An agentic browser harness: an AI side panel & orb in your own logged-in Chrome, Edge, Arc, Brave or even Firefox run by the agent/subscription you already use (Claude Code, Codex, Antigravity). Free, open source, no API key. Also an MCP server.
TypeScript agentic-browseragentic-browsingagy-cliaiai-agentai-browser-extensionai-toolsbrave-extensionbrowser-agentbrowser-automationbrowser-extensionbrowser-usebrowserskillchrome-extensionclaude-codecodexfirefox-addonmcp-servernanobrowserweb-automation
docs: DeepSeek Harness (beta) is in every list of agents, with how it signs in, keeps its own dsh home, is kept to the browser and is refused to an older extension, and the Mac app looks for all nine
Shahriarfb3eaac08d58
diff --git a/.github/ISSUE_TEMPLATE/bug_report.yml b/.github/ISSUE_TEMPLATE/bug_report.ymlindex 159f952..9a8676b 100644--- a/.github/ISSUE_TEMPLATE/bug_report.yml+++ b/.github/ISSUE_TEMPLATE/bug_report.yml@@ -42,7 +42,7 @@ body:attributes:label: Agentdescription: Which agent CLI runs your side panel, if the bug involves a run.- placeholder: claude / codex / agy / vibe / grok / cursor-agent / qwen / opencode, and its version+ placeholder: claude / codex / agy / vibe / grok / cursor-agent / qwen / opencode / dsh, and its version- type: inputid: environmentattributes:diff --git a/README.md b/README.mdindex cfcd9f5..80ba7e9 100644--- a/README.md+++ b/README.md@@ -38,7 +38,8 @@ Also an optional [MCP server](docs/guide/mcp-clients.md), so any MCP client can<a href="https://docs.x.ai/build/overview"><img src="docs/assets/agents/grok.svg" width="40" height="40" alt="Grok Build (beta)" title="Grok Build (beta)"></a> <a href="https://cursor.com/docs/cli/overview"><img src="docs/assets/agents/cursor.svg" width="40" height="40" alt="Cursor CLI (beta)" title="Cursor CLI (beta)"></a> <a href="https://qwenlm.github.io/qwen-code-docs/en/"><img src="docs/assets/agents/qwen.svg" width="40" height="40" alt="Qwen Code (beta)" title="Qwen Code (beta)"></a> - <a href="https://opencode.ai/docs/"><img src="docs/assets/agents/opencode.svg" width="40" height="40" alt="OpenCode (beta)" title="OpenCode (beta)"></a>+ <a href="https://opencode.ai/docs/"><img src="docs/assets/agents/opencode.svg" width="40" height="40" alt="OpenCode (beta)" title="OpenCode (beta)"></a> + <a href="https://deepseek-harness.github.io/deepseek-harness/"><img src="docs/assets/agents/deepseek.svg" width="40" height="40" alt="DeepSeek Harness (beta)" title="DeepSeek Harness (beta)"></a></p></div>diff --git a/SECURITY.md b/SECURITY.mdindex 8194d64..7d5d02f 100644--- a/SECURITY.md+++ b/SECURITY.md@@ -52,8 +52,8 @@ other liability arising from the software or its use. In plain language, that methe action; deciding is yours. Actions taken in your browser (purchases, submissions,messages, deletions) are actions taken on your accounts, by you.- **The agent is not part of this project.** Browsentic spawns whichever agent CLI you chose- (Claude Code, Codex, Antigravity, Mistral Vibe, Grok Build, Cursor CLI, Qwen Code or OpenCode)- and contains each as well as its own flags allow, but their behaviour, their terms, and what+ (Claude Code, Codex, Antigravity, Mistral Vibe, Grok Build, Cursor CLI, Qwen Code, OpenCode or+ DeepSeek Harness) and contains each as well as its own flags allow, but their behaviour, their terms, and whatthey do with your prompts are governed by their vendors, not by this repository.- **Automating a website may breach that website's terms of service.** Whether a site permitsautomated interaction is between you and the site.diff --git a/docs/README.md b/docs/README.mdindex 30dedc2..0edd5b2 100644--- a/docs/README.md+++ b/docs/README.md@@ -19,7 +19,7 @@ For running Browsentic on your own machine.| [Pair](guide/pair.md) | Connecting a browser to Browsentic Bridge with a single-use code || [First run](guide/first-run.md) | A tour of the side panel and your first instruction || [Features](guide/features/) | One page per capability: what it does and when to use it |-| [Choosing an agent](guide/agents.md) | The eight agent CLIs the side panel can run on |+| [Choosing an agent](guide/agents.md) | The nine agent CLIs the side panel can run on || [MCP clients](guide/mcp-clients.md) | Optional: drive the same browser from Claude Code, Cursor, Zed or Gemini CLI || [Configuration](guide/configuration.md) | Every key in `~/.browsentic/config.json` || [Approvals](guide/approvals.md) | Which actions wait for your approval, and how to change that |diff --git a/docs/assets/agents/deepseek.svg b/docs/assets/agents/deepseek.svgnew file mode 100644index 0000000..5a42558--- /dev/null+++ b/docs/assets/agents/deepseek.svg@@ -0,0 +1 @@+<svg xmlns="http://www.w3.org/2000/svg" viewBox="0 0 64 64" width="64" height="64"><rect x="0.5" y="0.5" width="63" height="63" rx="14" fill="#24292f" stroke="#3d444d"/><svg x="14" y="14" width="36" height="36" viewBox="0 -8 28 28"><path fill="#ffffff" d="M0 0H5L8 3V9L5 12H0ZM2 2V10H4.2L6 8.2V3.8L4.2 2ZM10 0H18V2H12V5H18V12H10V10H16V7H10ZM20 0H22V5H26V0H28V12H26V7H22V12H20Z"/></svg></svg>diff --git a/docs/guide/README.md b/docs/guide/README.mdindex efc8708..cb2d85e 100644--- a/docs/guide/README.md+++ b/docs/guide/README.md@@ -32,7 +32,7 @@ Three short steps, about five minutes in total, with no account, no API key and| | || --- | --- |-| [Choosing an agent](agents.md) | The eight agent CLIs the side panel can run on: switching, and what each needs |+| [Choosing an agent](agents.md) | The nine agent CLIs the side panel can run on: switching, and what each needs || [MCP clients](mcp-clients.md) | Optional: registering Browsentic with Claude Code, Cursor, Zed, Codex or Gemini CLI || [Configuration](configuration.md) | Every key in `~/.browsentic/config.json` || [Approvals](approvals.md) | Which actions wait for your approval, and how to tune that |diff --git a/docs/guide/agents.md b/docs/guide/agents.mdindex d620bed..70513ec 100644--- a/docs/guide/agents.md+++ b/docs/guide/agents.md@@ -1,7 +1,8 @@# Choosing an agent-The side panel runs an agent CLI you are already signed in to. Eight are supported: Claude Code,-Codex and Antigravity, plus Mistral Vibe, Grok Build, Cursor CLI, Qwen Code and OpenCode in beta.+The side panel runs an agent CLI you are already signed in to. Nine are supported: Claude Code,+Codex and Antigravity, plus Mistral Vibe, Grok Build, Cursor CLI, Qwen Code, OpenCode and DeepSeek+Harness in beta.This page covers picking one, how each is contained, and each one's setup and quirks.It applies to the side panel only. Driving Browsentic from another tool goes through the optional@@ -31,9 +32,10 @@ Where a CLI can list the models your account has, the select offers that list:| Cursor CLI | `cursor-agent models`; there are over 200, so the select becomes a filter, with Browsentic's picks first || OpenCode 2 | `opencode api --standalone model.list` on Browsentic's own OpenCode database, leaving out models that cannot call a tool and OpenCode Zen's free ones ([why](#opencode-is-in-beta)) |-Claude Code, Mistral Vibe, Qwen Code and OpenCode 1 have no such command, so the select offers a-short list shipped with Browsentic. For Claude Code that list is its aliases (`fable`, `opus`,-`sonnet`, `haiku`), which it resolves to the newest model of each family.+Claude Code, Mistral Vibe, Qwen Code, OpenCode 1 and DeepSeek Harness have no such command, so the+select offers a short list shipped with Browsentic. For Claude Code that list is its aliases (`fable`,+`opus`, `sonnet`, `haiku`), which it resolves to the newest model of each family. For DeepSeek+Harness it is `deepseek-flash` and `deepseek-v4-pro`.The Bridge reads each list in the background, so the agent picker never waits on one. A list iskept for six hours and read again at once when the CLI is updated, or for OpenCode 2 when its@@ -59,16 +61,16 @@ cannot resume each other's sessions, so the next instruction starts a new one.