Skip to content

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.

feat(agents): OpenCode 2 runs a turn over ACP once the browser's tools have reached it, with a sign-in and database of Browsentic's own, beside OpenCode 1 unchanged

Shahriar455b81da63d2
diff --git a/docs/guide/agents.md b/docs/guide/agents.md
index cf08b05..1fbea3f 100644
--- a/docs/guide/agents.md
+++ b/docs/guide/agents.md
@@ -268,40 +268,69 @@ and please report them.
### OpenCode is in beta
-The OpenCode runner was checked against `opencode` 1.18.32 with a stand-in model provider: every
+OpenCode ships in two lines, which both install an `opencode` command: OpenCode 1 (`npm i -g
+opencode-ai`, what OpenCode's own docs install) and OpenCode 2 (`npm i -g @opencode/cli`, which also
+installs `opencode2`). Browsentic reads which one it has from `opencode --version` and runs each its
+own way. The runner was checked against 1.18.32 and 2.0.32 with a stand-in model provider: every
flag and setting it passes, the containment under a user config written to break it, the browser
tools reaching a run, a follow-up turn resuming, and each error OpenCode reports. A real model
driving a real page has not been tried yet, because OpenCode Zen's free models refuse a contained
run and no paid provider was signed in where it was built. Expect rough edges, and please report
them.
-- **Sign in to a provider first.** Run `opencode auth login`. Zen's free models answer only
- requests that carry OpenCode's own built-in tools. A Browsentic run carries only the browser's, so
- they refuse it with *free tier can only be used from within OpenCode*. Until OpenCode has a login,
- or a provider in its config such as a local model, the agent picker shows *needs setup*.
+What both lines share:
+
+- **Zen's free models refuse a Browsentic run.** They answer only requests that carry OpenCode's
+ own built-in tools. A Browsentic run carries only the browser's, so they refuse it with *free tier
+ can only be used from within OpenCode*. Sign in to a provider first; until then the agent picker
+ shows *needs setup*.
- **Models are `provider/model`**, spelled as `opencode models` lists them.
- **A key in the environment does not reach a run.** Only `OPENCODE_*` does; `ANTHROPIC_API_KEY`
- and the like are sealed away. `opencode auth login` keeps the key in OpenCode's own file instead,
- which a run reads as usual.
+ and the like are sealed away. Sign in instead, which keeps the key in OpenCode's own store.
- **Your OpenCode settings are read, and cannot widen a run.** Your providers, models and
`~/.config/opencode/AGENTS.md` apply. The run uses an agent Browsentic defines, whose rules deny
every tool not named for it, so your permission rules, custom tools and the MCP servers you set up
- in OpenCode itself are hidden from the model (those servers still start). External plugins and
- project config are off for the run.
+ in OpenCode itself are hidden from the model. Project config is off for the run.
- **Sharing is off.** A `"share": "auto"` of your own would publish every browsing session to a
public link, so a run switches sharing off.
-- **Run sessions are kept apart from yours.** Runs and one-shots go into a session store under
- `~/.browsentic`, not into `opencode session list`.
-- **Replies arrive a part at a time.** Like Vibe's, OpenCode's stream carries each finished part,
- not tokens.
+- **Replies arrive a part at a time**, not token by token.
- **A tool result is cut at 100 KB**, not OpenCode's own 50 KB, which a whole-page snapshot can
exceed. Past that, the prompt asks for smaller reads, because a run cannot open the saved copy
OpenCode points to.
-- **Effort is a variant.** `agents.opencode.effort` is passed as `--variant`, whose names each model
- defines; a name the model lacks is ignored.
- **The run is stopped if a local tool ran.** OpenCode reports every tool call, and one that ran
outside the browser ends the run with `AGENT_UNSAFE`.
+**OpenCode 1:**
+
+- **Sign in with `opencode auth login`.** A provider in your config, such as a local model, counts
+ too.
+- **Run sessions are kept apart from yours.** Runs and one-shots go into a session store under
+ `~/.browsentic`, not into `opencode session list`. External plugins are off for the run, and the
+ MCP servers you set up still start, with their tools hidden.
+- **Effort is a variant.** `agents.opencode.effort` is passed as `--variant`, whose names each model
+ defines; a name the model lacks is ignored.
+
+**OpenCode 2:**
+
+- **Browsentic has a sign-in of its own.** OpenCode 2 keeps sign-ins in the same database as its
+ sessions, and Browsentic runs keep a database apart from yours, so your own sign-in does not carry
+ over. The agent picker shows the one line to run, which signs that database in; on macOS and Linux
+ it is `OPENCODE_DB="$HOME/.browsentic/agents/opencode/opencode-2.db" opencode auth login --standalone`.
+ An `OPENCODE_API_KEY` in the environment, or a provider in your config, counts too.
+- **Your sessions and your background service are never touched.** Every run starts its own private
+ OpenCode server rather than the shared one, so a run is never added to your session list and your
+ service never picks one up.
+- **A turn waits for the browser tools.** OpenCode 2 connects MCP servers in the background, so a
+ turn is held over its Agent Client Protocol (`opencode acp`) and the message goes once the browser's
+ tools have reached it, which adds well under a second.
+- **Your plugins and other MCP servers do not load.** A policy in the run's config keeps them off;
+ one in your global or managed config outranks it.
+- **Effort is set only where the model offers it.** OpenCode 2 refuses a run asking for an effort
+ the model lacks, so `agents.opencode.effort` is applied when the model lists that level and
+ otherwise left out.
+- **Sessions do not carry over from OpenCode 1.** After switching lines, a conversation's next turn
+ starts a fresh OpenCode session.
+
---
## Per-agent settings
@@ -343,7 +372,7 @@ 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 is signed in to no provider: run `opencode auth login`. |
+| `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. |
| 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. |
@@ -359,11 +388,14 @@ Claude runner's settings.
| Qwen Code: *No auth type is selected* | Qwen has no provider configured, and its OAuth free tier has ended. Run `qwen` and use `/auth`, or export `OPENAI_API_KEY` with `OPENAI_BASE_URL`. |
| Qwen Code cannot find a key you have exported | Only `QWEN_*`, `DASHSCOPE_*`, `BAILIAN_*` and `OPENAI_*` reach a run; `ANTHROPIC_*` and `GEMINI_*` are sealed away. Point Qwen at one of the first four. |
| `AGENT_UNSAFE`: *Qwen Code registered …* or *loaded the MCP server …* | Qwen's own `init` line named a tool or a server Browsentic denied, so the run was stopped. Update Qwen Code and Browsentic, and report it if it persists. |
-| OpenCode: *free models refuse a run whose tools Browsentic has narrowed to the browser* | Zen's free tier serves only requests carrying OpenCode's own tools. Run `opencode auth login`, then pick that provider's model in the side panel's agent picker. |
+| OpenCode: *free models refuse a run whose tools Browsentic has narrowed to the browser* | Zen's free tier serves only requests carrying OpenCode's own tools. Sign in to a provider (`opencode auth login` on OpenCode 1, the line in the message on OpenCode 2), then pick that provider's model in the side panel's agent picker. |
+| OpenCode 2: *no model "…" here* | OpenCode 2 does not offer that model. Pick one of those the message lists, `provider/model`. |
+| OpenCode 2: *could not start Browsentic's browser tools* | The run's browser tools did not reach OpenCode within 20 seconds. Run `browsentic restart`, then send the message again. |
+| OpenCode: *changed version since Browsentic last checked it* | OpenCode 1 and 2 refuse each other's flags, which is what a swap while Browsentic runs looks like. Send the message again. |
| OpenCode: *could not start this turn* | Usually a model OpenCode does not know. Pick one as `opencode models` lists it, `provider/model`. |
| The model select says *built-in list*, with a reason | Browsentic could not read that CLI's own list, usually because it is signed out. Sign it in, then press *Recheck*. The shipped list works meanwhile. |
| 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. Run `opencode auth login` instead, which keeps the key in OpenCode's own file. |
+| 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. |
---
diff --git a/docs/guide/limits.md b/docs/guide/limits.md
index 04eb8bb..b64154b 100644
--- a/docs/guide/limits.md
+++ b/docs/guide/limits.md
@@ -67,7 +67,7 @@ contains it with whatever controls that CLI offers, and they are not equal:
| Grok Build (beta) | A per-run tool list, approvals that refuse anything not granted up front, and a sandbox that keeps its writes in its own folder; MCP servers you set up in Grok itself still load |
| 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; external plugins and project config are off, and MCP servers you set up in OpenCode itself still start, with their tools hidden |
+| 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 |
The environment is sealed for all of them: cloud keys, registry tokens and database URLs inherited
from your shell are removed before the spawn, keeping only what that agent needs to authenticate.
@@ -86,7 +86,8 @@ end to end yet, so the agent picker, the docs and the release notes mark them *b
- Qwen was written against its source, not a running binary (no provider was configured on the
machine it was built on), and it needs a provider configured before it will answer at all;
- OpenCode needs a provider signed in, because Zen's free models refuse a run narrowed to the
- browser, and its replies arrive a part at a time;
+ browser, and on OpenCode 2 that sign-in is one of Browsentic's own; its replies arrive a part at a
+ time;
- the other three agents are untouched: the hooks the five share are covered by the spawn and
drive tests.
diff --git a/docs/guide/troubleshooting.md b/docs/guide/troubleshooting.md
index 29e3242..ccf6e04 100644
--- a/docs/guide/troubleshooting.md
+++ b/docs/guide/troubleshooting.md
@@ -86,9 +86,13 @@ Those three answer most questions. The Bridge's log is at `~/.browsentic/daemon.
| Qwen Code fails with *No auth type is selected* | Qwen has no provider configured, and its OAuth free tier ended on 2026-04-15 | Run `qwen` and use `/auth`, or export `OPENAI_API_KEY` with `OPENAI_BASE_URL` |
| Qwen Code cannot see an API key you exported | Only `QWEN_*`, `DASHSCOPE_*`, `BAILIAN_*` and `OPENAI_*` reach a run; the rest are sealed away | Point Qwen at one of those four providers |
| `AGENT_UNSAFE`: *Qwen Code registered …* / *loaded the MCP server …* | Qwen's own startup line named a tool or a server Browsentic denied, so the run was stopped before the model saw it | Update Qwen Code and Browsentic, and report it if it persists |
-| OpenCode fails with *free models refuse a run whose tools Browsentic has narrowed to the browser* | OpenCode Zen's free tier serves only requests carrying OpenCode's own built-in tools | Run `opencode auth login`, then pick that provider's model in the side panel's agent picker |
+| OpenCode fails with *free models refuse a run whose tools Browsentic has narrowed to the browser* | OpenCode Zen's free tier serves only requests carrying OpenCode's own built-in tools | Sign in to a provider: `opencode auth login` on OpenCode 1, the line in the message on OpenCode 2. Then pick that provider's model in the side panel's agent picker |
+| OpenCode 2 needs setup though you are signed in to OpenCode | Browsentic's runs keep an OpenCode database of their own, apart from yours, and OpenCode 2 keeps sign-ins in it | Run the line the agent picker shows, which signs that database in |
+| OpenCode 2 fails with *no model "…" here* | OpenCode 2 does not offer that model to Browsentic's sign-in | Pick one of the models the message lists, `provider/model` |
+| OpenCode 2 fails with *could not start Browsentic's browser tools* | The run's browser tools did not reach OpenCode within 20 seconds | Run `browsentic restart`, then send the message again |
+| OpenCode fails with *changed version since Browsentic last checked it* | OpenCode 1 and 2 refuse each other's flags, which is what swapping one for the other while Browsentic runs looks like | Send the message again |
| 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 | Run `opencode auth login`, which keeps the key in OpenCode's own file |
+| 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 |
## Pages and tabs
diff --git a/docs/internals/agent-runs.md b/docs/internals/agent-runs.md
index a8b81fe..ad352f2 100644
--- a/docs/internals/agent-runs.md
+++ b/docs/internals/agent-runs.md
@@ -97,11 +97,11 @@ Each runner delivers the same five things, through whichever mechanism its CLI s
| | 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` |
-| 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 |
+| 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`) |
-| 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` |
+| 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 |
**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
diff --git a/docs/internals/guardrails.md b/docs/internals/guardrails.md
index c20f511..01e6956 100644
--- a/docs/internals/guardrails.md
+++ b/docs/internals/guardrails.md
@@ -369,7 +369,7 @@ type LocalTools = 'allowlist' | 'sandbox' | 'host'
| **Grok Build** | `allowlist` | A per-run built-in tool list, `--permission-mode dontAsk`, deny rules, and a kernel sandbox for writes. Reads are closed by the tool list and a `Read` deny rather than the sandbox, and MCP servers the user set up in Grok itself still load |
| **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** | `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 |
**Windows is less measured.** Cursor's kernel sandbox has no Windows backend, so only its deny rules
apply there. Codex's `sandbox_mode="read-only"` and Grok's `--sandbox workspace` are asked for as on
@@ -449,6 +449,43 @@ Three behaviours of OpenCode shaped the rest, each measured against 1.18.32:
tools hidden from the model. And the user's own `~/.config/opencode/AGENTS.md` is read, as Claude
Code reads the user's `CLAUDE.md`.
+#### OpenCode 2: one of two ways of being contained
+
+OpenCode 2 (`@opencode/cli`) installs the same `opencode` command as OpenCode 1 with other flags, config
+and storage, so OpenCode has two rule sets in `CONTAINMENT`, and **a plan meets one of them in full**
+or is held to the one it came closest to. Each line also refuses the other's flags at parsing (2.0.32:
+*Unrecognized flag: --pure*; 1.19.3 prints its usage, and refuses 2.x's config with *V2 permissions are
+not supported*), so a plan built for the wrong line never starts. What 2.x changed, measured against
+2.0.32:
+
+- **`opencode run` hands the turn to a shared background service** unless it is given
+ `--standalone`, which starts a private server that inherits the run's environment and ends with it.
+ Without it the run's config would be ignored, and a service not yet running would be started with
+ the run's environment and keep it. Every 2.x plan carries `--standalone`, or is `opencode acp`, which
+ always starts its own.
+- **Its rules are a list where the last match wins**, so a later allow could sit beside the deny-all
+ and a substring search would still find the deny. An `inspect` step parses `OPENCODE_CONFIG_CONTENT`
+ instead: the agent's first rule denies everything; every later allow names a `browsentic_*` tool, a
+ web tool, or (a one-shot) `read` on `tmp/*`; the browser's server is the only one, with
+ `codemode: false`; `integration.use` policies end on denying `plugin:*` and `mcp:*`; a conversation
+ starts in `browsentic-contained` (`default_agent`); sharing is off.
+- **Code Mode is on by default**: an MCP server's tools are listed behind one `execute` tool, which runs
+ a script with the network. The rules deny it, and the browser's server sets `codemode: false`.
+- **`run` moves to `$PWD` before its own working directory**, so `envIsCwd` requires `PWD` to name the
+ plan's folder. The project switch is `OPENCODE_CONFIG_PROJECT_DISABLE` (the old name is read only when
+ the new one is unset, so both are set), and `OPENCODE_DISABLE_SHARE` is gone; `"share": "disabled"`
+ remains.
+- **Sign-ins and sessions share one database.** Runs keep their own (`opencode-2.db` under
+ `~/.browsentic`), which the user signs in to once, so the user's background service never opens a
+ run's session, which it would otherwise resume after a run killed mid-turn.
+- **MCP servers connect in the background**, and each step's tools are fixed when it is sent, so a
+ prompt sent at once reaches a model with no browser tools. A held `opencode acp` turn sends the
+ message only once the daemon has answered the run's MCP server for its tool list, and checks that
+ the session is in Browsentic's agent before it does.
+
+**What remains on 2.x:** a policy in the user's global or managed config outranks the run's, so a user
+who allows a plugin there gets it loaded; its tools are still denied by the agent's rules.
+
`localTools` records which case each runner is in, and the note is logged once per run, so the weak
one is **visible in the log rather than assumed away**.
@@ -458,7 +495,7 @@ 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 turning
a 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. `Read` is deliberately left out of `task`'s deny list, because some tasks are handed a
+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.
The file analyst adds three things. A file is screened before anything spawns, so an archive, an
@@ -466,7 +503,7 @@ executable or an oversized file never reaches an agent. The one-shot is stopped
is in rather than left to exit, and its copy of the file is deleted. And the session is not kept
where 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.
+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.
@@ -500,9 +537,11 @@ agent another vendor's key is what sealing exists to prevent, so `check()` repor
`ANTHROPIC_API_KEY` set as *needs setup* rather than promising a login that cannot happen.
OpenCode reads a dozen providers' keys from the environment, and keeping all of them would undo
-sealing for it altogether. It keeps its own logins in a file instead (`opencode auth login` writes
-it), so only `OPENCODE_` survives, and `check()` counts that file, a Zen key, or a provider declared
-in its config as signed in.
+sealing for it altogether. It keeps its own logins instead (`opencode auth login` writes them), so
+only `OPENCODE_` survives. On OpenCode 1 `check()` counts its `auth.json`, a Zen key, or a provider
+declared in its config as signed in. On OpenCode 2 it asks `opencode auth list --standalone` against
+the runs' own database, in the environment a run gets, so a key counts exactly when a run would see
+it; `--standalone` keeps the user's background service from being started with that environment.
Sealing is not absolute for Qwen either. Qwen loads the first `.env` it finds walking up from its
working directory, then `~/.qwen/.env` and `~/.env`, for variables not already present, so a key
diff --git a/docs/reference/errors.md b/docs/reference/errors.md
index e14465e..c085881 100644
--- a/docs/reference/errors.md
+++ b/docs/reference/errors.md
@@ -78,9 +78,9 @@ 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 is signed in to no provider (`opencode auth login`) |
+| `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, 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_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 |
## Guardrails
diff --git a/src/daemon/agent/runner.ts b/src/daemon/agent/runner.ts
index 5b827ac..4524e92 100644
--- a/src/daemon/agent/runner.ts
+++ b/src/daemon/agent/runner.ts
@@ -6,7 +6,7 @@ import { RemoteBridge } from '../remote-bridge';
import { toolHost, type ToolHost } from '../tool-host';
import type { AgentConfig } from './config';
import { runJson, runStream, type JsonOptions } from './runners/drive';
-import { mcpServerFor, runnerFor } from './runners';
+import { mcpServerFor, runnerFor, versionOf } from './runners';
import { RunError, type RunOutcome } from './runners/types';
export { RunError };
@@ -24,6 +24,10 @@ export interface RunRequest {
sessionId: string | null;
/** The tools the run's MCP server will offer, by the names it lists them under. */
mcpTools: string[];
+ /** Settles once the run's MCP server has listed those tools to the CLI. */
+ toolsListed: () => Promise<void>;
+ /** The agent CLI's `--version` line, from the probe that found it ready. */
+ version?: string;
signal: AbortSignal;
emit: (event: RunEvent) => void;
}
@@ -46,6 +50,8 @@ export async function runInstruction(request: RunRequest): Promise<RunOutcome> {
mcp: mcpServerFor(request.runId, runner.limits),
mcpTools: request.mcpTools,
tools: tools.open,
+ toolsListed: request.toolsListed,
+ version: request.version,
},
request.signal,
request.emit,
@@ -75,14 +81,15 @@ function inProcessTools(runId: string, config: AgentConfig): { open: () => Promi
}
/** One-shot, no browser: summarize a file, name a conversation, read back a recording. */
-export function runAgentJson(
+export async function runAgentJson(
prompt: string,
config: AgentConfig,
signal: AbortSignal,
{ reads = false, image, ...options }: JsonOptions & { reads?: boolean; image?: string },
): Promise<string> {
const { runner, settings } = runnerFor(config);
- return runJson(runner, { prompt, settings, reads, image, workspace: runner.workspace('task') }, signal, options);
+ const version = await versionOf(runner.kind, settings.bin);
+ return runJson(runner, { prompt, settings, version, reads, image, workspace: runner.workspace('task') }, signal, options);
}
/** Where a one-shot's scratch files go — inside the agent's own workspace, so it is allowed to read them. */
diff --git a/src/daemon/agent/runners/fixtures/README.md b/src/daemon/agent/runners/fixtures/README.md
index 4e61a95..a43af45 100644
--- a/src/daemon/agent/runners/fixtures/README.md
+++ b/src/daemon/agent/runners/fixtures/README.md
@@ -10,7 +10,10 @@ What each agent CLI printed on stdout, one event per line, for the reader tests
commands emptied. OpenCode's were recorded through the runner's own plan with the model provider
replaced by a local stand-in, so the model's words are scripted and its URL is 127.0.0.1 — all but
`free-tier`, which the real Zen endpoint answered and which lost only its `cf-ray` and
- `cf-placement` headers; paths became `/Users/you`. Codex's `web-search`, `shell` and `two-calls`
+ `cf-placement` headers; paths became `/Users/you`. OpenCode 2.0.32's were recorded the same way,
+ from `opencode run --standalone` and from `opencode acp`; the `acp-` ones keep only what the CLI
+ printed, with its replies renumbered to the request ids the runner uses, and both `free-tier`
+ recordings were answered by the real Zen endpoint. Codex's `web-search`, `shell` and `two-calls`
and Claude Code's `mcp-failed` were recorded the same way, each saying so in its first line.
