@team-agent/installer 0.5.101 → 0.5.103

package/crates/team-agent/src/state/repository.rs

@@ -161,10 +161,6 @@ impl<'a> StateRepository<'a> {
         #[cfg(test)]
         test_support::before_commit();
         let team_key = match intent {
-            // Assignment's existing reconcile helper updates the root task
-            // array and the explicit Team array independently (legacy parity).
-            StateWriteIntent::McpAssignTask { .. } => None,
-            StateWriteIntent::McpUpdateStateNote { team_key } => team_key,
             StateWriteIntent::ResultCollection { owner_team_id } => owner_team_id,
             StateWriteIntent::ForkAgent { team_key, .. } => Some(team_key),
             _ => {
@@ -517,16 +513,6 @@ pub enum StateWriteIntent<'a> {
     CoordinatorConptyShim {
         team_key: Option<&'a str>,
     },
-    McpAssignTask {
-        team_key: Option<&'a str>,
-        task_id: &'a str,
-    },
-    McpLifecycleAgentOps {
-        team_key: Option<&'a str>,
-    },
-    McpUpdateStateNote {
-        team_key: Option<&'a str>,
-    },
     MessagingDeliveryState {
         owner_team_id: Option<&'a str>,
     },
@@ -537,9 +523,6 @@ pub enum StateWriteIntent<'a> {
     ResultCollection {
         owner_team_id: Option<&'a str>,
     },
-    SchedulerSuppression {
-        owner_team_id: Option<&'a str>,
-    },
     TaskRepair {
         team_key: Option<&'a str>,
     },
@@ -735,15 +718,6 @@ fn route_direct(
         // CoordinatorConptyShim -> root save
         // (coordinator/conpty_shim.rs:426/674).
         StateWriteIntent::CoordinatorConptyShim { .. } => helper_write_root(workspace, state),
-        // McpAssignTask uses the reapply variant today; direct save routes
-        // to the root helper for parity.
-        StateWriteIntent::McpAssignTask { .. } => helper_write_root(workspace, state),
-        // McpLifecycleAgentOps -> root save
-        // (mcp_server/lifecycle_tools/agent_ops.rs:385).
-        StateWriteIntent::McpLifecycleAgentOps { .. } => helper_write_root(workspace, state),
-        // McpUpdateStateNote -> root save
-        // (mcp_server/lifecycle_tools/state_status.rs:29/69).
-        StateWriteIntent::McpUpdateStateNote { .. } => helper_write_root(workspace, state),
         // MessagingDeliveryState direct writes cover team-scoped, root
         // fallback, and root (messaging/delivery.rs:2356/2358/2361).
         StateWriteIntent::MessagingDeliveryState { owner_team_id } => match owner_team_id {
@@ -766,8 +740,6 @@ fn route_direct(
             Some(_) => helper_write_team_scoped(workspace, state),
             None => helper_write_root(workspace, state),
         },
-        // SchedulerSuppression -> root save (messaging/scheduler.rs:259).
-        StateWriteIntent::SchedulerSuppression { .. } => helper_write_root(workspace, state),
         // TaskRepair -> `save_team_scoped_state` (cli/adapters.rs:675).
         StateWriteIntent::TaskRepair { .. } => helper_write_team_scoped(workspace, state),
         // FakeE2eSeed / SelfMigration are diagnostic seams. The current
@@ -798,9 +770,6 @@ fn reapply_scope(intent: &StateWriteIntent<'_>) -> ReapplyScope {
                 owner_team_id: Some(_),
             }
             | StateWriteIntent::CoordinatorTick { .. }
-            | StateWriteIntent::McpUpdateStateNote {
-                team_key: Some(_),
-            }
     ) {
         ReapplyScope::Team
     } else {
@@ -817,8 +786,7 @@ fn route_reapply<F>(
 where
     F: FnOnce(&mut Value),
 {
-    if matches!(&intent, StateWriteIntent::McpAssignTask { .. }
-        | StateWriteIntent::McpUpdateStateNote { .. } | StateWriteIntent::ResultCollection { .. })
+    if matches!(&intent, StateWriteIntent::ResultCollection { .. })
     {
         return StateRepository::new(workspace).commit(intent, reapply).map(|_| ());
     }

