Changelog

What shipped, release by release.

Newest first. Limits are kept with the section they apply to, and upgrade notes with the release that needs them.

release

v2.2.0

Sophon 2.2 records every workflow run and lets you build, test and re-run workflows in the Dashboard without the agent, replaces Thinking with Effort, gives the in-chat CLI pickers, themes and image paste, lets a device’s owner manage Sophon Node from the Dashboard, adds six Dashboard themes and agent packages, and removes Discussions.

Workflow runs

Mobile has no run viewer, run list or debug tools yet, and its Run still waits on the synchronous route, giving up after 15 seconds on a long or approval run while the run carries on. Runs in flight at a restart are marked Interrupted and are not resumed.

  • Durable runs: every run and every step attempt is stored in the database, listed per workflow and streamed live to the run’s user and the person who started it, so a run that went wrong can be read afterwards.
  • Run viewer: the Dashboard draws a run on the canvas, step by step, with each step’s input, output and effect and why a skipped step did not run; Run now starts the run in the background and opens the viewer.
  • Start, follow and cancel: POST /api/workflows/{id}/runs answers 202 with a run id, /api/workflow-runs lists and reads runs, and a run can be cancelled; the synchronous /execute stays and now also returns the runId.
  • v2.1 history carried over: history files from v2.1 are imported once, a minute after the first start, as finished runs marked Legacy. Legacy runs have no inputs, step timings or definition snapshot, so the viewer draws the current workflow.
  • Retention: finished runs are pruned after 30 days, keeping the newest 20 and at most 500 per workflow, adjustable with SOPHON__Workflows__Runs__RetentionDays, MinRunsPerWorkflow and MaxRunsPerWorkflow. Stored records are redacted by setting name and pattern, and outputs and inputs over 64 KiB are cut.
  • Run caps: on one Gateway at most 8 runs of one owner and 32 of one tenant execute at once and the rest wait as Pending; past 64 per owner or 256 per tenant a new run is refused with 429 too_many_runs.
  • Cost per step: LLM Prompt steps record their model, tokens and cost, and API Call response bodies are capped at 1 MB by default (maxResponseBytes, up to 10 MB).
  • Notifications that fire: outbound workflow.completed and workflow.failed webhooks are now published, and Mobile gets a push when an unattended run fails. There is no success push.

Workflow builder

The builder has no field yet for settings.maxConcurrency, a Loop’s continueOnIterationError or maxResponseBytes; set those in the workflow JSON. Browser Back and Forward can still leave unsaved changes without asking.

  • Readable step ids: every card shows its id, new steps are named after their kind (prompt, prompt_2, branch), labelling a fresh step renames it to match, and renaming an id rewrites every reference and edge after saying how many.
  • Reference autocomplete: typing {{ in an expression field lists the earlier steps, the trigger, the run input, variables and functions, Enter or Tab inserts the entry, and a step can be picked on the canvas.
  • Edges and switches you can edit: an edge-condition editor, switch cases added and renamed on the switch itself (a rename rewrites its edges), the trigger drawn on the canvas, and a docked properties panel that no longer covers the canvas.
  • Problems before Save: the Gateway’s issues show on the steps, Save is blocked on a reference to the step itself, a later step or a step without .output, and a skill the workflow sandbox preset blocks is marked Blocked.
  • Unsaved changes ask: Run with unsaved edits offers Save and run, Run saved version or Cancel, leaving the builder asks before discarding, and Undo also covers setting edits, moves and deletes.
  • Templates: email-summary is repaired, the shipped templates label their steps with readable ids, and a new Web Digest template fetches a page every day at 09:00 UTC, summarizes it and sends the digest to WebChat.

Workflow debugging

Re-run uses the workflow’s current definition and re-runs a loop whole. Runs imported from v2.1, step tests, sub-workflow runs and runs of a workflow whose owner changed cannot be re-run.

  • Test step: run one step with mock outputs, a recent run’s data or the pinned sample, and see its resolved inputs and any reference that did not resolve. A step that can change something outside Sophon (Send Message, file writes, code, an API Call other than GET, HEAD or OPTIONS) asks for a second click first.
  • Re-run and Re-run from here: start a finished run again from its start or from a chosen step, keeping the earlier steps that succeeded and have not changed since; a plan shows which steps are kept and which run again, and repeating a step with an effect asks first.
  • Pinned sample input: one sample per workflow, typed or taken from a run, at most 64 KB, for Run with sample and for step tests. Only the owner and Admins see it, and it is stored in plaintext in the workflow file.
  • Each step records its effect (read-only, idempotent, effectful or interactive) and the variables it wrote, shown in the run viewer.

Workflow engine and triggers

Loop iterations run one at a time and loops cannot nest. Inbound webhook registrations have no replay protection, and /webhooks is outside the Gateway’s rate limiter, so keep it behind TLS and rate-limit it at the reverse proxy.

  • Joins and skips: a step fed by both branches of an if/else now runs once, when every incoming connection is decided and at least one was taken, a step that does not run shows as skipped with the reason, and logic.merge outputs only the inputs that were taken.
  • Loops that loop: a Loop step repeats the steps on its body handle once per item or while its condition holds, then continues on done; a list longer than maxIterations fails the loop.
  • Opt-in concurrency: a workflow with a Parallel step or settings.maxConcurrency runs up to that many steps at once (1 to 8, 4 by default, capped by the host); every other workflow still runs one step at a time.
  • Strict references: in a workflow created in 2.2, an action step whose reference does not resolve fails with “Unresolved reference” instead of sending the literal text; wrap optional values in default(…). Workflows created earlier are unchanged.
  • Timeouts by step type: approvals wait 30 minutes, questions 5 minutes and a Wait step its own duration, a run is bounded at 24 hours, and a failed run’s error names the step that failed.
  • Event triggers fire: channel_message and email_received start runs as the workflow’s owner on channels that owner owns in the same tenant, and a signed request to an inbound webhook registration (Bearer, HMAC-SHA256 or a custom header) starts a webhook run. Trigger types that cannot fire are refused on save and left out of GET /api/workflows/trigger-types.
  • Approvals reach the owner: a workflow’s approval and question cards go only to the owner and same-tenant Admins, name the step that raised them and link to the run.

Workflows from the command line

  • sophon workflows validate <file> prints the issues a save would answer, one per line, and exits 1 when one is an error; nothing is saved.
  • sophon workflows runs, run and cancel-run list a workflow’s runs, show one run with each step’s inputs and output, and stop a run.
  • sophon workflows execute starts the run, follows it to the end and exits by its outcome: 0 succeeded, 1 failed, cancelled or interrupted, and 130 on Ctrl+C, when the run keeps going. --async returns the run id at once.

Effort

Effort sets what Sophon asks each provider for; how much a model reasons at a level is up to the model. Azure OpenAI gets only the sampling change with Sophon’s built-in api-version, plugin providers get no level, and Mobile’s Effort chip applies only to messages sent from that screen.

  • Effort replaces Thinking: Auto, Low, Medium, High, Extra high and Max replace Auto, Off, Fast and Full. A level is clamped to what the model accepts, levels a model does not accept are greyed out, and a model with no effort control says so.
  • Auto: picks Low, Medium or High for each turn from its kind and content, without an extra model call, and never above Sophon:Models:EffortAutoCeiling (default high). Unattended turns start at Low, interactive chat at Medium, and coding turns in a bound workspace and researcher sub-agents at High; background calls such as titles and summaries get a fixed level, and a workflow’s LLM Prompt step keeps the model’s default.
  • Set it like a model: effort is set per chat, agent, user and host in the same layers as model routing, stored with the conversation on the Gateway so every device sees it, and every reply shows the level it ran at and why.
  • Where to set it: the Effort select in the Dashboard’s model picker, /effort and -e/--effort on sophon, sophon chat and sophon exec, and an Effort chip on Mobile.
  • What each provider receives: Claude gets adaptive thinking with an effort level, also with tools, with its thinking sent back inside a tool loop; OpenAI gets reasoning_effort, or reasoning.effort on the Responses API for GPT-5.6, GPT-6 and the -pro and -codex models; Gemini gets thinking_level or a thinking budget; Bedrock, Copilot, Codex and the Claude Code CLI get their own fields; and OpenAI-compatible providers get the field their model’s catalog row names.
  • Custom endpoints opt in: a custom OpenAI-compatible server, LM Studio or a vLLM server sends no effort field until its provider entry in models.json sets "effortWire".
  • The reply ceiling follows the level: max_tokens is 16K up to Medium, 32K at High and 64K at Extra high and Max on streaming turns, never above the model’s own limit.
  • Anthropic SDK: the Anthropic provider now uses Anthropic’s official .NET SDK on the same API version, 2023-06-01, and its requests carry that SDK’s own User-Agent and X-Stainless headers.