- `<case>.hand-written.jsonl` was written from the reader code, for streams a contained run cannot
produce or a CLI that was not installed. Its first line says so.
diff --git a/src/daemon/agent/runners/fixtures/opencode/2.0.32-acp-cancelled.jsonl b/src/daemon/agent/runners/fixtures/opencode/2.0.32-acp-cancelled.jsonl
new file mode 100644
index 0000000..edbc015
--- /dev/null
+++ b/src/daemon/agent/runners/fixtures/opencode/2.0.32-acp-cancelled.jsonl
@@ -0,0 +1,9 @@
+# Printed by `opencode acp` 2.0.32 to a client speaking the runner's protocol, its request ids renumbered to the runner's own.
+{"jsonrpc":"2.0","id":1,"result":{"protocolVersion":1,"agentCapabilities":{"loadSession":true,"mcpCapabilities":{"http":true,"sse":false},"promptCapabilities":{"embeddedContext":true,"image":true},"sessionCapabilities":{"additionalDirectories":{},"close":{},"delete":{},"fork":{},"list":{},"resume":{}},"_meta":{"opencode/child-session-updates":true}},"authMethods":[{"description":"Run `opencode auth login` in the terminal","name":"Login with opencode","id":"opencode-login"}],"agentInfo":{"name":"OpenCode","version":"2.0.32"}}}
+{"jsonrpc":"2.0","id":3,"result":{"sessionId":"ses_eb66d8899ffe4YnGsxL9MeYX0k","configOptions":[{"id":"model","name":"Model","category":"model","type":"select","currentValue":"mock/m1","options":[{"value":"mock/m1","name":"mock/Mock One"},{"value":"opencode/beam-free","name":"opencode/Beam Free"},{"value":"opencode/big-pickle","name":"opencode/Big Pickle"},{"value":"opencode/deepseek-v4.1-flash-free","name":"opencode/DeepSeek V4.1 Flash Free"},{"value":"opencode/glm-5.3-flash-free","name":"opencode/GLM-5.3-Flash Free"},{"value":"opencode/longcat-2.5-preview-free","name":"opencode/LongCat 2.5 Preview Free"},{"value":"opencode/mimo-v2.6-flash-free","name":"opencode/MiMo-V2.6-Flash Free"},{"value":"opencode/nemotron-3-ultra-free","name":"opencode/Nemotron 3 Ultra Free"},{"value":"opencode/nemotron-3.5-lightning-free","name":"opencode/Nemotron 3.5 Lightning Free"},{"value":"opencode/nemotron-3.5-super-vl-free","name":"opencode/Nemotron 3.5 Super VL Free"},{"value":"opencode/space-bunny-free","name":"opencode/Space Bunny Free"}]},{"id":"effort","name":"Effort","description":"Available effort levels for this model","category":"thought_level","type":"select","currentValue":"default","options":[{"value":"low","name":"Low"},{"value":"high","name":"High"},{"value":"default","name":"Default"}]},{"id":"mode","name":"Session Mode","category":"mode","type":"select","currentValue":"browsentic-contained","options":[{"value":"browsentic-contained","name":"browsentic-contained"},{"value":"build","name":"Build","description":"The default agent. Executes tools based on configured permissions."},{"value":"plan","name":"Plan","description":"Read-only agent for exploring the codebase and planning work before implementation."}]}]}}
+{"jsonrpc":"2.0","method":"session/update","params":{"sessionId":"ses_eb66d8899ffe4YnGsxL9MeYX0k","update":{"sessionUpdate":"available_commands_update","availableCommands":[{"name":"init","description":"guided AGENTS.md setup"},{"name":"review","description":"review changes [commit|branch|pr], defaults to uncommitted"},{"name":"compact","description":"Compact the session"}]}}}
+{"jsonrpc":"2.0","method":"session/update","params":{"sessionId":"ses_eb66d8899ffe4YnGsxL9MeYX0k","update":{"sessionUpdate":"tool_call","toolCallId":"call_25_0","name":"browsentic_page_sleep","title":"browsentic_page_sleep","kind":"other","status":"pending","locations":[],"rawInput":{}}}}
+{"jsonrpc":"2.0","method":"session/update","params":{"sessionId":"ses_eb66d8899ffe4YnGsxL9MeYX0k","update":{"sessionUpdate":"tool_call_update","toolCallId":"call_25_0","status":"in_progress","kind":"other","title":"browsentic_page_sleep","locations":[],"rawInput":{"seconds":30}}}}
+{"jsonrpc":"2.0","method":"session/update","params":{"sessionId":"ses_eb66d8899ffe4YnGsxL9MeYX0k","update":{"sessionUpdate":"tool_call_update","toolCallId":"call_25_0","status":"failed","kind":"other","title":"browsentic_page_sleep","locations":[],"rawInput":{"seconds":30},"content":[{"type":"content","content":{"type":"text","text":"Tool execution interrupted"}}],"rawOutput":{"metadata":{},"error":"Tool execution interrupted"}}}}
+{"jsonrpc":"2.0","method":"session/update","params":{"sessionId":"ses_eb66d8899ffe4YnGsxL9MeYX0k","update":{"sessionUpdate":"usage_update","used":1230,"size":200000,"cost":{"amount":0,"currency":"USD"}}}}
+{"jsonrpc":"2.0","id":6,"result":{"stopReason":"cancelled","usage":{"inputTokens":1200,"outputTokens":30,"totalTokens":1230},"_meta":{}}}
diff --git a/src/daemon/agent/runners/fixtures/opencode/2.0.32-acp-free-tier.jsonl b/src/daemon/agent/runners/fixtures/opencode/2.0.32-acp-free-tier.jsonl
new file mode 100644
index 0000000..04a6548
--- /dev/null
+++ b/src/daemon/agent/runners/fixtures/opencode/2.0.32-acp-free-tier.jsonl
@@ -0,0 +1,5 @@
+# Printed by `opencode acp` 2.0.32 to a client speaking the runner's protocol, its request ids renumbered to the runner's own.
+{"jsonrpc":"2.0","id":1,"result":{"protocolVersion":1,"agentCapabilities":{"loadSession":true,"mcpCapabilities":{"http":true,"sse":false},"promptCapabilities":{"embeddedContext":true,"image":true},"sessionCapabilities":{"additionalDirectories":{},"close":{},"delete":{},"fork":{},"list":{},"resume":{}},"_meta":{"opencode/child-session-updates":true}},"authMethods":[{"description":"Run `opencode auth login` in the terminal","name":"Login with opencode","id":"opencode-login"}],"agentInfo":{"name":"OpenCode","version":"2.0.32"}}}
+{"jsonrpc":"2.0","id":3,"result":{"sessionId":"ses_eb66c8358ffea1P2YVfgTqXf8T","configOptions":[{"id":"model","name":"Model","category":"model","type":"select","currentValue":"opencode/big-pickle","options":[{"value":"mock/m1","name":"mock/Mock One"},{"value":"opencode/beam-free","name":"opencode/Beam Free"},{"value":"opencode/big-pickle","name":"opencode/Big Pickle"},{"value":"opencode/deepseek-v4.1-flash-free","name":"opencode/DeepSeek V4.1 Flash Free"},{"value":"opencode/glm-5.3-flash-free","name":"opencode/GLM-5.3-Flash Free"},{"value":"opencode/longcat-2.5-preview-free","name":"opencode/LongCat 2.5 Preview Free"},{"value":"opencode/mimo-v2.6-flash-free","name":"opencode/MiMo-V2.6-Flash Free"},{"value":"opencode/nemotron-3-ultra-free","name":"opencode/Nemotron 3 Ultra Free"},{"value":"opencode/nemotron-3.5-lightning-free","name":"opencode/Nemotron 3.5 Lightning Free"},{"value":"opencode/nemotron-3.5-super-vl-free","name":"opencode/Nemotron 3.5 Super VL Free"},{"value":"opencode/space-bunny-free","name":"opencode/Space Bunny Free"}]},{"id":"mode","name":"Session Mode","category":"mode","type":"select","currentValue":"browsentic-contained","options":[{"value":"browsentic-contained","name":"browsentic-contained"},{"value":"build","name":"Build","description":"The default agent. Executes tools based on configured permissions."},{"value":"plan","name":"Plan","description":"Read-only agent for exploring the codebase and planning work before implementation."}]}]}}
+{"jsonrpc":"2.0","method":"session/update","params":{"sessionId":"ses_eb66c8358ffea1P2YVfgTqXf8T","update":{"sessionUpdate":"available_commands_update","availableCommands":[{"name":"init","description":"guided AGENTS.md setup"},{"name":"review","description":"review changes [commit|branch|pr], defaults to uncommitted"},{"name":"compact","description":"Compact the session"}]}}}
+{"jsonrpc":"2.0","id":6,"error":{"code":-32000,"message":"Authentication required: provider authentication required","data":{}}}
diff --git a/src/daemon/agent/runners/fixtures/opencode/2.0.32-acp-refused-tool.jsonl b/src/daemon/agent/runners/fixtures/opencode/2.0.32-acp-refused-tool.jsonl
new file mode 100644
index 0000000..556829f
--- /dev/null
+++ b/src/daemon/agent/runners/fixtures/opencode/2.0.32-acp-refused-tool.jsonl
@@ -0,0 +1,10 @@
+# Printed by `opencode acp` 2.0.32 to a client speaking the runner's protocol, its request ids renumbered to the runner's own.
+{"jsonrpc":"2.0","id":1,"result":{"protocolVersion":1,"agentCapabilities":{"loadSession":true,"mcpCapabilities":{"http":true,"sse":false},"promptCapabilities":{"embeddedContext":true,"image":true},"sessionCapabilities":{"additionalDirectories":{},"close":{},"delete":{},"fork":{},"list":{},"resume":{}},"_meta":{"opencode/child-session-updates":true}},"authMethods":[{"description":"Run `opencode auth login` in the terminal","name":"Login with opencode","id":"opencode-login"}],"agentInfo":{"name":"OpenCode","version":"2.0.32"}}}
+{"jsonrpc":"2.0","id":3,"result":{"sessionId":"ses_eb66d50beffeuQA7eYDFVfYAvC","configOptions":[{"id":"model","name":"Model","category":"model","type":"select","currentValue":"mock/m1","options":[{"value":"mock/m1","name":"mock/Mock One"},{"value":"opencode/beam-free","name":"opencode/Beam Free"},{"value":"opencode/big-pickle","name":"opencode/Big Pickle"},{"value":"opencode/deepseek-v4.1-flash-free","name":"opencode/DeepSeek V4.1 Flash Free"},{"value":"opencode/glm-5.3-flash-free","name":"opencode/GLM-5.3-Flash Free"},{"value":"opencode/longcat-2.5-preview-free","name":"opencode/LongCat 2.5 Preview Free"},{"value":"opencode/mimo-v2.6-flash-free","name":"opencode/MiMo-V2.6-Flash Free"},{"value":"opencode/nemotron-3-ultra-free","name":"opencode/Nemotron 3 Ultra Free"},{"value":"opencode/nemotron-3.5-lightning-free","name":"opencode/Nemotron 3.5 Lightning Free"},{"value":"opencode/nemotron-3.5-super-vl-free","name":"opencode/Nemotron 3.5 Super VL Free"},{"value":"opencode/space-bunny-free","name":"opencode/Space Bunny Free"}]},{"id":"effort","name":"Effort","description":"Available effort levels for this model","category":"thought_level","type":"select","currentValue":"default","options":[{"value":"low","name":"Low"},{"value":"high","name":"High"},{"value":"default","name":"Default"}]},{"id":"mode","name":"Session Mode","category":"mode","type":"select","currentValue":"browsentic-contained","options":[{"value":"browsentic-contained","name":"browsentic-contained"},{"value":"build","name":"Build","description":"The default agent. Executes tools based on configured permissions."},{"value":"plan","name":"Plan","description":"Read-only agent for exploring the codebase and planning work before implementation."}]}]}}
+{"jsonrpc":"2.0","method":"session/update","params":{"sessionId":"ses_eb66d50beffeuQA7eYDFVfYAvC","update":{"sessionUpdate":"available_commands_update","availableCommands":[{"name":"init","description":"guided AGENTS.md setup"},{"name":"review","description":"review changes [commit|branch|pr], defaults to uncommitted"},{"name":"compact","description":"Compact the session"}]}}}
+{"jsonrpc":"2.0","method":"session/update","params":{"sessionId":"ses_eb66d50beffeuQA7eYDFVfYAvC","update":{"sessionUpdate":"tool_call","toolCallId":"call_27_0","name":"shell","title":"shell","kind":"execute","status":"pending","locations":[{"path":"/Users/you/.browsentic/agents/opencode/run/conversation-1"}],"rawInput":{"cwd":"/Users/you/.browsentic/agents/opencode/run/conversation-1"}}}}
+{"jsonrpc":"2.0","method":"session/update","params":{"sessionId":"ses_eb66d50beffeuQA7eYDFVfYAvC","update":{"sessionUpdate":"tool_call_update","toolCallId":"call_27_0","status":"in_progress","kind":"execute","title":"echo PWNED > /tmp/pwned-oc2","locations":[{"path":"/Users/you/.browsentic/agents/opencode/run/conversation-1"}],"rawInput":{"command":"echo PWNED > /tmp/pwned-oc2","cwd":"/Users/you/.browsentic/agents/opencode/run/conversation-1"}}}}
+{"jsonrpc":"2.0","method":"session/update","params":{"sessionId":"ses_eb66d50beffeuQA7eYDFVfYAvC","update":{"sessionUpdate":"tool_call_update","toolCallId":"call_27_0","status":"failed","kind":"execute","title":"echo PWNED > /tmp/pwned-oc2","locations":[{"path":"/Users/you/.browsentic/agents/opencode/run/conversation-1"}],"rawInput":{"command":"echo PWNED > /tmp/pwned-oc2","cwd":"/Users/you/.browsentic/agents/opencode/run/conversation-1"},"content":[{"type":"content","content":{"type":"text","text":"No tool named \"shell\" is currently available. Please use a tool from the available tool list."}}],"rawOutput":{"metadata":{},"error":"No tool named \"shell\" is currently available. Please use a tool from the available tool list."}}}}
+{"jsonrpc":"2.0","method":"session/update","params":{"sessionId":"ses_eb66d50beffeuQA7eYDFVfYAvC","update":{"sessionUpdate":"agent_message_chunk","messageId":"msg_14992b1e0001xV5fqQ2i1CUcxz","content":{"type":"text","text":"ok"}}}}
+{"jsonrpc":"2.0","method":"session/update","params":{"sessionId":"ses_eb66d50beffeuQA7eYDFVfYAvC","update":{"sessionUpdate":"usage_update","used":1230,"size":200000,"cost":{"amount":0,"currency":"USD"}}}}
+{"jsonrpc":"2.0","id":6,"result":{"stopReason":"end_turn","usage":{"inputTokens":2400,"outputTokens":60,"totalTokens":2460},"_meta":{}}}
diff --git a/src/daemon/agent/runners/fixtures/opencode/2.0.32-acp-resume-missing.jsonl b/src/daemon/agent/runners/fixtures/opencode/2.0.32-acp-resume-missing.jsonl
new file mode 100644
index 0000000..eb70fa1
--- /dev/null
+++ b/src/daemon/agent/runners/fixtures/opencode/2.0.32-acp-resume-missing.jsonl
@@ -0,0 +1,4 @@
+# Printed by `opencode acp` 2.0.32 to a client speaking the runner's protocol, its request ids renumbered to the runner's own.
+{"jsonrpc":"2.0","id":1,"result":{"protocolVersion":1,"agentCapabilities":{"loadSession":true,"mcpCapabilities":{"http":true,"sse":false},"promptCapabilities":{"embeddedContext":true,"image":true},"sessionCapabilities":{"additionalDirectories":{},"close":{},"delete":{},"fork":{},"list":{},"resume":{}},"_meta":{"opencode/child-session-updates":true}},"authMethods":[{"description":"Run `opencode auth login` in the terminal","name":"Login with opencode","id":"opencode-login"}],"agentInfo":{"name":"OpenCode","version":"2.0.32"}}}
+{"jsonrpc":"2.0","id":2,"error":{"code":-32602,"message":"Invalid params: session not found: ses_0000000000000000000000000","data":{"sessionId":"ses_0000000000000000000000000"}}}
+{"jsonrpc":"2.0","id":6,"error":{"code":-32602,"message":"Invalid params: session not found: ses_0000000000000000000000000","data":{"sessionId":"ses_0000000000000000000000000"}}}
diff --git a/src/daemon/agent/runners/fixtures/opencode/2.0.32-acp-shell-ran.jsonl b/src/daemon/agent/runners/fixtures/opencode/2.0.32-acp-shell-ran.jsonl
new file mode 100644
index 0000000..2beaae3
--- /dev/null
+++ b/src/daemon/agent/runners/fixtures/opencode/2.0.32-acp-shell-ran.jsonl
@@ -0,0 +1,11 @@
+# Printed by `opencode acp` 2.0.32 to a client speaking the runner's protocol, its request ids renumbered to the runner's own.
+{"jsonrpc":"2.0","id":1,"result":{"protocolVersion":1,"agentCapabilities":{"loadSession":true,"mcpCapabilities":{"http":true,"sse":false},"promptCapabilities":{"embeddedContext":true,"image":true},"sessionCapabilities":{"additionalDirectories":{},"close":{},"delete":{},"fork":{},"list":{},"resume":{}},"_meta":{"opencode/child-session-updates":true}},"authMethods":[{"description":"Run `opencode auth login` in the terminal","name":"Login with opencode","id":"opencode-login"}],"agentInfo":{"name":"OpenCode","version":"2.0.32"}}}
+{"jsonrpc":"2.0","id":3,"result":{"sessionId":"ses_eb66cc1acffe44TrkD7NaXqtba","configOptions":[{"id":"model","name":"Model","category":"model","type":"select","currentValue":"mock/m1","options":[{"value":"mock/m1","name":"mock/Mock One"},{"value":"opencode/beam-free","name":"opencode/Beam Free"},{"value":"opencode/big-pickle","name":"opencode/Big Pickle"},{"value":"opencode/deepseek-v4.1-flash-free","name":"opencode/DeepSeek V4.1 Flash Free"},{"value":"opencode/glm-5.3-flash-free","name":"opencode/GLM-5.3-Flash Free"},{"value":"opencode/longcat-2.5-preview-free","name":"opencode/LongCat 2.5 Preview Free"},{"value":"opencode/mimo-v2.6-flash-free","name":"opencode/MiMo-V2.6-Flash Free"},{"value":"opencode/nemotron-3-ultra-free","name":"opencode/Nemotron 3 Ultra Free"},{"value":"opencode/nemotron-3.5-lightning-free","name":"opencode/Nemotron 3.5 Lightning Free"},{"value":"opencode/nemotron-3.5-super-vl-free","name":"opencode/Nemotron 3.5 Super VL Free"},{"value":"opencode/space-bunny-free","name":"opencode/Space Bunny Free"}]},{"id":"effort","name":"Effort","description":"Available effort levels for this model","category":"thought_level","type":"select","currentValue":"default","options":[{"value":"low","name":"Low"},{"value":"high","name":"High"},{"value":"default","name":"Default"}]},{"id":"mode","name":"Session Mode","category":"mode","type":"select","currentValue":"browsentic-contained","options":[{"value":"browsentic-contained","name":"browsentic-contained"},{"value":"build","name":"Build","description":"The default agent. Executes tools based on configured permissions."},{"value":"plan","name":"Plan","description":"Read-only agent for exploring the codebase and planning work before implementation."}]}]}}
+{"jsonrpc":"2.0","method":"session/update","params":{"sessionId":"ses_eb66cc1acffe44TrkD7NaXqtba","update":{"sessionUpdate":"available_commands_update","availableCommands":[{"name":"init","description":"guided AGENTS.md setup"},{"name":"review","description":"review changes [commit|branch|pr], defaults to uncommitted"},{"name":"compact","description":"Compact the session"}]}}}
+{"jsonrpc":"2.0","method":"session/update","params":{"sessionId":"ses_eb66cc1acffe44TrkD7NaXqtba","update":{"sessionUpdate":"tool_call","toolCallId":"call_33_0","name":"shell","title":"shell","kind":"execute","status":"pending","locations":[{"path":"/Users/you/.browsentic/agents/opencode/run/conversation-1"}],"rawInput":{"cwd":"/Users/you/.browsentic/agents/opencode/run/conversation-1"}}}}
+{"jsonrpc":"2.0","method":"session/update","params":{"sessionId":"ses_eb66cc1acffe44TrkD7NaXqtba","update":{"sessionUpdate":"tool_call_update","toolCallId":"call_33_0","status":"in_progress","kind":"execute","title":"echo ran > /tmp/oc2-ran","locations":[{"path":"/Users/you/.browsentic/agents/opencode/run/conversation-1"}],"rawInput":{"command":"echo ran > /tmp/oc2-ran","cwd":"/Users/you/.browsentic/agents/opencode/run/conversation-1"}}}}
+{"jsonrpc":"2.0","method":"session/update","params":{"sessionId":"ses_eb66cc1acffe44TrkD7NaXqtba","update":{"sessionUpdate":"tool_call_update","toolCallId":"call_33_0","status":"in_progress","kind":"execute","title":"echo ran > /tmp/oc2-ran","locations":[{"path":"/Users/you/.browsentic/agents/opencode/run/conversation-1"}],"rawInput":{"command":"echo ran > /tmp/oc2-ran","cwd":"/Users/you/.browsentic/agents/opencode/run/conversation-1"}}}}
+{"jsonrpc":"2.0","method":"session/update","params":{"sessionId":"ses_eb66cc1acffe44TrkD7NaXqtba","update":{"sessionUpdate":"tool_call_update","toolCallId":"call_33_0","status":"completed","locations":[{"path":"/Users/you/.browsentic/agents/opencode/run/conversation-1"}],"content":[{"type":"content","content":{"type":"text","text":"(no output)"}}],"rawOutput":{"metadata":{"status":"completed","truncated":false,"exit":0}}}}}
+{"jsonrpc":"2.0","method":"session/update","params":{"sessionId":"ses_eb66cc1acffe44TrkD7NaXqtba","update":{"sessionUpdate":"agent_message_chunk","messageId":"msg_149934102002JSO4gw6pDuNXcH","content":{"type":"text","text":"ok"}}}}
+{"jsonrpc":"2.0","method":"session/update","params":{"sessionId":"ses_eb66cc1acffe44TrkD7NaXqtba","update":{"sessionUpdate":"usage_update","used":1230,"size":200000,"cost":{"amount":0,"currency":"USD"}}}}
+{"jsonrpc":"2.0","id":6,"result":{"stopReason":"end_turn","usage":{"inputTokens":2400,"outputTokens":60,"totalTokens":2460},"_meta":{}}}
diff --git a/src/daemon/agent/runners/fixtures/opencode/2.0.32-acp-turn.jsonl b/src/daemon/agent/runners/fixtures/opencode/2.0.32-acp-turn.jsonl
new file mode 100644
index 0000000..49509af
--- /dev/null
+++ b/src/daemon/agent/runners/fixtures/opencode/2.0.32-acp-turn.jsonl
@@ -0,0 +1,10 @@
+# Printed by `opencode acp` 2.0.32 to a client speaking the runner's protocol, its request ids renumbered to the runner's own.