## Supported agents-| | Claude Code | Codex | Antigravity | Mistral Vibe (beta) | Grok Build (beta) | Cursor CLI (beta) | Qwen Code (beta) | OpenCode (beta) |-| --- | --- | --- | --- | --- | --- | --- | --- | --- |-| Vendor | Anthropic | OpenAI | Google | Mistral AI | xAI | Anysphere | Alibaba | Anomaly |-| Binary | `claude` | `codex` | `agy` | `vibe` | `grok` | `cursor-agent` | `qwen` | `opencode` |-| Install | `npm i -g @anthropic-ai/claude-code` | `npm i -g @openai/codex` | [antigravity.google/docs/cli/install](https://antigravity.google/docs/cli/install) | `uv tool install mistral-vibe` | `curl -fsSL https://x.ai/cli/install.sh \| bash` | `curl https://cursor.com/install -fsS \| bash` | `npm i -g @qwen-code/qwen-code` | `npm i -g opencode-ai` |-| Default model | `sonnet`, the newest Sonnet | the CLI's own | the CLI's own | the CLI's own | the CLI's own | the CLI's own | the CLI's own | the CLI's own |-| Effort names | `low`…`max` | `low`…`xhigh` | `low`…`high` | none; set `thinking` in Vibe's own config | `low`…`xhigh` | none; put it in the model id, e.g. `claude-opus-4-8[effort=high]` | none; the model id is the only lever | a variant the model defines, e.g. `high` or `max` |-| Kept off your machine by | a per-run tool allowlist plus an explicit deny list | its shell and image viewer switched off, no MCP server but the browser's, and a read-only sandbox (`sandbox_mode="read-only"`) | its own permission rules | a per-run tool allowlist; its shell and file tools are never loaded | a per-run tool list, approvals that refuse anything not granted up front, and a kernel sandbox that keeps its writes in its own folder | per-run deny rules, where a deny beats every allow; a kernel sandbox is asked for too but not depended on | `--safe-mode`, which drops every setting of your own, plus deny rules for the shell, the disk and the tools that reach either | a per-run agent whose rules deny every tool not named for it, so the model is never offered the shell, the disk or another server's tools |+| | Claude Code | Codex | Antigravity | Mistral Vibe (beta) | Grok Build (beta) | Cursor CLI (beta) | Qwen Code (beta) | OpenCode (beta) | DeepSeek Harness (beta) |+| --- | --- | --- | --- | --- | --- | --- | --- | --- | --- |+| Vendor | Anthropic | OpenAI | Google | Mistral AI | xAI | Anysphere | Alibaba | Anomaly | DeepSeek |+| Binary | `claude` | `codex` | `agy` | `vibe` | `grok` | `cursor-agent` | `qwen` | `opencode` | `dsh` |+| Install | `npm i -g @anthropic-ai/claude-code` | `npm i -g @openai/codex` | [antigravity.google/docs/cli/install](https://antigravity.google/docs/cli/install) | `uv tool install mistral-vibe` | `curl -fsSL https://x.ai/cli/install.sh \| bash` | `curl https://cursor.com/install -fsS \| bash` | `npm i -g @qwen-code/qwen-code` | `npm i -g opencode-ai` | `npm i -g @deepseek-ai/dsh` |+| Default model | `sonnet`, the newest Sonnet | the CLI's own | the CLI's own | the CLI's own | the CLI's own | the CLI's own | the CLI's own | the CLI's own | `deepseek-flash` |+| Effort names | `low`…`max` | `low`…`xhigh` | `low`…`high` | none; set `thinking` in Vibe's own config | `low`…`xhigh` | none; put it in the model id, e.g. `claude-opus-4-8[effort=high]` | none; the model id is the only lever | a variant the model defines, e.g. `high` or `max` | `off`, `low`, `high`, `max` |+| Kept off your machine by | a per-run tool allowlist plus an explicit deny list | its shell and image viewer switched off, no MCP server but the browser's, and a read-only sandbox (`sandbox_mode="read-only"`) | its own permission rules | a per-run tool allowlist; its shell and file tools are never loaded | a per-run tool list, approvals that refuse anything not granted up front, and a kernel sandbox that keeps its writes in its own folder | per-run deny rules, where a deny beats every allow; a kernel sandbox is asked for too but not depended on | `--safe-mode`, which drops every setting of your own, plus deny rules for the shell, the disk and the tools that reach either | a per-run agent whose rules deny every tool not named for it, so the model is never offered the shell, the disk or another server's tools | a per-run patch that switches off every tool of dsh's own, so the model is offered the browser's tools and a to-do list only, inside dsh's read-only sandbox |-All eight get the same system prompt, the same browser tools and the same+All nine get the same system prompt, the same browser tools and the same[approval gate](approvals.md). Codex is handed the tools directly; the others reach them through a`browsentic` MCP server pointed back at the Bridge. What differs is how well each can be fenced offfrom the rest of your machine;@@ -341,6 +343,45 @@ What both lines share:---+### DeepSeek Harness is in beta++DeepSeek Harness (`dsh`) is DeepSeek's own open-source agent harness, released as a developer+preview. Browsentic's runner was checked against `dsh` 0.2.1-rc.2 with a real DeepSeek account: a+browser tool call, a screenshot and a follow-up turn resuming. A real side-panel run has not been+through it yet. Expect rough edges, and please report them.++- **Sign in first.** Run `dsh web` and sign in with a DeepSeek account, or set `DEEPSEEK_API_KEY`.+ Until then the agent picker shows *needs setup*. Browsentic needs dsh 0.2.1 or later.+- **Runs keep a dsh home of their own.** Browsentic starts dsh in+ `~/.browsentic/agents/deepseek/home` and borrows only your sign-in file, `~/.dsh/.credentials.yaml`+ (or `$DSH_HOME/.credentials.yaml`). Your own dsh patches, your profiles' plugins, your `AGENTS.md`+ and your skills never load into a run, and Browsentic's conversations do not appear in the+ `dsh web` history. The side panel's `/` picker still lists your skills from `~/.dsh/skills` and+ `~/.agents/skills`, because it reads them from disk itself.+- **Only `DEEPSEEK_*` reaches a run from the environment.** Every other key is sealed away, and+ `$DSH_HOME/.env` is not read by a run.+- **Every tool of dsh's own is switched off.** Each run carries a patch over dsh's plugin tree that+ switches off the shell, the files, skills, sub-agents, workflows and the web (unless the run is+ [mapping a site](features/site-maps.md)), so the model is offered the browser's tools and a to-do+ list only. dsh's read-only sandbox is on as well, and its telemetry is off. If one of its own tools+ runs anyway, the run is stopped with `AGENT_UNSAFE`.+- **A dsh update can be refused.** Browsentic checks dsh's plugin tree against the patch, and a+ release that leaves on a tool plugin Browsentic has not checked is reported as unusable until+ Browsentic is updated.+- **Two models, and only one reads pictures.** The agent picker offers `deepseek-flash` (the+ default) and `deepseek-v4-pro`. `deepseek-v4-pro` is text only, so a screenshot reaches it as a+ placeholder.+- **Effort is dsh's reasoning effort:** `off`, `low`, `high` or `max`.+- **A turn is held over ACP.** Browsentic runs `dsh --profile acp` and hands the session the+ browser's MCP server. If dsh refuses ACP, the turn runs through `--profile headless` instead.+ One-shots (titles, file reports, recordings) always run headless.+- **dsh gives up on a tool call after 60 seconds.** As on Cursor, an approval you have not answered+ by then stays on screen and the agent asks again under the same card, and waits inside a call are+ cut to fit. A result of up to 25,000 tokens reaches the model whole.+- **No token counts.** dsh's reply to a turn carries none, so the context card has none to show.++---+## Per-agent settingsIn `~/.browsentic/config.json`:@@ -356,7 +397,8 @@ In `~/.browsentic/config.json`:"grok": { "bin": "grok", "model": "grok-4.7" },"cursor": { "bin": "cursor-agent", "model": "composer-2.5" },"qwen": { "bin": "qwen", "model": "qwen3-coder-plus" },- "opencode": { "bin": "opencode", "model": "anthropic/claude-sonnet-5" }+ "opencode": { "bin": "opencode", "model": "anthropic/claude-sonnet-5" },+ "deepseek": { "bin": "dsh", "model": "deepseek-flash", "effort": "high" }}}```@@ -365,7 +407,7 @@ In `~/.browsentic/config.json`:| --- | --- | --- || `agent` | `claude` | Which CLI the side panel runs on. The agent picker writes this. || `agents.<name>.bin` | the CLI's own command name | Absolute path to the binary. Set this when the Bridge's `PATH` differs from your shell's, the usual cause of `AGENT_MISSING`. On Windows it may name the `.cmd` npm installed, or the `.exe` itself. |-| `agents.