Nagent

nagent

Spawn a bb thread the way the user launches them. Read this before any bb thread spawn.

Use bb-cli for status, tell, wait, inspect, automations. This skill is spawn-only.

Default worker: Codex gpt-5.6-sol, --reasoning-level high, --service-tier fast ("GPT 5.6 Sol High Fast"); bump to xhigh for larger refactors. Use this unless the user requests otherwise.

Hard rules

• Pass --project explicitly. The CLI does not infer it from the current thread.
• Never guess provider or model IDs. Lookup, then spawn.
• After spawn: do not bb thread open, --split, or focus the new thread unless the user asks. Just launch. Report the thread id. Stop.
• Do not poll, bb thread wait, or read logs unless the user asks.
• Prefer bare bb. If BB_CLI is set, "$BB_CLI" is fine.
• Check your own mode first. A child can never exceed the parent's permission mode. If this thread is not full, --permission-mode full is silently downgraded and the worker will ask for approvals.

Recipe

# 0. Pre-flight: must print "full". If not, STOP and tell the user
#    "this thread is sandboxed (<mode>); relaunch me in full or spawn from the UI".
scripts/permission-mode.sh "$BB_THREAD_ID"

bb project list --json
bb provider list --json
bb provider models <provider-id> --json

bb thread spawn --json \
  --project <project-id> \
  --new-environment worktree \
  --provider <provider-id> \
  --model <model-id> \
  --reasoning-level <level> \
  --permission-mode full \
  --title "Short title" \
  --prompt "..."

Add --service-tier fast only when the user says fast.

Parent / sidebar nest

--parent-self and --parent-thread are optional. Use them when it makes sense — not on every spawn.

--parent-self parents the new thread to the current thread (BB_THREAD_ID). The left sidebar then nests the child under this thread (indented, expandable), like a worker under a manager.

--parent-thread <id> parents it to a specific other thread. Do not combine it with --parent-self.

Omit both for a root thread — its own top-level sidebar row, not nested. This is what the user means by "not as your own subagent" or "standalone thread".

When it makes sense: this thread is coordinating the work, the user asked for a worker under this session, or the job is a clear subtask of this thread. Skip it for unrelated one-off work that should sit as its own top-level thread.

# add when it makes sense — not required
--parent-self

environmentId is often null in the spawn JSON. The worktree is still creating. That is fine.

Lookup

Project. Match name (e.g. DeepAPI) in bb project list --json. Use that id. Confirm sources[].path is the repo the user named. Do not hardcode project IDs.

Provider. From bb provider list --json:

Do not use acp-grok for "Grok via Cursor CLI". That is Grok Build, a different product.

Model + reasoning. From bb provider models <id> --json. Use the catalog id. Match the user's words to displayName and supportedReasoningEfforts[].description. Those two are not the same field.

If the user does not name a reasoning level, use that model's defaultReasoningEffort from the catalog. Do not pick extra high or max on your own. Example: Claude Code Fable 5 defaults to high.

Known mappings (re-fetch if spawn rejects them)

Cursor CLI — Grok 4.6 Extra High Fast

User: "Grok 4.6 extra high fast (via Cursor CLI subscription)"

--provider acp-cursor \
--model cursor-grok-4.6-medium \
--reasoning-level xhigh \
--service-tier fast

Traps:

• The model id stays cursor-grok-4.6-medium even for Extra High. Extra High is a reasoning value, not a different model id.
• On this catalog, high displays as Fast. xhigh displays as Extra High.
• "Extra High Fast" = xhigh plus --service-tier fast. Fast-the-reasoning-label and Fast-the-service-tier are different flags.
• Cursor ACP has no auto permission mode. Use full or accept-edits.
• Do not pass Cursor Task-tool slugs like cursor-grok-4.6-high-fast. Those are not bb model ids.

Codex — GPT 5.6 Sol Max

User: "GPT 5.6 Sol Max, via the Codex subscription"

--provider codex \
--model gpt-5.6-sol \
--reasoning-level max

Traps:

• Use native gpt-5.6-sol, not cursor/gpt-5.6-sol (opencodex routed through Cursor).
• Max is --reasoning-level max, not a model id suffix.
• Codex supports auto permission mode. Still use full when the worker must hit prod, network, or leave the sandbox.
• Pass --service-tier fast only if the user asks for fast on Codex too.

Worktree

--new-environment worktree is the bb-managed worktree. Do not create a git worktree by hand unless the user says so.

• New worktrees get tracked files only.
• Untracked files (.env) copy only if the repo has .worktreeinclude.
• Optional: --base-branch <branch>. Omit to use the project's default.
• Do not combine --machine with an existing --environment id.

For repos that must keep .env in worktrees, the repo needs .worktreeinclude plus optional .bb-env-setup.sh. Details: bb guide environments.

Prompt

The worker starts blind. It does not see this chat. Write the full brief into --prompt.

Include:

1. Objective — one sentence.
2. Constraints — read-only vs implement; no git push; no prod writes.
3. Skills / files to use — name them. Skills do not carry over. Example: prod reads → ~/.claude/skills/read-prod-database/SKILL.md.
4. Deliverable — what to return.
5. Validation — how it knows it is done.
6. Report back — concise report only, or diff + files changed.

Read-only jobs must say no code changes, no git writes, no database writes in the prompt, even if permission mode is full.

Use a heredoc so quotes survive:

--prompt "$(cat <<'EOF'
Objective: ...
Constraints: READ ONLY. ...
Deliverable: ...
Report back: ...
EOF
)"

Always pass --json and --title.

Permission mode

Default for investigation/build workers: full, with the prompt forbidding writes when the job is read-only.

Cursor fallback. bb does not pass --force to cursor-agent; in any non-full mode Cursor's own allowlist decides (~/.cursor/cli-config.json → permissions.allow). "Not in allowlist: git -C" means a subcommand-level entry like Shell(git status). Fix once: use Shell(git) (any git subcommand) plus common read-only tools (ls, find, stat, du, jq, sed, awk). Trap: a project-level .cursor/cli.json overrides the global file and brings prompts back.

After spawn

Verify the child actually got full (a sandboxed parent downgrades it silently):

scripts/permission-mode.sh <child-thread-id>   # must print "full"

If it prints anything else, say so loudly in the report. Do not report success.

Tell the user, then stop:

• thread id
• title
• project name
• provider + model + reasoning (+ fast if set)
• worktree, yes/no
• read-only vs implement

Do not dump the prompt. Do not open the thread. Do not wait for it.

Unblocking the worker

If bb reports the worker is blocked on a command, file change, or permission, unblock it yourself. Do not wait for the user. Quickly read WHAT it wants to run and WHY (bb thread interactions list <id> --json), and if it fits the brief, approve it (bb thread interactions approve <interactionId> <id>). Only escalate to the user when the action is destructive, touches prod, or falls outside the brief.

If the subagents you launched with this skill are BLOCKED, and waiting for some input or multi-choice select, make sure to UNBLOCK them yourself when it's obvious, trivial or safe -- only wait for the user when the next step is truly non-obvious.

Archive (only when the user asks)

Do not auto-archive after spawn.

bb thread archive <id> --json
bb thread stop <id>

Archive first, then stop. Archive also archives child threads. Stop frees the agent runtime; the thread history stays.

bb thread unarchive <id> --json
bb thread list --archived

Do not use bb thread delete unless the user wants it gone forever. Do not use --visibility hidden as a substitute for archive. Hidden is sidebar-only; archive is the lifecycle close-out.

Bulk, one worktree/environment: bb environment archive-threads <environment-id>.