+{"jsonrpc":"2.0","id":1,"result":{"protocolVersion":1,"agentCapabilities":{"loadSession":true,"mcpCapabilities":{"http":true,"sse":false},"promptCapabilities":{"embeddedContext":true,"image":true},"sessionCapabilities":{"additionalDirectories":{},"close":{},"delete":{},"fork":{},"list":{},"resume":{}},"_meta":{"opencode/child-session-updates":true}},"authMethods":[{"description":"Run `opencode auth login` in the terminal","name":"Login with opencode","id":"opencode-login"}],"agentInfo":{"name":"OpenCode","version":"2.0.32"}}}
+{"jsonrpc":"2.0","id":3,"result":{"sessionId":"ses_eb66f4f81ffeKv3ZUQKMU5Z0Jv","configOptions":[{"id":"model","name":"Model","category":"model","type":"select","currentValue":"mock/m1","options":[{"value":"mock/m1","name":"mock/Mock One"},{"value":"opencode/beam-free","name":"opencode/Beam Free"},{"value":"opencode/big-pickle","name":"opencode/Big Pickle"},{"value":"opencode/deepseek-v4.1-flash-free","name":"opencode/DeepSeek V4.1 Flash Free"},{"value":"opencode/glm-5.3-flash-free","name":"opencode/GLM-5.3-Flash Free"},{"value":"opencode/longcat-2.5-preview-free","name":"opencode/LongCat 2.5 Preview Free"},{"value":"opencode/mimo-v2.6-flash-free","name":"opencode/MiMo-V2.6-Flash Free"},{"value":"opencode/nemotron-3-ultra-free","name":"opencode/Nemotron 3 Ultra Free"},{"value":"opencode/nemotron-3.5-lightning-free","name":"opencode/Nemotron 3.5 Lightning Free"},{"value":"opencode/nemotron-3.5-super-vl-free","name":"opencode/Nemotron 3.5 Super VL Free"},{"value":"opencode/space-bunny-free","name":"opencode/Space Bunny Free"}]},{"id":"effort","name":"Effort","description":"Available effort levels for this model","category":"thought_level","type":"select","currentValue":"default","options":[{"value":"low","name":"Low"},{"value":"high","name":"High"},{"value":"default","name":"Default"}]},{"id":"mode","name":"Session Mode","category":"mode","type":"select","currentValue":"browsentic-contained","options":[{"value":"browsentic-contained","name":"browsentic-contained"},{"value":"build","name":"Build","description":"The default agent. Executes tools based on configured permissions."},{"value":"plan","name":"Plan","description":"Read-only agent for exploring the codebase and planning work before implementation."}]}]}}
+{"jsonrpc":"2.0","method":"session/update","params":{"sessionId":"ses_eb66f4f81ffeKv3ZUQKMU5Z0Jv","update":{"sessionUpdate":"available_commands_update","availableCommands":[{"name":"init","description":"guided AGENTS.md setup"},{"name":"review","description":"review changes [commit|branch|pr], defaults to uncommitted"},{"name":"compact","description":"Compact the session"}]}}}
+{"jsonrpc":"2.0","method":"session/update","params":{"sessionId":"ses_eb66f4f81ffeKv3ZUQKMU5Z0Jv","update":{"sessionUpdate":"tool_call","toolCallId":"call_8_0","name":"browsentic_page_getPageInfo","title":"browsentic_page_getPageInfo","kind":"other","status":"pending","locations":[],"rawInput":{}}}}
+{"jsonrpc":"2.0","method":"session/update","params":{"sessionId":"ses_eb66f4f81ffeKv3ZUQKMU5Z0Jv","update":{"sessionUpdate":"tool_call_update","toolCallId":"call_8_0","status":"in_progress","kind":"other","title":"browsentic_page_getPageInfo","locations":[],"rawInput":{}}}}
+{"jsonrpc":"2.0","method":"session/update","params":{"sessionId":"ses_eb66f4f81ffeKv3ZUQKMU5Z0Jv","update":{"sessionUpdate":"tool_call_update","toolCallId":"call_8_0","status":"completed","locations":[],"content":[{"type":"content","content":{"type":"text","text":"The page is Example Domain. run=a2"}}],"rawOutput":{"metadata":{"truncated":false}}}}}
+{"jsonrpc":"2.0","method":"session/update","params":{"sessionId":"ses_eb66f4f81ffeKv3ZUQKMU5Z0Jv","update":{"sessionUpdate":"agent_message_chunk","messageId":"msg_14990b1e1001V069vb27rxHxX8","content":{"type":"text","text":"The page is Example Domain."}}}}
+{"jsonrpc":"2.0","method":"session/update","params":{"sessionId":"ses_eb66f4f81ffeKv3ZUQKMU5Z0Jv","update":{"sessionUpdate":"usage_update","used":1230,"size":200000,"cost":{"amount":0,"currency":"USD"}}}}
+{"jsonrpc":"2.0","id":6,"result":{"stopReason":"end_turn","usage":{"inputTokens":2400,"outputTokens":60,"totalTokens":2460},"_meta":{}}}
diff --git a/src/daemon/agent/runners/fixtures/opencode/2.0.32-acp-web.jsonl b/src/daemon/agent/runners/fixtures/opencode/2.0.32-acp-web.jsonl
new file mode 100644
index 0000000..8349a3d
--- /dev/null
+++ b/src/daemon/agent/runners/fixtures/opencode/2.0.32-acp-web.jsonl
@@ -0,0 +1,10 @@
+# Printed by `opencode acp` 2.0.32 to a client speaking the runner's protocol, its request ids renumbered to the runner's own.
+{"jsonrpc":"2.0","id":1,"result":{"protocolVersion":1,"agentCapabilities":{"loadSession":true,"mcpCapabilities":{"http":true,"sse":false},"promptCapabilities":{"embeddedContext":true,"image":true},"sessionCapabilities":{"additionalDirectories":{},"close":{},"delete":{},"fork":{},"list":{},"resume":{}},"_meta":{"opencode/child-session-updates":true}},"authMethods":[{"description":"Run `opencode auth login` in the terminal","name":"Login with opencode","id":"opencode-login"}],"agentInfo":{"name":"OpenCode","version":"2.0.32"}}}
+{"jsonrpc":"2.0","id":3,"result":{"sessionId":"ses_eb66cbd2cffegIpmsrNtES31xS","configOptions":[{"id":"model","name":"Model","category":"model","type":"select","currentValue":"mock/m1","options":[{"value":"mock/m1","name":"mock/Mock One"},{"value":"opencode/beam-free","name":"opencode/Beam Free"},{"value":"opencode/big-pickle","name":"opencode/Big Pickle"},{"value":"opencode/deepseek-v4.1-flash-free","name":"opencode/DeepSeek V4.1 Flash Free"},{"value":"opencode/glm-5.3-flash-free","name":"opencode/GLM-5.3-Flash Free"},{"value":"opencode/longcat-2.5-preview-free","name":"opencode/LongCat 2.5 Preview Free"},{"value":"opencode/mimo-v2.6-flash-free","name":"opencode/MiMo-V2.6-Flash Free"},{"value":"opencode/nemotron-3-ultra-free","name":"opencode/Nemotron 3 Ultra Free"},{"value":"opencode/nemotron-3.5-lightning-free","name":"opencode/Nemotron 3.5 Lightning Free"},{"value":"opencode/nemotron-3.5-super-vl-free","name":"opencode/Nemotron 3.5 Super VL Free"},{"value":"opencode/space-bunny-free","name":"opencode/Space Bunny Free"}]},{"id":"effort","name":"Effort","description":"Available effort levels for this model","category":"thought_level","type":"select","currentValue":"default","options":[{"value":"low","name":"Low"},{"value":"high","name":"High"},{"value":"default","name":"Default"}]},{"id":"mode","name":"Session Mode","category":"mode","type":"select","currentValue":"browsentic-contained","options":[{"value":"browsentic-contained","name":"browsentic-contained"},{"value":"build","name":"Build","description":"The default agent. Executes tools based on configured permissions."},{"value":"plan","name":"Plan","description":"Read-only agent for exploring the codebase and planning work before implementation."}]}]}}
+{"jsonrpc":"2.0","method":"session/update","params":{"sessionId":"ses_eb66cbd2cffegIpmsrNtES31xS","update":{"sessionUpdate":"available_commands_update","availableCommands":[{"name":"init","description":"guided AGENTS.md setup"},{"name":"review","description":"review changes [commit|branch|pr], defaults to uncommitted"},{"name":"compact","description":"Compact the session"}]}}}
+{"jsonrpc":"2.0","method":"session/update","params":{"sessionId":"ses_eb66cbd2cffegIpmsrNtES31xS","update":{"sessionUpdate":"tool_call","toolCallId":"call_35_0","name":"webfetch","title":"webfetch","kind":"fetch","status":"pending","locations":[],"rawInput":{}}}}
+{"jsonrpc":"2.0","method":"session/update","params":{"sessionId":"ses_eb66cbd2cffegIpmsrNtES31xS","update":{"sessionUpdate":"tool_call_update","toolCallId":"call_35_0","status":"in_progress","kind":"fetch","title":"webfetch","locations":[],"rawInput":{"url":"http://127.0.0.1:47311/v1/models","format":"text"}}}}
+{"jsonrpc":"2.0","method":"session/update","params":{"sessionId":"ses_eb66cbd2cffegIpmsrNtES31xS","update":{"sessionUpdate":"tool_call_update","toolCallId":"call_35_0","status":"completed","locations":[],"content":[{"type":"content","content":{"type":"text","text":"{\"object\":\"list\",\"data\":[{\"id\":\"m1\",\"object\":\"model\"}]}"}}],"rawOutput":{"metadata":{"contentType":"application/json","truncated":false}}}}}
+{"jsonrpc":"2.0","method":"session/update","params":{"sessionId":"ses_eb66cbd2cffegIpmsrNtES31xS","update":{"sessionUpdate":"agent_message_chunk","messageId":"msg_149934580001mOR7jYJzfPutYh","content":{"type":"text","text":"fetched"}}}}
+{"jsonrpc":"2.0","method":"session/update","params":{"sessionId":"ses_eb66cbd2cffegIpmsrNtES31xS","update":{"sessionUpdate":"usage_update","used":1230,"size":200000,"cost":{"amount":0,"currency":"USD"}}}}
+{"jsonrpc":"2.0","id":6,"result":{"stopReason":"end_turn","usage":{"inputTokens":2400,"outputTokens":60,"totalTokens":2460},"_meta":{}}}
diff --git a/src/daemon/agent/runners/fixtures/opencode/2.0.32-auth-error.jsonl b/src/daemon/agent/runners/fixtures/opencode/2.0.32-auth-error.jsonl
new file mode 100644
index 0000000..e5f0649
--- /dev/null
+++ b/src/daemon/agent/runners/fixtures/opencode/2.0.32-auth-error.jsonl
@@ -0,0 +1,2 @@
+{"type":"step_start","timestamp":1792236626003,"sessionID":"ses_eb65f6048ffeEBcJWXYi1KZz7m","part":{"id":"prt_149a0a053001Day3ABxW9VimNK","sessionID":"ses_eb65f6048ffeEBcJWXYi1KZz7m","messageID":"msg_149a0a046001k8QZ22wBFsQFXg","type":"step-start"}}
+{"type":"error","timestamp":1792236626004,"sessionID":"ses_eb65f6048ffeEBcJWXYi1KZz7m","error":{"type":"provider.auth","message":"Invalid API key provided","status":401,"response":{"body":"{\"error\":{\"message\":\"Invalid API key provided\",\"type\":\"invalid_request_error\"}}"}}}
diff --git a/src/daemon/agent/runners/fixtures/opencode/2.0.32-free-tier.jsonl b/src/daemon/agent/runners/fixtures/opencode/2.0.32-free-tier.jsonl
new file mode 100644
index 0000000..ea58889
--- /dev/null
+++ b/src/daemon/agent/runners/fixtures/opencode/2.0.32-free-tier.jsonl
@@ -0,0 +1 @@
+{"type":"error","timestamp":1792236627562,"sessionID":"ses_eb65f5b53ffeVzWpc3WsmDAAY0","error":{"type":"provider.auth","message":"OpenCode's free tier can only be used from within OpenCode. Download it at https://opencode.ai and keep building.","status":403,"response":{"body":"{\"type\":\"error\",\"error\":{\"type\":\"FreeTierError\",\"message\":\"OpenCode's free tier can only be used from within OpenCode. Download it at https://opencode.ai and keep building.\"}}"}}}
diff --git a/src/daemon/agent/runners/fixtures/opencode/2.0.32-resumed.jsonl b/src/daemon/agent/runners/fixtures/opencode/2.0.32-resumed.jsonl
new file mode 100644
index 0000000..e720cea
--- /dev/null
+++ b/src/daemon/agent/runners/fixtures/opencode/2.0.32-resumed.jsonl
@@ -0,0 +1,2 @@
+{"type":"step_start","timestamp":1792236624669,"sessionID":"ses_eb65f6781ffezd6RAcEQYaQRJC","part":{"id":"prt_149a09b1d001qiOgtteVdQzYQ4","sessionID":"ses_eb65f6781ffezd6RAcEQYaQRJC","messageID":"msg_149a09b0f0011CxTK0bdUwyoJN","type":"step-start"}}
+{"type":"text","timestamp":1792236624676,"sessionID":"ses_eb65f6781ffezd6RAcEQYaQRJC","part":{"id":"prt_149a09b0f0011CxTK0bdUwyoJN_text-0","sessionID":"ses_eb65f6781ffezd6RAcEQYaQRJC","messageID":"msg_149a09b0f0011CxTK0bdUwyoJN","type":"text","text":"Still Example Domain.","time":{"start":1792236624658,"end":1792236624676}}}
diff --git a/src/daemon/agent/runners/fixtures/opencode/2.0.32-shell-ran.jsonl b/src/daemon/agent/runners/fixtures/opencode/2.0.32-shell-ran.jsonl
new file mode 100644
index 0000000..0233086
--- /dev/null
+++ b/src/daemon/agent/runners/fixtures/opencode/2.0.32-shell-ran.jsonl
@@ -0,0 +1,5 @@
+{"type":"step_start","timestamp":1792236625539,"sessionID":"ses_eb65f6216ffevIdD9JDA0OKCxm","part":{"id":"prt_149a09e83001m9FNYkfrHbpM0B","sessionID":"ses_eb65f6216ffevIdD9JDA0OKCxm","messageID":"msg_149a09e74001Bj8F2eUJaA04rg","type":"step-start"}}
+{"type":"tool_use","timestamp":1792236625565,"sessionID":"ses_eb65f6216ffevIdD9JDA0OKCxm","part":{"partID":"prt_149a09e84001nuv7NO8T6UOOqX","sessionID":"ses_eb65f6216ffevIdD9JDA0OKCxm","messageID":"msg_149a09e74001Bj8F2eUJaA04rg","type":"tool","id":"call_71_0","tool":"shell","state":{"status":"completed","input":{"command":"echo ran"},"output":"ran\n","title":"shell","metadata":{"metadata":{"status":"completed","truncated":false,"exit":0},"content":[{"type":"text","text":"ran\n"}],"providerCall":{"executed":false},"providerResult":{"executed":false},"rawInput":"{\"command\":\"echo ran\"}"},"time":{"start":1792236625540,"end":1792236625565}}}}
+{"type":"step_finish","timestamp":1792236625566,"sessionID":"ses_eb65f6216ffevIdD9JDA0OKCxm","part":{"id":"prt_149a09e9e0015QRI7vu5AYsBpz","sessionID":"ses_eb65f6216ffevIdD9JDA0OKCxm","messageID":"msg_149a09e74001Bj8F2eUJaA04rg","type":"step-finish","reason":"tool-calls","cost":0,"tokens":{"input":1200,"output":30,"reasoning":0,"cache":{"read":0,"write":0}}}}
+{"type":"step_start","timestamp":1792236625575,"sessionID":"ses_eb65f6216ffevIdD9JDA0OKCxm","part":{"id":"prt_149a09ea7001x4iRS6ovTjJtZG","sessionID":"ses_eb65f6216ffevIdD9JDA0OKCxm","messageID":"msg_149a09ea1001DFD6A15VN8BSjy","type":"step-start"}}
+{"type":"text","timestamp":1792236625593,"sessionID":"ses_eb65f6216ffevIdD9JDA0OKCxm","part":{"id":"prt_149a09ea1001DFD6A15VN8BSjy_text-0","sessionID":"ses_eb65f6216ffevIdD9JDA0OKCxm","messageID":"msg_149a09ea1001DFD6A15VN8BSjy","type":"text","text":"ok","time":{"start":1792236625572,"end":1792236625593}}}
diff --git a/src/daemon/agent/runners/fixtures/opencode/2.0.32-text-suffix.hand-written.jsonl b/src/daemon/agent/runners/fixtures/opencode/2.0.32-text-suffix.hand-written.jsonl
new file mode 100644
index 0000000..c0bf3a6
--- /dev/null
+++ b/src/daemon/agent/runners/fixtures/opencode/2.0.32-text-suffix.hand-written.jsonl
@@ -0,0 +1,6 @@
+# Written from OpenCode 2.0.32's run/noninteractive.ts: a text block printed when it ended, then the reconcile pass
+# sending the part of it the server had not printed yet, under the block's number, before a second block of its own.