<name>.model` | `sonnet` for Claude, otherwise the CLI's own default | Passed as `--model`. The picker's model select writes this. A value that starts with a dash or holds a space is ignored, so a typo cannot pass the CLI a flag. |+| `agents.<name>.model` | `sonnet` for Claude, `deepseek-flash` for DeepSeek Harness, otherwise the CLI's own default | Passed as `--model`. The picker's model select writes this. A value that starts with a dash or holds a space is ignored, so a typo cannot pass the CLI a flag. || `agents.<name>.effort` | unset | Passed as that CLI's reasoning-effort flag. A value the CLI does not accept is dropped rather than failing the run. |Changes apply to the next run: the config is re-read each time, so the Bridge needs no restart.@@ -380,7 +422,7 @@ Claude runner's settings.| Symptom | Fix || --- | --- || `AGENT_MISSING` | The Bridge's `PATH` differs from your shell's. Set `agents.<name>.bin` to an absolute path. |-| `AGENT_NEEDS_PERMISSION` | Antigravity has no rule for Browsentic's tools: press the button, or `browsentic agent fix antigravity`. Grok Build is not signed in: run `grok login`. Qwen Code has no model provider: run `qwen` and use `/auth`. OpenCode 1 is signed in to no provider: run `opencode auth login`. OpenCode 2 has no sign-in of its own for Browsentic: run the line the picker shows. |+| `AGENT_NEEDS_PERMISSION` | Antigravity has no rule for Browsentic's tools: press the button, or `browsentic agent fix antigravity`. Grok Build is not signed in: run `grok login`. Qwen Code has no model provider: run `qwen` and use `/auth`. OpenCode 1 is signed in to no provider: run `opencode auth login`. OpenCode 2 has no sign-in of its own for Browsentic: run the line the picker shows. DeepSeek Harness is signed in to no DeepSeek account and no `DEEPSEEK_API_KEY` is set: run `dsh web`. || Codex: "not logged in" | The Bridge inherits no session. Run `codex login`, then retry. || Codex answers about the page without opening it, or from a web search | Update Browsentic. Codex hides the browser tools until the model searches for them, and an older Browsentic left Codex's own web search switched on, which the model reached for first. || Mistral Vibe: a follow-up turn says *this agent run is no longer active* | Update Browsentic, then start a new conversation. An older one gave each turn its own folder, and Vibe keeps re-reading the first turn's, so a conversation begun before the update stays broken. |@@ -407,6 +449,10 @@ Claude runner's settings.| A model shows *not listed* | You pinned it, and the CLI no longer lists it. It is still passed to the CLI; pick a listed one if runs start failing. || OpenCode cannot find a key you have exported | Only `OPENCODE_*` reaches a run. Sign in instead (`opencode auth login` on OpenCode 1, the picker's line on OpenCode 2), which keeps the key in OpenCode's own store. || `AGENT_UNSAFE`: *OpenCode ran its own … tool* | A tool outside the browser ran despite the run's rules, so the run was stopped. Update OpenCode and Browsentic, and report it. |+| DeepSeek Harness: *not signed in* | Runs use a dsh home of their own and borrow only `~/.dsh/.credentials.yaml`. Run `dsh web` and sign in, or set `DEEPSEEK_API_KEY` where the Bridge starts. |+| `AGENT_UNSAFE`: *DeepSeek Harness ran its own … tool* | A tool outside the browser ran despite the run's patch, so the run was stopped. Update DeepSeek Harness and Browsentic, and report it. |+| DeepSeek Harness is unusable after a dsh update | The new release's plugin tree has a tool plugin Browsentic has not checked. Update Browsentic, or install the dsh it was checked against. |+| DeepSeek Harness: a screenshot reads as *[image unavailable]* | `deepseek-v4-pro` has no image input. Pick `deepseek-flash`. |---diff --git a/docs/guide/configuration.md b/docs/guide/configuration.mdindex 9b1a324..fef1590 100644--- a/docs/guide/configuration.md+++ b/docs/guide/configuration.md@@ -55,9 +55,9 @@ None of these keys is required.| Key | Default | Notes || --- | --- | --- |-| `agent` | `claude` | Which CLI the side panel runs on: `claude`, `codex`, `antigravity`, `vibe`, `grok`, `cursor`, `qwen` or `opencode` |+| `agent` | `claude` | Which CLI the side panel runs on: `claude`, `codex`, `antigravity`, `vibe`, `grok`, `cursor`, `qwen`, `opencode` or `deepseek` || `agents.<name>.bin` | the CLI's own command name | Absolute path to the binary |-| `agents.<name>.model` | `sonnet` for Claude, else the CLI's default | Passed as `--model`; the agent picker's model select writes it. One starting with a dash or holding a space is ignored |+| `agents.<name>.model` | `sonnet` for Claude, `deepseek-flash` for DeepSeek Harness, else the CLI's default | Passed as `--model`; the agent picker's model select writes it. One starting with a dash or holding a space is ignored || `agents.<name>.effort` | unset | That CLI's reasoning-effort flag; an unaccepted value is dropped || `agents.codex.transport` | unset (the app-server) | `"exec"` runs Codex through `codex exec` instead of its app-server; see [agents.md § Codex](agents.md#codex-gets-the-browser-tools-as-its-own) |diff --git a/docs/guide/features/conversations.md b/docs/guide/features/conversations.mdindex df8bf14..ca904f1 100644--- a/docs/guide/features/conversations.md+++ b/docs/guide/features/conversations.md@@ -110,8 +110,8 @@ can resume where it left off.The card also shows token counts from the agent CLI's own reporting: roughly how many tokens themodel's context window held after its last reply, and how many it generated on the last run. Claude-Code, Codex, Grok Build, Cursor CLI, Qwen Code and OpenCode report these; Antigravity and Mistral-Vibe do not, and the card says so.+Code, Codex, Grok Build, Cursor CLI, Qwen Code and OpenCode report these; Antigravity, Mistral Vibe+and DeepSeek Harness do not, and the card says so.Browsentic builds the card in the browser without waking an agent, so it also works mid-run. Itdescribes what this browser holds; the agent's own working memory lives with the agent.diff --git a/docs/guide/features/page-actions.md b/docs/guide/features/page-actions.mdindex 036712c..4093b2b 100644--- a/docs/guide/features/page-actions.md+++ b/docs/guide/features/page-actions.md@@ -4,10 +4,10 @@ The 52 page tools the agent uses to read, click, type, move between tabs and scrnot name them in an instruction, since the agent picks the tool, but knowing what exists tells youwhat is worth asking for.-On Claude Code, OpenCode, Mistral Vibe and Codex (unless it runs through `codex exec`), the less-common ones (captchas, theming, diagnostics, monitors and timers, downloads, dragging) stay out of-the agent's tool list until a page calls for them, because those agents would otherwise send the-model every tool's description on every turn. The agent opens their guide, or Codex searches for+On Claude Code, OpenCode, Mistral Vibe, DeepSeek Harness and Codex (unless it runs through+`codex exec`), the less common ones (captchas, theming, diagnostics, monitors and timers,+downloads, dragging) stay out of the agent's tool list until a page calls for them, because those+agents would otherwise send the model every tool's description on every turn. The agent opens their guide, or Codex searches forthem, when it needs one, and they then work exactly as the others do.Every parameter is in [reference/tools.md](../../reference/tools.md).diff --git a/docs/guide/features/skills.md b/docs/guide/features/skills.mdindex 0296ca6..27d1bbd 100644--- a/docs/guide/features/skills.md+++ b/docs/guide/features/skills.md@@ -109,6 +109,7 @@ The `/` picker also lists the skills the active agent CLI keeps for itself:| Cursor CLI | `~/.cursor/skills/`, `~/.agents/skills/` || Qwen Code | `~/.qwen/skills/`, `~/.agents/skills/` || OpenCode | `~/.config/opencode/skills/` (and `skill/`), `~/.opencode/skills/`, `~/.agents/skills/`, `~/.claude/skills/` |+| DeepSeek Harness | `~/.dsh/skills/` (`$DSH_HOME/skills/`), `~/.agents/skills/` (`$DSH_AGENTS_HOME/skills/`) |Picking one attaches it to that message: Browsentic Bridge reads the file when it starts the run andappends it to the system prompt, clearly marked, with a note that browser tools are all the run has.diff --git a/docs/guide/install.md b/docs/guide/install.mdindex 2477b47..dea3ca5 