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@team-agent/installer",
-  "version": "0.5.101",
+  "version": "0.5.103",
   "description": "npx installer for Team Agent",
   "keywords": [
     "codex",
@@ -20,9 +20,9 @@
     "team-agent-installer": "npm/install.mjs"
   },
   "optionalDependencies": {
-    "@team-agent/cli-darwin-arm64": "0.5.101",
-    "@team-agent/cli-darwin-x64": "0.5.101",
-    "@team-agent/cli-linux-x64": "0.5.101"
+    "@team-agent/cli-darwin-arm64": "0.5.103",
+    "@team-agent/cli-darwin-x64": "0.5.103",
+    "@team-agent/cli-linux-x64": "0.5.103"
   },
   "scripts": {
     "postinstall": "node npm/bincheck.mjs",

package/skills/team-agent/SKILL.md

@@ -12,14 +12,14 @@ team-agent --version
 team-agent --help
 team-agent doctor --help
 ```
-**Launch** from a tmux-addressable pane: `team-agent claude` or `team-agent codex`, then `team-agent quick-start .team/current`. Do not start a real team from a naked terminal; workers use independent background windows on the workspace-scoped tmux server.
+**Launch** from a tmux-addressable pane: `team-agent claude` or `team-agent codex`, then `team-agent quick-start .team/current`. For a plain-text team directory, use `team-agent quick-start <plain-text-team-dir>`. Do not start a real team from a naked terminal; workers use independent background windows on the workspace-scoped tmux server.
 **Operate**
 - Dispatch: `team-agent send TO MESSAGE` (positional TO; `--watch-result` is deprecated). After success, do not poll with `sleep` / `status` / `inbox` / `collect`. TO has two co-equal logical forms: an in-team short name (`team-agent send reviewer "..."`) and a fully qualified `<workspace>::<team>/<agent>`. Use the qualified form across workspaces.
 - Inspect: `team-agent status` / `status --json`; `ok: true` plus `ready: false` is not a crash.
 - Lifecycle: `restart .` resumes a stopped team; `add-agent NAME --role-file FILE` adds or `--force` recreates one worker; `shutdown --workspace .` stops. Do not shutdown the team to add a worker.
-- Roles: every `agents/*.md` must declare boolean `dangerously_skip_permissions`. Never rewrite user-supplied model ids or read `.env` files.
+- Documents: `TEAM.md` and `agents/*.md` may be plain text/Markdown with no frontmatter; the whole team body is its objective and the whole role body is the prompt. Default worker id is the filename stem, provider `pi`, model `openai-codex/gpt-6-luna`, effort `max`, and bypass enabled. Valid optional frontmatter can override supported metadata.
 - On a structured failure `action`, run that action first, then stop. Do not guess flags. `coordinator.session_missing` is self-healing; re-check `status --json`, do not shutdown.
-The current user-facing agent is the leader (orchestrate only). Workers call `report_result` exactly once. Nested teams: `references/team-in-team.md`.
+The current user-facing agent is the leader (orchestrate only). Worker MCP exposes exactly `send_message`, `report_result`, and optional `get_team_status`; lifecycle management stays on the authorized native CLI. Workers call `report_result` exactly once. Nested teams: `references/team-in-team.md`.
 ## Provider Capability Matrix
 See `references/team-agent-operator.md` for Claude / Codex / Copilot / Gemini / fake. `quick-start` / `restart` / `claim-leader` JSON is compact by default; pass `--detail` only for internal diagnostics.
 | Provider | Resume | Turn-state detection | Per-worker model override | Team Agent fork |
@@ -30,8 +30,8 @@ See `references/team-agent-operator.md` for Claude / Codex / Copilot / Gemini /
 Grok / `cursor_agent` have no JSONL turn-state reader (classify → Unknown). Pi `fork-agent SOURCE --as TARGET` requires a captured Pi subscription source and reports both session ids plus the target backing path; it does not mutate the source session.