+{"type":"step_start","timestamp":1792235333924,"sessionID":"ses_eb6731d5effeUY30mRea760dHx","part":{"id":"prt_1498ce924001bOBF5r3EVFsXsL","sessionID":"ses_eb6731d5effeUY30mRea760dHx","messageID":"msg_1498ce95f0012chcB9TYehfVSO","type":"step-start"}}
+{"type":"text","timestamp":1792235334004,"sessionID":"ses_eb6731d5effeUY30mRea760dHx","part":{"id":"prt_1498ce966001syNZNzhOevipLx","sessionID":"ses_eb6731d5effeUY30mRea760dHx","messageID":"msg_1498ce95f0012chcB9TYehfVSO","type":"text","text":"The page is Exam","time":{"start":1792235333987,"end":1792235334004}}}
+{"type":"text","timestamp":1792235334010,"sessionID":"ses_eb6731d5effeUY30mRea760dHx","part":{"id":"prt_1498ce95f0012chcB9TYehfVSO_text-0","sessionID":"ses_eb6731d5effeUY30mRea760dHx","messageID":"msg_1498ce95f0012chcB9TYehfVSO","type":"text","text":"ple Domain.","time":{"start":1792235333987,"end":1792235334010}}}
+{"type":"text","timestamp":1792235334011,"sessionID":"ses_eb6731d5effeUY30mRea760dHx","part":{"id":"prt_1498ce95f0012chcB9TYehfVSO_text-1","sessionID":"ses_eb6731d5effeUY30mRea760dHx","messageID":"msg_1498ce95f0012chcB9TYehfVSO","type":"text","text":"It has one link.","time":{"start":1792235333987,"end":1792235334011}}}
diff --git a/src/daemon/agent/runners/fixtures/opencode/2.0.32-turn.jsonl b/src/daemon/agent/runners/fixtures/opencode/2.0.32-turn.jsonl
new file mode 100644
index 0000000..b993499
--- /dev/null
+++ b/src/daemon/agent/runners/fixtures/opencode/2.0.32-turn.jsonl
@@ -0,0 +1,8 @@
+{"type":"step_start","timestamp":1792236624162,"sessionID":"ses_eb65f6781ffezd6RAcEQYaQRJC","part":{"id":"prt_149a099220016pojBSegVVUxRY","sessionID":"ses_eb65f6781ffezd6RAcEQYaQRJC","messageID":"msg_149a09911001oFzEdQRTAatTNs","type":"step-start"}}
+{"type":"tool_use","timestamp":1792236624169,"sessionID":"ses_eb65f6781ffezd6RAcEQYaQRJC","part":{"partID":"prt_149a09923001owa8bh4k0xx6K0","sessionID":"ses_eb65f6781ffezd6RAcEQYaQRJC","messageID":"msg_149a09911001oFzEdQRTAatTNs","type":"tool","id":"call_64_0","tool":"browsentic_page_getPageInfo","state":{"status":"error","input":{},"error":"No tool named \"browsentic_page_getPageInfo\" is currently available. Please use a tool from the available tool list.","metadata":{"providerCall":{"executed":false},"providerResult":{"executed":false},"rawInput":"{}"},"time":{"start":1792236624163,"end":1792236624169}}}}
+{"type":"step_finish","timestamp":1792236624170,"sessionID":"ses_eb65f6781ffezd6RAcEQYaQRJC","part":{"id":"prt_149a0992a001nlWOnPWTfJfSae","sessionID":"ses_eb65f6781ffezd6RAcEQYaQRJC","messageID":"msg_149a09911001oFzEdQRTAatTNs","type":"step-finish","reason":"tool-calls","cost":0,"tokens":{"input":1200,"output":30,"reasoning":0,"cache":{"read":0,"write":0}}}}
+{"type":"step_start","timestamp":1792236624179,"sessionID":"ses_eb65f6781ffezd6RAcEQYaQRJC","part":{"id":"prt_149a09933001lTBXVB3CkKfx6k","sessionID":"ses_eb65f6781ffezd6RAcEQYaQRJC","messageID":"msg_149a0992d001Xoq4jE7tpp0Zst","type":"step-start"}}
+{"type":"tool_use","timestamp":1792236624194,"sessionID":"ses_eb65f6781ffezd6RAcEQYaQRJC","part":{"partID":"prt_149a099340010SULFHnIqCiuJ4","sessionID":"ses_eb65f6781ffezd6RAcEQYaQRJC","messageID":"msg_149a0992d001Xoq4jE7tpp0Zst","type":"tool","id":"call_65_0","tool":"browsentic_page_getPageInfo","state":{"status":"error","input":{},"error":"No tool named \"browsentic_page_getPageInfo\" is currently available. Please use a tool from the available tool list.","metadata":{"providerCall":{"executed":false},"providerResult":{"executed":false},"rawInput":"{}"},"time":{"start":1792236624180,"end":1792236624194}}}}
+{"type":"step_finish","timestamp":1792236624194,"sessionID":"ses_eb65f6781ffezd6RAcEQYaQRJC","part":{"id":"prt_149a0994200204sIDlzRYko2UW","sessionID":"ses_eb65f6781ffezd6RAcEQYaQRJC","messageID":"msg_149a0992d001Xoq4jE7tpp0Zst","type":"step-finish","reason":"tool-calls","cost":0,"tokens":{"input":1200,"output":30,"reasoning":0,"cache":{"read":0,"write":0}}}}
+{"type":"step_start","timestamp":1792236624204,"sessionID":"ses_eb65f6781ffezd6RAcEQYaQRJC","part":{"id":"prt_149a0994c0019kWtSqITEsQDJ4","sessionID":"ses_eb65f6781ffezd6RAcEQYaQRJC","messageID":"msg_149a09944001xpl7xA8NlQYxMV","type":"step-start"}}
+{"type":"text","timestamp":1792236624220,"sessionID":"ses_eb65f6781ffezd6RAcEQYaQRJC","part":{"id":"prt_149a09944001xpl7xA8NlQYxMV_text-0","sessionID":"ses_eb65f6781ffezd6RAcEQYaQRJC","messageID":"msg_149a09944001xpl7xA8NlQYxMV","type":"text","text":"The page is Example Domain.","time":{"start":1792236624199,"end":1792236624220}}}
diff --git a/src/daemon/agent/runners/fixtures/opencode/2.0.32-unknown-model.jsonl b/src/daemon/agent/runners/fixtures/opencode/2.0.32-unknown-model.jsonl
new file mode 100644
index 0000000..720c97b
--- /dev/null
+++ b/src/daemon/agent/runners/fixtures/opencode/2.0.32-unknown-model.jsonl
@@ -0,0 +1 @@
+{"type":"error","timestamp":1792236626408,"sessionID":"ses_eb65f5ea0ffe90TeacFut3SDFQ","error":{"type":"provider.no-route","message":"Model unavailable: mock/zzz"}}
diff --git a/src/daemon/agent/runners/fixtures/opencode/2.0.32-unknown-variant.jsonl b/src/daemon/agent/runners/fixtures/opencode/2.0.32-unknown-variant.jsonl
new file mode 100644
index 0000000..06d620e
--- /dev/null
+++ b/src/daemon/agent/runners/fixtures/opencode/2.0.32-unknown-variant.jsonl
@@ -0,0 +1 @@
+{"type":"error","timestamp":1792236626831,"sessionID":"ses_eb65f5cfdfferjtVdyVO1Kvc48","error":{"type":"provider.no-route","message":"Variant unavailable for mock/m1: ludicrous"}}
diff --git a/src/daemon/agent/runners/fixtures/opencode/2.0.32-web.jsonl b/src/daemon/agent/runners/fixtures/opencode/2.0.32-web.jsonl
new file mode 100644
index 0000000..6e3ba25
--- /dev/null
+++ b/src/daemon/agent/runners/fixtures/opencode/2.0.32-web.jsonl
@@ -0,0 +1,5 @@
+{"type":"step_start","timestamp":1792236625093,"sessionID":"ses_eb65f63d4ffeQOecQz9KH1awbO","part":{"id":"prt_149a09cc5001fektVB5S6fZz7u","sessionID":"ses_eb65f63d4ffeQOecQz9KH1awbO","messageID":"msg_149a09cb6001peUC4hzEvy6CID","type":"step-start"}}
+{"type":"tool_use","timestamp":1792236625103,"sessionID":"ses_eb65f63d4ffeQOecQz9KH1awbO","part":{"partID":"prt_149a09cc600147siaCEb765p6t","sessionID":"ses_eb65f63d4ffeQOecQz9KH1awbO","messageID":"msg_149a09cb6001peUC4hzEvy6CID","type":"tool","id":"call_68_0","tool":"webfetch","state":{"status":"completed","input":{"url":"http://127.0.0.1:47311/v1/models","format":"text"},"output":"{\"object\":\"list\",\"data\":[{\"id\":\"m1\",\"object\":\"model\"}]}","title":"webfetch","metadata":{"metadata":{"contentType":"application/json","truncated":false},"content":[{"type":"text","text":"{\"object\":\"list\",\"data\":[{\"id\":\"m1\",\"object\":\"model\"}]}"}],"providerCall":{"executed":false},"providerResult":{"executed":false},"rawInput":"{\"url\":\"http://127.0.0.1:47311/v1/models\",\"format\":\"text\"}"},"time":{"start":1792236625094,"end":1792236625103}}}}
+{"type":"step_finish","timestamp":1792236625103,"sessionID":"ses_eb65f63d4ffeQOecQz9KH1awbO","part":{"id":"prt_149a09ccf002oDYP9ozidAOrkg","sessionID":"ses_eb65f63d4ffeQOecQz9KH1awbO","messageID":"msg_149a09cb6001peUC4hzEvy6CID","type":"step-finish","reason":"tool-calls","cost":0,"tokens":{"input":1200,"output":30,"reasoning":0,"cache":{"read":0,"write":0}}}}
+{"type":"step_start","timestamp":1792236625113,"sessionID":"ses_eb65f63d4ffeQOecQz9KH1awbO","part":{"id":"prt_149a09cd9001Yc894NMU7fuKHP","sessionID":"ses_eb65f63d4ffeQOecQz9KH1awbO","messageID":"msg_149a09cd2001ro1t6tOUHlwU3w","type":"step-start"}}
+{"type":"text","timestamp":1792236625127,"sessionID":"ses_eb65f63d4ffeQOecQz9KH1awbO","part":{"id":"prt_149a09cd2001ro1t6tOUHlwU3w_text-0","sessionID":"ses_eb65f63d4ffeQOecQz9KH1awbO","messageID":"msg_149a09cd2001ro1t6tOUHlwU3w","type":"text","text":"Fetched.","time":{"start":1792236625110,"end":1792236625127}}}
diff --git a/src/daemon/agent/runners/fixtures/support.ts b/src/daemon/agent/runners/fixtures/support.ts
index 7fffa78..4470561 100644
--- a/src/daemon/agent/runners/fixtures/support.ts
+++ b/src/daemon/agent/runners/fixtures/support.ts
@@ -86,6 +86,7 @@ export const streamContext = (settings: AgentSettings, overrides: Partial<Stream
mcp,
mcpTools: ['page_getPageInfo', 'page_clickElement', 'browsentic_status'],
tools: () => Promise.reject(new Error('no browser tools in a runner test')),
+ toolsListed: () => Promise.resolve(),
...overrides,
});
diff --git a/src/daemon/agent/runners/index.test.ts b/src/daemon/agent/runners/index.test.ts
index cce3f33..8a1883e 100644
--- a/src/daemon/agent/runners/index.test.ts
+++ b/src/daemon/agent/runners/index.test.ts
@@ -7,7 +7,7 @@ import { stateDir } from '../../lockfile';
import type { AgentConfig, AgentSettings } from '../config';
import { MCP_RULE, settingsPath } from './antigravity';
import { stubCli } from './fixtures/support';
-import { agentState, grantRunner, mcpServerFor, refreshModelLists, RUNNERS, runnerFor } from './index';
+import { agentState, grantRunner, mcpServerFor, refreshModelLists, RUNNERS, runnerFor, versionOf } from './index';
const bin = join(stateDir, 'stub-bin');
const probes = join(stateDir, 'probes.log');
@@ -83,6 +83,7 @@ describe('readiness probes', () => {
beforeAll(() => {
mkdirSync(bin, { recursive: true });
for (const name of ['claude', 'claude-next', 'codex', 'agy', 'vibe', 'grok', 'cursor-agent', 'qwen', 'opencode']) stub(name);
+ stub('opencode2', { prints: 'opencode v2.0.32' });
stub('codex-expired', { prints: 'licence expired', exit: 3 });
});
@@ -154,6 +155,23 @@ describe('readiness probes', () => {
expect([withLogin?.ready, (await runner('grok', configWith({})))?.ready]).toEqual([true, true]);
});
+ // Its sign-ins moved into a database of Browsentic's own, which only a 2.x check knows to look in.
+ test('the version an agent printed is the one its readiness is checked for', async () => {
+ vi.stubEnv('OPENCODE_API_KEY', '');
+ rmSync(join(stateDir, 'agents', 'opencode'), { recursive: true, force: true });
+ const status = await runner('opencode', configWith({ opencode: { bin: join(bin, 'opencode2') } }));
+ expect(status).toMatchObject({ version: 'opencode v2.0.32', ready: false, problem: { fix: expect.stringContaining('auth login --standalone') } });
+ });
+
+ test('a one-shot is built for the version last probed, and a binary never probed is probed once', async () => {
+ await agentState(configWith({ opencode: { bin: join(bin, 'opencode2') } }), { refresh: true });
+ expect(await versionOf('opencode', join(bin, 'opencode2'))).toBe('opencode v2.0.32');
+ rmSync(probes, { force: true });
+ expect(await versionOf('opencode', join(bin, 'opencode'))).toBe('opencode 1.0.0');
+ expect(await versionOf('opencode', join(bin, 'opencode'))).toBe('opencode 1.0.0');
+ expect(readFileSync(probes, 'utf8')).toBe('opencode\n');
+ });
+
test('probes are reused for half a minute, even when only the active agent changed', async () => {
await agentState(configWith({}), { refresh: true });
const again = await agentState(configWith({}, 'antigravity'));
diff --git a/src/daemon/agent/runners/index.ts b/src/daemon/agent/runners/index.ts
index f2d4eb3..6eebb11 100644
--- a/src/daemon/agent/runners/index.ts
+++ b/src/daemon/agent/runners/index.ts
@@ -68,6 +68,19 @@ const PROBE_TTL_MS = 30_000;
let cached: { at: number; signature: string; state: AgentState } | null = null;
+/** Each agent's `--version` line as last probed, which outlives the cache a refresh or a grant clears. */
+const probed = new Map<AgentKind, { bin: string; version?: string }>();
+
+/** The version a plan is built for: the one last probed for this bin, or a fresh probe when there is none. */
+export async function versionOf(kind: AgentKind, bin: string): Promise<string | undefined> {
+ const known = probed.get(kind);
+ if (known?.bin === bin) return known.version;
+ const found = await version(bin, RUNNERS[kind].versionArgs);
+ if (!found.ok) return undefined;
+ probed.set(kind, { bin, version: found.detail });
+ return found.detail;
+}
+
export async function agentState(config: AgentConfig, { refresh = false } = {}): Promise<AgentState> {
const signature = JSON.stringify(config.agents);
if (!refresh && cached && cached.signature === signature && Date.now() - cached.at < PROBE_TTL_MS) {
@@ -154,7 +167,8 @@ async function probe(runner: Runner, settings: AgentSettings): Promise<RunnerSta
};
}
- const problem = (await runner.check?.(settings)) ?? null;
+ probed.set(runner.kind, { bin: settings.bin, version: found.detail });
+ const problem = (await runner.check?.(settings, found.detail)) ?? null;
return {
kind: runner.kind,
bin: settings.bin,
diff --git a/src/daemon/agent/runners/opencode-acp.test.ts b/src/daemon/agent/runners/opencode-acp.test.ts
new file mode 100644
index 0000000..ed24d3b
--- /dev/null
+++ b/src/daemon/agent/runners/opencode-acp.test.ts
@@ -0,0 +1,219 @@
+import { afterEach, beforeEach, describe, expect, test, vi } from 'vitest';
+import { opencodeRunner } from './opencode';
+import { streamContext, transcript, type Call } from './fixtures/support';
+import type { Conversation, StreamContext, StreamSink } from './types';
+
+const V2 = 'opencode v2.0.32';
+const SESSION = 'ses_eb66f4f81ffeKv3ZUQKMU5Z0Jv';
+
+interface Written {
+ id?: number;
+ method?: string;
+ params?: Record<string, unknown>;
+ result?: unknown;
+ error?: { code: number };
+}
+
+const frames = (name: string) => transcript('opencode', `2.0.32-acp-${name}.jsonl`);
+const answerTo = (name: string, id: number) => frames(name).find((line) => (JSON.parse(line) as Written).id === id) as string;
+const notices = (name: string) => frames(name).filter((line) => (JSON.parse(line) as Written).method === 'session/update');
+
+/** The session the `turn` recording opened, as a different `mode` or model would have it. */
+function opened(change: (options: { id: string; currentValue: string; options: { value: string }[] }[]) => void, sessionId = SESSION): string {
+ const message = JSON.parse(answerTo('turn', 3)) as { result: { sessionId: string; configOptions: Parameters<typeof change>[0] } };
+ change(message.result.configOptions);
+ message.result.sessionId = sessionId;
+ return JSON.stringify(message);
+}
+
+function holding(overrides: Partial<StreamContext> = {}) {
+ let listTools = () => {};
+ const listed = new Promise<void>((resolve) => (listTools = resolve));
+ const written: Written[] = [];
+ const calls: Call[] = [];
+ const record =
+ (signal: keyof StreamSink) =>
+ (...args: unknown[]) =>
+ void calls.push([signal, ...args]);
+ const sink: StreamSink = {
+ text: record('text'),
+ tool: record('tool'),
+ toolResult: record('toolResult'),
+ session: record('session'),
+ usage: record('usage'),
+ done: record('done'),
+ fail: record('fail'),
+ };
+ const io = { write: (message: unknown) => void written.push(message as Written) };
+ const conversation = opencodeRunner.converse?.(
+ streamContext({ bin: 'opencode' }, { version: V2, toolsListed: () => listed, ...overrides }),
+ ) as Conversation;
+ conversation.open(io);
+ return {
+ conversation,
+ written,
+ calls,
+ listTools,
+ feed: (...lines: string[]) => lines.forEach((line) => conversation.read(line, sink, io)),
+ asked: (method: string) => written.filter((message) => message.method === method),
+ failed: () => calls.filter(([signal]) => signal === 'fail'),
+ };
+}
+
+beforeEach(() => vi.useFakeTimers());
+afterEach(() => vi.useRealTimers());
+
+describe('opening a session', () => {
+ test('a new session is asked for with no MCP server of its own, which the config already names', () => {
+ const turn = holding();
+ turn.feed(answerTo('turn', 1));
+ expect(turn.asked('session/new')).toEqual([
+ { jsonrpc: '2.0', id: 3, method: 'session/new', params: { cwd: turn.conversation.plan.cwd, mcpServers: [] } },
+ ]);
+ });
+
+ // `session/load` would replay the whole history to the panel first.
+ test('a held session is resumed, and one this database never had is opened afresh', () => {
+ const turn = holding({ sessionId: 'ses_0000000000000000000000000' });
+ turn.feed(answerTo('turn', 1));
+ expect(turn.asked('session/resume')[0].params).toMatchObject({ sessionId: 'ses_0000000000000000000000000', mcpServers: [] });
+ turn.feed(answerTo('resume-missing', 2));
+ expect(turn.asked('session/new')).toHaveLength(1);
+ expect(turn.failed()).toEqual([]);
+ });
+
+ test('a session opened in any agent but the contained one stops the run', () => {
+ const turn = holding();
+ const inBuild = opened((options) => {
+ for (const option of options) if (option.id === 'mode') option.currentValue = 'build';
+ });
+ turn.feed(answerTo('turn', 1), inBuild);
+ expect(turn.failed()).toEqual([['fail', 'AGENT_UNSAFE', expect.stringContaining('"build" agent')]]);
+ expect(turn.conversation.holding).toBe(false);
+ });
+
+ test('an OpenCode with no ACP server is turned down, so later turns go through run', () => {
+ const turn = holding({ settings: { bin: '/opt/opencode-no-acp' } });
+ turn.feed(JSON.stringify({ jsonrpc: '2.0', id: 1, error: { code: -32601, message: 'Method not found' } }));
+ expect(turn.failed()).toEqual([['fail', 'CONVERSATION_REFUSED', expect.stringContaining('Method not found')]]);
+ expect(opencodeRunner.converse?.(streamContext({ bin: '/opt/opencode-no-acp' }, { version: V2 }))).toBeNull();
+ });
+});
+
+describe('waiting for the browser tools', () => {
+ // OpenCode connects MCP servers in the background and fixes each step's tools as it is sent.
+ test('the prompt goes only once the run’s tools have reached OpenCode and it has had time to list them', async () => {
+ const turn = holding();
+ turn.feed(answerTo('turn', 1), answerTo('turn', 3));
+ expect(turn.calls).toEqual([['session', SESSION]]);
+ await vi.advanceTimersByTimeAsync(5_000);
+ expect(turn.asked('session/prompt')).toEqual([]);
+ turn.listTools();
+ await vi.advanceTimersByTimeAsync(599);
+ expect(turn.asked('session/prompt')).toEqual([]);
+ await vi.advanceTimersByTimeAsync(1);
+ expect(turn.asked('session/prompt')).toEqual([
+ { jsonrpc: '2.0', id: 6, method: 'session/prompt', params: { sessionId: SESSION, prompt: [{ type: 'text', text: 'what does this page cost' }] } },
+ ]);
+ });
+
+ test('tools that never arrive end the turn rather than leave the model without them', async () => {
+ const turn = holding();
+ turn.feed(answerTo('turn', 1), answerTo('turn', 3));
+ await vi.advanceTimersByTimeAsync(20_000);
+ expect(turn.failed()).toEqual([['fail', 'AGENT_FAILED', expect.stringContaining("could not start Browsentic's browser tools")]]);
+ expect(turn.asked('session/prompt')).toEqual([]);
+ });
+});
+
+describe('choosing the model and the effort', () => {
+ // A model OpenCode does not know is swapped for another without a word.
+ test('a pinned model it offers is asked for by name, and one it does not stops the turn with some it does', () => {
+ const pinned = holding({ settings: { bin: 'opencode', model: 'opencode/big-pickle' } });
+ pinned.feed(answerTo('turn', 1), answerTo('turn', 3));
+ expect(pinned.asked('session/set_config_option')).toEqual([
+ { jsonrpc: '2.0', id: 4, method: 'session/set_config_option', params: { sessionId: SESSION, configId: 'model', value: 'opencode/big-pickle' } },
+ ]);
+ const missing = holding({ settings: { bin: 'opencode', model: 'nowhere/model-x' } });
+ missing.feed(answerTo('turn', 1), answerTo('turn', 3));
+ expect(missing.failed()).toEqual([['fail', 'AGENT_FAILED', expect.stringMatching(/no model "nowhere\/model-x".*It offers mock\/m1, /)]]);
+ });
+
+ test('the model it already has is not asked for again', () => {
+ const turn = holding({ settings: { bin: 'opencode', model: 'mock/m1' } });
+ turn.feed(answerTo('turn', 1), answerTo('turn', 3));
+ expect(turn.asked('session/set_config_option')).toEqual([]);
+ });
+
+ // 2.x refuses a run asking for an effort the model lacks, so only one it lists is asked for.
