claude-code-provider

# Claude Code provider Read settings with `bb plugin config provider-claude-code`; change a declared key with `bb plugin config provider-claude-code set <key> <value>`. - `subagentsDisabled` defaults to `false`. Enable with `bb plugin config provider-claude-code set subagentsDisabled true` to remove native Agent, Task, and ListAgents tools from the model context. Changes restart Claude before the next turn, preserving the conversation. A deny hook also blocks native delegation during an active turn after disabling it. - `disable1MContext` defaults to `false`. Enable with `bb plugin config provider-claude-code set disable1MContext true` to set `CLAUDE_CODE_DISABLE_1M_CONTEXT=1` (`0` when off). Changes restart the thread's Claude process before its next turn, preserving context. - `chromeEnabled` defaults to `false`. It starts Claude Code with `--chrome` for Claude in Chrome tools. The host needs the extension and a claude.ai login. A change restarts the thread's Claude process before its next turn, preserving context. - `artifactsEnabled` defaults to `false`. Enable with `bb plugin config provider-claude-code set artifactsEnabled true` to set `CLAUDE_CODE_ARTIFACT=1`, which gives threads Claude Code's `Artifact`, `ArtifactComments`, and `ArtifactData` tools for reading `claude.ai/artifact/` links and publishing pages. Claude Code withholds them from SDK sessions unless opted in. Claude may publish private pages to claude.ai without asking. The claude.ai account's plan and organization policy still decide whether the tools load. When off, bb leaves an inherited `CLAUDE_CODE_ARTIFACT` (shell, environment, or `~/.claude/settings.json` `env`) untouched. A change restarts the thread's Claude process before its next turn, preserving context. - Select the Fast service tier in the model picker or pass `--service-tier fast` to `bb thread spawn` for supported Opus models. Use `bb thread tell --service-tier fast` to change a thread on its next turn; `default` turns it off. Fast mode requires eligible Claude access; subscription accounts need usage credits. It is billed at premium rates. The provider passes the selection as a session-scoped SDK setting, so it does not change the user's Claude defaults. - `sandboxEnabled` defaults to `true`. In Accept Edits and Approve for me modes bb runs Claude Code's Bash commands in its sandbox. Set it to `false` to use Claude Code's own command approvals and sandbox settings instead. A change restarts the thread's Claude process before its next turn, preserving context. - `fullAccessApprovalsEnabled` defaults to `false`, so Full Access approves every request Claude Code sends, and unattended threads never wait. Set it to `true` to show Claude Code's own prompts in Full Access: `permissions.ask` rules, PreToolUse hooks that answer `ask`, and its built-in safety checks (such as `rm` on a possibly-empty variable). It applies from the thread's next turn without a restart. When bb runs as root, Full Access always approves, because Claude Code refuses `bypassPermissions` there. - bb passes only `BB_CLAUDE_CODE_EXECUTABLE` and `CLAUDE_CODE_OAUTH_TOKEN` to the CLI. Mint the token with `claude setup-token` for machines with no interactive login. - Claude Code's own slash commands, including bundled skills such as `/simplify` and `/code-review`, appear in the composer's `/` menu, `bb project commands`, and `bb thread commands`. The provider reads them from Claude Code for the project folder without sending a prompt. With **No project** selected, it uses an empty catalog folder on the selected machine so global commands appear before a workspace exists. Each thread refreshes them when a turn starts. Commands for settings bb manages (such as `/model`, `/config`, `/fast`), terminal-only commands (such as `/doctor`), and bb's own skills are left out. - Structured plan, message editing, and compaction are supported through the corresponding `bb thread` commands. Unlisted model IDs are accepted by the provider; verify actual availability on the target host. Inspect the thread and provider state after a change; do not restart unrelated threads or change settings merely to answer a question. Automatic context breakdowns run in a separate, non-persisted fork of the saved conversation. Reporting never sends prompts or control requests to the live Claude session. A timeout, replacement, or thread stop closes the reporter; the live conversation does not wait for reporting or restart because of it.