100644--- a/docs/guide/install.md+++ b/docs/guide/install.md@@ -25,7 +25,7 @@ Setup opens the Chrome Web Store in every Chromium browser, Edge included, and t| **Browser** | Chrome, Edge, Brave, Arc, Vivaldi, Opera or another Chromium browser, or Firefox 140 or newer | || **System** | macOS, Windows 10 or 11 ([experimental](limits.md#windows-is-experimental)), or Linux | || **Node** | 20 or newer, only for `npx`: the [Mac app](mac-app.md) and the [Windows app](windows-app.md) bring their own | `node --version` |-| **Agent** | One of [Claude Code](https://claude.com/claude-code), [Codex](https://developers.openai.com/codex/cli), [Antigravity](https://antigravity.google/docs/cli/install), [Mistral Vibe](https://github.com/mistralai/mistral-vibe) (beta), [Grok Build](https://docs.x.ai/build/overview) (beta), [Cursor CLI](https://cursor.com/docs/cli/overview) (beta), [Qwen Code](https://qwenlm.github.io/qwen-code-docs/en/) (beta) or [OpenCode](https://opencode.ai/docs/cli/) (beta) on your `PATH`, logged in | `claude --version`, `codex --version`, `agy --version`, `vibe --version`, `grok --version`, `cursor-agent --version`, `qwen --version`, `opencode --version` |+| **Agent** | One of [Claude Code](https://claude.com/claude-code), [Codex](https://developers.openai.com/codex/cli), [Antigravity](https://antigravity.google/docs/cli/install), [Mistral Vibe](https://github.com/mistralai/mistral-vibe) (beta), [Grok Build](https://docs.x.ai/build/overview) (beta), [Cursor CLI](https://cursor.com/docs/cli/overview) (beta), [Qwen Code](https://qwenlm.github.io/qwen-code-docs/en/) (beta), [OpenCode](https://opencode.ai/docs/cli/) (beta) or [DeepSeek Harness](https://deepseek-harness.github.io/deepseek-harness/) (beta) on your `PATH`, logged in | `claude --version`, `codex --version`, `agy --version`, `vibe --version`, `grok --version`, `cursor-agent --version`, `qwen --version`, `opencode --version`, `dsh --version` |Before you start:diff --git a/docs/guide/limits.md b/docs/guide/limits.mdindex f809f39..cbeea75 100644--- a/docs/guide/limits.md+++ b/docs/guide/limits.md@@ -74,16 +74,18 @@ contains it with whatever controls that CLI offers, and they are not equal:| Cursor CLI (beta) | Per-run deny rules, where a deny beats every allow including your own; a kernel sandbox is asked for too, but it has no Windows backend and is not depended on || Qwen Code (beta) | `--safe-mode` drops every setting of your own (hooks, extensions, bundled skills, MCP servers, permission rules), and deny rules close the shell, the disk and the tools that reach either; because that mode also disables Qwen's own fail-closed tool allowlist, the run's startup line is read back and anything unexpected stops it || OpenCode (beta) | A per-run agent whose rules open on a deny for every tool, applied after your own rules, so the model is offered only the browser; project config is off. On OpenCode 1 external plugins are off, and MCP servers you set up in OpenCode itself still start, with their tools hidden. On OpenCode 2 a policy keeps your plugins and other MCP servers from loading, though one in your global or managed config outranks it, and each run has a private server and database of its own |+| DeepSeek Harness (beta) | A per-run patch over its plugin tree switches off every tool of its own (the shell, the files, skills, sub-agents, workflows, and the web except when mapping a site), so the model is offered the browser's tools and a to-do list only; runs keep a dsh home of Browsentic's own, so your patches, plugins, `AGENTS.md` and skills never load, and dsh's read-only sandbox is on |The environment is sealed for all of them: cloud keys, registry tokens and database URLs inheritedfrom your shell are removed before the spawn, keeping only what that agent needs to authenticate.Details are in [internals/guardrails.md](../internals/guardrails.md#spawn-containment).-## Five of the agents are beta+## Six of the agents are beta-Mistral Vibe, Grok Build, Cursor CLI, Qwen Code and OpenCode were built by reading each CLI and-checking its flags, its containment and its error texts, but none has carried a whole conversation-end to end yet, so the agent picker, the docs and the release notes mark them *beta*. In practice:+Mistral Vibe, Grok Build, Cursor CLI, Qwen Code, OpenCode and DeepSeek Harness were built by reading+each CLI and checking its flags, its containment and its error texts, but none has carried a whole+conversation end to end yet, so the agent picker, the docs and the release notes mark them *beta*.+In practice:- a run that fails says why, with the command that fixes it, rather than hanging;- Vibe's replies arrive a message at a time rather than streaming, and it reports no token counts;@@ -94,7 +96,9 @@ end to end yet, so the agent picker, the docs and the release notes mark them *b- OpenCode needs a provider signed in, because Zen's free models refuse a run narrowed to thebrowser, and on OpenCode 2 that sign-in is one of Browsentic's own; its replies arrive a part at atime;-- the other three agents are untouched: the hooks the five share are covered by the spawn and+- DeepSeek Harness was checked with a real DeepSeek account but not yet through a real side-panel+ run; it gives up on a tool call after 60 seconds, as Cursor does, and it reports no token counts;+- the other three agents are untouched: the hooks the six share are covered by the spawn anddrive tests.[Choosing an agent](agents.md) has the per-agent detail. Please report what you find.diff --git a/docs/guide/mac-app.md b/docs/guide/mac-app.mdindex 4cc2d77..c881583 100644--- a/docs/guide/mac-app.md+++ b/docs/guide/mac-app.md@@ -47,7 +47,7 @@ The app opens on five checks and runs them automatically:| **Node.js runtime** | `node` 20 or newer is on your `PATH` | **Install** downloads the current LTS from nodejs.org into `~/.browsentic/runtime/node`, verified against its published SHA-256. No Homebrew, no password || **Browsentic Bridge** | `~/.browsentic/cli` holds the version this app carries | **Install** copies it there, writes the `browsentic` launcher to `~/.browsentic/bin`, and registers it with your browsers so they can start it || **A browser** | Chrome, Brave, Edge, Arc, Vivaldi, Opera, Chromium or Firefox is installed | **Get Chrome** opens the download page. Advisory: it does not block you |-| **An AI agent** | One of `claude`, `codex`, `agy`, `vibe`, `grok`, `cursor-agent`, `qwen` or `opencode` is on your `PATH` | **Install Claude Code** runs `npm i -g @anthropic-ai/claude-code`. Advisory |+| **An AI agent** | One of `claude`, `codex`, `agy`, `vibe`, `grok`, `cursor-agent`, `qwen`, `opencode` or `dsh` is on your `PATH` | **Install Claude Code** runs `npm i -g @anthropic-ai/claude-code`. Advisory |**Set up everything** installs Node and the Bridge in one click. The browser, the extension and theagent are your choice, so each has its own button. Once everything required is in place the appdiff --git a/docs/guide/troubleshooting.md b/docs/guide/troubleshooting.mdindex c3ffb1d..9bb0517 100644--- a/docs/guide/troubleshooting.md+++ b/docs/guide/troubleshooting.md@@ -104,6 +104,10 @@ Those three answer most questions. The Bridge's log is at `~/.browsentic/daemon.