+ test('an effort the model offers is set, and one it lacks is left out', async () => {
+ const high = holding({ settings: { bin: 'opencode', effort: 'high' } });
+ high.feed(answerTo('turn', 1), answerTo('turn', 3));
+ expect(high.asked('session/set_config_option').map((message) => message.params)).toEqual([{ sessionId: SESSION, configId: 'effort', value: 'high' }]);
+ const unknown = holding({ settings: { bin: 'opencode', effort: 'max' } });
+ unknown.feed(answerTo('turn', 1), answerTo('turn', 3));
+ unknown.listTools();
+ await vi.advanceTimersByTimeAsync(600);
+ expect([unknown.asked('session/set_config_option'), unknown.asked('session/prompt').length]).toEqual([[], 1]);
+ });
+});
+
+describe('a turn', () => {
+ const through = (name: string) => {
+ const turn = holding();
+ turn.feed(answerTo(name, 1), answerTo(name, 3), ...notices(name));
+ const end = answerTo(name, 6);
+ if (end) turn.feed(end);
+ return turn.calls.filter(([signal]) => signal !== 'session');
+ };
+
+ test('says its words, leaves the browser call to the daemon, and ends with what it spent', () => {
+ expect(through('turn')).toEqual([
+ ['text', 'The page is Example Domain.'],
+ ['usage', { contextTokens: 1230, outputTokens: 60 }],
+ ['done', 'end_turn'],
+ ]);
+ });
+
+ test("a research run's web fetch is drawn and closed", () => {
+ expect(through('web').filter(([signal]) => signal === 'tool' || signal === 'toolResult')).toEqual([
+ ['tool', 'call_35_0', 'webfetch'],
+ ['toolResult', 'call_35_0', true],
+ ]);
+ });
+
+ test('a tool the model was never offered is its mistake, and a built-in that ran stops the run', () => {
+ expect(through('refused-tool').filter(([signal]) => signal === 'fail')).toEqual([]);
+ expect(through('shell-ran').filter(([signal]) => signal === 'fail')).toEqual([['fail', 'AGENT_UNSAFE', expect.stringContaining('its own shell tool')]]);
+ });
+
+ test('a turn stopped from Browsentic is cancelled, not finished', () => {
+ expect(through('cancelled').at(-1)).toEqual(['fail', 'CANCELLED', 'Run cancelled.']);
+ });
+
+ // Recorded against the real Zen endpoint with no sign-in: ACP passes on only "Authentication required".
+ test('a model with no sign-in says how to sign Browsentic in', () => {
+ expect(through('free-tier')).toEqual([['fail', 'AGENT_FAILED', expect.stringMatching(/narrowed to the browser.*OPENCODE_DB=.*auth login --standalone/)]]);
+ });
+
+ test("another session's updates are not this turn's", () => {
+ const turn = holding();
+ turn.feed(answerTo('turn', 1), answerTo('turn', 3));
+ turn.feed(JSON.stringify({ jsonrpc: '2.0', method: 'session/update', params: { sessionId: 'ses_other', update: { sessionUpdate: 'agent_message_chunk', content: { type: 'text', text: 'not ours' } } } }));
+ expect(turn.calls).toEqual([['session', SESSION]]);
+ });
+});
+
+describe('what OpenCode asks of Browsentic', () => {
+ test('a permission is never granted, and nothing else it asks is answered', () => {
+ const turn = holding();
+ turn.feed(
+ JSON.stringify({ jsonrpc: '2.0', id: 'p1', method: 'session/request_permission', params: { sessionId: SESSION, options: [] } }),
+ JSON.stringify({ jsonrpc: '2.0', id: 'f1', method: 'fs/read_text_file', params: { path: '/etc/hosts' } }),
+ );
+ expect(turn.written.slice(-2)).toEqual([
+ { jsonrpc: '2.0', id: 'p1', result: { outcome: { outcome: 'cancelled' } } },
+ { jsonrpc: '2.0', id: 'f1', error: { code: -32601, message: 'Browsentic does not answer fs/read_text_file' } },
+ ]);
+ });
+});
diff --git a/src/daemon/agent/runners/opencode-acp.ts b/src/daemon/agent/runners/opencode-acp.ts
new file mode 100644
index 0000000..643603a
--- /dev/null
+++ b/src/daemon/agent/runners/opencode-acp.ts
@@ -0,0 +1,227 @@
+import { MCP_SERVER_NAME } from './claude';
+import { CONVERSATION_REFUSED, type Conversation, type ConversationIO, type Plan, type StreamContext, type StreamSink } from './types';
+import { browserToolsDidNotStart } from './util';
+
+const ASK = { initialize: 1, session: 2, fresh: 3, model: 4, effort: 5, prompt: 6 } as const;
+
+/** JSON-RPC's own code for a method the agent does not know: an OpenCode without this protocol, not a failed turn. */
+const UNKNOWN = -32601;
+
+/** ACP's answer to a parameter it cannot use, which for `session/resume` is a session this database never held. */
+const INVALID = -32602;
+
+/**
+ * OpenCode connects MCP servers in the background and reads its tool list afresh at each step, so a
+ * prompt sent at once is answered by a model offered no browser tools. A run's tools are listed to
+ * OpenCode when its server starts, and OpenCode registers them within a 100 ms debounce; this waits
+ * out that debounce with room to spare. Measured on 2.0.32: 120 ms was always enough, 60 ms never.
+ */
+const SETTLE_MS = 600;
+
+/** How long a run's browser tools may take to reach OpenCode before the turn is given up. */
+const LISTED_WITHIN_MS = 20_000;
+
+const WEB_TOOLS = ['webfetch', 'websearch'];
+
+interface Message {
+ id?: number | string;
+ method?: string;
+ params?: { sessionId?: string; update?: Update };
+ result?: { sessionId?: string; configOptions?: ConfigOption[]; stopReason?: string; usage?: { outputTokens?: number } };
+ error?: { code?: number; message?: string };
+}
+
+interface ConfigOption {
+ id?: string;
+ currentValue?: string;
+ options?: { value?: string }[];
+}
+
+interface Update {
+ sessionUpdate?: string;
+ toolCallId?: string;
+ name?: string;
+ title?: string;
+ status?: string;
+ content?: { type?: string; text?: string };
+ used?: number;
+}
+
+export interface AcpTurn {
+ plan: Plan;
+ /** The agent the config defines, which every session has to be in. */
+ agent: string;
+ declined: (reason: string) => void;
+ explain: (message: string | undefined) => string;
+ stoppedFor: (tool: string) => string;
+}
+
+/**
+ * One turn with OpenCode 2 over the Agent Client Protocol, `opencode acp`, which keeps its own server
+ * for as long as the turn holds it. The run's config comes from its environment as it would for
+ * `opencode run`, but the prompt goes only once the browser tools have reached the session. A session
+ * found in any agent but the run's own is refused, and a built-in that runs stops the run.
+ */
+export function acpTurn(context: StreamContext, turn: AcpTurn): Conversation {
+ let session: string | undefined;
+ let holding = false;
+ let contextTokens = 0;
+ const calls = new Map<string, string>();
+
+ const refuse = (sink: StreamSink, reason: string) => {
+ turn.declined(reason);
+ sink.fail(CONVERSATION_REFUSED, reason);
+ };
+
+ const ask = (io: ConversationIO, id: number, method: string, params: Record<string, unknown>) => io.write({ jsonrpc: '2.0', id, method, params });
+
+ const openSession = (io: ConversationIO, resuming: boolean) =>
+ resuming
+ ? ask(io, ASK.session, 'session/resume', { sessionId: context.sessionId, cwd: turn.plan.cwd, mcpServers: [] })
+ : ask(io, ASK.fresh, 'session/new', { cwd: turn.plan.cwd, mcpServers: [] });
+
+ const option = (options: ConfigOption[] | undefined, id: string) => options?.find((candidate) => candidate.id === id);
+ const offers = (choice: ConfigOption | undefined, value: string) => choice?.options?.some((offered) => offered.value === value) === true;
+
+ const prompt = (io: ConversationIO, sink: StreamSink) => {
+ const late = new Promise<'late'>((resolve) => setTimeout(() => resolve('late'), LISTED_WITHIN_MS).unref());
+ void Promise.race([context.toolsListed(), late]).then((listed) => {
+ if (listed === 'late') return sink.fail('AGENT_FAILED', browserToolsDidNotStart('OpenCode'));
+ setTimeout(() => ask(io, ASK.prompt, 'session/prompt', { sessionId: session, prompt: [{ type: 'text', text: context.instruction }] }), SETTLE_MS);
+ });
+ };
+
+ // OpenCode lists each level of effort only for a model that has it, and refuses a run asking for one it lacks.
+ const chooseEffort = (options: ConfigOption[] | undefined, io: ConversationIO, sink: StreamSink) => {
+ const effort = context.settings.effort;
+ const choice = option(options, 'effort');
+ if (effort && choice?.currentValue !== effort && offers(choice, effort)) {
+ return ask(io, ASK.effort, 'session/set_config_option', { sessionId: session, configId: 'effort', value: effort });
+ }
+ return prompt(io, sink);
+ };
+
+ // A model OpenCode does not know is swapped for another without a word, so a pinned one is asked for by name.
+ const chooseModel = (options: ConfigOption[] | undefined, io: ConversationIO, sink: StreamSink) => {
+ const model = context.settings.model;
+ const choice = option(options, 'model');
+ if (!model || choice?.currentValue === model) return chooseEffort(options, io, sink);
+ if (!offers(choice, model)) {
+ const some = (choice?.options ?? []).flatMap((offered) => (offered.value ? [offered.value] : [])).slice(0, 6);
+ return sink.fail(
+ 'AGENT_FAILED',
+ `OpenCode offers no model "${model}" here.${some.length ? ` It offers ${some.join(', ')}, among others.` : ''} ` +
+ 'Pick one in the Browsentic side panel, spelled provider/model, then try again.',
+ );
+ }
+ return ask(io, ASK.model, 'session/set_config_option', { sessionId: session, configId: 'model', value: model });
+ };
+
+ const onSession = (message: Message, io: ConversationIO, sink: StreamSink) => {
+ const options = message.result?.configOptions;
+ const mode = option(options, 'mode')?.currentValue;
+ if (mode !== turn.agent) {
+ return sink.fail('AGENT_UNSAFE', `OpenCode opened this session as its "${mode ?? 'unnamed'}" agent rather than Browsentic's, so the run was stopped.`);
+ }
+ session = message.result?.sessionId ?? context.sessionId ?? undefined;
+ if (!session) return sink.fail('AGENT_FAILED', 'OpenCode opened a session without naming it.');
+ holding = true;
+ sink.session(session);
+ chooseModel(options, io, sink);
+ };
+
+ const onResponse = (message: Message, io: ConversationIO, sink: StreamSink) => {
+ const failed = message.error;
+ switch (message.id) {
+ case ASK.initialize:
+ if (failed) return refuse(sink, `its ACP server would not start: ${failed.message}`);
+ return openSession(io, context.sessionId !== null);
+
+ case ASK.session:
+ case ASK.fresh:
+ if (failed?.code === INVALID && message.id === ASK.session) return openSession(io, false);
+ if (failed) {
+ const reason = `OpenCode would not ${message.id === ASK.session ? 'resume' : 'start'} the session: ${failed.message}`;
+ return failed.code === UNKNOWN ? refuse(sink, reason) : sink.fail('AGENT_FAILED', turn.explain(failed.message));
+ }
+ return onSession(message, io, sink);
+
+ case ASK.model:
+ if (failed) return sink.fail('AGENT_FAILED', turn.explain(failed.message));
+ return chooseEffort(message.result?.configOptions, io, sink);
+
+ case ASK.effort:
+ return prompt(io, sink);
+
+ case ASK.prompt: {
+ if (failed) return sink.fail('AGENT_FAILED', turn.explain(failed.message));
+ const output = message.result?.usage?.outputTokens;
+ if (output !== undefined) sink.usage({ contextTokens, outputTokens: output });
+ // Stopped from Browsentic, OpenCode ends the prompt itself before it exits.
+ if (message.result?.stopReason === 'cancelled') return sink.fail('CANCELLED', 'Run cancelled.');
+ return sink.done(message.result?.stopReason ?? 'end_turn');
+ }
+ }
+ };
+
+ const onToolCall = (update: Update, sink: StreamSink) => {
+ const id = update.toolCallId;
+ if (!id) return;
+ const name = update.name ?? calls.get(id) ?? update.title ?? '';
+ if (!calls.has(id)) {
+ calls.set(id, name);
+ if (WEB_TOOLS.includes(name)) sink.tool(id, name);
+ }
+ if (update.status !== 'completed' && update.status !== 'failed') return;
+ if (WEB_TOOLS.includes(name)) return sink.toolResult(id, update.status === 'completed');
+ // A refused call is the model reaching for a tool it was never offered; one that ran means the rules did not hold.
+ if (update.status === 'completed' && !name.startsWith(`${MCP_SERVER_NAME}_`)) sink.fail('AGENT_UNSAFE', turn.stoppedFor(name));
+ };
+
+ const onUpdate = (update: Update | undefined, sink: StreamSink) => {
+ switch (update?.sessionUpdate) {
+ case 'agent_message_chunk':
+ if (update.content?.type === 'text' && update.content.text) sink.text(update.content.text);
+ return;
+ case 'tool_call':
+ case 'tool_call_update':
+ return onToolCall(update, sink);
+ case 'usage_update':
+ if (typeof update.used === 'number') contextTokens = update.used;
+ return;
+ }
+ };
+
+ return {
+ plan: turn.plan,
+ get holding() {
+ return holding;
+ },
+ declined: turn.declined,
+ open(io) {
+ ask(io, ASK.initialize, 'initialize', {
+ protocolVersion: 1,
+ clientCapabilities: { fs: { readTextFile: false, writeTextFile: false }, terminal: false },
+ clientInfo: { name: 'browsentic', title: 'Browsentic', version: '1' },
+ });
+ },
+ read(line, sink, io) {
+ let message: Message;
+ try {
+ message = JSON.parse(line) as Message;
+ } catch {
+ return;
+ }
+ if (message.method && message.id !== undefined) {
+ // Nobody is there to approve anything, and a denied tool is never asked about.
+ if (message.method === 'session/request_permission') return io.write({ jsonrpc: '2.0', id: message.id, result: { outcome: { outcome: 'cancelled' } } });
+ return io.write({ jsonrpc: '2.0', id: message.id, error: { code: UNKNOWN, message: `Browsentic does not answer ${message.method}` } });
+ }
+ if (message.method === 'session/update') {
+ if (session && message.params?.sessionId === session) onUpdate(message.params.update, sink);
+ return;
+ }
+ if (!message.method) onResponse(message, io, sink);
+ },
+ };
+}
diff --git a/src/daemon/agent/runners/opencode.test.ts b/src/daemon/agent/runners/opencode.test.ts
index af87863..ba1d2c3 100644
--- a/src/daemon/agent/runners/opencode.test.ts
+++ b/src/daemon/agent/runners/opencode.test.ts
@@ -1,9 +1,9 @@
-import { mkdirSync, rmSync, writeFileSync } from 'node:fs';
-import { join, parse, relative, sep } from 'node:path';
+import { existsSync, mkdirSync, readFileSync, rmSync, writeFileSync } from 'node:fs';
+import { dirname, join, parse, relative, sep } from 'node:path';
import { afterEach, beforeEach, describe, expect, test, vi } from 'vitest';
import { stateDir } from '../../lockfile';
-import { opencodeRunner } from './opencode';
-import { jsonContext, readThrough, shown, streamContext, transcript, valueOf } from './fixtures/support';
+import { isOpenCode2, opencodeRunner, signInLine } from './opencode';
+import { jsonContext, readThrough, shown, streamContext, stubCli, transcript, valueOf } from './fixtures/support';
import type { Plan } from './types';
const settings = { bin: 'opencode' };
@@ -320,3 +320,295 @@ test('an XDG_CONFIG_HOME of the user’s own is honoured', () => {
vi.stubEnv('XDG_CONFIG_HOME', '/elsewhere');
expect(opencodeRunner.skillDirs?.()?.[0]).toBe(join('/elsewhere', 'opencode', 'skills'));
});
+
+const V2 = 'opencode v2.0.32';
+
+describe('which OpenCode it is', () => {
+ test.each([
+ ['opencode v2.0.32', true],
+ ['opencode v10.1.0', true],
+ ['1.19.3', false],
+ ['1.18.32', false],
+ [undefined, false],
+ // A notice printed ahead of the version is no version, and 1.x's plan fails safe on 2.x's parser.
+ ['Update available: opencode v2.1.0', false],
+ ])('%s is 2.x: %s', (version, expected) => {
+ expect(isOpenCode2(version)).toBe(expected);
+ });
+
+ test('1.x holds no conversation, so its turns stay on run', () => {
+ expect(opencodeRunner.converse?.(streamContext(settings, { version: '1.19.3' }))).toBeNull();
+ });
+});
+
+interface Config2 {
+ share: string;
+ default_agent: string;
+ instructions?: string[];
+ policies: { action: string; resource: string; effect: string }[];
+ mcp: { servers: Record<string, { type: string; command: string[]; environment: Record<string, string>; codemode: boolean; timeout: { execution: number } }> };
+ agents: Record<string, { disabled?: boolean; permissions?: { action: string; resource: string; effect: string }[] }>;
+}
+const config2Of = (plan: Plan) => JSON.parse(plan.env?.OPENCODE_CONFIG_CONTENT ?? 'null') as Config2;
+const rules2Of = (plan: Plan) => config2Of(plan).agents['browsentic-contained'].permissions ?? [];
+const conversationOf = (overrides: Parameters<typeof streamContext>[1] = {}) => {
+ const conversation = opencodeRunner.converse?.(streamContext(settings, { version: V2, ...overrides }));
+ if (!conversation) throw new Error('OpenCode 2 held no conversation');
+ return conversation;
+};
+const turn2 = (overrides: Parameters<typeof streamContext>[1] = {}) => conversationOf(overrides).plan;
+const fallback2 = (overrides: Parameters<typeof streamContext>[1] = {}) => opencodeRunner.stream(streamContext(settings, { version: V2, ...overrides }));
+const task2 = (overrides: Parameters<typeof jsonContext>[1] = {}) => opencodeRunner.json(jsonContext(settings, { version: V2, ...overrides }));
+
+describe('an OpenCode 2 turn', () => {
+ test('is one `opencode acp` held for the turn, its config in the environment and its prompt in a file', () => {
+ const plan = shown(turn2());
+ expect({ ...plan, env: { ...plan.env, OPENCODE_CONFIG_CONTENT: '<config>' } }).toMatchInlineSnapshot(`
+ {
+ "args": [
+ "acp",
+ ],
+ "cwd": "<state>/agents/opencode/run/conversation-1",
+ "env": {
+ "BROWSENTIC_AGENT_RUN": "run-1",
+ "OPENCODE_CONFIG_CONTENT": "<config>",
+ "OPENCODE_CONFIG_PROJECT_DISABLE": "1",
+ "OPENCODE_DB": "<state>/agents/opencode/opencode-2.db",
+ "OPENCODE_DISABLE_AUTOUPDATE": "1",
+ "OPENCODE_DISABLE_CLAUDE_CODE": "1",
+ "OPENCODE_DISABLE_PROJECT_CONFIG": "1",
+ "PWD": "<state>/agents/opencode/run/conversation-1",
+ },
+ "files": [
+ {
+ "content": "You are Browsentic.
+
+ A tool result over 100 KB comes back cut, and the saved copy OpenCode points to cannot be opened in this run, so ask for less at a time — \`page_getPageInfo\` with a small \`maxPerKind\`, \`page_extractText\` with the cursor it hands back.
+ ",
+ "path": "instructions.md",
+ },
+ ],
+ }
+ `);
+ });
+
+ // 2.x reads `$PWD` before its own working directory, and the user's sessions are in a database Browsentic never opens.
+ test('it runs in its conversation folder, against a database of its own that 1.x never shares', () => {
+ const plan = turn2();
+ expect([plan.env?.PWD, plan.env?.OPENCODE_DB]).toEqual([plan.cwd, join(stateDir, 'agents', 'opencode', 'opencode-2.db')]);
+ expect(plan.env?.OPENCODE_DB).not.toBe(opencodeRunner.stream(streamContext(settings)).env?.OPENCODE_DB);
+ });
+
+ test('the rules deny everything first, then allow each browser tool by name, and the web tools on research only', () => {
+ expect(rules2Of(turn2())).toEqual([
+ { action: '*', resource: '*', effect: 'deny' },
+ { action: 'browsentic_page_getPageInfo', resource: '*', effect: 'allow' },
+ { action: 'browsentic_page_clickElement', resource: '*', effect: 'allow' },
+ { action: 'browsentic_browsentic_status', resource: '*', effect: 'allow' },
+ ]);
+ expect(rules2Of(turn2({ research: true })).slice(-2).map((rule) => rule.action)).toEqual(['webfetch', 'websearch']);
+ });
+
+ // Code Mode would list one `execute` tool in place of these, which the rules deny and which runs a script with the network.