Sophon Code CLI

Allow all tools is off unless the host turns on Sophon:ApprovalPolicy:AllowAllMode. It covers only interactive turns in a trusted, bound workspace, needs the Admin role on Pro and Enterprise, and ends for good in a session once it reads untrusted content.

  • Pickers: a bare /agent, /resume, /permissions, /mode, /trust, /export and /effort, and on the classic and pinned chat /model and /save, open a selection dialog instead of printing or needing an argument.
  • Fewer commands: ten duplicated or unneeded commands are removed (/provider, /quit, /select, /tokens, /info, /tasks, /approvals, /agents, /sessions and /auto-approve); for this release a removed name prints its replacement.
  • Themes: sophon-code is the new default, with sophon-code-light, sophon-code-dimmed, sophon-code-high-contrast, catppuccin-mocha and dracula added for ten in all, and /theme and the saved theme now work on every surface.
  • Composer keys: Ctrl or Alt with Left and Right move by word, Ctrl+Backspace and Alt+Backspace delete the previous word, Ctrl+Delete and Alt+D the next, Ctrl+A, Ctrl+E and Ctrl+K work on the current line, and an emoji or accented letter counts as one character.
  • Image paste: Ctrl+V or Alt+V pastes a clipboard image into the composer (Alt+V in Windows Terminal), dragging an image file in attaches it, and attached images now reach vision models as pixels, up to Sophon:Documents:InlineImageMaxBytes (5 MiB by default).
  • Full-screen layout: an input band, a context line with the folder, branch, uncommitted changes, pull request, context and cost, and a footer with the permission mode and approval policy; ? on an empty composer opens the key help, and a new session opens on a Getting started welcome.
  • Allow all tools: a host opt-in permission mode, entered from /permissions, in which tool calls that would ask run without a card for up to 120 minutes by default (Sophon:ApprovalPolicy:AllowAllTtlMinutes). Plan cards, workflow approvals, agent deletion and plugin installs still ask, and the Dashboard shows the mode as a red chip whose only action is Leave allow all.
  • Workspace git status: the context line marks uncommitted changes and, when gh is installed and signed in as the Gateway’s OS user, shows the branch’s pull request; turn the lookup off with Sophon:Workspace:PullRequestLookup=false.

Sophon Node settings and confirmations

Managing folders and confirmations from the Dashboard needs both the Gateway and Sophon Node 2.2.0 or later; renaming works with any node. Mobile cannot edit node settings, and a headless Linux machine or a Windows session-0 service cannot show the confirmation dialog.

  • Settings tab: a device’s owner renames it, manages the folders the agent may use and the actions the device asks before, from the Dashboard, and changes apply live on the node without a restart, or at its next connect if it is offline.
  • Confirmations off, for a while: the owner can turn confirmations off for 1 hour, 8 hours, 24 hours, 7 days or until turned back on; the device list shows a Confirmations off chip, both sides audit it, and the computer shows a notification.
  • Owner loosens, admins tighten: only the owner can loosen a setting, admins can tighten one or turn confirmations back on, and Sophon:Nodes:AllowConfirmationsOff=false turns the feature off for the whole Gateway.
  • A lock at the computer: sophon-node lock stops Sophon from turning confirmations off or widening folder access there, sophon-node unlock removes it, sophon-node status shows the confirmations, the lock and the settings managed from Sophon, and sophon-node settings reset returns the node to its config.json.
  • A hardened dialog: on Windows the dialog receives its text on standard input, so quotes, $(…) and backticks are shown as typed and never run as PowerShell; Enter now refuses, a Stop closes an open dialog, and mouse and keyboard commands are refused while a dialog is open.
  • Folder checks: folders managed from Sophon are checked on the node, which refuses whole drives, the home folder, hidden home folders, AppData, Library, system folders and network, device and short-name paths, with the reason shown in the Dashboard.
  • Timeouts cancel: when the agent stops waiting for a node command, the Gateway cancels it on the device and an open confirmation dialog closes, so a late approval no longer runs it.

Themes

  • Six more Dashboard themes: Sophon Code, Sophon Code Light, Sophon Code Dimmed, Sophon Code High Contrast, Catppuccin Mocha and Dracula, for eleven in all, grouped in Settings → Appearance as Sophon Core, Sophon Code, Community and Automatic (System).
  • Sophon Core names: Daylight is now Core Light, Indigo Core is Core Dark and Deep Void is Core Extra Dark on every client, with no colour change; Graphite keeps its name.
  • The CLI ids follow: core-light, core-dark and core-extra-dark, with daylight, indigo-core and deep-void accepted permanently.
  • Nothing saved changes meaning: the Dashboard keeps its stored theme values and its default stays Core Dark. The names are shared, but each client keeps its own colours.

Agent packages

A package agent stays third-party for its whole life: no standing auto-approval covers its turns, its heartbeat auto-approves up to Low, and the router never picks it on its own. Publishing agents needs the marketplace’s agent release in production.

  • A third package type: agents, made of persona files and a manifest with no code, join skills and plugins in the Marketplace, installed from the marketplace or from a .sophon-agent file.
  • One preview everywhere: the Dashboard install dialog, sophon agents install and the chat card show the same preview (persona files, the injection scan, required skills and tools, the settings the agent will get), and an install creates the agent with its heartbeat, auto-approve, planning and any model pin off.
  • Your edits survive updates: an edited agent pauses its updates until you review a three-way diff, and agents update one at a time, never through Update all.
  • Found when needed: the model can search the marketplace for an agent when it needs a specialist and asks before installing one, on a High-risk card that can be approved only in the Dashboard, after opening its full preview.
  • Withdrawn versions stop: when the installed version is revoked or the package removed, the agent stops until you keep using it or take another version, and Sophon sends one security notification.
  • Export and publish: any agent can be exported as a .sophon-agent file and published through the linked marketplace account; sophon agents search, install, updates, update, export and publish, and sophon dev new, validate, build, install and publish, understand agent projects.

Removed and deprecated

  • Discussions removed: the multi-agent debate feature is gone, with no replacement: its Dashboard and Mobile screens, the discussion.run tool, the action.discussion workflow node, sophon discuss, its push category and its webhook events. Past discussion sessions stay as ordinary chats, and <dataDir>/discussions stays on disk, unread.
  • /api/discussions/* answers 410 feature_removed, except four list routes that answer [] until v2.3.0, and the CancelDiscussion hub method and Discussion* events are gone.
  • Thinking removed: /thinking, --thinking and the cli.json defaultThinking are replaced by /effort, -e/--effort and defaultEffort.
  • CLI slash commands removed: /provider, /quit, /select, /tokens, /info, /tasks, /approvals, /agents, /sessions and /auto-approve, replaced by /model, /exit, F2, /context, the /agent and /resume pickers, /permissions and the sophon tasks and sophon approvals commands.
  • Deprecated: the WorkflowProgress hub event, removed in v2.3.0 (listen to WorkflowRunUpdated), and legacy agent.json dev projects, which sophon dev build and sophon dev publish refuse.

Security

  • Workflows scoped to tenant and owner: workflow routes, the agent’s workflow tools and live run events are tenant-scoped, non-Admins can edit, run and delete only their own workflows, and every run executes as the workflow’s owner with the owner’s credentials, whoever started it.
  • Workflow writes need a role: on Pro and Enterprise, Viewers lose workflow create, edit and run, and new users get Viewer by default.
  • Confined effects: the file node stays inside <dataDir>/workflow-files/<tenant>/<user>/ and its create no longer overwrites, Send message uses only the owner’s or a same-tenant Admin’s channels, and a sub-workflow must have the run’s owner.
  • Substitutions encoded: values put into URLs, headers, JSON bodies and skill parameters are encoded for where they land, and code steps refuse {{ }} references on webhook and channel runs.
  • Untrusted text marked: fetched and trigger text is wrapped in <external_content> markers inside prompts and flagged on approval cards, and under Block a step whose fetched output trips the prompt-injection detector fails.
  • Agents in their tenant: every /api/agents/{id} route checks the caller’s tenant, delegation and agent tools no longer reach another tenant’s agent, and file tools refuse reads and writes under <dataDir>/agents, marketplace, vault and security. Shell tools that a person approves are not covered.
  • Node files protected: filesystem commands can no longer reach ~/.sophon-node, network, device, short-name and NTFS stream paths are refused, and config.json is readable by its owner only after each write on macOS and Linux.
  • Rotate token and Revoke now take the device’s current connection out of command delivery before they answer.