| OpenCode fails with *could not start this turn* | Usually a model OpenCode does not know | Pick a model exactly as `opencode models` lists it, `provider/model` || OpenCode cannot see an API key you exported | Only `OPENCODE_*` reaches a run; the rest are sealed away | Sign in instead (`opencode auth login` on OpenCode 1, the agent picker's line on OpenCode 2), which keeps the key in OpenCode's own store || `AGENT_UNSAFE`: *OpenCode ran its own … tool* | A tool outside the browser ran despite the run's rules, so the run was stopped | Update OpenCode and Browsentic, and report it |+| DeepSeek Harness says it is not signed in | Runs use a dsh home of their own and borrow only your sign-in file, `~/.dsh/.credentials.yaml` | Run `dsh web` and sign in, or set `DEEPSEEK_API_KEY` where the Bridge starts |+| `AGENT_UNSAFE`: *DeepSeek Harness ran its own … tool* | A tool outside the browser ran despite the run's patch, so the run was stopped | Update DeepSeek Harness and Browsentic, and report it |+| DeepSeek Harness is unusable after a dsh update | The new release's plugin tree has a tool plugin Browsentic has not checked | Update Browsentic, or install the dsh it was checked against |+| A DeepSeek Harness screenshot reads as *[image unavailable]* | `deepseek-v4-pro` has no image input | Pick `deepseek-flash` |## Pages and tabsdiff --git a/docs/guide/windows-app.md b/docs/guide/windows-app.mdindex 5d8f1c9..7a4091f 100644--- a/docs/guide/windows-app.md+++ b/docs/guide/windows-app.md@@ -46,7 +46,7 @@ The app opens on five checks and runs them automatically:| **Node.js runtime** | `node` 20 or newer is on your `PATH` | **Install** downloads the current LTS from nodejs.org into `%USERPROFILE%\.browsentic\runtime\node`, verified against its published SHA-256. No administrator || **Browsentic Bridge** | `%USERPROFILE%\.browsentic\cli` holds the version this app carries | **Install** copies it there, writes the `browsentic.exe` launcher to `%USERPROFILE%\.browsentic\bin`, and registers it with your browsers so they can start it || **A browser** | Chrome, Edge, Brave, Vivaldi, Opera, Chromium or Firefox is installed | **Get Chrome** opens the download page. Advisory: it does not block you |-| **An AI agent** | One of `claude`, `codex`, `agy`, `vibe`, `grok`, `cursor-agent`, `qwen` or `opencode` is on your `PATH` | **Install Claude Code** runs `npm install --global @anthropic-ai/claude-code`. Advisory |+| **An AI agent** | One of `claude`, `codex`, `agy`, `vibe`, `grok`, `cursor-agent`, `qwen`, `opencode` or `dsh` is on your `PATH` | **Install Claude Code** runs `npm install --global @anthropic-ai/claude-code`. Advisory |**Set up everything** installs Node and the Bridge in one click. The browser, the extension and theagent are your choice, so each has its own button. Once everything required is in place the appdiff --git a/docs/internals/agent-runs.md b/docs/internals/agent-runs.mdindex 1625aaf..e9a66e9 100644--- a/docs/internals/agent-runs.md+++ b/docs/internals/agent-runs.md@@ -95,13 +95,13 @@ an agent CLI is spawned or stopped (see [Windows](#windows)).Each runner delivers the same five things, through whichever mechanism its CLI supports:-| | Claude Code | Codex | Antigravity | Mistral Vibe | Grok Build | Cursor CLI | Qwen Code | OpenCode |-| --- | --- | --- | --- | --- | --- | --- | --- | --- |-| Run | `claude -p --output-format stream-json`, the message on stdin | `codex app-server`, held over stdin for the turn; `codex exec --json` when it is refused, when `agents.codex.transport` is `"exec"`, and for one-shots | `agy -p --output-format stream-json` | `vibe --prompt … --output streaming --trust --agent ask` | `grok -p --output-format streaming-json` | `cursor-agent -p --output-format stream-json --stream-partial-output` | `qwen -p --output-format stream-json --include-partial-messages` | `opencode run --format json --pure --agent browsentic-contained`; on OpenCode 2, `opencode acp` held over stdin for the turn, the message sent once the browser's tools have reached it (`opencode run --standalone` when it is refused, and for one-shots) |-| MCP server | `--mcp-config` + `--strict-mcp-config` | none: the tools go to `thread/start` as `dynamicTools`, and each of the user's servers is switched off in the thread's config. `exec`: `-c mcp_servers.browsentic.*` under `--ignore-user-config`, with `default_tools_approval_mode="approve"`, because headless Codex refuses any MCP call it would have prompted for | `.agents/mcp_config.json` in its cwd | `.vibe/config.toml` in a folder per conversation, every tool granted by name | `.grok/config.toml` in its cwd, loaded with `GROK_FOLDER_TRUST=0`, and approved with `--allow MCPTool(browsentic__*)` | `.cursor/mcp.json` in a folder per conversation, approved each turn with `cursor-agent mcp enable browsentic` (a set-up step the plan carries), allowed as `Mcp(browsentic:*)` with every other server denied by name and every plugin's as `Mcp(plugin-*:*)` | `--mcp-config` + `--allowed-mcp-server-names`, under `--safe-mode`, which drops the ones on disk | `mcp.browsentic` in `OPENCODE_CONFIG_CONTENT`, each tool allowed by name on the run's own agent; on OpenCode 2, `mcp.servers.browsentic` with `codemode: false` |-| System prompt | `--append-system-prompt-file`, a file per run, kept from the first turn ([below](#a-prompt-the-session-keeps)) | `developerInstructions` on `thread/start` (`-c developer_instructions` on `exec`), kept from the first turn | `AGENTS.md` in its cwd | `AGENTS.md` beside its config | `--rules` | `AGENTS.md` in its cwd, kept from the first turn | `--append-system-prompt` | a file in a folder per conversation, named in the config's `instructions` (never the config itself, which OpenCode expands `{file:…}` in) |-| Follow-up turns | `--resume <session>` | `thread/resume` (`exec resume <thread>`); either resumes a thread the other began | `--conversation <id>` | `--resume <session>`, re-read from the folder the session began in | `--session-id <uuid>` names it, `--resume <uuid>` continues it, in a folder named after it | `--resume <session>` | `--session-id <uuid>` names it, `--resume <uuid>` continues it | `--session <id>`, from a session store of Browsentic's own (`OPENCODE_DB`); on OpenCode 2, `session/resume`, from a database of its own that also holds Browsentic's sign-in |-| Kept off the machine by | `--allowedTools` + `--disallowedTools` | the user's MCP servers off (`--ignore-user-config` on `exec`), `-c sandbox_mode="read-only"`, `-c approval_policy="never"`, `-c features.shell_tool=false`, `-c features.view_image=false`, `-c features.multi_agent=false` | its own permission rules | `--enabled-tools`, which is an allowlist | `--tools`, `--permission-mode dontAsk`, `--deny`, `--sandbox workspace` | deny rules in `.cursor/cli.json`, plus `--sandbox enabled` | `--safe-mode`, `--exclude-tools`, `--approval-mode default` | an agent whose permission opens on `"*": "deny"`, `--pure`, `OPENCODE_DISABLE_PROJECT_CONFIG`, `OPENCODE_DISABLE_SHARE`; on OpenCode 2, an agent whose `permissions` open on a deny for everything, a private server (`--standalone`, or `acp`'s own), `integration.use` policies denying `plugin:*` and `mcp:*`, and the session's agent checked before the message goes |+| | Claude Code | Codex | Antigravity | Mistral Vibe | Grok Build | Cursor CLI | Qwen Code | OpenCode | DeepSeek Harness |+| --- | --- | --- | --- | --- | --- | --- | --- | --- | --- |+| Run | `claude -p --output-format stream-json`, the message on stdin | `codex app-server`, held over stdin for the turn; `codex exec --json` when it is refused, when `agents.codex.transport` is `"exec"`, and for one-shots | `agy -p --output-format stream-json` | `vibe --prompt … --output streaming --trust --agent ask` | `grok -p --output-format streaming-json` | `cursor-agent -p --output-format stream-json --stream-partial-output` | `qwen -p --output-format stream-json --include-partial-messages` | `opencode run --format json --pure --agent browsentic-contained`; on OpenCode 2, `opencode acp` held over stdin for the turn, the message sent once the browser's tools have reached it (`opencode run --standalone` when it is refused, and for one-shots) | `dsh --profile acp --patch <file>`, held over stdin for the turn; `dsh --profile headless --patch <file> --json` when ACP is refused, and for one-shots |+| MCP server | `--mcp-config` + `--strict-mcp-config` | none: the tools go to `thread/start` as `dynamicTools`, and each of the user's servers is switched off in the thread's config. `exec`: `-c mcp_servers.browsentic.