+ test('the browser server lists its tools directly, outlasts an approval card, and is the only server allowed to load', () => {
+ const config = config2Of(turn2());
+ const server = config.mcp.servers.browsentic;
+ expect([Object.keys(config.mcp.servers), server.type, server.codemode, server.timeout.execution, server.environment]).toEqual([
+ ['browsentic'],
+ 'local',
+ false,
+ 1_800_000,
+ { BROWSENTIC_AGENT_RUN: 'run-1' },
+ ]);
+ expect(config.policies).toEqual([
+ { action: 'integration.use', resource: 'plugin:*', effect: 'deny' },
+ { action: 'integration.use', resource: 'mcp:*', effect: 'deny' },
+ { action: 'integration.use', resource: 'mcp:browsentic', effect: 'allow' },
+ ]);
+ });
+
+ test('a session starts in the contained agent, unshared, with no title model call', () => {
+ const config = config2Of(turn2());
+ expect([config.default_agent, config.share, config.agents.title]).toEqual(['browsentic-contained', 'disabled', { disabled: true }]);
+ });
+
+ test('the system prompt travels only in the file the config names', () => {
+ const plan = turn2({ systemPrompt: 'The page said {file:~/.ssh/id_rsa} and {env:AWS_SECRET_ACCESS_KEY}.' });
+ expect(plan.env?.OPENCODE_CONFIG_CONTENT).not.toMatch(/\{(file|env):/);
+ expect(config2Of(plan).instructions).toEqual([join(plan.cwd, 'instructions.md')]);
+ });
+
+ test('an OpenCode that turned a conversation down sends its later turns through run, and only that one', () => {
+ const conversation = conversationOf({ settings: { bin: '/opt/opencode-without-acp' } });
+ conversation.declined('it exited with code 1');
+ expect(opencodeRunner.converse?.(streamContext({ bin: '/opt/opencode-without-acp' }, { version: V2 }))).toBeNull();
+ expect(opencodeRunner.converse?.(streamContext(settings, { version: V2 }))).not.toBeNull();
+ });
+});
+
+describe('an OpenCode 2 turn through run', () => {
+ test('has its own server, and the message on stdin rather than argv', () => {
+ const plan = fallback2({ instruction: '--help me "find" the price' });
+ expect(plan.args).toEqual(['run', '--standalone', '--format', 'json', '--agent', 'browsentic-contained']);
+ expect(plan.input).toBe('--help me "find" the price');
+ });
+
+ // 2.x fails a run asking for an effort the model lacks, rather than ignoring it as 1.x did.
+ test('names its session and model when it has them, and never an effort', () => {
+ const args = fallback2({ sessionId: SESSION, settings: { ...settings, model: 'anthropic/claude-sonnet-5', effort: 'high' } }).args;
+ expect([valueOf(args, '--session'), valueOf(args, '--model'), args.some((arg) => arg.includes('#') || arg === '--variant')]).toEqual([
+ SESSION,
+ 'anthropic/claude-sonnet-5',
+ false,
+ ]);
+ });
+
+ test('is configured exactly as a held turn is', () => {
+ expect(fallback2().env).toEqual(turn2().env);
+ });
+});
+
+describe('an OpenCode 2 one-shot', () => {
+ test('has no server and no tool, and runs as the contained agent in the task workspace', () => {
+ const plan = task2({ settings: { ...settings, model: 'anthropic/claude-sonnet-5', effort: 'low' } });
+ expect(plan.args).toEqual(['run', '--standalone', '--format', 'json', '--agent', 'browsentic-contained', '--model', 'anthropic/claude-sonnet-5']);
+ expect([plan.input, plan.cwd, plan.env?.PWD]).toEqual(['summarize this', opencodeRunner.workspace('task'), opencodeRunner.workspace('task')]);
+ expect([config2Of(plan).mcp.servers, rules2Of(plan)]).toEqual([{}, [{ action: '*', resource: '*', effect: 'deny' }]]);
+ });
+
+ // Measured: tmp/* opens the scratch folder by either path, and /tmp, /etc/hosts and a ../ are an external directory the deny covers.
+ test('a one-shot handed a file may read its scratch folder and nothing else', () => {
+ expect(rules2Of(task2({ reads: true })).slice(1)).toEqual([{ action: 'read', resource: 'tmp/*', effect: 'allow' }]);
+ });
+});
+
+describe('reading an OpenCode 2 stream', () => {
+ const read2 = (name: string, only?: Parameters<typeof readThrough>[2]) => readThrough(opencodeRunner, transcript('opencode', name), only);
+
+ // The first call reached for a tool before the run's server had connected, which a held turn waits out.
+ test('a turn says its text and what it spent, and a call the model was not yet offered is no breach', () => {
+ expect(read2('2.0.32-turn.jsonl').filter(([signal]) => signal !== 'session')).toEqual([
+ ['usage', { contextTokens: 1230, outputTokens: 30 }],
+ ['usage', { contextTokens: 1230, outputTokens: 60 }],
+ ['text', 'The page is Example Domain.'],
+ ]);
+ });
+
+ test('every line names the session, and a resumed turn names the same one', () => {
+ const turn = new Set(read2('2.0.32-turn.jsonl', 'session').map(([, id]) => id));
+ expect(turn.size).toBe(1);
+ expect(new Set(read2('2.0.32-resumed.jsonl', 'session').map(([, id]) => id))).toEqual(turn);
+ });
+
+ test("a research run's web fetch is drawn by the call's own id", () => {
+ expect(read2('2.0.32-web.jsonl', 'tool')).toEqual([['tool', expect.stringMatching(/^call_/), 'webfetch']]);
+ });
+
+ test('a built-in that ran stops the run', () => {
+ expect(read2('2.0.32-shell-ran.jsonl', 'fail')).toEqual([['fail', 'AGENT_UNSAFE', expect.stringContaining('its own shell tool')]]);
+ });
+
+ test("a text block finished after it was first printed goes on where it stopped, and the next one starts a paragraph", () => {
+ expect(read2('2.0.32-text-suffix.hand-written.jsonl', 'text')).toEqual([
+ ['text', 'The page is Exam'],
+ ['text', 'ple Domain.'],
+ ['text', '\n\nIt has one link.'],
+ ]);
+ expect(opencodeRunner.answer(transcript('opencode', '2.0.32-text-suffix.hand-written.jsonl').join('\n'))).toEqual({
+ text: 'The page is Example Domain.\n\nIt has one link.',
+ });
+ });
+
+ test('a refused key says what the provider said, and how to sign Browsentic in', () => {
+ const [[, code, message]] = read2('2.0.32-auth-error.jsonl', 'fail');
+ expect([code, message]).toEqual(['AGENT_FAILED', expect.stringMatching(/^Invalid API key provided/)]);
+ expect(message).toContain('auth login --standalone');
+ });
+
+ // Recorded against the real Zen endpoint, which still serves its free models only to OpenCode's own tools.
+ test("Zen's free tier refusing a contained run says to sign Browsentic in to a provider", () => {
+ const [[, , message]] = read2('2.0.32-free-tier.jsonl', 'fail');
+ expect(message).toEqual(expect.stringContaining('narrowed to the browser'));
+ expect(message).toContain(`OPENCODE_DB=`);
+ });
+
+ test('a model, or an effort, OpenCode does not have says which to change', () => {
+ expect(read2('2.0.32-unknown-model.jsonl', 'fail')[0][2]).toContain('provider/model');
+ expect(read2('2.0.32-unknown-variant.jsonl', 'fail')[0][2]).toContain('effort');
+ });
+});
+
+describe('what an OpenCode 2 says when it cannot start', () => {
+ // The other line's flags are what a swap under a running Browsentic looks like.
+ test('a flag of the other line asks for the message again', () => {
+ expect(opencodeRunner.hint?.('ERROR\n Unrecognized flag: --pure in command opencode run')).toContain('changed version');
+ expect(opencodeRunner.hint?.('Configuration is invalid ↳ V2 permissions are not supported by OpenCode V1.')).toContain('changed version');
+ });
+});
+
+describe('OpenCode 2 readiness', () => {
+ const db = join(stateDir, 'agents', 'opencode', 'opencode-2.db');
+ const bins = join(stateDir, 'opencode-bins');
+ let seq = 0;
+ /** A stand-in `opencode` that answers `auth list` with `stdout`, and keeps how it was asked. */
+ const answering = (stdout: string, code = 0) => {
+ mkdirSync(bins, { recursive: true });
+ const asked = join(bins, `asked-${++seq}.json`);
+ const bin = stubCli(
+ join(bins, `opencode-${seq}`),
+ `require('node:fs').writeFileSync(${JSON.stringify(asked)}, JSON.stringify({ args: process.argv.slice(2), db: process.env.OPENCODE_DB, pwd: process.env.PWD }));\n` +
+ `process.stdout.write(${JSON.stringify(stdout)});\nprocess.exit(${code});\n`,
+ );
+ return { bin, asked: () => JSON.parse(readFileSync(asked, 'utf8')) as { args: string[]; db: string; pwd: string } };
+ };
+ const problem = (bin = 'opencode') => opencodeRunner.check?.({ bin }, V2);
+ const created = () => {
+ mkdirSync(dirname(db), { recursive: true });
+ writeFileSync(db, '');
+ };
+
+ beforeEach(() => rmSync(dirname(db), { recursive: true, force: true }));
+ afterEach(() => rmSync(bins, { recursive: true, force: true }));
+
+ test('with no database of its own it needs signing in, which it says how to do, and the database has a folder to go in', async () => {
+ expect(await problem()).toEqual({ code: 'AGENT_NEEDS_PERMISSION', message: expect.stringContaining('OpenCode 2'), fix: signInLine('opencode') });
+ expect(existsSync(dirname(db))).toBe(true);
+ });
+
+ // Without --standalone, `auth list` would start the user's background service with this environment, on this database.
+ test("the database is asked through OpenCode's own server, in Browsentic's environment", async () => {
+ created();
+ const { bin, asked } = answering('[]');
+ expect((await problem(bin))?.code).toBe('AGENT_NEEDS_PERMISSION');
+ expect(asked()).toEqual({ args: ['auth', 'list', '--standalone', '--format', 'json'], db, pwd: opencodeRunner.workspace('task') });
+ }, 20_000);
+
+ test('a stored sign-in or a working key is enough, and one that needs signing in again is not', async () => {
+ created();
+ expect(await problem(answering(JSON.stringify([{ id: 'anthropic', connections: [{ type: 'credential', label: 'Anthropic' }] }])).bin)).toBeNull();
+ const stale = [{ id: 'opencode', connections: [{ type: 'env', name: 'OPENCODE_API_KEY', status: { status: 'needs_auth' } }] }];
+ expect((await problem(answering(JSON.stringify(stale)).bin))?.code).toBe('AGENT_NEEDS_PERMISSION');
+ }, 20_000);
+
+ // A picker that wrongly says "needs setup" blocks a working agent; the run's own error catches the opposite case.
+ test('an answer it cannot read is taken as signed in', async () => {
+ created();
+ expect(await problem(answering('Standalone server exited before reporting readiness', 1).bin)).toBeNull();
+ expect(await problem(answering('not json').bin)).toBeNull();
+ }, 20_000);
+
+ test('a Zen key in the environment, or a provider of the user’s own config, is enough without asking', async () => {
+ vi.stubEnv('OPENCODE_API_KEY', 'zen-key');
+ expect(await problem('/nonexistent/opencode')).toBeNull();
+ vi.stubEnv('OPENCODE_API_KEY', '');
+ mkdirSync(join(home, '.config', 'opencode'), { recursive: true });
+ writeFileSync(join(home, '.config', 'opencode', 'opencode.jsonc'), '{ "providers": { "ollama": {} } }');
+ expect(await problem('/nonexistent/opencode')).toBeNull();
+ });
+
+ test('the sign-in line points OpenCode at that database, quoted for the shell it runs in', () => {
+ expect(signInLine('opencode', 'darwin')).toBe(`OPENCODE_DB='${db}' opencode auth login --standalone`);
+ expect(signInLine("/Users/o'neil/bin/opencode 2", 'darwin')).toBe(`OPENCODE_DB='${db}' '/Users/o'\\''neil/bin/opencode 2' auth login --standalone`);
+ expect(signInLine('C:\\Program Files\\opencode.cmd', 'win32')).toBe(`$env:OPENCODE_DB='${db}'; & 'C:\\Program Files\\opencode.cmd' auth login --standalone`);
+ });
+});
diff --git a/src/daemon/agent/runners/opencode.ts b/src/daemon/agent/runners/opencode.ts
index bd54845..8a1e796 100644
--- a/src/daemon/agent/runners/opencode.ts
+++ b/src/daemon/agent/runners/opencode.ts
@@ -1,10 +1,14 @@
import { randomUUID } from 'node:crypto';
-import { readFileSync } from 'node:fs';
+import { mkdirSync, readFileSync, statSync } from 'node:fs';
import { homedir } from 'node:os';
import { dirname, join, relative, sep } from 'node:path';
import type { AgentProblem } from '@/lib/agents/catalog';
import { stateDir } from '../../lockfile';
+import type { AgentSettings } from '../config';
import { MCP_SERVER_NAME } from './claude';
+import { spawnCli, stopTree, type CliProcess } from './command';
+import { childEnv } from './drive';
+import { acpTurn } from './opencode-acp';
import { conversationDir, effortOf, MCP_CALL_TIMEOUT_MS, parseJsonLine, sweepRunDirs } from './util';
import type { JsonContext, McpServer, Plan, Runner, RunMode, StreamContext, StreamReader, StreamSink } from './types';
@@ -18,6 +22,9 @@ import type { JsonContext, McpServer, Plan, Runner, RunMode, StreamContext, Stre
* the system prompt is a file the config points at, and instruction files are read verbatim.
*
* `opencode run` waits for stdin to close and appends it to the message; drive.ts spawns it ignored.
+ *
+ * OpenCode 2 is a second line of releases (`@opencode/cli`, beside 1's `opencode-ai`) with its own
+ * flags, config, database and protocol; which one a plan is for comes from the `--version` probed.
*/
/**
@@ -38,26 +45,72 @@ const TRUNCATION = `A tool result over ${RESULT_BYTES / 1000} KB comes back cut,
/** A user's `share: auto` would publish every browsing session; project config would load from folders above the workspace. */
const SEALED = { OPENCODE_DISABLE_PROJECT_CONFIG: '1', OPENCODE_DISABLE_CLAUDE_CODE: '1', OPENCODE_DISABLE_SHARE: '1' };
+/** OpenCode 2 renamed the project switch, reading the old name only when the new one is unset, and dropped the share one. */
+const SEALED_2 = {
+ OPENCODE_CONFIG_PROJECT_DISABLE: '1',
+ OPENCODE_DISABLE_PROJECT_CONFIG: '1',
+ OPENCODE_DISABLE_CLAUDE_CODE: '1',
+ OPENCODE_DISABLE_AUTOUPDATE: '1',
+};
+
/** Every folder outside git is one project to OpenCode, so runs kept in its own store would fill the user's session list. */
const sessionsDb = () => join(stateDir, 'agents', 'opencode', 'sessions.db');
+/**
+ * OpenCode 2 keeps its logins in the same database as its sessions, and migrates one 1 cannot open
+ * again. Runs keep their own, so the user's background service never sees one, and it holds the
+ * sign-in the user gave Browsentic.
+ */
+const database2 = () => join(stateDir, 'agents', 'opencode', 'opencode-2.db');
+
const configHome = () => join(process.env.XDG_CONFIG_HOME || join(homedir(), '.config'), 'opencode');
const INSTALL = 'npm i -g opencode-ai';
+/** `opencode v2.0.32` from 2.x, a bare `1.19.3` from 1.x. */
+export const isOpenCode2 = (version?: string): boolean => Number(/^(?:opencode\s+)?v?(\d+)\./i.exec(version?.trim() ?? '')?.[1] ?? 0) >= 2;
+
+/** Bins whose OpenCode would not hold a conversation, whose turns go through `opencode run` until the daemon restarts. */
+const noConversation = new Set<string>();
+
type Permission = Record<string, string | Record<string, string>>;
+interface Rule {
+ action: string;
+ resource: string;
+ effect: 'allow' | 'deny';
+}
+
+const rule = (action: string, effect: Rule['effect'] = 'allow', resource = '*'): Rule => ({ action, resource, effect });
+
+/**
+ * Plugins run inside OpenCode, and another MCP server of the user's would start beside the run's own.
+ * A policy in the user's global or managed config outranks these, which the containment note says.
+ */
+const POLICIES = [
+ { action: 'integration.use', resource: 'plugin:*', effect: 'deny' },
+ { action: 'integration.use', resource: 'mcp:*', effect: 'deny' },
+ { action: 'integration.use', resource: `mcp:${MCP_SERVER_NAME}`, effect: 'allow' },
+];
+
interface Part {
type?: string;
text?: string;
tool?: string;
callID?: string;
+ id?: string;
+ messageID?: string;
state?: { status?: string };
tokens?: { input?: number; output?: number; reasoning?: number; cache?: { read?: number; write?: number } };
}
+/** 1.x names an error and nests what the provider said; 2.x types it, `provider.auth` and the like. */
interface Failure {
name?: string;
+ type?: string;
+ message?: string;
+ status?: number;
+ response?: { body?: string };
data?: { message?: string; statusCode?: number; responseBody?: string };
}
@@ -86,6 +139,7 @@ export const opencodeRunner: Runner = {
],
stream(context: StreamContext): Plan {
+ if (isOpenCode2(context.version)) return standalone(context);
const { settings, research } = context;
const base = this.workspace('run');
sweepRunDirs(base);
@@ -117,9 +171,26 @@ export const opencodeRunner: Runner = {
};
},
+ /**
+ * OpenCode 2's `run` sends its first request before the run's MCP server has connected, and that
+ * step's tool list is fixed when it is sent, so the turn is held over ACP until the tools are there.
+ */
+ converse(context: StreamContext) {
+ const { bin } = context.settings;
+ if (!isOpenCode2(context.version) || noConversation.has(bin)) return null;
+ return acpTurn(context, {
+ plan: { ...turn2(context), args: ['acp'] },
+ agent: AGENT,
+ declined: () => noConversation.add(bin),
+ explain: (message) => explain({ message }, bin),
+ stoppedFor: (tool) => stoppedFor(tool),
+ });
+ },
+
reader(): StreamReader {
let spoke = false;
let generated = 0;
+ const continues = continuation();
const report = (tokens: NonNullable<Part['tokens']>, sink: StreamSink) => {
const made = (tokens.output ?? 0) + (tokens.reasoning ?? 0);
@@ -135,23 +206,21 @@ export const opencodeRunner: Runner = {
const part = event.part;
switch (event.type) {
- case 'text':
+ case 'text': {
if (!part?.text) return;
- sink.text(spoke ? `\n\n${part.text}` : part.text);
+ const goesOn = continues(part);
+ sink.text(spoke && !goesOn ? `\n\n${part.text}` : part.text);
spoke = true;
return;
+ }
case 'tool_use': {
const name = part?.tool;
if (!name || ownTool(name)) return;
- if (WEB_TOOLS.includes(name)) return sink.tool(part.callID ?? randomUUID(), name);
+ if (WEB_TOOLS.includes(name)) return sink.tool(part.callID ?? part.id ?? randomUUID(), name);
// A refused call is the model reaching for a tool it was never offered; one that ran means the ruleset did not hold.
if (part.state?.status !== 'completed') return;
- return sink.fail(
- 'AGENT_UNSAFE',
- `OpenCode ran its own ${name} tool in this run, which Browsentic denied, so the run was stopped. ` +
- 'Update OpenCode and Browsentic; if it persists, please report it.',
- );
+ return sink.fail('AGENT_UNSAFE', stoppedFor(name));
}
case 'step_finish':
@@ -167,6 +236,20 @@ export const opencodeRunner: Runner = {
json(context: JsonContext): Plan {
const { settings, reads } = context;
const cwd = this.workspace('task');
+ if (isOpenCode2(context.version)) {
+ return {
+ cwd,
+ env: {
+ ...SEALED_2,
+ PWD: cwd,
+ OPENCODE_DB: database2(),
+ // A path inside the folder a run starts in is matched as seen from it, and anything outside is an external directory the deny covers.
+ OPENCODE_CONFIG_CONTENT: config2({ servers: {}, allowed: reads ? [rule('read', 'allow', 'tmp/*')] : [] }),
+ },
+ args: [...invocation2(), ...(settings.model ? ['--model', settings.model] : [])],
+ input: context.prompt,
+ };
+ }
const permission: Permission = { '*': 'deny', ...(reads ? { read: scratchReads(cwd) } : {}) };
return {
cwd,
@@ -182,11 +265,15 @@ export const opencodeRunner: Runner = {
answer(stdout: string) {
let said: string[] = [];
+ const continues = continuation();
for (const line of stdout.split('\n')) {
const event = parseJsonLine<Event>(line);
if (event?.type === 'error') return { error: explain(event.error) };
if (event?.type === 'step_start') said = [];
- if (event?.type === 'text' && event.part?.text) said.push(event.part.text);
+ if (event?.type === 'text' && event.part?.text) {
+ if (continues(event.part) && said.length) said[said.length - 1] += event.part.text;
+ else said.push(event.part.text);
+ }
}
return { text: said.length ? said.join('\n\n') : undefined };
},
@@ -195,10 +282,15 @@ export const opencodeRunner: Runner = {
if (/Unknown arguments?|Not enough arguments|Invalid values/i.test(stderrTail)) {
return `Your OpenCode does not understand the flags Browsentic uses. Update it (${INSTALL}), then try again. (${stderrTail.trim()})`;
}
+ // Each release line refuses the other's flags, which happens when OpenCode is swapped while Browsentic is running.
