imbue-mngr-claude

Claude agent type plugin for mngr.

Provides the claude, code-guardian, and fixme-fairy agent types.

Shared CLAUDE_CONFIG_DIR (local agents)

By default, every Claude agent gets its own per-agent config directory under $MNGR_AGENT_STATE_DIR/plugin/claude/anthropic/ (populated by copying/symlinking from ~/.claude/). Set use_env_config_dir = true on the agent type config to have local Claude agents share the user's $CLAUDE_CONFIG_DIR instead:

[agent_types.claude]
use_env_config_dir = true

When enabled:

• $CLAUDE_CONFIG_DIR must be set in the parent shell; mngr errors out otherwise.
• Only local hosts are supported.
• mngr never writes to the user's Claude config (no trust additions, no dialog dismissal, no per-agent settings.json, no keychain provisioning). The user is responsible for one-time interactive claude setup (trust the work_dir, complete onboarding, log in).
• Other sync/override/auto-dismiss fields on the agent config are silently ignored since shared mode has no per-agent dir to write into.

Approximate response streaming (streaming_snapshot_interval_seconds)

Set streaming_snapshot_interval_seconds on the agent type config to get an approximate, live view of Claude's in-progress assistant text:

[agent_types.claude]
streaming_snapshot_interval_seconds = 0.25

When the value is > 0, a background watcher periodically captures the agent's tmux pane, reverse-maps the rendered assistant text back into markdown, and writes it to $MNGR_AGENT_STATE_DIR/plugin/claude/stream_buffer. When the value is <= 0 (the default), the watcher is neither provisioned nor run.

stream_buffer format:

• Line 1: the id (uuid) of the last complete assistant message (empty string if none yet), so consumers can tell a genuinely-new streaming message apart from leftover text right after a message finished.
• Lines 2+: the in-progress assistant text, reverse-mapped to markdown.

Notes:

• This is best-effort and approximate. It reconstructs bold/italic, inline code, links, blockquotes, lists, code blocks, and tables from the terminal rendering; heading levels and code-block languages are not recoverable.
• The body is strict-append within a message (snapshots are overlap-stitched) and is emptied when the agent goes idle. A table is held back until it stops changing across polls, so it appears once rendered rather than row-by-row.
• The watcher is a Python script; the agent host must have python3 available (provisioning fails fast if it does not).