 ## Provider Prep
 ### Cursor provider notes
-Frontmatter: `provider: cursor_agent` (not `cursor`; launcher verb `team-agent cursor`), `auth_mode: subscription`, `name:` required, plus `role:`, `tools:`, and boolean `dangerously_skip_permissions:`. Subscription needs no `profile`.
-`model:` is required. The flag stays on argv; same-family catalog ids can change pane chrome, while unknown ids silent-fallback without warning. Pick a catalog id; after spawn, `capture-pane` once for chrome. The role field and pane chrome do not prove the live model.
+To opt into Cursor, use optional frontmatter with `provider: cursor_agent` (not `cursor`; launcher verb `team-agent cursor`) and `auth_mode: subscription`; no profile or bypass field is required. `agent_id` (legacy `name`) and `role` can be omitted, but Cursor still requires an explicit `model:`.
+The model flag stays on argv; same-family catalog ids can change pane chrome, while unknown ids silent-fallback without warning. Pick a catalog id; after spawn, `capture-pane` once for chrome. Pane chrome does not prove the live model.
 One `cursor_agent` per workspace; a second seat fail-closes because `.cursor/mcp.json` is directory-scoped and overwrites `TEAM_AGENT_ID`. Several Cursor seats need separate workspace directories. Same seat, fresh context: `reset-agent --discard-session`.
 `clone-agent` copies the source role but starts fresh. Runtime role replacement: `clone-agent` → `stop-agent` → `remove-agent --confirm` (deletes `.team/dynamic-role-files/`) → write role file → `add-agent --role-file` → dispatch.
 - Restart emits `--resume <chatId>` when `store.db` / `meta.json` exist; the gate does not read chat text. Persist anything that must survive restart.

package/skills/team-agent/command-coverage.json

@@ -261,6 +261,13 @@
         }
       }
     },
+    {
+      "command": "team-agent quick-start <plain-text-team-dir>",
+      "bucket": "declared_gap",
+      "covered": false,
+      "owner": "runtime-owner:lifecycle",
+      "plan": "Run exact argv team-agent quick-start <plain-text-team-dir> in an isolated workspace with plain-text TEAM.md and agents/<stem>.md; bind successful compilation, objective, filename-derived id, and default provider/model/effort/bypass in the generated spec."
+    },
     {
       "command": "team-agent remove-agent <agent> --workspace . --confirm",
       "bucket": "declared_gap",

package/skills/team-agent/references/team-agent-operator.md

@@ -45,7 +45,7 @@ team-agent claude
 
 Pass provider flags after the provider name, for example `team-agent codex --dangerously-bypass-approvals-and-sandbox`. Run `team-agent quick-start` from the leader's current tmux pane. Do not start a real team from a naked terminal that Team Agent cannot address through tmux.
 
-**0.5.66:** leader `--dangerously-*` flags no longer inherit bypass to workers. Per-worker bypass is the role field `dangerously_skip_permissions` (see Permissions).
+Leader `--dangerously-*` flags do not determine worker bypass. Worker bypass defaults to enabled in compiled role specs; no bypass field needs to be written (see Document Format and Permissions).
 
 ## Leader Role
 
@@ -55,48 +55,22 @@ When the user has been communicating in Chinese throughout the conversation, all
 
 ## Minimal Copy-Paste Team
 
-`dangerously_skip_permissions` is **required** on every role doc (boolean). `TEAM.md` `dangerous_auto_approve` is **not** the 0.5.66 bypass switch (see Permissions). Values under `provider_models:` below are **fill-ins for omitted role `model`**, not a validity catalog; the values are configuration fallbacks.
+`TEAM.md` and `agents/*.md` are plain text/Markdown; YAML frontmatter is optional. The whole `TEAM.md` body becomes the team objective, and the whole role body becomes the worker prompt. A default role uses its filename stem as its id, provider `pi`, model `openai-codex/gpt-6-luna`, effort `max`, and bypass enabled. Valid optional frontmatter may still override supported metadata.
 