+ if (/Unrecognized flag/i.test(stderrTail) || /V2 permissions are not supported/i.test(stderrTail)) {
+ return `OpenCode changed version since Browsentic last checked it. Send the message again. (${oneLine(stderrTail)})`;
+ }
return null;
},
- async check(): Promise<AgentProblem | null> {
+ async check(settings: AgentSettings, version?: string): Promise<AgentProblem | null> {
+ if (isOpenCode2(version)) return check2(settings);
if (process.env.OPENCODE_API_KEY || process.env.OPENCODE_AUTH_CONTENT || signedIn() || declaresProvider()) return null;
return {
code: 'AGENT_NEEDS_PERMISSION',
@@ -211,8 +303,32 @@ export const opencodeRunner: Runner = {
const invocation = () => ['run', '--format', 'json', '--pure', '--agent', AGENT];
+/** No message on argv: 2.x quotes any word with a space in it, and appends what follows `--` a second time. */
+const invocation2 = () => ['run', '--standalone', '--format', 'json', '--agent', AGENT];
+
const ownTool = (name: string) => name.startsWith(`${MCP_SERVER_NAME}_`);
+/**
+ * Whether a text part goes on with one already said. 2.x ends a turn by sending whatever of a text
+ * block it had not yet printed, under that block's number in its message — `prt_<message>_text-<n>` —
+ * and a block printed in part before goes on where it stopped. 1.x's parts never carry the number.
+ */
+function continuation(): (part: Part) => boolean {
+ const blocks = new Map<string, number>();
+ return (part) => {
+ const message = part.messageID ?? '';
+ const shown = blocks.get(message) ?? 0;
+ const numbered = /_text-(\d+)$/.exec(part.id ?? '');
+ if (numbered && Number(numbered[1]) < shown) return true;
+ blocks.set(message, numbered ? Number(numbered[1]) + 1 : shown + 1);
+ return false;
+ };
+}
+
+const stoppedFor = (tool: string) =>
+ `OpenCode ran its own ${tool} tool in this run, which Browsentic denied, so the run was stopped. ` +
+ 'Update OpenCode and Browsentic; if it persists, please report it.';
+
const server = (mcp: McpServer) => ({
type: 'local',
command: [mcp.command, ...mcp.args],
@@ -221,6 +337,15 @@ const server = (mcp: McpServer) => ({
timeout: MCP_CALL_TIMEOUT_MS,
});
+/** Code Mode would hide the tools behind an `execute` tool the ruleset denies, and that tool is a script with the network. */
+const server2 = (mcp: McpServer) => ({
+ type: 'local',
+ command: [mcp.command, ...mcp.args],
+ environment: mcp.env,
+ codemode: false,
+ timeout: { execution: MCP_CALL_TIMEOUT_MS },
+});
+
/** Title generation is a model call per new session that the panel already makes for itself. */
function config({ mcp, instructions, permission }: { mcp: Record<string, unknown>; instructions?: string[]; permission: Permission }): string {
return JSON.stringify({
@@ -234,6 +359,60 @@ function config({ mcp, instructions, permission }: { mcp: Record<string, unknown
});
}
+/** 2.x's own shape: the rules are a list where the last match wins, and an ACP session starts in `default_agent`. */
+function config2({ servers, instructions, allowed }: { servers: Record<string, unknown>; instructions?: string[]; allowed: Rule[] }): string {
+ return JSON.stringify({
+ share: 'disabled',
+ update: 'disable',
+ snapshots: false,
+ tool_output: { max_bytes: RESULT_BYTES, max_lines: RESULT_BYTES },
+ ...(instructions ? { instructions } : {}),
+ default_agent: AGENT,
+ mcp: { servers },
+ policies: POLICIES,
+ agents: { title: { disabled: true }, [AGENT]: { mode: 'primary', permissions: [rule('*', 'deny'), ...allowed] } },
+ });
+}
+
+/** What a 2.x turn starts from, however it is then spoken to. 2.x moves to `$PWD` before its own working directory. */
+function turn2(context: StreamContext): Omit<Plan, 'args'> {
+ const base = opencodeRunner.workspace('run');
+ sweepRunDirs(base);
+ const cwd = conversationDir(base, context.conversation ?? context.runId);
+ const allowed = [...context.mcpTools.map((tool) => `${MCP_SERVER_NAME}_${tool}`), ...(context.research ? WEB_TOOLS : [])];
+ return {
+ cwd,
+ env: {
+ ...SEALED_2,
+ PWD: cwd,
+ OPENCODE_DB: database2(),
+ OPENCODE_CONFIG_CONTENT: config2({
+ servers: { [MCP_SERVER_NAME]: server2(context.mcp) },
+ instructions: [join(cwd, INSTRUCTIONS)],
+ allowed: allowed.map((action) => rule(action)),
+ }),
+ BROWSENTIC_AGENT_RUN: context.runId,
+ },
+ files: [{ path: INSTRUCTIONS, content: `${context.systemPrompt.trim()}\n\n${TRUNCATION}\n` }],
+ };
+}
+
+/**
+ * A 2.x turn through `opencode run`, for an OpenCode that would not hold a conversation. Its first step
+ * may not see the browser tools yet. No effort: 2.x refuses the run when the model lacks that variant.
+ */
+function standalone(context: StreamContext): Plan {
+ return {
+ ...turn2(context),
+ args: [
+ ...invocation2(),
+ ...(context.sessionId ? ['--session', context.sessionId] : []),
+ ...(context.settings.model ? ['--model', context.settings.model] : []),
+ ],
+ input: context.instruction,
+ };
+}
+
/**
* OpenCode matches a read against its path relative to the project root — `/` outside git, the
* repository when the home directory is one — so the scratch folder is allowed as seen from every
@@ -251,24 +430,53 @@ function scratchReads(workspace: string): Record<string, string> {
}
}
-function explain(failure: Failure | undefined): string {
- const message = oneLine(failure?.data?.message ?? failure?.name ?? 'OpenCode reported an error');
- const status = failure?.data?.statusCode;
+/** The command that signs Browsentic's own OpenCode 2 database in, in the user's shell. */
+export function signInLine(bin: string, platform: NodeJS.Platform = process.platform): string {
+ const db = database2();
+ if (platform === 'win32') return `$env:OPENCODE_DB='${db.replaceAll("'", "''")}'; & '${bin.replaceAll("'", "''")}' auth login --standalone`;
+ const quote = (value: string) => `'${value.replaceAll("'", `'\\''`)}'`;
+ return `OPENCODE_DB=${quote(db)} ${/^[\w./-]+$/.test(bin) ? bin : quote(bin)} auth login --standalone`;
+}
+
+function explain(failure: Failure | undefined, bin = 'opencode'): string {
+ const message = oneLine(failure?.data?.message ?? failure?.message ?? failure?.name ?? failure?.type ?? 'OpenCode reported an error');
+ const status = failure?.data?.statusCode ?? failure?.status;
+ const body = failure?.data?.responseBody ?? failure?.response?.body ?? '';
+ // 1.x names its errors and nests what the provider said; 2.x types them, and over ACP passes on only a message.
+ const v1 = failure?.name !== undefined || failure?.data !== undefined;
// Zen serves its free models only to requests carrying OpenCode's own built-in tools.
- if (failure?.data?.responseBody?.includes('FreeTierError')) {
+ if (v1 && body.includes('FreeTierError')) {
return (
`OpenCode Zen's free models refuse a run whose tools Browsentic has narrowed to the browser. Run "opencode auth login" ` +
`to sign in to a provider, then pick one of its models in the Browsentic side panel. (${message})`
);
}
- if (failure?.name === 'ProviderAuthError' || status === 401 || status === 403) {
+ if (!v1 && body.includes('FreeTierError')) {
+ return (
+ `OpenCode Zen's free models refuse a run whose tools Browsentic has narrowed to the browser. Sign Browsentic's OpenCode in ` +
+ `to a provider with ${signInLine(bin)}, then pick one of its models in the side panel. (${message})`
+ );
+ }
+ // Over ACP, 2.x says only this, for a provider with no sign-in and for Zen's free tier alike.
+ if (!v1 && /authentication required/i.test(message)) {
+ return (
+ `OpenCode has no sign-in for this model in Browsentic, and OpenCode Zen's free models refuse a run whose tools Browsentic ` +
+ `has narrowed to the browser. Sign Browsentic's OpenCode in to a provider with ${signInLine(bin)}, then pick one of its models in the side panel. (${message})`
+ );
+ }
+ if (failure?.name === 'ProviderAuthError' || failure?.type === 'provider.auth' || status === 401 || status === 403) {
+ if (!v1) return `${message} If that is a login problem, sign Browsentic's OpenCode in with ${signInLine(bin)}.`;
return (
`${message} If that is a login problem, run "opencode auth login" — a key kept only in an environment ` +
'variable is not passed to a Browsentic run.'
);
}
- if (status === 429) return `The provider behind OpenCode is rate-limiting this account. Wait and try again. (${message})`;
- if (failure?.name === 'UnknownError') {
+ if (status === 429 || failure?.type === 'provider.rate-limit') return `The provider behind OpenCode is rate-limiting this account. Wait and try again. (${message})`;
+ if (failure?.type === 'provider.quota') return `The provider behind OpenCode says this account is out of credit or quota. Top it up, or pick another provider's model. (${message})`;
+ if (/variant unavailable/i.test(message)) {
+ return `The model OpenCode is using has no such level of effort. Clear the effort in Browsentic's settings, or pick another, then try again. (${message})`;
+ }
+ if (failure?.name === 'UnknownError' || /model unavailable|invalid model reference/i.test(message)) {
return (
`OpenCode could not start this turn, most often because it does not know the model. Pick one in the ` +
`Browsentic side panel as "opencode models" lists it — provider/model — then try again. (${message})`
@@ -291,12 +499,101 @@ function signedIn(): boolean {
}
/** A provider the user configured by hand — a local model needs no login. */
-function declaresProvider(): boolean {
+function declaresProvider(pattern = /"provider"\s*:/): boolean {
return ['opencode.json', 'opencode.jsonc', 'config.json'].some((name) => {
try {
- return /"provider"\s*:/.test(readFileSync(join(configHome(), name), 'utf8'));
+ return pattern.test(readFileSync(join(configHome(), name), 'utf8'));
} catch {
return false;
}
});
}
+
+const AUTH_TIMEOUT_MS = 5_000;
+
+/** One answer per state of the database, and one question at a time. */
+let lastAnswer: { key: string; problem: AgentProblem | null } | undefined;
+let asking: Promise<AgentProblem | null> | undefined;
+
+/**
+ * OpenCode 2's sign-ins live in its database, so its own `auth list` is asked, against Browsentic's
+ * database and in the environment a run gets: an exported key shows there exactly when a run sees it.
+ * `--standalone` keeps the user's background service out of it — without it that service would be
+ * started with this environment and run on Browsentic's database. Anything unexpected reads as signed
+ * in, because a picker that wrongly says *needs setup* blocks a working agent.
+ */
+async function check2(settings: AgentSettings): Promise<AgentProblem | null> {
+ if (process.env.OPENCODE_API_KEY || declaresProvider(/"providers?"\s*:/)) return null;
+ const db = database2();
+ // The sign-in line points OpenCode at this file, and its server will not start in a folder that is not there.
+ mkdirSync(dirname(db), { recursive: true });
+ const stamped = stamp(db);
+ if (!stamped) return needsSignIn(settings.bin);
+ const key = `${settings.bin}|${stamped}|${stamp(`${db}-wal`)}`;
+ if (lastAnswer?.key === key) return lastAnswer.problem;
+ asking ??= listSignIns(settings.bin)
+ .then((signedIn) => {
+ const problem = signedIn ? null : needsSignIn(settings.bin);
+ lastAnswer = { key, problem };
+ return problem;
+ })
+ .finally(() => (asking = undefined));
+ return asking;
+}
+
+const stamp = (path: string) => {
+ try {
+ const { mtimeMs, size } = statSync(path);
+ return `${mtimeMs}:${size}`;
+ } catch {
+ return '';
+ }
+};
+
+const needsSignIn = (bin: string): AgentProblem => ({
+ code: 'AGENT_NEEDS_PERMISSION',
+ message:
+ "This OpenCode 2 has no sign-in of its own for Browsentic, which keeps one apart from yours, and OpenCode Zen's free models refuse a run whose tools Browsentic has narrowed to the browser.",
+ fix: signInLine(bin),
+});
+
+interface Integration {
+ connections?: { status?: { status?: string } }[];
+}
+
+function listSignIns(bin: string): Promise<boolean> {
+ const cwd = opencodeRunner.workspace('task');
+ mkdirSync(cwd, { recursive: true });
+ let child: CliProcess;
+ try {
+ child = spawnCli(bin, ['auth', 'list', '--standalone', '--format', 'json'], {
+ cwd,
+ env: { ...childEnv('opencode'), ...SEALED_2, PWD: cwd, OPENCODE_DB: database2(), OPENCODE_CONFIG_CONTENT: JSON.stringify({ policies: POLICIES }) },
+ });
+ } catch {
+ return Promise.resolve(true);
+ }
+ return new Promise((resolve) => {
+ let output = '';
+ let settled = false;
+ const done = (value: boolean) => {
+ if (settled) return;
+ settled = true;
+ clearTimeout(timer);
+ resolve(value);
+ };
+ const timer = setTimeout(() => {
+ stopTree(child, 'SIGKILL');
+ done(true);
+ }, AUTH_TIMEOUT_MS);
+ timer.unref();
+ child.stdout.on('data', (chunk: Buffer) => void (output += chunk.toString()));
+ child.stderr.on('data', () => {});
+ child.on('error', () => done(true));
+ child.on('close', (code) => {
+ const integrations = code === 0 ? parseJsonLine<Integration[]>(output.trim()) : null;
+ if (!Array.isArray(integrations)) return done(true);
+ done(integrations.some((integration) => integration.connections?.some((connection) => !connection.status || connection.status.status === 'connected')));
+ });
+ });
+}
diff --git a/src/daemon/agent/runners/types.ts b/src/daemon/agent/runners/types.ts
index 859a43f..3f1f36d 100644
--- a/src/daemon/agent/runners/types.ts
+++ b/src/daemon/agent/runners/types.ts
@@ -57,11 +57,16 @@ export interface StreamContext {
mcpTools: string[];
/** The run's browser tools served in this process, for a CLI that takes tools other than over MCP. */
tools: () => Promise<ToolHost>;
+ /** Settles once the run's MCP server has handed its tool list to the CLI, for a CLI that does not wait for it. */
+ toolsListed: () => Promise<void>;
+ /** The CLI's `--version` line as last probed, for a runner whose CLI changed its flags between releases. */
+ version?: string;
}
export interface JsonContext {
prompt: string;
settings: AgentSettings;
+ version?: string;
workspace: string;
/** The prompt names a file in the workspace that the agent has to open. */
reads: boolean;
@@ -150,8 +155,8 @@ export interface Runner {
answer(stdout: string): { text?: string; error?: string };
/** Turns a stderr tail into something more useful than "exited with code 1". */
hint?(stderrTail: string): string | null;
- /** Anything beyond "the binary is installed" that has to be true before a run. */
- check?(settings: AgentSettings): Promise<AgentProblem | null>;
+ /** Anything beyond "the binary is installed" that has to be true before a run, for the `--version` line just probed. */
+ check?(settings: AgentSettings, version?: string): Promise<AgentProblem | null>;
/** Repairs what `check` reported, when the user asks for it. */
grant?(): Promise<AgentProblem | null>;
}
diff --git a/src/daemon/agent/service-tools-listed.test.ts b/src/daemon/agent/service-tools-listed.test.ts
new file mode 100644
index 0000000..f990207
--- /dev/null
+++ b/src/daemon/agent/service-tools-listed.test.ts
@@ -0,0 +1,50 @@
+import { mkdirSync, readFileSync, rmSync, writeFileSync } from 'node:fs';
+import { dirname, join } from 'node:path';
+import { afterEach, describe, expect, test, vi } from 'vitest';
+import type { RunEvent } from '@/lib/actions/protocol';
+import { failure } from '@/lib/actions/protocol';
+import { configPath } from './config';
+import { stubCli } from './runners/fixtures/support';
+import { AgentSession } from './service';
+
+// OpenCode 2 connects MCP servers in the background and fixes each step's tool list as it sends it.
+describe('a turn that waits for its tools', () => {
+ afterEach(() => vi.unstubAllEnvs());
+
+ test('an OpenCode 2 turn is prompted only once its MCP server has listed the tools to it', async () => {
+ vi.stubEnv('OPENCODE_API_KEY', 'zen-key');
+ mkdirSync(dirname(configPath), { recursive: true });
+ const heard = join(dirname(configPath), 'acp-heard.log');
+ rmSync(heard, { force: true });
+ const bin = stubCli(
+ join(dirname(configPath), 'opencode-acp'),
+ `const fs = require('node:fs');
+if (process.argv[2] === '--version') { console.log('opencode v2.0.32'); process.exit(0); }
+const send = (message) => process.stdout.write(JSON.stringify(message) + '\\n');
+require('node:readline').createInterface({ input: process.stdin }).on('line', (line) => {
+ const message = JSON.parse(line);
+ fs.appendFileSync(${JSON.stringify(heard)}, message.method + '\\n');
+ if (message.method === 'initialize') send({ jsonrpc: '2.0', id: message.id, result: { protocolVersion: 1 } });
+ if (message.method === 'session/new') send({ jsonrpc: '2.0', id: message.id, result: { sessionId: 'ses_1', configOptions: [{ id: 'mode', currentValue: 'browsentic-contained' }] } });
+ if (message.method === 'session/prompt') send({ jsonrpc: '2.0', id: message.id, result: { stopReason: 'end_turn' } });
+});`,
+ );
+ writeFileSync(configPath, JSON.stringify({ agent: 'opencode', agents: { opencode: { bin } } }));
+ const events: RunEvent[] = [];
+ const waiting = new AgentSession({
+ invoke: async () => failure('EXTENSION_OFFLINE', 'no browser'),
+ emit: (_runId: string, event: RunEvent) => void events.push(event),
+ draft: () => {},
+ actionNames: () => [],
+ });
+ waiting.handle({ t: 'instruct', id: 'r1', text: 'what is on this page', context: { sessionId: 's1' } });
+ const said = () => readFileSync(heard, 'utf8').trim().split('\n');
+ await vi.waitFor(() => expect(said()).toContain('session/new'), { timeout: 10_000 });
+ await new Promise((resolve) => setTimeout(resolve, 300));
+ expect(said()).toEqual(['initialize', 'session/new']);
+
+ waiting.described('r1');
+ await vi.waitFor(() => expect(events.some((event) => event.kind === 'done')).toBe(true), { timeout: 5_000 });
+ expect(said()).toEqual(['initialize', 'session/new', 'session/prompt']);
+ }, 20_000);
+});
diff --git a/src/daemon/agent/service.ts b/src/daemon/agent/service.ts
index c74fd68..e8cb7fc 100644
--- a/src/daemon/agent/service.ts
+++ b/src/daemon/agent/service.ts
@@ -109,6 +109,9 @@ interface ActiveRun {
limits?: CallLimits;
/** An approval still on the user's screen after its call was handed back, until the agent calls again. */
parked?: Waiting;
+ /** Settles when the run's MCP server first lists its tools to the agent. */
+ listed: Promise<void>;
+ markListed: () => void;
}
const FOCUS_SHOT_TOOL_NAME = toolNameFor(FOCUS_SHOT_ACTION);
@@ -195,6 +198,11 @@ export class AgentSession {
return this.runs.has(runId);
}
+ /** The run's MCP server has just been told what its tools are, which a CLI that does not wait for them needs to know. */
+ described(runId: string): void {
+ this.runs.get(runId)?.markListed();
+ }
+
/**
* What a run's tool list should hold. A tool the run would only be refused is left
* out: its schema is re-sent on every turn, and listing it invites the call.
