Bug: foreground one-shot subagents (workflow/ralph/run_in_background:false) fail on any tool call when Agent Teams is mounted

[NingYuanLee]

Bug: foreground one-shot subagents (workflow/ralph/subagent with run_in_background: false) fail on any tool call when Agent Teams is mounted

Summary

When the experimental Agent Teams layer is mounted (host rows from dsh-experimental-agent-team-profile), any foreground, one-shot child that calls a tool fails its whole run — including workflow agent(), ralph rounds, and plain subagent with run_in_background: false. Pure-text replies succeed; background/continuable children (default background subagent, Agent Teams teammates) are unaffected.

Observed symptoms:

• foreground subagent calling todo_write/read → tool result subagent run failed (stopReason error, empty output)
• workflow agent() requiring any tool → resolves null (child failed)
• ralph round requiring any tool → Ralph round N child failed before producing a structured report

Root cause

A window between a one-shot child's creation and the appending of its subagent/descriptor event:

1. The one-shot driver (packages/subagent/subagent-in-process-driver/src/index.ts:81-91, attachDescriptorAppend) appends the subagent/descriptor event only at the child's first pre-step, after the prompt assembly has run (agent/pre-step → await next() → append).
2. agent/created fires during publication, before that first pre-step. At that moment the roster's tryMembership finds no descriptor event in the child's own events, so TeamRoster.tryMembership falls through to the "implicit root Team" fallback (packages/experimental/agent-team/src/roster.ts) and classifies the one-shot child as a Team Lead.
3. dsh-experimental-tool-agent-team (maybeInstall, tool-agent-team/src/index.ts:404-409) therefore installs the Team policy section + ten Team tools into that child's scope.
4. When the child calls any tool, the turn reaches a second prompt assembly (step ≥ 2). The team:policy section text calls ctx.agentTeams.membership(agent); by then the descriptor event has been appended, so tryMembership now correctly returns undefined and membership() throws TEAM_NOT_MEMBER.
5. The assembly exception converges the turn with kind: 'error' → stopReason: 'error' → foreground caller sees null/subagent run failed.

Session-log evidence from a failing child (tool/call → tool/result ok → turn/end reason {"kind":"error","error":{"message":"agent \"…\" is not a member of an active Agent Team","code":"UNKNOWN"}}, no second step, no assistant text).

Continuation children (background subagents, teammates) are immune because their descriptor is seeded before publication (descriptor-seed.ts used by continuation.ts), so agent/created observers classify them correctly.

Proposed fix

Classify provider-owned children by the durable session header (origin: 'subagent', written at creation by childSessionMeta) rather than by the deferred log event alone. Minimal change in roster.ts subagentDescriptor():

return agent.session.header.origin === 'subagent'
  || foldSubagentDescriptor(agent.session.ownEvents()) !== undefined

plus a regression test asserting a provider child is not treated as a Team Lead before its descriptor event exists.

Reproduction

Deployment: source checkout, web profile with @deepseek-ai/dsh-experimental-agent-team-profile (+ -web-profile) bundles; default agent preset. Steps: run the foreground subagent tool with run_in_background: false and ask it to call read or todo_write; or run a workflow script whose agent() prompt requires reading a file; or a ralph objective requiring a tool.

Environment

• dsh source checkout (apps/cli web profile), Node 24
• experimental Agent Teams profile layers mounted
• agent preset with delegation tools