*` under `--ignore-user-config`, with `default_tools_approval_mode="approve"`, because headless Codex refuses any MCP call it would have prompted for | `.agents/mcp_config.json` in its cwd | `.vibe/config.toml` in a folder per conversation, every tool granted by name | `.grok/config.toml` in its cwd, loaded with `GROK_FOLDER_TRUST=0`, and approved with `--allow MCPTool(browsentic__*)` | `.cursor/mcp.json` in a folder per conversation, approved each turn with `cursor-agent mcp enable browsentic` (a set-up step the plan carries), allowed as `Mcp(browsentic:*)` with every other server denied by name and every plugin's as `Mcp(plugin-*:*)` | `--mcp-config` + `--allowed-mcp-server-names`, under `--safe-mode`, which drops the ones on disk | `mcp.browsentic` in `OPENCODE_CONFIG_CONTENT`, each tool allowed by name on the run's own agent; on OpenCode 2, `mcp.servers.browsentic` with `codemode: false` | handed to the ACP session in `session/new` and `session/resume`; on headless, a row the patch inserts |+| System prompt | `--append-system-prompt-file`, a file per run, kept from the first turn ([below](#a-prompt-the-session-keeps)) | `developerInstructions` on `thread/start` (`-c developer_instructions` on `exec`), kept from the first turn | `AGENTS.md` in its cwd | `AGENTS.md` beside its config | `--rules` | `AGENTS.md` in its cwd, kept from the first turn | `--append-system-prompt` | a file in a folder per conversation, named in the config's `instructions` (never the config itself, which OpenCode expands `{file:…}` in) | `AGENTS.md` in a folder per conversation, which the patch marks as the project root so no folder above is read; re-read on resume |+| Follow-up turns | `--resume <session>` | `thread/resume` (`exec resume <thread>`); either resumes a thread the other began | `--conversation <id>` | `--resume <session>`, re-read from the folder the session began in | `--session-id <uuid>` names it, `--resume <uuid>` continues it, in a folder named after it | `--resume <session>` | `--session-id <uuid>` names it, `--resume <uuid>` continues it | `--session <id>`, from a session store of Browsentic's own (`OPENCODE_DB`); on OpenCode 2, `session/resume`, from a database of its own that also holds Browsentic's sign-in | `session/resume` (`--session-id <id>` on headless), in a dsh home of Browsentic's own |+| Kept off the machine by | `--allowedTools` + `--disallowedTools` | the user's MCP servers off (`--ignore-user-config` on `exec`), `-c sandbox_mode="read-only"`, `-c approval_policy="never"`, `-c features.shell_tool=false`, `-c features.view_image=false`, `-c features.multi_agent=false` | its own permission rules | `--enabled-tools`, which is an allowlist | `--tools`, `--permission-mode dontAsk`, `--deny`, `--sandbox workspace` | deny rules in `.cursor/cli.json`, plus `--sandbox enabled` | `--safe-mode`, `--exclude-tools`, `--approval-mode default` | an agent whose permission opens on `"*": "deny"`, `--pure`, `OPENCODE_DISABLE_PROJECT_CONFIG`, `OPENCODE_DISABLE_SHARE`; on OpenCode 2, an agent whose `permissions` open on a deny for everything, a private server (`--standalone`, or `acp`'s own), `integration.use` policies denying `plugin:*` and `mcp:*`, and the session's agent checked before the message goes | a patch (written as JSON) that switches off every plugin row giving the model a tool of its own, `DSH_PERMISSION_MODE=read-only`, `DSH_TELEMETRY_MODE=DISABLED`, and `DSH_HOME` and `DSH_AGENTS_HOME` inside Browsentic's folder |**Three CLIs do not put the browser tools in the model's list, and the prompt tells each one so.**Cursor hands the model a tool's schema only through its own `GetMcpTools`, by pattern or by name, so@@ -118,9 +118,9 @@ Codex also cuts any tool result at 10,000 tokens by default, a limit its model c`// @exec: {"max_output_tokens": 25000}`; only the prompt can ask for that, so it does. It alsoasks for page reads in pieces.-**Four CLIs would send every tool they are given on every turn, so their runs list only the core.**+**Five CLIs would send every tool they are given on every turn, so their runs list only the core.**`Runner.delivery(settings)` says how the rest reach the model. Claude Code (whose own `ToolSearch`-is denied), Mistral Vibe and OpenCode answer `on-demand`. For their runs `offerFor` names every+is denied), Mistral Vibe, OpenCode and DeepSeek Harness answer `on-demand`. For their runs `offerFor` names every[grouped action](registry.md) in `onDemand`, which the describe reply carries beside `withheld`with the delivery, and the tool host leaves those off the list and adds `browsentic_use`. The agent finds them through`browsentic_openSkill`, whose description and the server's instructions carry one line per group,@@ -151,7 +151,8 @@ looked up, and a second level of lookup would only lose the schema-checked call.**A CLI with limits on a tool call declares them** (`Runner.limits`), and the daemon works insidethem so the run does not have to discover them. Cursor abandons a call at 60 s (the MCP SDK'sdefault request timeout, which it never overrides) and puts text over 40,000 bytes in a file the run-cannot read (measured on 2026.09.18). So for its runs:+cannot read (measured on 2026.09.18). DeepSeek Harness abandons a call at 60 s as well (measured on+0.2.1-rc.2), and declares only that limit. So for their runs:- an approval still unanswered 10 s before the limit is *parked*: the call returns `APPROVAL_PENDING`with no `toolResult`, so the card stays up, and the same action with the same input rejoins it@@ -160,7 +161,7 @@ cannot read (measured on 2026.09.18). So for its runs:- a wait inside a call (`page_awaitMonitor`, `page_pickElement`, `page_solveCaptcha`,`page_captureDownload`, and any `timeoutMs` on `page_runCode`, `page_callSiteTool` or`page_waitForElement`) is cut to what is left of the call;-- the run's MCP server is started with `BROWSENTIC_RESULT_BYTES`, refuses a result whose text is+- on Cursor, the run's MCP server is started with `BROWSENTIC_RESULT_BYTES`, refuses a result whose text isover it with `RESULT_TOO_LARGE` rather than cutting it, and asks `page_extractText` for groups of aquarter that many characters.@@ -262,6 +263,11 @@ The app-server interface is marked experimental. Codex's own editor extensions utranscript in `fixtures/codex/0.155.1-app-server-turn.jsonl` is what 0.155.1 printed, and`codex app-server generate-ts --experimental --out <dir>` prints the current shapes.+OpenCode 2 and DeepSeek Harness are held the same way over the Agent Client Protocol, through+[`acp.ts`](../../src/daemon/agent/runners/acp.ts): `session/new` or `session/resume` with the+browser's MCP server, then `session/prompt`. A refused ACP turn falls back the same way, to the+CLI's headless run.+### ReadinessBefore a run starts, `agentState()` probes each CLI (`--version`, plus any extra readiness check)diff --git a/docs/internals/guardrails.md b/docs/internals/guardrails.mdindex 01e6956..e591c93 100644--- a/docs/internals/guardrails.md+++ b/docs/internals/guardrails.md@@ -370,6 +370,7 @@ type LocalTools = 'allowlist' | 'sandbox' | 'host'| **Qwen Code** | `allowlist` | `--allowed-tools mcp__browsentic` is what a run may call without being asked, and a headless turn refuses everything it would otherwise have prompted for. `--safe-mode` is what makes the rest hold: it drops the user's hooks, extensions, bundled skills, `settings.json` MCP servers, `.mcp.json` and permission rules, while keeping `--mcp-config`, an explicit per-invocation argument rather than ambient state. Its cost is that it also **silently ignores `--core-tools`**, the fail-closed allowlist over Qwen's twenty-one core tools, so the built-ins are closed with `--exclude-tools` deny rules instead. A deny list cannot be fail-closed, so the `init` line is read back: it names every tool and MCP server that actually registered, and the reader ends the run with `AGENT_UNSAFE` on anything Browsentic did not ask for, before the model has spoken || **Cursor CLI** | `allowlist` | Deny rules in a project `.cursor/cli.json`, where **a deny beats every allow**, including the user's own. Measured against 2026.09.18: a headless run asked to `echo` a marker got `permissionDenied`, twice, and gave up. `--sandbox enabled` is asked for as well but not depended on, and it has no Windows backend. This is the first runner whose containment lives in a file rather than in argv, which is why `vetPlan` checks file *content* and not only that the file was written. `--trust` is **required**, not forbidden: headless Cursor refuses to start in a folder nobody trusted, and the folder is one Browsentic created and wrote every file in; trusting it grants no tool permission of its own. The project's MCP server starts only once approved, keyed to its exact config, so each turn's plan carries one set-up