Fixes

  • E-mail channel: each unread message is delivered once instead of on every poll, and once more after a restart; mail sent from the channel’s own address never starts a workflow.
  • The email-summary template now passes its summary to the send step, tests the branch for matches and searches for “action required” in the subject.
  • Workflow builder: an edge whose condition is the empty string is kept instead of becoming unconditional on the next save, and Save keeps a switched-off workflow switched off instead of switching it on.
  • Sophon Node: on macOS and Linux a stopped or timed-out command now ends its helper process and everything it started, where a stopped shell command used to keep running; "consentRequiredCommands": null loads the defaults instead of stopping the node; and sophon-node pair exits 1 when it fails.
  • sophon agents soul <id> --edit and --from-file work again.
  • CLI: on the pinned and classic chat a multi-line paste arrives as one edit instead of being sent line by line.

Upgrade notes

  • Back up the database and the data directory before upgrading: the first start of the v2.2.0 Gateway applies two migrations, AddWorkflowRuns and AddNodeSettings. They only add tables and indexes, but there is no down-migration, so the way back to 2.1.x is restoring that backup. The backup also keeps <dataDir>/discussions, which 2.2.0 no longer reads.
  • Upgrade every client: the Gateway with its Dashboard image first, then the CLI, Sophon Node and Mobile. Until a v2.1.x client is upgraded, its explicit Thinking pick, including a CLI defaultThinking of off, fast or full, overrides every stored effort, and a Dashboard or Mobile app that does not see effort.v1 shows no effort control.
  • Discussions: let running discussions finish or cancel them with sophon discuss cancel before upgrading the Dashboard and the CLI, since no client can cancel one afterwards. Replace every action.discussion workflow step (the builder’s Replace with LLM Prompt), rewrite prompts and personas that call discussion.run, and expect webhook registrations for discussion.completed or discussion.failed never to fire again.
  • Workflow runs in flight at the upgrade show as Interrupted and are not resumed: re-run them by hand, and run one Gateway per database. v2.1 history is imported once, a minute after the first start; on Pro, first delete <dataDir>/workflows/history/<id>/ for any id another tenant or owner re-created, or create history/.imported-v1.json to skip the import. Insights workflow counts change.
  • Every workflow run is now stored, about 1,440 a day for a one-minute schedule, and pruned after 30 days, keeping the newest 20 and at most 500 per workflow. To keep more, set SOPHON__Workflows__Runs__RetentionDays, MinRunsPerWorkflow or MaxRunsPerWorkflow before upgrading, and raise MaxRunningRunsPerUser on Personal if you run more than 8 at once.
  • Pro and Enterprise: Viewers lose workflow create, edit and run, and new users get Viewer by default (grant Operator, or set SOPHON__Tenant__DefaultRole=Operator). Workflows are owner-only for non-Admins, so transfer ownerless and departed users’ workflows with POST /api/workflows/{id}/owner. Every run executes as the owner, run visibility narrows to the run’s user, its requester and same-tenant Admins, and a workflow webhook hears only runs its creator may read.
  • Saves are validated: workflow ids outside the allowed characters, unknown node types, and new steps with invalid ids or references to steps that do not exist are refused. Trigger types that cannot fire (file_change, calendar_event, agent_request) never fire and are refused on new saves, so change them.
  • Edge conditions are normalized on any save, the enable switch included: True becomes true and Approved becomes approved, so an if/else or approval branch that never ran starts running. Check what such a branch does before you save. A step fed by both if/else branches now runs once, logic.merge outputs only its taken inputs, and graph exports with a re-approval gate need re-exporting.
  • Loops and branches: re-connect every Loop step (the steps to repeat on body, the next step on done), since existing Loop steps fail until re-connected; a workflow with a Parallel step now runs branches concurrently, so connect order-sensitive steps or set settings.maxConcurrency to 1; and new workflows fail an unresolved reference instead of sending literal text.
  • Workflow steps: move file-node outputs into <dataDir>/workflow-files/<tenant>/<user>/, recreate channels Send message uses under the owner or an Admin, have code steps on webhook and channel runs read SOPHON_TRIGGER_DATA, SOPHON_INPUT_DATA or SOPHON_NODE_DATA, raise maxResponseBytes for API Call bodies over 1 MB and timeoutMs on skill steps whose tool asks for approval, and check workflows you had switched off, since Save no longer switches them on.
  • Triggers and notifications: keep one enabled channel config per type across all tenants, set routing.debounceEnabled to false on an e-mail channel that starts workflows, and keep /webhooks/in behind TLS with an allowlist and a rate limit. Outbound workflow.completed and workflow.failed webhooks and failure pushes start firing: check the registrations and pause schedules that fail often.
  • Custom clients: register a handler for WorkflowRunUpdated; WorkflowProgress reaches only the run’s user and requester and is removed in v2.3.0. Accept new members such as runId and issues, handle 400 workflow_invalid, 403 not_workflow_owner, 409 effect_confirmation_required and 429 too_many_runs, read step results by nodeId, and expect sophon workflows execute to exit by the run’s outcome. /api/discussions/* answers 410 feature_removed.
  • Effort shifts cost: unattended turns and background calls now run at Low, interactive chat at Medium and coding turns at High, so models that did not reason unless asked now do on some turns, and models with a higher default reason less. Set an explicit level on the agent or host, or lower Sophon:Models:EffortAutoCeiling, to hold it down. OpenAI reasoning models no longer receive temperature or top_p, and Claude models drop temperature while they think.
  • Egress rules: OpenAI’s newest models (GPT-5.6, GPT-6, the -pro and -codex models) now call /v1/responses on api.openai.com, and Anthropic API-key requests carry the official SDK’s User-Agent and X-Stainless-* headers. A custom OpenAI-compatible endpoint sends no effort field until its models.json provider entry sets "effortWire".
  • Effort in scripts and clients: /thinking is /effort, sophon exec --thinking exits 1 (use -e/--effort), and cli.json defaultThinking is read as defaultEffort (off → Low, fast → Medium, full → High) until sophon config set defaultEffort rewrites it. Headless NDJSON and every reply’s routing gain effort and effortSource, an unknown level in SendMessage is refused with EFFORT_INVALID, and LlmRequest.Thinking is gone for in-process hooks.
  • CLI commands: /provider, /quit, /select, /tokens, /info, /tasks, /approvals, /agents, /sessions and /auto-approve are removed, so update scripts that pipe them into sophon (/provider <id> <model> becomes /model <id>:<model>). Bare /permissions, /export, /mode and /trust open dialogs; pass an argument where you expect printed output. The default theme is now sophon-code: run sophon config set theme core-dark to keep the old look.
  • CLI terminal: Ctrl+H now deletes the previous word, so set a terminal that sends ^H for Backspace to send DEL; image paste on Linux needs wl-clipboard or xclip on PATH; and the full-screen chat tints the terminal background (opt out with SOPHON_CLI_TERMINAL_BACKGROUND=0). Images attached in chat are stored base64 in session events up to 5 MiB each, so set Sophon:Documents:InlineImageMaxBytes to 0 to keep the text-only behaviour.
  • Allow all tools and git status: before turning on Sophon:ApprovalPolicy:AllowAllMode, update the CLI and the Dashboard with the Gateway, and on Pro keep the Admin role narrow. When gh is signed in as the Gateway’s OS user, the Gateway looks up pull requests with that one GitHub identity for every trusted workspace; set Sophon:Workspace:PullRequestLookup=false to stop it.
  • Sophon Node devices: the confirmation dialog’s default is now No (Reject on Linux), mouse and keyboard commands are refused while a dialog is open, and on macOS and Linux a stopped command ends everything it started. Rename a filesystemRoots key in config.json to fileSystemRoots, run chmod 600 ~/.sophon-node/config.json, and replace network, short-name (PROGRA~1) and NTFS stream paths, which are refused. A node running as a LocalSystem service reads the system profile, so lock that one.
  • Node settings from the Dashboard: an admin who does not own a device can no longer loosen its settings, app.launch with arguments is admin-only on the manual Send command route, and Sophon:Nodes:AllowConfirmationsOff=false forbids turning confirmations off. When the agent stops waiting the command is cancelled on the device, so raise timeoutSeconds for longer work; after Rotate token, put the new token in config.json and restart the node; after downgrading the Gateway, run sophon-node settings reset on a managed device.
  • Themes: the CLI saves new choices as core-light, core-dark and core-extra-dark and prints those ids. A v2.1.x CLI does not know them, so to share cli.json with one, save the old id (daylight, indigo-core or deep-void), which 2.2.0 still accepts and never rewrites.
  • Agent packages: update DevStudio to 1.1.0 or later with the CLI, migrate legacy agent.json dev projects with sophon dev new agent (build and publish refuse them), start the Gateway before sophon dev install of an agent project, and approve agent installs in the Dashboard, since chat channels, Mobile and the CLI can only reject them. Re-link the marketplace account to publish from Sophon.
  • Agents and tenants: every chat’s prompt cache misses once after the upgrade on a Gateway with the marketplace on; a workflow that named another tenant’s agent stops working, so recreate the agent in that tenant; turning on heartbeat autoApprove or allowPlanning needs ManageSkills; and an agent or workflow that edited SOUL.md through fs.write must use agent.update or the persona routes, since file tools refuse Sophon’s data folders.
Full upgrade steps

release

v2.1.2

Sophon 2.1.2 fixes every request to Claude Sonnet 5.5, Opus 5.5 and Haiku 5.5, and to the other Claude models that refuse a temperature, failing with a 400.