 ```bash
 mkdir -p .team/current/agents
-team-agent profile init codex-default --auth-mode subscription --workspace .
 cat > .team/current/TEAM.md <<'EOF'
----
-name: demo-team
-objective: One worker handles bounded tasks and reports through Team Agent MCP.
-dangerous_auto_approve: false
-fast: false
-provider_models:
-  codex: gpt-5.5
-  claude: claude-sonnet-4-6
-  claude_code: claude-sonnet-4-6
----
-
-Team config only. This is not a worker role.
+A small team for bounded implementation tasks. Workers report progress to the leader and return a concise result.
 EOF
 cat > .team/current/agents/coder.md <<'EOF'
----
-name: coder
-role: Implementation Worker
-provider: codex
-auth_mode: subscription
-profile: codex-default
-dangerously_skip_permissions: false
-tools:
-  - fs_read
-  - fs_list
-  - fs_write
-  - execute_bash
-  - mcp_team
-  - provider_builtin
----
-
-Handle one bounded task at a time. Send progress to leader only when needed. Final completion must call report_result exactly once; MCP fills task ids and result envelope fields.
+# Implementation Worker
+
+Handle one bounded task at a time. Send progress or blockers to the leader with `send_message`, and report completion exactly once with `report_result`.
 EOF
 team-agent quick-start .team/current
 ```
 
-YAML lists must be block style. Use `tools:\n  - fs_read`; do not use `tools: [fs_read, mcp_team]`.
+The MCP surface is exactly `send_message`, `report_result`, and `get_team_status`; status is optional, not a prerequisite. Agent/team lifecycle management remains on the separately authorized native CLI.
 
 Workers always run headlessly in independent windows on the per-workspace tmux server. No GUI display backend is configured in `TEAM.md`.
 
@@ -132,7 +106,7 @@ Run `claim-leader` only from the **leader** pane. From a worker pane it refuses
 | `fake` (testing only) | no | no | n/a | no |
 
 Notes:
-- Per-worker model override means a role-doc `model:` value wins over `TEAM.md` `provider_models.<provider>` at **compile** time; subscription defaults still fill blanks when there is no profile-deferred null.
+- A plain-text/default role uses Pi with `openai-codex/gpt-6-luna` and `max` effort; valid optional role frontmatter can select a provider and override supported model settings. For provider paths using `provider_models`, an explicit role `model:` takes precedence.
 - Copilot fork copies the source session into an isolated `COPILOT_HOME` and rekeys its SQLite session references atomically. Missing or incomplete backing fails closed; it never falls back to a fresh spawn.
 - Copilot phase-1 idle/turn detection is intentionally Unknown; tick emits a single explicit `provider.classify.unsupported` event per state change (P4 dedup), never a silent default.
 - `pi` `fork-agent SOURCE --as TARGET` requires a captured Pi subscription source. It copies the complete JSONL history into a distinct target session with a new id and exact `--session <path>` resume; it does not alter the source file or fall back to a fresh session.
@@ -141,7 +115,7 @@ Notes:
 
 ### Subscription auth (Codex / Claude account login)
 
-Before workers can use a subscription provider, create a named subscription profile in the workspace and reference it from role docs:
+To opt a worker into Codex or Claude subscription providers, create a named subscription profile in the workspace and reference it from optional role frontmatter:
 
 ```bash
 team-agent profile init codex-default --auth-mode subscription --workspace .
@@ -174,7 +148,7 @@ codex debug models
 
 That command is the Codex catalog. This skill does **not** keep a model id list. If `codex` is not on PATH, the lookup is unavailable; do not invent slugs.
 