step, `cursor-agent mcp enable browsentic`, run in its folder before the turn; `vetPlan` allows that step and no other, for a run and never a task, and `--approve-mcps`, which approves every server the user has, stays forbidden. Every plugin's MCP server loads in every run under `plugin-<plugin>-<server>`, so `Mcp(plugin-*:*)` is denied beside the user's own servers by name, and an answer from any server but `browsentic` ends the run with `AGENT_UNSAFE` || **OpenCode** | `allowlist` | The run is an agent the config defines, whose permission ruleset opens on `"*": "deny"` and then allows each browser tool by name. OpenCode applies the last rule that matches and puts an agent's rules after the user's, so every other tool (built-in, custom, another MCP server's, or one a later release adds) is denied, and a tool denied outright is never offered to the model. `--pure` keeps plugins out, which run inside OpenCode and can answer its permission prompts. MCP servers the user set up in OpenCode itself still start, with their tools hidden. OpenCode 2 is held to a second rule set, below |+| **DeepSeek Harness** | `allowlist` | A per-run patch over dsh's plugin tree switches off every row that gives the model a tool of its own: the shell, files and file search, skills, sub-agents, workflows and scripted tool calls, goals, plan mode, its commands and user questions, MCP resources, and the web unless the run is researching. The model is offered the browser's tools and a to-do list only. Runs keep a dsh home of Browsentic's own, so the user's patches, profiles' plugins, `AGENTS.md` and skills never compose in. `DSH_PERMISSION_MODE=read-only` puts dsh's sandbox under whatever might still run, and a built-in tool that runs anyway ends the run with `AGENT_UNSAFE`. Detail below |**Windows is less measured.** Cursor's kernel sandbox has no Windows backend, so only its deny rulesapply there. Codex's `sandbox_mode="read-only"` and Grok's `--sandbox workspace` are asked for as on@@ -486,6 +487,34 @@ not supported*), so a plan built for the wrong line never starts. What 2.x chang**What remains on 2.x:** a policy in the user's global or managed config outranks the run's, so a userwho allows a plugin there gets it loaded; its tools are still denied by the agent's rules.+#### DeepSeek Harness: a patch over its plugin tree++dsh is a tree of plugins, each a row of a profile, so a run's containment is a patch over that tree.+The patch is written as JSON, which dsh reads as YAML unchanged and which can carry no `!!js`+expression. A held turn is `dsh --profile acp --patch <file>`, with the browser's MCP server handed+to the session. A one-shot is `dsh --profile headless --patch <file> --json` with no server at all,+and a turn ACP refused runs the same way with the browser's server mounted by the patch. Before+every spawn `vetPlan()` parses the patch rather than searching it:++- every required switch-off is present, each a row that switches its plugin off and does nothing+ else;+- nothing is switched on but the sign-in rows;+- every other row is a known override: the credentials row pointed at a `.credentials.yaml`, the+ route (a DeepSeek provider, a model, an effort dsh accepts), workspace instructions kept to the+ run's own folder, the persona, the result ceiling, a one-shot's read-only `read`, and on a headless+ run the browser's own server;+- `DSH_HOME` and `DSH_AGENTS_HOME` lie inside Browsentic's own folder, and+ `DSH_PERMISSION_MODE=read-only` and `DSH_TELEMETRY_MODE=DISABLED` are set.++A patch switches off only the rows it names, so a dsh release that adds a tool plugin would hand it+to the model. At readiness Browsentic composes the installed dsh's plugin tree with the patch and+reports a release that leaves an unvetted tool plugin on as `AGENT_UNUSABLE`, as it does a dsh older+than 0.2.1. The reader checks the rest: a built-in tool that runs anyway stops the run with+`AGENT_UNSAFE`.++**What remains:** a one-shot that reads a file gets dsh's `read` tool alone, read-only, which can+read any file the user can, as Codex's shell can in the same case.+`localTools` records which case each runner is in, and the note is logged once per run, so the weakone is **visible in the log rather than assumed away**.@@ -495,17 +524,20 @@ one is **visible in the log rather than assumed away**.analyst reading an attached file, the captcha analyst looking at one round of a challenge, or turninga raw recording trace into steps. That is asserted by requiring `{"mcpServers":{}}` (or`mcp_servers={}`, or Grok's `--deny MCPTool`) in its argv, or OpenCode's `"mcp":{}` and `"*":"deny"`-in its config (on OpenCode 2, no servers and a deny-all agent). `Read` is deliberately left out of `task`'s deny list, because some tasks are handed a-file in the scratch workspace.+in its config (on OpenCode 2, no servers and a deny-all agent), or a DeepSeek Harness patch that+mounts no server. `Read` is deliberately left out of `task`'s deny list, because some tasks are+handed a file in the scratch workspace.The file analyst adds three things. A file is screened before anything spawns, so an archive, anexecutable or an oversized file never reaches an agent. The one-shot is stopped the moment its answeris in rather than left to exit, and its copy of the file is deleted. And the session is not keptwhere the CLI allows that: Claude Code runs a task with `--no-session-persistence` and Codex with`--ephemeral`. OpenCode has no such switch, but its runs and tasks both go into a session store-under `~/.browsentic` (`OPENCODE_DB`) rather than the user's, one per line of releases.-Antigravity, Mistral Vibe, Grok Build, Cursor CLI and Qwen Code have neither, so a task's session-stays in that CLI's own history, inside the per-mode task workspace it was started in.+under `~/.browsentic` (`OPENCODE_DB`) rather than the user's, one per line of releases. DeepSeek+Harness has none either, but its runs and tasks keep their sessions in dsh homes under+`~/.browsentic`. Antigravity, Mistral Vibe, Grok Build, Cursor CLI and Qwen Code have neither, so a+task's session stays in that CLI's own history, inside the per-mode task workspace it was started+in.### Sealing the environment@@ -528,6 +560,7 @@ Each agent keeps only the prefixes it needs to authenticate:| Cursor CLI | `CURSOR_` || Qwen Code | `QWEN_`, `DASHSCOPE_`, `BAILIAN_`, `OPENAI_` || OpenCode | `OPENCODE_` |+| DeepSeek Harness | `DEEPSEEK_` |Qwen Code is the awkward one. It authenticates through whichever provider it is pointed at, and thedocumented mainstream setup is `OPENAI_API_KEY` with `OPENAI_BASE_URL` at a DashScope endpoint, so@@ -543,11 +576,15 @@ declared in its config as signed in. On OpenCode 2 it asks `opencode auth list -the runs' own database, in the environment a run gets, so a key counts exactly when a run would seeit; `--standalone` keeps the user's background service from being started with that environment.+DeepSeek Harness signs in through the file `dsh web` writes, which the run's patch points to, so only+`DEEPSEEK_` survives, for `DEEPSEEK_API_KEY`. A run's `DSH_` variables are the ones Browsentic sets,+and `$DSH_HOME/.env` is not read. `DSH_` and `DEEPSEEK_` are sealed away from every other agent.+Sealing is not absolute for Qwen either. Qwen loads the first `.env` it finds walking up from itsworking directory, then `~/.qwen/.env` and `~/.env`, for variables not already present, so a key`sealEnv` stripped can come back from disk. It reaches the model provider rather than the model,because the shell and file tools are denied, but the guarantee is weaker there than for the other-seven.+eight.Then there are **federated** cases, where a flag turns another prefix into the agent's owncredentials. Claude Code on Bedrock authenticates with `AWS_*`, so sealing it would lock the agentdiff --git a/docs/internals/transport.md b/docs/internals/transport.mdindex 78ba31e..c4a3560 100644--- a/docs/internals/transport.md+++ b/docs/internals/transport.md@@ -173,7 +173,7 @@ dialling. It never deletes the stored key; only pairing again or `disconnect` re## Protocol version-Both sides compile in `SOCKET_PROTOCOL_VERSION` (currently **24**). They used to have to match+Both sides compile in `SOCKET_PROTOCOL_VERSION` (currently **25**). They used to have to matchexactly. A store copy updates when its browser decides to, not when the daemon does, so theextension and the daemon are routinely a version apart, and the daemon accepts a **window**:@@ -193,6 +193,14 @@ a change neither side can do without raises a minimum. A refusal must never strav0.8.0 store build unpairs on any refusal and does not retry, so no daemon may refuse protocol 22until a store build that retries has replaced it.