Fixes

On the adaptive models the reasoning trace stays empty, and a model that does not think by default is not asked to think when tools are attached; v2.2.0’s Effort changes both. Haiku 5.5 on Bedrock and OpenRouter has no catalog row, so its cost is recorded as $0.

  • Claude 5.5 works again: requests to Sonnet 5.5, Opus 5.5 and Haiku 5.5 failed with “temperature is deprecated for this model”, and so did Opus 4.7, Opus 4.8, Opus 5, Sonnet 5 and Fable 5.x whenever they were the provider’s default model.
  • No sampling or thinking budgets on adaptive models: Opus 4.7 and later, every 5.x model, Fable, Mythos and any Claude id Sophon cannot read get no temperature, and Thinking Fast and Full become adaptive thinking at effort medium and high. Older Claude models get the same request as before.
  • Models that think by default (Opus 5 and 5.5, Sonnet 5 and 5.5, Haiku 5.5, Fable and Mythos): Thinking Off sends effort low, and small background calls send effort low with 1,024 more tokens of room so a title or summary is not cut off mid-thought. Bedrock and GitHub Copilot follow the same rules.
  • Catalog: new rows for claude-opus-5-5, claude-sonnet-5-5 and claude-haiku-5-5, with 1M context and 128K output, and Bedrock rows for the Opus and Sonnet 5.5 global ids. Without a catalog row, 5.x models now get 1M context and 128K output instead of 200K and 8K.

release

v2.1.1

Sophon 2.1.1 lets the assistant create and update agents again when its model fills in every model-routing field, and reads hand-edited agent model policies without their placeholder values.

Fixes

  • agent.create and agent.update: a model that filled in every field of the tools’ modelRouting (inherit with a profile, empty pin and role fields, an allowFallback) was refused, so the assistant could not create an agent; the tools now drop what the mode never reads and list what they ignored in the result and the approval preview.
  • Allow fallback through the agent tools is kept only when the policy pins a model of its own or it matches the agent’s current value, and the localOnly profile without auto is refused with a message saying what to send instead.
  • Hand-edited agents: an agent.json whose modelRouting carries a profile without auto, a pin without pinned or empty role fields now reads without them, so the agent’s Models tab and sophon agents models can save it again, and a stray profile no longer becomes a chat’s Auto profile.
  • Linux installer: the .deb package accepts libicu78, so it installs on Ubuntu 26.04.

release

v2.1.0

Sophon 2.1 lets you choose which model handles each kind of work, gives installs without Qdrant a built-in vector store with local-first embeddings, turns the CLI into a coding agent bound to your project folder, and lets a stopped plan be resumed, retried or continued.

Model routing

Upgraded chats stay on Default, which still runs the chat on the default provider’s model as v2.0 did, with the changes listed in the upgrade notes. Mobile has no routing controls yet.

  • Default, Auto or a pinned model: each chat runs on Default (the default provider’s model, as before), on Auto, which picks a model for each job from the models you have listed, or on a model you pin, set from the model chip in the Dashboard or with /model in the CLI.
  • Five roles: every model call is routed as orchestrator (the chat turn), executor (sub-agents and plan steps that write or run), researcher (read-only sub-agents and document Q&A), utility (titles, compaction and other background chores) or vision, so a frontier model can run the conversation while cheaper or local models do the rest, and each role can be overridden on its own.
  • Auto profiles: Balanced (the default), Quality, Economy, Local-first and Local-only map each role to a model tier; Auto uses only models with tool calling for the tool roles and enough context for each role, and price breaks ties rather than leading the ranking.
  • Four scopes: set routing for one chat, an agent, yourself (“My default”) or, as an admin, the whole host, where only Auto and role settings apply and nothing can be pinned; sophon models resolve shows which layer won for each role.
  • Pins stay put: a pinned model never moves to another provider unless you turn on Allow fallback; without it, Sophon retries on the same provider and then fails the turn with a message saying it did not switch, and a pin whose provider was removed or disabled stops that chat with “model is not available” until you change it.
  • Failover that follows the error: only rate limits, overloads, server errors, timeouts, connection failures, budget stops, context overflows and, under Auto, a missing model fail over, while authentication errors, bad requests and unrecognized errors fail the turn; failover tries the same model on another provider of the same type, then an equivalent model at the same tier, never another provider’s default model, and a reply that fell back says so.
  • Local only means local: Auto · Local-only, a local pin without fallback, or the host Local only setting keeps every model call, including sub-agents and background chores, off the cloud. Memory embedding follows the host embedder, so pick a local embedder too.
  • Provider management: exactly one default provider, with Make default now beating Priority, and one default model per provider; providers can be enabled and disabled, each model’s context window, tier and roles can be edited, Discover lists the models installed on an Ollama, LM Studio or OpenAI-compatible server, and each local provider serves one request at a time by default.
  • From the CLI: /model pins a model or switches to auto, role, fallback, reset and save; -m/--model works on sophon, sophon chat and sophon exec; and sophon models gains set-default, enable, disable, discover, roles and resolve, among others. /provider remains as a legacy alias.
  • New installs: the setup wizard offers Auto, preselecting Local-first when a listed local model has tool calling and a declared context window and Balanced otherwise, with “Default (as before)” as the other choice.

Embeddings and semantic memory

Semantic recall needs an embedder: a local one (Ollama or LM Studio with an embedding model) or a cloud provider an admin has consented to; until then, search uses keywords. The embedded vector store is single-node, so multi-replica deployments need Qdrant, which is also recommended above about 100,000 vectors.

  • Built-in vector store: installs without Qdrant (Personal, native and Desktop) now get an embedded SQLite vector store in the data directory, so semantic search no longer needs a separate service.
  • Local first: Sophon looks for a local embedding model on Ollama, LM Studio or a local OpenAI-compatible server and prefers it over a cloud provider, and with Ollama configured but no embedding model you can install nomic-embed-text (about 274 MB, stays on your machine) from Settings → Models → Embeddings or with sophon models embeddings install-local.
  • Consent before cloud embedding: a cloud embedder that hasn’t produced vectors yet sends nothing until an admin consents, after seeing how many memories, daily logs and documents would be sent, and its first full index waits for a confirmed start. A local-only chat searches by keyword instead of calling a cloud embedder, but the host embedder still embeds that chat’s memories later.
  • No more vector wipes: each embedder writes to its own collections, so switching embedders never mixes or deletes vectors and switching back reuses the old index, and Sophon never switches embedders on its own (a better local option is offered as a suggestion).
  • Embedding trouble doesn’t break chat: a slow or failing embedder or vector store no longer fails chat turns, document search or memory search; after 8 seconds Sophon falls back to keyword results, and the embeddings status shows what actually works instead of a green Active.
  • Self-healing index: a background reindex builds a new embedder’s index, refills a deleted or restored vector store and catches up on memory written during an outage, on its own for a local embedder; admins get test, rebuild, and reindex with progress and cancel in the Dashboard and under sophon models embeddings, and can delete an old embedder’s collections from the Dashboard.
  • Embedding calls now appear in Usage under the source “embeddings” and count against provider budgets.

Coding agent in the CLI

Workspace binding needs a Gateway on the same machine, reached over localhost and running as your OS user; remote and Docker Gateways can’t bind. The shell is a limit, not a sandbox: a command can still leave the folder. External coding agents (Claude Code, Codex, Copilot) are confined only by their working directory, and /diff and /rewind don’t see their edits.