-Role docs may omit `model` for non-grok subscription workers; compile then fills from `TEAM.md` `provider_models` or leaves `null` when a profile is present (see compatible_api). **Do not treat those fill-ins as a validity catalog.** **Grok roles cannot omit `model`** — the internal compiler fails closed.
+A plain-text/default role compiles as provider `pi` with model `openai-codex/gpt-6-luna` and effort `max`. If valid optional frontmatter selects another provider, its established provider-specific model/profile resolution applies; Grok and Cursor roles retain their explicit model requirements. **Do not treat provider/team fill-ins as a validity catalog.**
 
 Claude: run `claude auth status`; if missing, run `claude auth login`. Team Agent stores Claude worker sessions by passing `--session-id` and resumes with `--resume`.
 Use `provider: claude` or `provider: claude_code` for Claude workers.
@@ -250,7 +224,7 @@ Observed (exit 1): `usage error: invalid --auth-mode: not-a-mode`. That is the l
 
 Team Agent loads the profile during quick-start, launch, restart, and start-agent. Compatible API workers inherit the current shell proxy/CA environment by default. Claude compatible API workers use Team Agent managed `CLAUDE_CONFIG_DIR` so user-level Claude subscription settings cannot re-inject Anthropic proxy variables into third-party API sessions. If quick-start reports an ambient proxy blocker, do not silently unset proxy for the whole team; tell the user to choose one path: fix that proxy for `BASE_URL`, put `HTTPS_PROXY=`/`HTTP_PROXY=` in the profile, or put `PROXY_MODE=direct` in the profile to bypass proxy only for that worker. Subscription workers keep their native provider settings and environment. Startup runs a redacted smoke check for compatible API profiles before worker windows are created, so a bad URL/key/model or proxy/base URL connectivity failure is reported to the leader command instead of producing idle workers.
 