+### Protocol 25: an agent the extension does not know++An extension looks every agent up by its kind, so one built before an agent was added cannot show+it. `AGENT_PROTOCOL` names the first protocol whose extension knows each agent added since the window+opened; for DeepSeek Harness that is 25. An extension below it is never sent that agent: its agent+list leaves it out, and while it is the chosen agent a run started there is refused with+`AGENT_UNUSABLE`, *DeepSeek Harness needs a newer Browsentic extension*.+### Protocol 24: notices for the desktop appProtocol 24 lets a browser hand a finished monitor, scheduled task or reminder to the Browsentic app,diff --git a/docs/reference/cli.md b/docs/reference/cli.mdindex 6572ad1..dd5dc1c 100644--- a/docs/reference/cli.md+++ b/docs/reference/cli.md@@ -91,7 +91,7 @@ See [guide/pair.md](../guide/pair.md).| Command | Does || --- | --- || `browsentic agent` | Show which agent runs the side panel, and which are installed |-| `browsentic agent <name>` | Switch to `claude`, `codex`, `antigravity`, `vibe`, `grok`, `cursor`, `qwen` or `opencode` |+| `browsentic agent <name>` | Switch to `claude`, `codex`, `antigravity`, `vibe`, `grok`, `cursor`, `qwen`, `opencode` or `deepseek` || `browsentic agent fix <name>` | Let Browsentic fix what that agent still needs || `browsentic agent model <name> [model]` | Pin that agent’s model in `config.json`; omit the model to go back to the CLI’s own default |diff --git a/docs/reference/errors.md b/docs/reference/errors.mdindex af133d0..63bb1c6 100644--- a/docs/reference/errors.md+++ b/docs/reference/errors.md@@ -80,9 +80,9 @@ run, and the message carries the *fix*. In the **Origin** column, Daemon is Brow| Code | Origin | Meaning and next move || --- | --- | --- || `AGENT_MISSING` | Runner | The chosen agent's binary is not on the *daemon's* `PATH`. Set `agents.<name>.bin` to an absolute path |-| `AGENT_NEEDS_PERMISSION` | Runner | Antigravity has no rule allowing Browsentic's MCP tools (`browsentic agent fix antigravity`), Grok Build or Cursor CLI is not signed in (`grok login`, `cursor-agent login`), Qwen Code has no model provider configured (`qwen`, then `/auth`), or OpenCode 1 is signed in to no provider (`opencode auth login`), OpenCode 2 has no sign-in of Browsentic's own (the line the agent picker shows) |-| `AGENT_UNUSABLE` | Runner | The CLI is present but cannot run, usually because it is too old for the flags Browsentic passes. On Windows, also a batch file that is not an npm or pnpm shim: set `agents.<name>.bin` to the program it runs |-| `AGENT_UNSAFE` | Runner | The run was stopped before the agent could act on the machine: its plan had lost its containment, which is a bug in Browsentic, or the CLI itself said it had not applied it (Grok Build offering tools Browsentic never asks for, Qwen Code reporting a denied tool or a second MCP server as registered, OpenCode reporting that a tool outside the browser ran or opening a session in an agent other than Browsentic's, Codex running a shell command or changing a file with both switched off, or Cursor CLI getting an answer from an MCP server the run denies). Update both, and report it |+| `AGENT_NEEDS_PERMISSION` | Runner | Antigravity has no rule allowing Browsentic's MCP tools (`browsentic agent fix antigravity`), Grok Build or Cursor CLI is not signed in (`grok login`, `cursor-agent login`), Qwen Code has no model provider configured (`qwen`, then `/auth`), or OpenCode 1 is signed in to no provider (`opencode auth login`), OpenCode 2 has no sign-in of Browsentic's own (the line the agent picker shows), or DeepSeek Harness is signed in to no DeepSeek account and no `DEEPSEEK_API_KEY` is set (`dsh web`) |+| `AGENT_UNUSABLE` | Runner | The CLI is present but cannot run, usually because it is too old for the flags Browsentic passes. Also a DeepSeek Harness older than 0.2.1, or one whose plugin tree leaves a tool on that Browsentic has not checked. On Windows, also a batch file that is not an npm or pnpm shim: set `agents.<name>.bin` to the program it runs. A run from an extension too old to know the chosen agent is refused the same way, with *needs a newer Browsentic extension* |+| `AGENT_UNSAFE` | Runner | The run was stopped before the agent could act on the machine: its plan had lost its containment, which is a bug in Browsentic, or the CLI itself said it had not applied it (Grok Build offering tools Browsentic never asks for, Qwen Code reporting a denied tool or a second MCP server as registered, OpenCode reporting that a tool outside the browser ran or opening a session in an agent other than Browsentic's, Codex running a shell command or changing a file with both switched off, Cursor CLI getting an answer from an MCP server the run denies, or DeepSeek Harness reporting that one of its own tools ran). Update both, and report it |## Guardrailsdiff --git a/docs/reference/tools.md b/docs/reference/tools.mdindex 2fd94f1..f9365de 100644--- a/docs/reference/tools.md+++ b/docs/reference/tools.md@@ -86,7 +86,7 @@ only the tools this caller is offered and names the rest; one with nothing to of### browsentic_useRun a page tool that is not in this caller's tool list, by name. A side-panel run on Claude Code,-OpenCode or Mistral Vibe lists only the core tools, because those CLIs send the description of every+OpenCode, Mistral Vibe or DeepSeek Harness lists only the core tools, because those CLIs send the description of everytool they are given on every turn. The tools of the guides above are left off its list, and theagent calls them through `browsentic_use` once a guide, a skill or a result has named one. The callgoes the way the tool's own call would: the same approvals, guardrails, blocked sites and steps indiff --git a/src/daemon/README.md b/src/daemon/README.mdindex 36f7206..1c69689 100644--- a/src/daemon/README.md+++ b/src/daemon/README.md@@ -45,7 +45,7 @@ The extension updates itself from its store, and the two do not have to be the s- Node.js 20 or newer- Chrome, Edge, Brave, Arc, Vivaldi, Opera or another Chromium browser, or Firefox 140 or newer- One agent CLI on your `PATH`, logged in: `claude`, `codex`, `agy`, `vibe`, `grok`, `cursor-agent`,- `qwen` or `opencode`+ `qwen`, `opencode` or `dsh`## Commandsdiff --git a/src/mac/Sources/Browsentic/App/AppModel.swift b/src/mac/Sources/Browsentic/App/AppModel.swiftindex 99231e4..c4ccb7b 100644--- a/src/mac/Sources/Browsentic/App/AppModel.swift+++ b/src/mac/Sources/Browsentic/App/AppModel.swift@@ -182,7 +182,7 @@ final class AppModel: ObservableObject {case .agent:let found = AgentProbe.installed()return found.isEmpty- ? .advisory("None on your PATH. The side panel needs Claude Code, Codex, Antigravity, Mistral Vibe or Grok Build.")+ ? .advisory("None on your PATH. The side panel needs one, such as Claude Code or Codex."): .passed(found.joined(separator: ", "))}}diff --git a/src/mac/Sources/Browsentic/Core/Preflight.swift b/src/mac/Sources/Browsentic/Core/Preflight.swiftindex bbc4311..e91ec0f 100644--- a/src/mac/Sources/Browsentic/Core/Preflight.swift+++ b/src/mac/Sources/Browsentic/Core/Preflight.swift@@ -88,7 +88,17 @@ struct Browser: Identifiable, Equatable {}enum AgentProbe {- static let bins = [("Claude Code", "claude"), ("Codex", "codex"), ("Antigravity", "agy"), ("Mistral Vibe", "vibe"), ("Grok Build", "grok")]+ static let bins: [(String, String)] = [+ ("Claude Code", "claude"),+ ("Codex", "codex"),+ ("Antigravity", "agy"),+ ("Mistral Vibe", "vibe"),+ ("Grok Build", "grok"),+ ("Cursor CLI", "cursor-agent"),+ ("Qwen Code", "qwen"),+ ("OpenCode", "opencode"),+ ("DeepSeek Harness", "dsh"),+ ]static func installed() -> [String] {bins.filter { Shell.which($0.1) != nil }.map(\.0)