  • Bound to your project: run sophon in a repository and the chat binds to that folder (the nearest one holding .git), asks once whether to trust it, and in a trusted folder loads AGENTS.md (or CLAUDE.md) and sends the git branch and changed files with each turn; turn binding off with SOPHON_CLI_WORKSPACE=0 or "workspaceBinding": false in cli.json.
  • Coding tools: bound turns get file tools confined to the folder, including the new fs.edit, fs.multi_edit, fs.apply_patch (used by OpenAI-family models) and a regex fs.search, plus a shell that runs inside the folder with a filtered environment and streams its output live.
  • Every change asks: in a bound session every edit and shell command asks for approval whatever the risk threshold, unless Accept edits or an always-allow rule covers it. Edit cards show the diff and shell cards the command. Rejecting with feedback lets the agent keep going with your note; a plain reject, or a card that times out, now ends the turn.
  • Permission modes: Default (ask before edits), Accept edits and Plan (read-only), cycled with Shift+Tab in the full-screen and classic chat or set with /permissions; Accept edits covers file edits inside a trusted folder only, while shells, deletes and sensitive files such as manifests, lockfiles, CI and build files, git hooks and AGENTS.md always ask.
  • Review and undo: every change made with the file edit tools is recorded before it’s written, so /diff [session|turn|uncommitted] shows what changed and /rewind [n] restores the files after showing the diff (it does not rewind the conversation). Shell commands, deletes, moves and copies are not recorded, so /diff doesn’t show them and /rewind can’t undo them. Esc stops a turn and keeps finished work, and /init asks the agent to write AGENTS.md.
  • Plan mode: only read-only tools run, for the agent and all its sub-agents, and the agent proposes a plan you can run here, run as a graph, edit, send back with feedback or reject. Plan mode is not a network control: web search, page fetches and read-only HTTP requests still run. On Personal, only the CLI can approve a plan that leaves plan mode; the Dashboard can reject it.
  • Checklist: the agent can keep a todo checklist, shown as a band in the full-screen chat (Ctrl+T or /todos) and as a Checklist panel in the Dashboard.
  • Sub-agents you can see and stop: each sub-agent shows its current tool, last line of output, changed files, tokens and cost; /subagents lists them, /subagents watch N follows one live in the full-screen chat, /subagents stop N stops one while the parent keeps going, and the Dashboard shows a live agent tree with Stop. There is no new way to define your own sub-agents.

Headless runs and standing rules

  • Headless runs: sophon -p "<prompt>" (or sophon exec) runs one prompt with no one at the keyboard and prints text, a single JSON result or NDJSON events (--output-format stream-json), exiting 0 on success, 1 on failure, 2 when something needed approval and was denied, and 130 when interrupted.
  • Headless never approves: every approval card is rejected with a note and every question is skipped; --trust-workspace is the only way to trust the folder for a run, --allow pre-approves a tool or a command prefix such as --allow "os.execute:dotnet test", and -c/--continue resumes this repository’s last session.
  • Always allow: in a trusted folder, an approval card can offer to always allow read-only network tools (page fetches and GET requests) or a command prefix of up to three words, such as dotnet test; rules expire after 8 hours by default, a session holds at most 32, and /permissions rules and /permissions revoke list and remove them.
  • Rules stay narrow: shells, interpreters, curl, docker, git push, package installs and chained commands are never remembered, network and command rules stop applying for the rest of a session once it reads untrusted content, and MCP tool calls in a bound session always ask.

Plan recovery

Covers chat plans run by Sophon Graph; Classic plans and delegated task graphs are not included. Mobile shows a paused plan but has no Resume or Retry controls.

  • Pause instead of cancel: a graceful Gateway stop, or a turn reaching its time limit, now pauses running chat plans instead of canceling them, and Resume picks up where they stopped; a step that was running starts over and usually doesn’t use up a retry (up to two interruptions per step by default).
  • Why it stopped: plan cards in the Dashboard and the CLI say why a plan stopped (Sophon restarted, the turn time limit, the budget with the amounts, you stopped it, or a newer message replaced it) and name the next step.
  • Retry and continue: Retry failed steps re-runs only the work that didn’t finish, and a plan stopped by its budget now offers Continue with a new budget instead of ending as failed; each starts a new run that keeps every completed step, after a preview of what runs again and a warning about side effects.
  • From the CLI and the chat: /retry resumes, retries or continues depending on the plan’s state, /plan shows the session’s latest plan after you reopen it, sophon plans gains list, show, resume and retry, and you can ask the agent to retry the plan, which always asks for approval first.

Usage and cost

Cost covers this session’s own turns; sub-agents and background calls outside a turn, such as compaction and titles, are not included.

  • Context and cost as you go: the full-screen CLI status line shows how full the context is and what the session has cost so far (for example ctx 11% · $0.13), with the model, workspace and branch, and the new /context command prints the details on every surface.
  • Whole-turn totals: a turn’s token count now sums every model call in it, not just the last one, and records its cost, so session stats and Insights show higher totals than before for the same work.

Providers and subscriptions

  • Claude providers follow your Default: Anthropic, Claude subscription and Claude Code providers now run the provider’s configured Default model instead of a fixed claude-sonnet-4-6 (or, for Claude Code, the sonnet alias).
  • One Claude Code login: when the Gateway refreshes a Claude Code token, or you log in from the Dashboard, it writes the credential back without logging Claude Code out. It keeps the fields Claude Code needs, replaces the file in one step, and no longer damages the macOS Keychain entry. The Keychain path has not been tested on a Mac.
  • Azure OpenAI sends each model to its own deployment.
  • Local vision and tools: tagged local model ids such as qwen3-vl:8b now take their family’s vision and tool support from the catalog, so a local vision model can receive images directly.
  • sophon claude-code send works again and accepts --working-directory and --model, and its help now says plainly that Claude Code runs with bypassPermissions on the Gateway host unless the coding-agent mode is Container. The working directory is sent as given and is not confined by the Gateway.

Security

  • Browser origin check: browser requests to the realtime hubs, to MCP and to any API write are refused (403) unless they come from the Gateway’s own address, the built-in development origins or an origin listed in Sophon:Security:AllowedOrigins; the CLI, Sophon Node and scripts send no Origin and are unaffected.
  • Tighter Personal auto-admin: the no-login admin on Personal now also requires a localhost host name (or one listed in Sophon:Security:PersonalHosts) and no forwarding headers, which blocks DNS-rebinding pages and same-machine proxies that forward other people’s traffic with forwarding headers. A same-machine proxy that adds no forwarding headers is still not detected.
  • Ownership checks: approval decisions, task cancel and every session route now check the caller’s tenant or ownership, channel approval replies decide only their own tenant’s approvals, and sub-agent sessions are read-only.
  • Safer shells and paths: os.execute kills the whole process tree on timeout or cancel and drops secret-looking variables (provider keys, tokens, Sophon config overrides) from its environment. This is a list of known names and patterns, so a secret under another name still gets through. Path checks catch UNC and device paths after normalization, and code.build and code.test validate their arguments and run without a shell.
  • Hidden characters shown: diffs and approval previews in the CLI and the Dashboard, including voice approvals, show control, bidi and zero-width characters as visible markers, and untrusted Markdown in plans never loads remote images.
  • Sandbox presets that work: os.execute_sandboxed is now gated by the sandbox presets (a name mismatch let every preset allow it), and os.execute is denied under the standard and hardened presets.
  • Backups leave out the edit journal and the vector cache.
  • Patched dependencies: Scriban 7.2.2, MimeKit 4.16.0, SharpCompress 0.48.0, AngleSharp 1.5.0, the SignalR client 10.0.13 and System.Security.Cryptography.Pkcs 10.0.13. ImageSharp stays on 3.1.12, which has six known advisories, because newer versions need a license key.

Fixes

  • Dashboard: long chats render faster, mounting the newest 200 entries with “Show earlier” for the rest; the composer stays blocked while any approval is still open; and code blocks in replies are syntax-highlighted.
  • CLI: the full-screen chat detects truecolor on Windows Terminal and Windows consoles and maps colors to the nearest match on 256-color terminals, /rename uses the real session id, and a resumed session shows only approvals that are still live.
  • Diff tools: code.git diff emits standard unified diffs, and patch_apply checks every hunk against the file and writes nothing when one doesn’t match.
  • Planning: the check that decides whether a request needs a plan matches whole words, so “authentication” no longer counts as “then”.
  • Provider health turns to Error only after two failed checks in a row, and saving models.json on Windows retries over brief file locks, so provider changes no longer fail at random.
  • The Gateway exposes Retry-After to cross-origin clients, so the desktop app waits the right time after a rate limit.
  • Mobile: a paused plan shows as Paused instead of a stuck tracker, and the notification for a stopped plan opens its chat.