-## How role-doc frontmatter becomes live
+## How role documents become live
 
 ```bash
 team-agent restart --help
@@ -276,72 +250,15 @@ To apply a changed role file to an **existing** worker name, use `add-agent --fo
 
 Do **not** shutdown/restart the whole team just to add a **new** worker (that drops other workers' resumable sessions). Use `add-agent` without `--force` for a new name.
 
-## Permissions (0.5.66)
-
-There is **no** `team-agent permission-modes` command.
-
-```bash
-team-agent permission-modes
-```
-
-Observed (exit 1): `invalid choice: 'permission-modes'`.
-
-`team-agent --help` does **not** list `permission_mode`, `TEAM_AGENT_LEADER_BYPASS`, or `dangerous_auto_approve`. Use `team-agent validate "$d" --json` with an existing spec file or team directory; it does not enumerate permission modes.
-
-### Live control: role `dangerously_skip_permissions`
-
-Required boolean on every role doc. Missing or non-bool fails the internal compiler; this is checked again before lifecycle launch.
-
-For an existing spec file, the retained schema check is:
-
-```bash
-team-agent validate "$d" --json
-```
-
-With a team-directory input, `validate` internally compiles `TEAM.md` and `agents/*.md` and validates role-document completeness; with a file input, it runs schema structural validation. Neither input writes or exports a spec file. The internal compiler rejects non-bool (`dangerously_skip_permissions: bypass`) and accepts only boolean **`true`** or **`false`**.
-
-`true` is the per-worker bypass opt-in. `false` is the default you should copy unless the user explicitly wants prompts skipped.
+## Document Format and Permissions
 
-### `permission_mode` (historical, not consumed)
+`TEAM.md` and `agents/*.md` accept ordinary text/Markdown without required YAML frontmatter. A valid optional frontmatter map may set supported metadata; absent, malformed, unterminated, or non-object frontmatter is treated as part of the document body, not a compile error. The entire TEAM body becomes `objective`; the role body is injected as the prompt.
 
-A role may still contain `permission_mode`. Compile **does not** validate it and **does not** copy it into the compiled spec.
+For a role with no overrides, the filename stem supplies its id, and defaults are provider `pi`, model `openai-codex/gpt-6-luna`, effort `max`, and `dangerously_skip_permissions: true`. A valid `agent_id` (or legacy `name`) may override the filename-derived id. Supported valid frontmatter overrides remain available for existing roles; no permission field needs to be written. TEAM defaults are the workspace directory name, a 2-second tick, `leader_centric` communication, and `team-<workspace-name>` session.
 
-Observed: `permission_mode: bypass` and `permission_mode: not-a-mode` both compiled `ok: true` (exit 0); the spec agent block had `dangerously_skip_permissions: false` and **no** `permission_mode` key.
-
-Do not set `permission_mode: bypass` expecting 0.5.66 bypass. Use `dangerously_skip_permissions: true`.
-
-### `TEAM.md` `dangerous_auto_approve`
-
-The demo `TEAM.md` still shows `dangerous_auto_approve: false`. Compile of a team with `dangerous_auto_approve: true` produced a spec **without** that key on the agent (only `dangerously_skip_permissions`). Treat it as leftover team YAML, not the bypass switch.
-
-`dangerous_auto_approve` as a **tool sentinel** is internal (adapters). Operators do not configure it as a tools list item.
-
-### `TEAM_AGENT_LEADER_BYPASS`
-
-Not in `team-agent --help`. Not a user-facing config key. Do not export it by hand. Worker env injection was **not** live-spawned on this gauge.
-
-### Leader argv bypass
-
-Leader `--dangerously-*` is detected only as a **warning**, not as worker bypass (`detect_bypass_flag_in_argv` — "只做检测,不做行为决策"). Declare `dangerously_skip_permissions` per role.
-
-### Priority — 未找到单一裁决点
-
-An external report ranked `permission_mode` > `TEAM_AGENT_LEADER_BYPASS` > `TEAM.md` `dangerous_auto_approve`. **That ranking is not in this tree.** Grep of `crates/` found **no** function that takes those three keys and picks a winner (search: `permission_mode.*TEAM_AGENT_LEADER_BYPASS` and the reverse: zero hits).
-
-What the sources actually do (reader can re-open these files):
-
-| Symbol | File | What it does |
-|---|---|---|
-| `required_dangerously_skip_permissions` | `crates/team-agent/src/compiler.rs` | Compile **requires** role `dangerously_skip_permissions` bool. This is the 0.5.66 worker bypass source. |
-| `resolved_tool_strings_for_command` | `crates/team-agent/src/lifecycle/worker_command_context.rs` | `true` appends tool sentinel `dangerous_auto_approve`; comment: no longer consumes team/runtime/leader `DangerousApproval`. |
-| `provider_bypass_flag` | `crates/team-agent/src/provider/bypass_flags.rs` | **唯一权威**: role `dangerously_skip_permissions: true` decides whether to add a provider bypass argv flag. Maps provider → flag string only. |
-| spec allowed key `permission_mode` | `crates/team-agent/src/model/spec.rs` | Historical; **not consumed** by compiler (comment: 0.5.66 起不再被 compiler 消费); the key is absent from the compiled agent spec. |
-| `apply_mcp_auto_approval_env` | `crates/team-agent/src/lifecycle/launch/worker_env.rs` | Writes `TEAM_AGENT_LEADER_BYPASS` `1`/`0` from `DangerousApproval` when source is `LeaderProcess`. Separate from compile. |
-| `worker_spawn_env` test | `crates/team-agent/src/layout/worker_env.rs` | **Strips** inherited `TEAM_AGENT_LEADER_BYPASS` from parent env. |
-| `claude_dangerous_auto_approve` (and siblings) | `crates/team-agent/src/provider/adapters/*.rs` | Adapters look for tool name `dangerous_auto_approve`, not `TEAM.md`. |
-
-**Do not copy a three-row priority table.** The live operator knob on this gauge is role `dangerously_skip_permissions`. The three names from the report still exist as strings; they are not one ranked merge.
+The generated spec still carries a boolean bypass value. Missing or malformed role metadata defaults to `true`; an explicit valid boolean is preserved. Provider adapters translate that value using the provider's own mechanism. Do not copy a Codex-only bypass flag into Pi or another provider's launch command.
 
+`permission_mode`, `dangerous_auto_approve`, and leader environment switches are not required document fields. Keep provider-native bypass behavior inside the adapter; do not ask users to add a bypass toggle to `TEAM.md` or a plain-text role.
 ## Ignore vs handle (status / alerts)
 
