Skip to main content

Team members

One Server hosts many workspaces, and each workspace is one team member with its own checkout and its own orchestration. The member kind fixes what that workspace's agent may read, write, drive, and raise. This page is the contract for that matrix and for the in-process agent PM reviewers run on.

Keep in view

Read

GET /api/workspaces/{id}/agent/member

Change

PUT /api/workspaces/{id}/agent/member with { "kind": "Author" | "DevReviewer" | "PmReviewer" }

Default

Author

Definition

TeamMemberGrants in the Server Agents slice

Matrix​

KindReads repositoryWrites repositoryStarts its workspaceDrives browserSubmits reviewsFixture
AuthoryesyesyesyesnoACP coding CLI
DevRevieweryesnoyesnoyesACP coding CLI
PmReviewernonono, the Server starts ityesyesin-process

Authors and dev reviewers work on code, so they run an ACP coding CLI against their checkout: Claude, Codex, or Cursor today. PM reviewers have nothing to do with code and skip that layer.

Each kind also has a default unattended role: authors take picked issues and end with a proposal, reviewers take proposals awaiting their review and end with a verdict. The Server commits, pushes, and posts for them under the member's own forge account; see Team pipeline.

A review raises threads another member must address, so only reviewers submit one, through submit_review on /mcp/tasks. Authors still post plain proposal comments, and address threads through enqueue_review_fix_commit.

Enforcement​

Every enforcement point reads TeamMemberGrants; none restates the rule.

  • Binding. Binding, scheduling, or resuming an agent of the other fixture is refused with a typed error naming the member kind.
  • Permission policy. PUT .../agent/policy refuses an allow rule outside the grant. A member that cannot write may allow only named Execute commands, because an unscoped command can write the worktree. A PM may allow only Network.
  • Forge reviews. ForgeService.SubmitReviewAsync refuses an Author workspace; POST .../forge/reviews answers 403. REST, an in-process agent, and any later caller meet the same check.
  • MCP. A session pinned to a workspace by a workspace claim is narrowed to the member's operation permissions after the token, workspace, and role narrowings. Each narrowing only removes. An ACP agent is pinned the same way through the session credential its offered MCP servers carry, which also carries the member's permissions for the endpoint floors.
  • Changing kind. Refused while autonomy is active, while the bound agent is the wrong fixture, or while a stored policy rule exceeds the new grant.
  • Linux. Every ACP agent starts inside bubblewrap through IAgentSandbox, so the grant holds even for an agent that never asks permission. See below.

Agent sandbox​

The Server runs inside its Nix shell, which supplies bubblewrap. BubblewrapAgentSandboxProvider wraps each ACP launch, and AgentSandboxLayout decides the view from the member's grant. The view is an allowlist; anything not listed is absent.

PathAuthorDev reviewer
/usr, /etc, /lib*, /bin, /optread-onlyread-only
/nixread-only, daemon socket writableread-only, daemon socket writable
Server PATH entries, capability shells, DOTNET_ROOT, the agent's packageread-onlyread-only
Agent CLI home under the data directorywritablewritable
Its clonewritableread-only
Its .git (both directories of a linked worktree)read-onlyread-only
The rest of the data directory, other clones, the service account's homeabsentabsent

A single-user Nix install has no daemon, so there /nix is writable for nix shell. The Server decides that from the daemon's socket file, not its directory, which installers leave on single-user hosts too.

The agent is executed by the path found on PATH, not the symlink's target, and the target's directory is mounted beside it. nix-shell is a link to the multi-call nix binary, which reads its mode from argv[0], so following the link ran capability shells as plain nix. The agent also gets a private /tmp, its own PID, IPC, and UTS namespaces, and no capabilities. Its environment is cleared to a short allowlist (locale, proxy, certificate, and Nix variables, PATH, DOTNET_ROOT) plus what the Server sets for it, so AGENTUP_ADMIN_PASSWORD and other Server secrets never reach it.

A Linux host without bubblewrap fails the session start with that cause. bubblewrap is Linux only; on macOS the agent starts with the cleared environment but its filesystem is not confined, so the clone-level guarantee holds only on a Linux Server. Running the agent with the Server's own permissions instead would let a reviewer write the clone. In a container, bubblewrap needs unprivileged user namespaces; Ubuntu 24.04 hosts may need kernel.apparmor_restrict_unprivileged_userns=0.

In-process agent​

The ollama agent runs inside the Server against an Ollama-compatible /api/chat endpoint with tool calling. It is listed only when AGENTUP_OLLAMA_MODEL is set. AGENTUP_OLLAMA_URL defaults to http://127.0.0.1:11434. Use a vision-capable model so screenshots carry information.

It answers ACP v1 (initialize, session/new, session/prompt, session/cancel), so scheduling, the event stream, permission decisions, the loop kernel, and both clients treat it like any other agent. It has no process, no filesystem, and no shell.

Its tools come from the Server's own /mcp/browser and /mcp/orchestration over loopback, through LoopbackMcpClientProvider. AGENTUP_INPROCESS_MCP_URL overrides the dialed listener. InProcessAgentToolService narrows them three ways, each only removing:

  1. Tools that name their workspace through workspaceId. That parameter and worktreePath are removed from the schema the model sees and supplied by the host.
  2. Tools whose McpToolPermissionMap permission the member grant holds.
  3. The one workspace the agent belongs to.

Two host tools sit beside them; they are not MCP tools. Submit review goes through the forge review check and records ReviewSubmitted for the role termination. Complete task ends the loop task.

Every tool call except Complete task is first sent as ACP session/request_permission with tool kind fetch, which DeclaredPermissionPolicy classifies as Network. A PM's autonomy policy is therefore one Network allow rule, and without a policy a person approves each call.

PM review loop​

A PM reviewer is an autonomous workspace with role trigger ProposalAwaitingReview and termination ReviewSubmitted. Its prompt names the proposal id, title, and URL. An unmet termination after Completed still stops as blocked:awaiting-input.

{
"prompt": "Review the running product against the proposal. Report what a user would notice.",
"tools": [],
"trigger": "ProposalAwaitingReview",
"termination": "ReviewSubmitted"
}

Tests​

TeamMemberGrantsTests holds the matrix. AutonomyHttpTests covers the member routes, fixture refusal, and policy refusal. ForgeServiceTests covers the review gate. Provider tests cover the chat request shape, loopback MCP sessions, schema pinning, and the ACP emulation with a scripted model.

Next in this slice

Sign-in

Subscription sign-in transports for ACP coding CLIs.