Upgrade notes

  • Back up before upgrading: the first start of the v2.1.0 Gateway applies four database migrations automatically. They only add a table for tool-call history and columns for plan retries, approval previews and why a turn stopped, with no data rewritten, but there is no automatic downgrade path.
  • Upgrade the Gateway together with its Dashboard image first, then the CLI, then Mobile; installed v2.0 Mobile builds show a paused plan as Executing.
  • Embedded vector store: installs without Qdrant now create vectors/vectors.db in the data directory. It is a cache that backups leave out and the background reindex rebuilds; set Sophon:Vectors:Provider=none to stay keyword-only. The coding edit journal in checkpoints/ (up to 1 GiB) is also left out of backups.
  • Qdrant: Qdrant:Endpoint is the gRPC address, so a URL without a port now means 6334 and an endpoint that names :6333 (the REST port) only logs a warning, so change it to :6334; credentials in the URL stop the boot (use the new Qdrant:ApiKey); the Helm chart now points at 6334. Collections are now named per embedder (memory__ and documents__ plus eight hex characters); the first start aliases your v2.0 collections and deletes nothing, but tools that read Qdrant directly must use the new names.
  • Cloud embedders: one that hasn’t written vectors yet embeds nothing until an admin consents, on new Pro and Enterprise installs too; a v2.0 Qdrant install keeps its existing embedder with no new consent step.
  • Browser origins: a Dashboard or other browser client served from a different origin gets 403 origin_not_allowed on hubs and API writes until you add it to Sophon:Security:AllowedOrigins.
  • Reverse proxies must forward the Host header with its port (nginx $http_host, not $host), and the Dashboard image must be rebuilt or re-pulled because its nginx template changed. A proxy on the same machine as a Personal Gateway needs Sophon:Proxy:Enabled with 127.0.0.1 and ::1 in KnownProxies and must send X-Forwarded-For, or every request through it gets 401. A same-machine proxy that adds no forwarding headers and passes Host: localhost through is not detected and gives every client the no-login admin, so never expose one beyond the machine.
  • Personal by host name: a browser on the Gateway’s own machine that opens it by host name or LAN IP instead of localhost no longer gets the no-login admin until that name is listed in Sophon:Security:PersonalHosts. Other devices on the LAN never got it and still don’t.
  • Model routing: upgraded chats stay on Default, but Default changed. The default provider now beats Priority and exactly one provider is the default (the first start keeps the first enabled provider by priority), authentication, bad-request and unrecognized errors no longer fail over, failover never lands on another provider’s default model, a pinned model no longer falls back unless Allow fallback is on, and Claude providers run their Default model, so check each provider’s Default model before upgrading. Background chores such as titles and compaction move to the default provider’s model tagged utility or routine when one exists, and images go to the first listed model with vision.
  • Local providers serve one request at a time (Sophon:Models:LocalMaxConcurrent, 0 turns the limit off), and provider health turns to Error only after two failed checks in a row (Sophon:HealthMonitor:FailureThreshold, 1 for the old behavior).
  • MaxRunTokens counts the whole turn: run token usage now sums every model call in a turn, so delegation graphs reach Sophon:GraphRuntime:MaxRunTokens (default 1,500,000) sooner, and session stats and Insights totals rise.
  • os.execute and the sandbox presets: os.execute now carries the process_execute capability, so workflow runs (the hardened preset) and the standard preset deny it with SANDBOX_DENIED, while cron, heartbeat and webhook turns keep it; os.execute_sandboxed is now denied under standard, automation and hardened. Adding process_execute to Sophon:Sandbox:ToolPolicies:AllowCapabilities allows it again in every preset, not only workflows.
  • Auto-approve: a person’s /auto-approve no longer covers shells, coding agents, skills or tools that send data out, and cron and webhook runs are refused in a session bound to a workspace.
  • Planning is now set with Sophon:AgentExecution:PlanningMode (default Auto) and its trigger is stricter, so plans appear less often; Always is not the v2.0 behavior, since it plans every turn, including cron, webhook and sub-agent turns.
  • Custom clients and scripts: the SignalR event SessionModelOverrideChanged is replaced by SessionModelRoutingChanged, a budget stop arrives as plan_interrupted instead of plan_failed, /api/graph-runs and document-embed errors now return an error code with a separate message, and a failed sophon memory reindex exits 1 (re-embedding is sophon memory reindex --vectors). See the API reference.
  • Before any downgrade to v2.0, set Sophon:GraphRuntime:PauseOnStop to false in config/appsettings.user.json while the Gateway runs (an environment variable needs a restart, and that restart pauses plans again), then resume or cancel every paused plan (sophon plans list --status Interrupted): v2.0 can’t read a plan paused by a restart or the turn time limit, and one such plan stops all run recovery.
  • Claude Code credentials: on macOS, if an older Gateway already broke your Keychain entry, log in once more from the Dashboard or with /login in Claude Code; in Container mode on Linux, give the runner user read access to <data dir>/config/claude/.credentials.json in a way that survives the file being replaced, such as a shared primary group or a setgid folder.
Full upgrade steps

major

v2.0.0

Sophon 2.0 rebuilds the voice engine behind the Voice page, the chat voice ribbon and the Command Bridge; Sophon Graph becomes the default way approved plans run; the CLI opens full screen; and Sophon Node gains Control Run.

Voice

Available on every tier. With no speech provider configured, voice uses your browser’s speech recognition and voice (Chrome or Edge), and the browser may send audio to its vendor; admins add speech providers in Voice settings.

  • One voice engine: the Voice page, the in-chat voice ribbon and the Command Bridge keep their familiar look but now run on one rebuilt engine, so listening, playback and interrupting follow the same rules on all three.
  • Speech that starts early (needs a text-to-speech provider): with a provider configured, Sophon starts speaking a reply sentence by sentence while it is still writing the rest, and a one-sentence reply is spoken once it’s complete; without a provider, the browser’s voice reads the finished reply.
  • Hands-free that works: Conversation mode now runs end to end (speak, pause, hear the answer), with the mic re-opening only after the reply has finished playing, and Sophon detects the pause itself, so hands-free works with any supported speech-to-text provider.
  • Voice turns are regular tasks: a voice request runs as an agent task on your chat session, queued and approval-gated the same way as a typed message and shown live in the chat thread, with a smaller, voice-focused tool set.
  • Interrupt with the mic: press the mic while Sophon is thinking or speaking to silence it and talk; the interrupted reply isn’t spoken later but still appears in the chat, and work already under way keeps running.
  • Approvals by voice follow risk: on the Voice page, the Command Bridge and Sophon Mobile, Low- and Medium-risk actions take a plain “yes”, High- and Critical-risk actions need “yes, approve”, a “yes” that also says “no”, “wait” or “don’t” is asked again, and a plain “no” rejects; prompts are read in the browser’s or phone’s built-in voice, not your provider voice.
  • Questions by voice: when an agent asks a question with options, the Voice page and the Command Bridge read it out with numbered choices, and you answer by saying the number (one to four) or the option’s name, or by tapping it; option names with symbols can be said as words, such as “C sharp”.
  • No guessing, no crosstalk: an unclear answer is asked again at most twice before the mic closes and the prompt waits for a tap, negating an option (“don’t send”) never selects it, and voice only speaks prompts that belong to its own conversation.
  • Voice that stays with the conversation: voice started from a chat talks on that chat, and switching chats ends it on the old one and restarts it, listening, on the new one instead of speaking into the old one; every open tab attached to the same voice conversation hears the reply, and after a dropped connection you hear the rest of the reply from the moment you reconnect, with the full reply in the chat.
  • Command Bridge voice bar: the hands-free bar listens once you arm it and switches itself off with the reason shown after silence, inactivity or an error, and while an approval or question is waiting the mic opens so you can answer out loud, and a separate Stop button turns voice off without answering, leaving the prompt on its chat card.
  • Provider health you can see: configured speech providers are health-checked in the background, speech provider requests have time limits so a provider that stops responding shows up as a named error instead of a stuck conversation, and when voice can’t use a provider the Voice page says whether speech isn’t set up or which provider is failing; the notice appears only when no working provider of that kind is left, and a provider is flagged only after repeated failed checks.
  • Progress lines and speakable replies (needs a text-to-speech provider): during long, multi-step tasks Sophon can speak a short progress line after a quiet stretch (a single long step can stay quiet for its whole run), and a streamed reply is spoken without Markdown formatting, with links read as their text or site name; code and tables are best read in the chat.
  • Fewer silent failures: a voice request that fails is spoken or shown instead of going quiet, the microphone is released every time you leave voice, and with server speech-to-text a slow transcript of something said earlier is much less likely to be taken as your answer or as a new request.
  • Voice from the command line: the new sophon voice status command shows the health of every configured speech provider and whether hands-free is available, sophon voice runtime start now works with --agent and --session, and sophon voice settings set no longer resets your personal listening settings.

Sophon Graph

Graph supports a single Gateway instance; Classic, the previous step-by-step engine, remains a supported choice.

  • Sophon Graph, on by default: approved multi-step plans now run as durable graphs, with steps that don’t depend on each other running in parallel, each step running in its own isolated session, and progress saved to the database as the run goes.
  • Graph or Classic, your choice: switch a single conversation between Graph and Classic, set your own default on the new Settings → Execution page, or use /mode in the CLI, while admins set parallelism and the cost cap there or turn Graph off for the whole workspace, which puts everyone on Classic.
  • Safe restarts: if the Gateway restarts mid-run, finished plan steps are never re-run, and with default settings a run that was mid-step pauses, notifies you and waits for you to resume or cancel it from the plan card in the Dashboard, where plans and their Resume and Cancel controls now come back correctly after a page reload.
  • Self-correcting plans: when a step keeps failing, Sophon retries it with a suggested alternative approach and then revises the failed part of the plan; revised plans show a version badge and wait for your approval if they raise the plan’s risk, bring in a new kind of tool or add steps that may have side effects, and with default settings every revision asks.
  • Plan budgets: Graph-mode plans run under a cost cap and a token limit, and when a run goes over, Sophon stops starting new steps once the running ones finish and tells you in the chat which limit was hit; the limit is checked between steps, so a run can overshoot by what its running steps cost, and a run stopped by its budget shows as failed in the Dashboard, with no Resume control.
  • Dependent delegation: with Graph on, tasks handed to sub-agents can depend on each other, so they run in ordered waves that pass their results forward; the cost cap doesn’t cover these waves.
  • Save a plan as a workflow: turn a completed or failed Graph-mode plan into an editable workflow in the visual builder, with a notes dialog listing anything that couldn’t carry over. It is a one-way export.