@@ -464,6 +472,8 @@ export class AgentSession {
extraHosts: config.guardrails?.hosts,
pinTab: true,
});
+ let markListed = () => {};
+ const listed = new Promise<void>((resolve) => (markListed = resolve));
const run: ActiveRun = {
id: runId,
sessionId,
@@ -480,6 +490,8 @@ export class AgentSession {
approved: new Set(),
scheduled: context?.task !== undefined,
limits: RUNNERS[config.agent].limits,
+ listed,
+ markListed,
};
this.runs.set(runId, run);
if (run.liveTools && sessionId) this.codeListed.add(sessionId);
@@ -577,6 +589,8 @@ export class AgentSession {
config: run.config,
sessionId: resuming,
mcpTools: agentRunToolNames(this.deps.actionNames()),
+ toolsListed: () => run.listed,
+ version: runner.version,
signal: run.abort.signal,
emit,
});
diff --git a/src/daemon/daemon.ts b/src/daemon/daemon.ts
index cb758c1..5f389bc 100644
--- a/src/daemon/daemon.ts
+++ b/src/daemon/daemon.ts
@@ -781,7 +781,8 @@ export async function startDaemon({ version, idleExit = true, android = androidO
function describeFor(runId: string | undefined, binding: Binding): Described {
const running = runId ? [...agents].find(([, agent]) => agent.owns(runId)) : undefined;
const offer = running && runId ? running[1].offerFor(runId) : null;
- if (offer && running) {
+ if (offer && running && runId) {
+ running[1].described(runId);
const withheld = (name: string) => offer.withheld.includes(name);
return {
tools: running[0].tools.filter(({ name }) => !withheld(name)),
diff --git a/src/daemon/guardrails/spawn.test.ts b/src/daemon/guardrails/spawn.test.ts
index 2e22ad8..8457122 100644
--- a/src/daemon/guardrails/spawn.test.ts
+++ b/src/daemon/guardrails/spawn.test.ts
@@ -8,7 +8,7 @@ import { CONTAINMENT, sealEnv, sealedAway, vetPlan, within, type SpawnMode } fro
const settingsFor = (kind: AgentKind) => ({ bin: AGENTS[kind].bin });
-const contextFor = (kind: AgentKind, research: boolean): StreamContext => ({
+const contextFor = (kind: AgentKind, research: boolean, version?: string): StreamContext => ({
runId: 'run-1',
conversation: 'conversation-1',
instruction: 'what does this page cost',
@@ -20,11 +20,13 @@ const contextFor = (kind: AgentKind, research: boolean): StreamContext => ({
mcp: mcpServerFor('run-1'),
mcpTools: ['page_getPageInfo', 'page_clickElement', 'browsentic_status'],
tools: () => Promise.reject(new Error('no browser tools in a containment test')),
+ toolsListed: () => Promise.resolve(),
+ version,
});
-const planOf = (kind: AgentKind, mode: SpawnMode, { research = false, reads = false } = {}): Plan => {
- if (mode === 'task') return RUNNERS[kind].json({ prompt: 'summarize this', settings: settingsFor(kind), workspace: stateDir, reads });
- const context = contextFor(kind, research);
+const planOf = (kind: AgentKind, mode: SpawnMode, { research = false, reads = false, version = undefined as string | undefined } = {}): Plan => {
+ if (mode === 'task') return RUNNERS[kind].json({ prompt: 'summarize this', settings: settingsFor(kind), version, workspace: stateDir, reads });
+ const context = contextFor(kind, research, version);
if (mode === 'run') return RUNNERS[kind].stream(context);
const conversation = RUNNERS[kind].converse?.(context);
if (!conversation) throw new Error(`${kind} holds no conversation`);
@@ -493,6 +495,94 @@ describe('spawn containment', () => {
expect(vetPlan('opencode', 'run', { ...opencodeRun, files: [] }, stateDir)).toHaveLength(1);
});
+ describe('OpenCode 2', () => {
+ const version = 'opencode v2.0.32';
+ const turn = planOf('opencode', 'conversation', { version });
+ const fallback = planOf('opencode', 'run', { version });
+ const oneShot = planOf('opencode', 'task', { version });
+ interface Statement {
+ action: string;
+ resource: string;
+ effect: string;
+ }
+ const agentRules = (config: Record<string, unknown>) =>
+ (config.agents as Record<string, { permissions: Statement[] }>)['browsentic-contained'].permissions;
+
+ test('its turn, the plain run it falls back to, and its one-shots are contained, on research and reading a file too', () => {
+ for (const [mode, plan] of [
+ ['conversation', turn],
+ ['conversation', planOf('opencode', 'conversation', { version, research: true })],
+ ['run', fallback],
+ ['task', oneShot],
+ ['task', planOf('opencode', 'task', { version, reads: true })],
+ ] as const) {
+ expect(vetPlan('opencode', mode, plan, stateDir)).toEqual([]);
+ }
+ });
+
+ // Without it, `run` hands the turn to the user's background service, which starts with this environment if it is not running.
+ test('a run without its own server is caught', () => {
+ expect(vetPlan('opencode', 'run', without(fallback, '--standalone'), stateDir)).not.toEqual([]);
+ expect(vetPlan('opencode', 'task', without(oneShot, '--standalone'), stateDir)).not.toEqual([]);
+ });
+
+ test('a turn whose folder is not the one OpenCode moves to is caught', () => {
+ expect(vetPlan('opencode', 'conversation', withEnv(turn, 'PWD', '/'), stateDir)).toHaveLength(1);
+ expect(vetPlan('opencode', 'task', withEnv(oneShot, 'PWD'), stateDir)).toHaveLength(1);
+ });
+
+ // A list keeps the deny-all where it was, so only reading the rules shows a later allow.
+ test('an allow added after the deny-all, or a deny-all that went, is caught', () => {
+ const shell = reconfigured(turn, (config) => void agentRules(config).push({ action: 'shell', resource: '*', effect: 'allow' }));
+ expect(vetPlan('opencode', 'conversation', shell, stateDir)).toEqual([expect.stringContaining('allow shell')]);
+ expect(vetPlan('opencode', 'conversation', reconfigured(turn, (config) => void agentRules(config).shift()), stateDir)).not.toEqual([]);
+ });
+
+ test('a one-shot reading anything but its scratch folder is caught', () => {
+ const reads = planOf('opencode', 'task', { version, reads: true });
+ const everywhere = reconfigured(reads, (config) => void (agentRules(config)[1].resource = '*'));
+ expect(vetPlan('opencode', 'task', everywhere, stateDir)).toHaveLength(1);
+ });
+
+ // Code Mode lists one `execute` tool in place of the server's, and it runs a script with the network.
+ test("browser tools behind Code Mode, or a one-shot given a server, are caught", () => {
+ const behind = reconfigured(turn, (config) => void ((config.mcp as { servers: Record<string, { codemode?: boolean }> }).servers.browsentic.codemode = true));
+ expect(vetPlan('opencode', 'conversation', behind, stateDir)).toHaveLength(1);
+ const reaching = reconfigured(oneShot, (config) => void (config.mcp = { servers: { browsentic: { type: 'local' } } }));
+ expect(vetPlan('opencode', 'task', reaching, stateDir)).toHaveLength(1);
+ });
+
+ test("letting the user's plugins or other MCP servers load is caught", () => {
+ const plugins = reconfigured(turn, (config) => void (config.policies as Statement[]).push({ action: 'integration.use', resource: 'plugin:*', effect: 'allow' }));
+ expect(vetPlan('opencode', 'conversation', plugins, stateDir)).toHaveLength(1);
+ expect(vetPlan('opencode', 'conversation', reconfigured(turn, (config) => void delete config.policies), stateDir)).toHaveLength(2);
+ });
+
+ test('a session started in another agent, or shared, is caught', () => {
+ expect(vetPlan('opencode', 'conversation', reconfigured(turn, (config) => void (config.default_agent = 'build')), stateDir)).toHaveLength(1);
+ expect(vetPlan('opencode', 'conversation', reconfigured(turn, (config) => void (config.share = 'auto')), stateDir)).toHaveLength(1);
+ });
+
+ test('project config switched back on under either name is caught', () => {
+ expect(vetPlan('opencode', 'conversation', withEnv(turn, 'OPENCODE_CONFIG_PROJECT_DISABLE', '0'), stateDir)).toHaveLength(1);
+ expect(vetPlan('opencode', 'conversation', withEnv(turn, 'OPENCODE_DISABLE_PROJECT_CONFIG'), stateDir)).toHaveLength(1);
+ });
+
+ // Each line refuses the other's flags, but the vet holds a plan to one line's rules in full.
+ test("one line's flags with the other line's config are caught", () => {
+ const v1 = planOf('opencode', 'run');
+ expect(vetPlan('opencode', 'run', { ...fallback, env: v1.env }, stateDir)).not.toEqual([]);
+ expect(vetPlan('opencode', 'run', { ...v1, env: fallback.env }, stateDir)).not.toEqual([]);
+ });
+
+ test('a turn sent to another server, handed a file, or given a model it could read as a flag is caught', () => {
+ expect(vetPlan('opencode', 'run', plus(fallback, '--server', 'http://127.0.0.1:4096'), stateDir)).toHaveLength(1);
+ expect(vetPlan('opencode', 'run', plus(fallback, '--file', '/etc/hosts'), stateDir)).toHaveLength(1);
+ expect(vetPlan('opencode', 'run', plus(fallback, '--agent=build'), stateDir)).toHaveLength(1);
+ expect(vetPlan('opencode', 'run', plus(fallback, '-m', '--auto'), stateDir)).not.toEqual([]);
+ });
+ });
+
test('running outside the state dir is caught', () => {
expect(vetPlan('antigravity', 'run', { ...agyRun, cwd: '/tmp/anywhere' }, stateDir)).toHaveLength(1);
});
diff --git a/src/daemon/guardrails/spawn.ts b/src/daemon/guardrails/spawn.ts
index 523de6d..f68af42 100644
--- a/src/daemon/guardrails/spawn.ts
+++ b/src/daemon/guardrails/spawn.ts
@@ -86,8 +86,15 @@ interface Requirements {
readonly fileContains?: Readonly<Record<string, readonly string[]>>;
/** The only set-up commands the plan may run before its turn, each exactly as written here. */
readonly prepare?: readonly (readonly string[])[];
+ /** Variables that must name the plan's own working directory, for a CLI that moves to one before using its cwd. */
+ readonly envIsCwd?: readonly string[];
+ /** What this shape cannot say, read out of the plan itself: each answer is a clause the label goes in front of. */
+ readonly inspect?: (plan: SpawnPlan) => string[];
}
+/** One way of being contained, or several a CLI's releases each need, of which a plan has to meet one in full. */
+type Rules = Requirements | readonly Requirements[];
+
interface Containment {
readonly localTools: LocalTools;
/** Env prefixes this agent needs to authenticate, kept when the rest is sealed. */
@@ -105,10 +112,10 @@ interface Containment {
readonly forbidden?: readonly RegExp[];
/** Why this runner is only as contained as it is. Logged once per run. */
readonly note: string;
- readonly run: Requirements;
- readonly task: Requirements;
+ readonly run: Rules;
+ readonly task: Rules;
/** Only a runner that holds conversations has these; any other refuses to start one. */
- readonly conversation?: Requirements;
+ readonly conversation?: Rules;
}
/**
@@ -160,6 +167,96 @@ const OPENCODE_SEALED = { OPENCODE_DISABLE_PROJECT_CONFIG: '1', OPENCODE_DISABLE
/** The ruleset that denies every tool not named after it, and the switch that keeps a session off opncd.ai. */
const OPENCODE_CONFIG = ['"*":"deny"', '"share":"disabled"'];
+/**
+ * OpenCode 2 renamed the project switch, keeping the old name only as a fallback the new one overrides
+ * when set at all, and reads a global config folder `OPENCODE_DISABLE_PROJECT_CONFIG` never reaches.
+ */
+const OPENCODE2_SEALED = {
+ OPENCODE_CONFIG_PROJECT_DISABLE: '1',
+ OPENCODE_DISABLE_PROJECT_CONFIG: '1',
+ OPENCODE_DISABLE_CLAUDE_CODE: '1',
+ OPENCODE_DISABLE_AUTOUPDATE: '1',
+};
+
+const OPENCODE_AGENT = 'browsentic-contained';
+
+const OPENCODE_WEB = ['webfetch', 'websearch'];
+
+interface OpenCodeStatement {
+ action?: unknown;
+ resource?: unknown;
+ effect?: unknown;
+}
+
+interface OpenCode2Config {
+ share?: unknown;
+ default_agent?: unknown;
+ policies?: OpenCodeStatement[];
+ agents?: Record<string, { permissions?: OpenCodeStatement[] } | undefined>;
+ mcp?: { servers?: Record<string, { codemode?: unknown } | undefined> };
+}
+
+/**
+ * OpenCode 2's config read rather than searched: its rules are a list, where an allow added after the
+ * deny would sit beside it and still match. A run's agent opens on denying everything and then allows
+ * the browser's tools, the web's on research, and a one-shot's scratch reads, nothing else; the
+ * browser's server is the only one and lists its tools directly, outside Code Mode's `execute`; and
+ * policies keep the user's plugins and other MCP servers from loading at all.
+ */
+const openCode2 =
+ (mode: SpawnMode) =>
+ (plan: SpawnPlan): string[] => {
+ let config: OpenCode2Config;
+ try {
+ config = JSON.parse(plan.env?.OPENCODE_CONFIG_CONTENT ?? '') as OpenCode2Config;
+ } catch {
+ return ['is spawned without a readable OPENCODE_CONFIG_CONTENT'];
+ }
+ const problems: string[] = [];
+ if (config.share !== 'disabled') problems.push('is spawned with sharing left on');
+ if (mode === 'conversation' && config.default_agent !== OPENCODE_AGENT) problems.push(`starts its sessions in an agent other than ${OPENCODE_AGENT}`);
+
+ const [first, ...rest] = config.agents?.[OPENCODE_AGENT]?.permissions ?? [];
+ if (first?.action !== '*' || first.resource !== '*' || first.effect !== 'deny') {
+ problems.push(`gives ${OPENCODE_AGENT} rules that do not open on denying everything`);
+ }
+ const mayAllow = ({ action, resource }: OpenCodeStatement) =>
+ mode === 'task'
+ ? action === 'read' && typeof resource === 'string' && resource !== '*'
+ : typeof action === 'string' && (action.startsWith('browsentic_') || OPENCODE_WEB.includes(action));
+ for (const rule of rest) {
+ if (rule.effect !== 'deny' && !mayAllow(rule)) problems.push(`lets ${OPENCODE_AGENT} ${String(rule.effect)} ${String(rule.action)} on ${String(rule.resource)}`);
+ }
+
+ const servers = config.mcp?.servers ?? {};
+ if (mode === 'task' && Object.keys(servers).length) problems.push('gives a one-shot an MCP server');
+ if (mode !== 'task' && (Object.keys(servers).join() !== 'browsentic' || servers.browsentic?.codemode !== false)) {
+ problems.push("serves the browser's tools other than directly, from its one server");
+ }
+
+ const policies = config.policies ?? [];
+ for (const resource of ['plugin:*', 'mcp:*']) {
+ const last = policies.filter((policy) => policy.action === 'integration.use' && policy.resource === resource).at(-1);
+ if (last?.effect !== 'deny') problems.push(`does not keep ${resource} from loading`);
+ }
+ return problems;
+ };
+
+const openCode2Rules = (mode: SpawnMode): Requirements => ({
+ required: mode === 'conversation' ? ['acp'] : ['--standalone'],
+ pairs:
+ mode === 'conversation'
+ ? []
+ : [
+ ['--agent', OPENCODE_AGENT],
+ ['--format', 'json'],
+ ],
+ env: OPENCODE2_SEALED,
+ envIsCwd: ['PWD'],
+ inspect: openCode2(mode),
+ files: mode === 'task' ? [] : ['instructions.md'],
+});
+
export const CONTAINMENT: Record<AgentKind, Containment> = {
claude: {
localTools: 'allowlist',
@@ -378,35 +475,43 @@ export const CONTAINMENT: Record<AgentKind, Containment> = {
localTools: 'allowlist',
// A key in ANTHROPIC_* or OPENAI_* would be one of a dozen providers OpenCode reads; `opencode auth login` keeps them in a file.
keepsEnv: ['OPENCODE_'],
- // `--auto` approves whatever is not denied, `--attach` runs the turn in a server started with someone
- // else's config, `--dir` moves it out of the workspace, and `--share` publishes it.
- forbidden: [/^--auto$/i, /^--attach/i, /^--dir/i, /^--share$/i],
+ // `--auto` approves whatever is not denied, `--attach` and `--server` run the turn in a server started with someone
+ // else's config, `--dir` moves it out of the workspace, `--share` publishes it, and `--file` hands the model a file.
+ forbidden: [/^--auto$/i, /^--attach/i, /^--server/i, /^--dir/i, /^--share$/i, /^--file/i, /^--(agent|format|model)=/i],
note:
- 'a per-run permission ruleset opening on "*": "deny", which OpenCode applies after the user’s own rules, ' +
+ 'a per-run permission ruleset opening on denying everything, which OpenCode applies after the user’s own rules, ' +
'so the model is offered no tool that was not named — built-in, custom, or another MCP server’s; ' +
- 'external plugins and project config stay off, but MCP servers the user gave OpenCode itself still start, with their tools hidden',
- run: {
- // Plugins run inside OpenCode and can add tools or answer its permission prompts.
- required: ['--pure'],
- pairs: [
- ['--agent', 'browsentic-contained'],
- ['--format', 'json'],
- ],
- env: OPENCODE_SEALED,
- envContains: { OPENCODE_CONFIG_CONTENT: OPENCODE_CONFIG },
- files: ['instructions.md'],
- },
- task: {
- required: ['--pure'],
- pairs: [
- ['--agent', 'browsentic-contained'],
- ['--format', 'json'],
- ],
- env: OPENCODE_SEALED,
- // An empty map adds no server of ours, so a one-shot has no way to the browser.
- envContains: { OPENCODE_CONFIG_CONTENT: [...OPENCODE_CONFIG, '"mcp":{}'] },
- files: [],
- },
+ 'on 1.x external plugins and project config stay off, but MCP servers the user gave OpenCode itself still start, with their tools hidden; ' +
+ 'on 2.x a policy keeps the user’s plugins and other MCP servers from loading, though one in their global or managed config outranks it',
+ run: [
+ {
+ // Plugins run inside OpenCode and can add tools or answer its permission prompts.
+ required: ['--pure'],
+ pairs: [
+ ['--agent', OPENCODE_AGENT],
+ ['--format', 'json'],
+ ],
+ env: OPENCODE_SEALED,
+ envContains: { OPENCODE_CONFIG_CONTENT: OPENCODE_CONFIG },
+ files: ['instructions.md'],
+ },
+ openCode2Rules('run'),
+ ],
+ task: [
+ {
+ required: ['--pure'],
+ pairs: [
+ ['--agent', OPENCODE_AGENT],
+ ['--format', 'json'],
+ ],
+ env: OPENCODE_SEALED,
+ // An empty map adds no server of ours, so a one-shot has no way to the browser.
+ envContains: { OPENCODE_CONFIG_CONTENT: [...OPENCODE_CONFIG, '"mcp":{}'] },
+ files: [],
+ },
+ openCode2Rules('task'),
+ ],
+ conversation: openCode2Rules('conversation'),
},
};
@@ -419,6 +524,30 @@ export function vetPlan(kind: AgentKind, mode: SpawnMode, plan: SpawnPlan, home:
const rules = CONTAINMENT[kind][mode];
const label = `${AGENTS[kind].label} (${mode})`;
if (!rules) return [`${label} has no containment written for it.`];
+ // A plan meets one way of being contained in full, or is held to the way it came closest to.
+ const shortfalls = alternativesOf(rules).map((alternative) => unmet(alternative, plan, label));
+ const problems = [...(shortfalls.find((shortfall) => shortfall.length === 0) ?? shortfalls.reduce((a, b) => (b.length < a.length ? b : a)))];
+
+ const banned = [...FORBIDDEN, ...(CONTAINMENT[kind].forbidden ?? [])];
+ for (const arg of plan.args) {
+ if (banned.some((pattern) => pattern.test(arg))) {
+ problems.push(`${label} is spawned with ${arg}, which disables its own containment.`);
+ }
+ }
+
+ for (const model of [...everyValueOf(plan.args, '--model'), ...everyValueOf(plan.args, '-m')]) {
+ if (!isModelId(model)) problems.push(`${label} is spawned with the model ${JSON.stringify(model)}, which it could read as a flag.`);
+ }
+
+ if (!within(plan.cwd, home)) problems.push(`${label} would run in ${plan.cwd}, which is outside ${home}.`);
+
+ return problems;
+}
+
+const alternativesOf = (rules: Rules): readonly Requirements[] => ('required' in rules ? [rules] : rules);
+
+/** What a plan lacks of one way of being contained. */
+function unmet(rules: Requirements, plan: SpawnPlan, label: string): string[] {
const problems: string[] = [];
for (const arg of rules.required) {
@@ -449,6 +578,10 @@ export function vetPlan(kind: AgentKind, mode: SpawnMode, plan: SpawnPlan, home:
if (plan.env?.[name] !== value) problems.push(`${label} is spawned without ${name}=${value}.`);
}
+ for (const name of rules.envIsCwd ?? []) {
+ if (plan.env?.[name] !== plan.cwd) problems.push(`${label} is spawned with a ${name} other than the folder it runs in.`);
+ }
+
for (const [name, needles] of Object.entries(rules.envContains ?? {})) {
const missing = needles.filter((needle) => !plan.env?.[name]?.includes(needle));
if (missing.length) problems.push(`${label} is spawned with a ${name} missing ${missing.join(', ')}.`);
@@ -474,18 +607,7 @@ export function vetPlan(kind: AgentKind, mode: SpawnMode, plan: SpawnPlan, home:
if (!allowed) problems.push(`${label} would run "${command.join(' ')}" before its turn, which is not a set-up step it is allowed.`);
}
- const banned = [...FORBIDDEN, ...(CONTAINMENT[kind].forbidden ?? [])];
- for (const arg of plan.args) {
- if (banned.some((pattern) => pattern.test(arg))) {
- problems.push(`${label} is spawned with ${arg}, which disables its own containment.`);
- }
- }
-
- for (const model of everyValueOf(plan.args, '--model')) {
- if (!isModelId(model)) problems.push(`${label} is spawned with the model ${JSON.stringify(model)}, which it could read as a flag.`);
- }
-
- if (!within(plan.cwd, home)) problems.push(`${label} would run in ${plan.cwd}, which is outside ${home}.`);
+ for (const clause of rules.inspect?.(plan) ?? []) problems.push(`${label} ${clause}.`);
return problems;
}