.2.5

JAN 29, 2026

User-facing improvements including task cancellation display, new AI model support (GLM 4.7, MiniMax M2.1, Devstral 2), and various UI/UX enhancements. Backend migration of cli-agent to Bun and HTTPS everywhere. ECR image tag: 595-cde1c958

added

Display when a task is cancelled by the user

improved

Removed crontab dependency

added

Added support for GLM 4.7, MiniMax M2.1, and Devstral 2

improved

Speed up task and job pickup by agents after creation

improved

Prune Docker images during deployment

improved

Various UI/UX improvements in Command Center

improved

Various Infra-CLI improvements and bugfixes

added

Integrated AWS Bedrock + Azure drivers for Sopra Steria

improved

Finalized migration of cli-agent to the monorepo and to Bun

improved

Enable HTTPS everywhere

v4.1.3

JAN 10, 2026

Hotfix release addressing critical authentication edge case and improving SDK stability.

fixed

Critical auth token refresh failure during SSO federation handoff

improved

Python SDK error messages now include actionable remediation steps

fixed

Race condition in concurrent pipeline execution resolved

$ subscribe --feed=changelog

NEVER MISS AN UPDATE

Get notified when we ship new features, fixes, and improvements.

SUBSCRIBE_TO_UPDATES

VIEW_FEATURES

// AVERAGE RESPONSE TIME: < 24H — TRANSPARENT BY DEFAULT

CHANGELOG

NEVER MISS AN UPDATE