Sophon CLI

  • Full-screen terminal chat: sophon now opens a full-screen chat by default, with its own scrollback, a live status header and footer, streamed Markdown with live code highlighting, and a resume card on exit so you can always pick the session back up.
  • Approval and question cards in the terminal: approvals show risk, parameters, a scrollable preview and a live countdown, with one-key Approve, Reject, Skip, and Edit where the tool allows it; questions show numbered options and also accept a typed free-text answer, and an approval that goes unanswered is treated as rejected.
  • Dialogs and selection: /theme previews the whole screen live before you commit, /resume opens a session picker, /plan shows the full plan as an overlay while a progress row keeps the running step visible, Shift-drag or F2 selects transcript text, and typing / lists every command that works in the current view.
  • Token usage: each reply in the full-screen chat ends with a usage line (agent, model, tokens), the footer keeps the running total and /tokens shows this session’s breakdown, and nothing is shown until the provider reports usage.
  • Automatic fallback: small windows, redirected output and the Dashboard’s built-in terminal get the pinned-input or classic prompt, an active Windows screen reader (or SOPHON_CLI_ACCESSIBLE=1) gets the linear classic prompt, SOPHON_CLI_FULLSCREEN=0 turns full screen off, and sophon doctor --reset-terminal recovers a stuck terminal.

Sophon Node

Desktop actions run through a paired Sophon Node and the approval gate; activity or screenshots the agent reads are sent to your model provider for that turn and stay in the conversation.

  • Control Run: the agent drives a paired desktop in short batches of actions that stay on one machine and stop at the first failure, and a batch whose actions all succeed ends with a fresh screenshot in the Dashboard’s live panel.
  • Screenshots to the model, on Claude only: with Claude models reached through the Anthropic API or a directly connected Claude subscription, the latest screenshot from each desktop step goes to the model as an image, with the screen geometry it needs to map clicks back to real pixels; other providers, including the Claude Code CLI, Bedrock and Copilot, get a text description.
  • Plan-level desktop consent: in an interactive chat, approving a desktop action also covers later desktop actions up to that risk level in the same conversation for a limited time; that consent never covers Critical actions and a Gateway restart clears it, shell commands always ask again, and workflows, scheduled runs and MCP calls ask every time.
  • Stop and fail fast: the Agent Drive panel in Dashboard chat now has a Stop control that ends the agent’s turn so no further actions reach the machine (an action already running may finish first), and a run aimed at a machine that has stopped checking in fails at once with a “machine is offline” message.
  • Windows UI grounding: on Windows the agent can read the real buttons, text boxes and menu items inside windows, with their names and on-screen positions, so it can find the control it needs instead of guessing from pixels; password fields are never returned, and macOS and Linux keep their existing, more limited support.
  • File access on your own machine: with a separate opt-in permission, the agent can list, read and write text files only inside folders you allow (Desktop, Documents and Downloads by default), binary files are refused, each read or write is size-capped, and writing a file is a high-risk action that needs your approval (in an interactive chat, one approval can cover later writes in the same conversation for a limited time).
  • Activity history (opt-in): a device you enable records which app and window you’re working in to a local database on that machine, skipping common password managers and private-browsing windows (matched by app name and window title, both editable) before anything is written, and once the device has the separate activity-history permission (not granted by default) and is online, the Dashboard lets you browse it by day, filter it, delete single entries and change recording settings without restarting the node.
  • Activity screenshots (separate opt-in): the device can save a screenshot as an ordinary image file each time a new activity entry starts, by default kept for less time than the text history and within a disk cap, with Compact, Balanced and Sharp quality presets; when you ask what you were working on, a Claude model can look at the saved screenshot, while other providers see only the text timeline.

Approvals

  • One prompt, before the action: tools that need approval now ask once, before they run, instead of twice, and a rejected action is reported back to the agent as rejected (a shell command on a paired desktop can still show two prompts at the default threshold).
  • Approvals on messaging channels: approval requests raised in a conversation on a channel such as Telegram or WhatsApp are now sent to that channel, where you answer with “approve <id>” or “reject <id>”, and replies are accepted only from the chat the request was delivered to, though anyone in that chat can answer; after a Gateway restart, a pending approval is checked against the channel’s default chat.
  • Plan-step and sub-agent requests in your main chat: approvals and questions raised inside a plan step or by a sub-agent now appear in your main conversation in the Dashboard, Mobile and the CLI and in the messaging channel it came from, marked with the step they came from; these requests wait in memory, so a Gateway restart fails that step attempt, an approval left unanswered is rejected, and an unanswered question lets the step continue on its best judgment.
  • Live approval policy: threshold and per-tool changes saved in the Dashboard take effect immediately with no Gateway restart, security settings saves are atomic, and the approval quiet-hours controls, which were never enforced, are removed.
  • Resolved everywhere, recorded once: answer an approval by voice, on a chat card, in another tab or in the mobile app and any other open chat card or voice prompt for it closes on its own, cards on the Dashboard and Mobile now say Approved, Rejected, Cancelled or Timed out, and a timed-out or cancelled approval can never later become an approval.

Prompt-injection guard

The guard matches known instruction phrasing; it is a filter, not a security boundary.

  • Off, Warn or Block: an admin chooses the prompt-injection mode in the Dashboard under Settings → Security, Warn is the default, and Block stops a turn only when third-party content (a tool result, a webhook payload, or a channel message from a group chat or from someone other than you) matches a high-severity pattern, never for your own direct messages or memory.
  • Injection notices: a detection appears as a card in the conversation on Dashboard, CLI and Mobile, stays after a reload, is recorded in the audit log, and alerts you when it stops a scheduled run.
  • Blocked steps stay stopped: a plan step the guard stops fails with its explanation and is not retried or re-planned around, and a stopped sub-agent reports a failure instead of passing its refusal on as a finished result.

Marketplace

  • Updates panel: check installed packages for updates, update one, or update all from Settings → Marketplace.
  • Automatic updates (opt-in, off by default): when an admin turns them on, the regular update check also installs new package versions through the same verified install path, and rolls back if an update fails.

Models and cost

  • Expanded model catalog: Settings → Models now covers 181 models from 18 provider types, showing each exact model ID (copyable) and its aliases, with a provider filter, and models the catalog lists as vision-capable are now used for image understanding and document OCR.
  • Cost tracking for more models: usage of models newly added to the catalog, previously recorded at $0, now carries an estimated cost based on list prices, so spending budgets count it and may trigger where they did not before.

Chat, sessions and operations

  • Long conversations stay current: the model’s context is now built from the most recent part of a long session instead of its oldest messages, and a repeated short reply like “yes” no longer causes the current message to be skipped.
  • Faster back-to-back turns: a turn now completes as soon as the answer is delivered, memory indexing runs afterwards off the reply path so your next message isn’t held up, and stuck tasks are marked as lost automatically.
  • Unified chat cards: plans, approvals, questions, live desktop and browser panels now share one design across all four themes, risk badges show the real level, and a locked session can no longer be toggled by a stray click.
  • Agent activity groups: plan-step and sub-agent sessions now sit in an expandable “Agent activity” group under their conversation in the Dashboard sidebar and update live, failed or interrupted plan-step attempts are archived out of the group (kept for audit), and deleting the conversation deletes its activity too.
  • Sturdier storage and diagnostics: the SQLite database now uses write-ahead logging with a busy timeout so reads no longer block behind writes, each conversation event gets a unique sequence number even when events are written at the same time, and the Dashboard’s runtime, log and doctor pages no longer fail on Windows while the Gateway writes its log.
  • Metrics for operators: Graph run counts (started, completed, paused and resumed, plus replans and budget stops) and, when a text-to-speech provider speaks a reply, time to first audio and per-provider text-to-speech latency and failures are exported on the Gateway’s authenticated Prometheus metrics endpoint.

Developer tools

  • Scriptable CLI: global flags such as --json, --format and --gateway-url can now follow any subcommand, sophon dev validate --json and sophon dev build --json each print a single JSON document including the built package’s exact file list, and failing sophon dev commands now exit with a non-zero code.
  • Safer packages: sophon dev build and sophon dev install no longer include .env files (including .env.example), a .venv folder, or editor folders such as .vscode and .idea, and packages built on Windows now install on macOS and Linux.
  • Working skill templates and Marketplace pre-checks: sophon dev new skill now creates Python and C# skills that run in the sandbox without changes and come with tests, and sophon dev validate warns before you upload about skill manifest problems the Marketplace’s publishing rules flag, such as a missing author, license, README or LICENSE file.