 ```bash
@@ -524,7 +441,7 @@ provider or treats a diagnostic example as a command approval.
 
 `restart` takes one workspace argument. It preserves each worker's original provider. If a verified provider session exists, the worker resumes (`codex resume <id>` or `claude --resume <id>`). Claude sessions are considered resumable only after the provider has written a project transcript for that session; a freshly opened blank Claude window is not recorded as recovered context. If the stored id is stale, the runtime first tries to repair it from verified transcript history. If a stored session cannot be verified or repaired, restart fails closed instead of silently losing context; use `team-agent restart . --allow-fresh` only when the user explicitly accepts a fresh worker context. If multiple stopped teams in the same workspace have restart context, plain `team-agent restart .` fails and lists candidates; rerun with `--team <session_name_or_team_name>`. If no prior session id exists, that worker starts fresh and the event log records `restart.fresh_spawn`. Claude resume must run from the original cwd and the same provider transcript root; Team Agent stores `spawn_cwd` and compatible-API `claude_projects_root` for that.
 
-`restart` `--help` does not mention rereading `agents/*.md`. See **How role-doc frontmatter becomes live**.
+`restart` `--help` does not mention rereading `agents/*.md`. See **How role documents become live**.
 
 Startup trust prompts are handled by the runtime/coordinator with bounded probes; do not wait on raw worker screens or manually press Enter for routine startup trust prompts.
 
@@ -536,20 +453,9 @@ To add a new worker to a running team, write the role doc and run **one command*
 
 ```bash
 cat > .team/current/agents/reviewer.md <<'EOF'
----
-name: reviewer
-role: Code Reviewer
-provider: codex
-auth_mode: subscription
-profile: codex-default
-dangerously_skip_permissions: false
-tools:
-  - fs_read
-  - fs_list
-  - mcp_team
----
-
-Review changed files and report findings to leader.
+# Code Reviewer
+
+Review changed files and report findings to the leader. Use `send_message` for progress and `report_result` exactly once for completion.
 EOF
 team-agent add-agent reviewer --role-file .team/current/agents/reviewer.md --workspace .
 ```
@@ -581,6 +487,8 @@ Removing a worker at runtime is the symmetric `team-agent remove-agent <agent> -
 