Upgrade notes

  • Back up before upgrading: the first start of the v2.0.0 Gateway applies five database migrations automatically, and there is no automatic downgrade path.
  • Sophon Graph is on by default: to keep the previous step-by-step engine, choose Classic per conversation, as your default in Settings → Execution or with /mode in the CLI, or have an admin turn Graph off for the workspace.
  • The CLI opens full screen by default: set SOPHON_CLI_FULLSCREEN=0, or "fullScreen": false in cli.json, to keep the pinned-input chat.
  • Prompt-injection detection now defaults to Warn, which shows a notice card in the conversation; installs that had turned detection off stay off.
  • Spending budgets may trigger where they did not before, because usage of newly cataloged models is no longer recorded at $0.
  • Skills that declare the deprecated “proxy” credential mode now receive the stored credential inside their sandbox, the same as “injected”.
  • Node voice routes return 404 until an operator turns them on, and once on, the wake route requires the node’s voice permission, is rate-limited per node and rejects over-long transcripts.
  • Paired devices: the agent can reach only devices owned by the requesting user, launching, closing or focusing an app now asks for approval, the ungated single-command passthrough tool is removed, and desktop command results use camelCase fields with the app-launch parameter named “name”.
  • Custom realtime clients: the voice status event now only means a reply finished (it is no longer sent when speech-to-text starts), transcripts carry an utterance id, and approvals now broadcast a resolved event with outcomes approved, rejected, timed out or cancelled. See the realtime API reference.
  • APIs: the session list returns only top-level conversations by default (request view=all for the previous flat list), and deciding an approval that is already decided returns the recorded outcome instead of overwriting it, while unrecognized outcome values are rejected.
  • Failing sophon dev commands now exit with a non-zero code, so scripts and CI no longer treat a broken project as a success.
Full upgrade steps

release

v1.15.0

  • Sophon Plugins: extend the platform with out-of-process plugins, adding new messaging channels and document formats as local processes, hot-started and stopped from Settings → Plugins with health checks and automatic restart (off by default, admin allowlist)
  • Memory Graph v2: a force-directed graph explorer with agent colors, kind filters, date windows, and a co-mention overlay, plus Mermaid diagram export from the Dashboard, chat, agent tools, and API
  • Memory that maintains itself: scheduled consolidation promotes daily-log facts into long-term memory, superseded facts drop out of default views, and near-duplicate entities can be reviewed and merged from the Dashboard
  • Ask-the-user v2: agents ask structured multi-question forms with tappable options on Dashboard, CLI, and Mobile, and numbered-reply questions on messaging channels; timeouts are explicit, never a silently assumed answer
  • Plan approval modes: interactive sessions still gate multi-step plans, while heartbeat, cron, webhook, and sub-agent runs proceed automatically, configurable per session or host-wide
  • Documents v2: grounded Q&A with citations. Ask a single document or your whole library and get answers that cite their sources, or an honest "no answer" when the documents don't contain one
  • Rich in-app document viewer for PDFs, spreadsheets, DOCX, images, audio, Markdown, Mermaid, and code, plus save-by-URL ingestion and automatic ingestion of files sent over WhatsApp, Telegram, and Slack
  • Version-preserving document updates with per-version download, an async processing pipeline with a durable queue, and OCR for scanned PDFs
  • Claude subscription providers: run Sophon on your Claude Pro/Max plan, with direct chat (anthropic-subscription), text-only via the Claude Code CLI (claude-code), or full tool-using agent runs with no per-token cost (claude-code-agent)
  • Canvas v3: per-frame version history with a stepper, a frame toolbar with width presets, and library documents that open directly on the canvas
  • Backup from the Dashboard: create, monitor, and delete full data-directory backups with live progress
  • MCP per-user identity: MCP tokens are now bound to a Sophon user, with per-user memory, document, and agent resources; regenerate tokens bound to a real user after upgrading
  • Parallel delegation: agent.delegate can fan out 2–5 sub-agent tasks in one call, and its model and budget overrides now apply

release

v1.14.0

  • Sophon Marketplace: browse, search, and install community skills and plugins straight from the Dashboard (/skills) and CLI, every install human-approved with an up-front permission preview
  • Verified installs: marketplace downloads are SHA-256-checked, safely extracted, and hot-registered with no restart; plugins unload and update cleanly
  • Device linking and reviews: link a Sophon instance to your marketplace account and leave ratings and written reviews from the Dashboard
  • Leaner bundle: Jira, Confluence, and Gmail skills now ship via the marketplace instead of being bundled (google → google-gmail)
  • Refreshed model catalog: added the Claude 5 family, GPT-5.6 / 5.4, and Gemini 3.x
  • Memory: keyword search across stored memories (portable across databases), a vector scope contract, and memory-write for scoped sub-agents
  • Agent rebinding: switch the agent on an existing conversation from the Dashboard picker or the CLI /agent command; the session adopts it on resume and switch
  • Codex reasoning replay: encrypted provider reasoning is captured and replayed across agentic-loop iterations, with prompt-cache keys for cheaper multi-turn runs
  • More accurate usage and cost accounting: cache-token counting across Codex, Claude Code CLI, Bedrock Converse, Gemini implicit cache, and Copilot

release

v1.11.0

  • Heartbeat autonomy: agents run standing checklists on a schedule with noop suppression, failure backoff, and a Medium-risk auto-approve ceiling
  • Real-time speech-to-text: streaming server-side transcription with live interim results on Dashboard and Mobile (Deepgram, OpenAI, Azure, Google, ElevenLabs, Sophon Managed Speech)
  • Model catalog: capability and pricing matrix for 20+ bundled models, hidden-model controls, and per-session model selection via Dashboard picker or CLI /provider
  • Group chat controls: per-channel policies (mention-only, allowlist, all-messages, disabled) with mention, reply, quote, and thread triggers across six platforms
  • Thread-aware sessions: Slack threads, Discord forum posts, Telegram topics, and more each get an isolated session with automatic expiry and cleanup
  • Inbound message debouncing: rapid-fire messages coalesce into a single agent turn, with per-channel tuning
  • Reply and quote context: quoted messages are captured and shown to the agent on Telegram, Discord, Slack, Mattermost, and Matrix
  • Media to text: inbound images, audio, and video are automatically transcribed or captioned so any model can work with them
  • PowerPoint generation: build complete themed .pptx decks in one call, with 10 layouts, 6 themes, speaker notes, and deck editing operations
  • Mobile Material 3 redesign: Daylight, Indigo Core, and AMOLED themes, Material You dynamic color (Android 12+), edge-to-edge layout, predictive back
  • Mobile chat reliability: per-session stream buffers, mid-task reconnection with stream reconciliation, and approval recovery
  • CLI v2 by default: pinned input zone, inline slash and @-mention pickers, live subagent status band, and four terminal themes
  • New CLI chat commands: /provider (per-session model override), /thinking, /auto-approve, /canvas; thinking and auto-approve settings persist across sessions
  • Session-scoped realtime events: per-session badges and approvals across Dashboard, Mobile, CLI, and Desktop, with no more cross-device noise
  • Hardened multi-tenant realtime isolation: all live events are tenant-scoped on shared gateways
  • Production hardening: automatic credential migration into the vault, channel config rollback on failure, SSRF protections, and configurable rate limits

launch

v1.0.0

  • Initial release of Sophon
  • Multi-agent runtime with LLM-based routing
  • 12+ channel adapters (WhatsApp, Telegram, Slack, Discord, etc.)
  • Visual workflow builder with 20+ node types
  • Memory system with agent-scoped isolation
  • Document processing pipeline (PDF, DOCX, XLSX, images, audio)
  • Secure code sandbox (Docker + gVisor, process fallback)
  • Credential vault with OAuth 2.1 + PKCE
  • Human approval gates with risk classification
  • React 19 Dashboard with 18+ pages
  • Electron desktop app (Windows, macOS, Linux)
  • Three tiers: Personal (free), Pro, Enterprise
  • Enterprise features: SSO, RBAC, audit logs, multi-tenancy

beta

v0.9.0-beta

  • Agent-scoped memory with database storage
  • Channel file attachment support
  • Process sandbox fallback for non-Docker environments
  • Workflow expression engine with recursive descent parser
  • Dynamic service connections with manifest-driven configuration

alpha

v0.8.0-alpha

  • Core agent runtime and session management
  • Basic channel adapters (WebChat, Telegram)
  • Memory engine with short-term and long-term storage
  • Skill engine with bundled skills
  • SQLite database for Personal tier