 ## Worker Protocol
 
+The Team Agent MCP surface is exactly `send_message`, `report_result`, and optional `get_team_status`. Status is self-inspection only, never a prerequisite; lifecycle management remains on the authorized native CLI.
+
 Workers normally do not run nested Team Agent teams. When the user or leader explicitly asks for a child team, follow `references/team-in-team.md`; otherwise workers only provide the target and content for progress, and a short completion summary at the end:
 
 ```text
@@ -594,7 +502,7 @@ team_orchestrator.report_result(summary="short completion", status="success", te
 
 For typed orchestration traffic, both `send_message` and `report_result` accept `presentation={"sink":"leader|casefile|silent","class":"message|progress|stage_result|stage_pass|bounce|blocking|final_review|timeout","case_id":"optional-case"}`. If the object is present, `sink` and `class` are required and unknown values fail closed. `casefile` and `silent` are durable-only, not deletion. The fixed critical classes `stage_pass`, `bounce`, `blocking`, `final_review`, and `timeout` always appear on the leader screen even when another sink is requested. Routing uses the typed class, never words in the content or summary.
 
-Do not pass `sender`, `task_id`, `requires_ack`, `schema_version`, or `agent_id` unless doing a low-level compatibility diagnostic. The MCP runtime fills those fields and keeps delivery metadata in runtime state and event logs. If provider env loses the worker id, MCP infers it from active task/message state and falls back to an explicit `unknown` sender instead of treating the worker as leader.
+Do not pass `sender`, `task_id`, `schema_version`, or `agent_id` unless doing a low-level compatibility diagnostic. The MCP runtime fills those fields and keeps delivery metadata in runtime state and event logs. If provider env loses the worker id, MCP infers it from active task/message state and falls back to an explicit `unknown` sender instead of treating the worker as leader.
 
 Message targets are team-scoped. Use `leader`, another teammate agent id, or `*` for all other team members. The runtime excludes the sender from `*` broadcasts and never scans unrelated terminal windows for recipients.
 
@@ -630,7 +538,7 @@ Examples of `action` observed on this gauge:
 
 Do not retry with changed flags. Do not inspect source code or private runtime state. Do not operate tmux directly except when the user asks for a manual diagnostic. Do not answer provider approval prompts for the user.
 
-If `quick-start` reports `tmux session already exists`, treat it as a team-name collision. The existing session may be an active team; do not terminate it and do not suggest `shutdown` as the normal fix. Change `name:` in `TEAM.md` so the next launch uses a different tmux session name, then run `team-agent quick-start .team/current` again.
+If `quick-start` reports `tmux session already exists`, treat it as a team-name collision. The existing session may be an active team; do not terminate it and do not suggest `shutdown` as the normal fix. Choose a distinct workspace folder name or set an optional `name:` override in valid `TEAM.md` frontmatter, then retry `team-agent quick-start .team/current`.
 
 Known Team Agent control-plane MCP prompts such as `team_orchestrator.report_result` and `team_orchestrator.send_message` are handled by the coordinator. It uses session-scoped approval, verifies the prompt cleared, retries boundedly, and logs the result. Do not ask the user to approve those routine internal prompts.
 

package/skills/team-agent/references/team-in-team.md

@@ -9,7 +9,7 @@ This reference is for explicitly requested nested Team Agent teams. Use it only
 - Never edit, delete, or reuse files under the parent team's `.team/current`.
 - Never shut down a team just because `.team/current` is already occupied. In a nested setup, that is probably the parent team; shutting it down can kill the main node.
 - A child team needs its own workspace directory. The role-doc directory is not the child workspace.
-- MCP tools are team-scoped. A worker must not try to widen `team_orchestrator.send_message` to another team.
+- The MCP surface is exactly `send_message`, `report_result`, and optional `get_team_status`; it is team-scoped. A worker must not try to widen `send_message` to another team. Lifecycle management uses the authorized native CLI.
 
 ## Child Workspace
 
@@ -34,29 +34,14 @@ Then create the child team's files inside the child workspace:
 
 ```bash
 cat > .team/current/TEAM.md <<'EOF'
----
-name: child-review-team
-objective: Run a bounded child-team task and report to the parent worker.
-dangerous_auto_approve: false
-fast: false
----
-
-Child team config only.
+A bounded child team for reviewing the requested files and reporting findings.
 EOF
 
 cat > .team/current/agents/reviewer.md <<'EOF'
----
-name: reviewer
-role: Independent Reviewer
-provider: codex
-tools:
-  - fs_read
-  - fs_list
-  - mcp_team
----
+# Independent Reviewer
 
 Review only the files and instructions provided by the child-team leader.
-Report findings with team_orchestrator.report_result exactly once.
+Use `send_message` for progress and report findings with `report_result` exactly once.
 EOF
 
 team-agent quick-start .team/current