imbue-mngr-robinhood 0.1.4

Drop-in replacement for `claude -p` that runs through mngr create/message/transcript

imbue-mngr-robinhood

robinhood is a drop-in replacement for claude -p that's implemented on top of mngr, i.e., you can use your Max / Pro subscription

Install

# install mngr (if not already installed)
curl -fsSL https://raw.githubusercontent.com/imbue-ai/mngr/main/scripts/install.sh | bash

# either select the robinhood plugin while installing above, or run:
uv tool install imbue-mngr-robinhood

Usage

# Single prompt, text output
mngr robinhood "summarize this repo"

# Pipe stdin in
cat error.log | mngr robinhood "explain this"

# Structured JSON output (claude-native shape; cost/usage fields zeroed)
mngr robinhood "summarize this repo" --output-format json

# Live event stream
mngr robinhood "explain recursion" --output-format stream-json --verbose

# Multi-turn via stream-json input
printf '%s\n%s\n' \
  '{"type":"user","message":{"role":"user","content":"hi"}}' \
  '{"type":"user","message":{"role":"user","content":"and again"}}' \
  | mngr robinhood --input-format stream-json --output-format stream-json

# Live (approximate) streaming of the response as it is produced
mngr robinhood --output-format stream-json --include-partial-messages "tell me a long story"
mngr robinhood --stream-plain-text "tell me a long story"

The mngr robinhood command takes the same arguments as the regular claude CLI, always behaves as if -p/--print was passed, and routes the prompt through a fresh, ephemeral mngr claude agent. The agent runs in-place in the current directory, processes the prompt (or stream of prompts), and is destroyed when the command exits.

Streaming the response

mngr robinhood can surface an approximate live view of the response, sourced from the agent's stream_buffer (the tmux-based response stream; see the imbue-mngr-claude README). Two opt-in flags enable it:

• --include-partial-messages (requires --output-format stream-json): emits claude-native stream_event / text_delta events as the response is produced, followed by the authoritative assistant message from the transcript.
• --stream-plain-text (text output, the default): streams the response text to stdout incrementally and suppresses the trailing full-text dump to avoid duplication.

When either flag is set, robinhood enables the streaming watcher on the spawned agent and defaults the model to sonnet (so fast mode is off and streaming is observable); a user-passed --model still takes precedence. The streamed text is best-effort: the result envelope (and, in stream-json, the final assistant message) remain the source of truth.

tmux window size

The streamed response is reverse-mapped from the spawned agent's rendered tmux pane, so the pane width determines where lines hard-wrap in the output. mngr robinhood therefore creates its agent in a large, pinned window by default (2048 columns x 256 rows, resize policy manual) so the streamed text is not chopped at a narrow width. Override with:

• --tmux-width <columns> (default 2048)
• --tmux-height <rows> (default 256)
• --tmux-window-size manual|latest|largest|smallest (default manual; manual pins the window to the given size and never resizes it)

These flags are consumed by the wrapper and not forwarded to the spawned claude; an invalid value exits with code 2.

Flags not supported in v1

The following claude flags are explicitly rejected (exit code 2):

• --fallback-model
• --max-budget-usd
• --no-session-persistence
• --include-hook-events
• -c / --continue
• -r / --resume
• --session-id

Every other claude flag is forwarded verbatim to the spawned agent.

Alternatives

1. robinhood is Claude-specific. For a better way of doing the same thing that is generic across agents, just use mngr directly: mngr create --message "some prompt" will create an agent and send a prompt to it. You can use mngr transcript to see the outputs, or mngr wait to wait for it to finish.
2. Use claude with an API key

mngr-backed Agent SDK (experimental)

This package also exposes a drop-in re-implementation of the Claude Agent SDK Python surface that is backed by mngr instead of a directly-spawned claude subprocess. Import query / ClaudeAgentOptions / ClaudeSDKClient (and the session functions) from imbue.mngr_robinhood.agent_sdk instead of from claude_agent_sdk; every type is re-exported verbatim, so isinstance checks and field shapes are identical. Each session is a robinhood-prefixed mngr claude agent, driven through the in-process mngr API and read back from its native transcript.

Supported control surfaces and how they map onto mngr:

• can_use_tool + hooks — served by a local HTTP bridge: the agent is launched with a --settings file whose hook commands POST each event to the bridge, which runs the in-process Python callback and returns claude's hook JSON (allow / deny / updated_input); denials surface in ResultMessage.permission_denials.
• interrupt() — stops the agent mid-turn; the response stream ends at a ResultMessage and the next query() restarts-with-resume.
• set_model / set_permission_mode — rewrite the agent's stored launch command with the new configuration and restart it on the resumed session.
• get_server_info() — runs a one-shot claude stream-json probe for the real commands / output style, cached per session.
• total_cost_usd — computed from per-turn token usage times a per-model price table (approximate).
• include_partial_messages -> StreamEvent — when set, the agent's tmux pane is watched and the reconstructed assistant text is wrapped as the claude-native partial-event sequence (message_start -> content_block_delta(text_delta)* -> message_stop). Approximate (text is reconstructed from the rendered pane, not claude's token-level deltas) and best-effort; the authoritative AssistantMessage still arrives from the transcript, and usage/total_cost_usd remain on the final ResultMessage.

Limitations

Documented limitations (real-SDK-only):

• fork_session raises AgentSdkNotImplementedError. claude's --fork-session does not assign a new session id when driven interactively over an adopted, resumed session (the forked turn is written under the source id), so a faithful fork cannot be produced on this transport.
• The agent is not hermetic from the host's claude config -- it must load real settings to authenticate -- whereas the real SDK with setting_sources=[] is hermetic.
• Lots more, see: here

Why so many limitations?

Honestly the Claude Agent SDK is kind of a mess, and I don't have a particular reason to make it work perfectly. There will always be some trade-offs--you'll have to decide how, exactly, to make them for your own use case. Consider this implementation more of an example of how you could do it, or a starting point, rather than a finished product.

Instead of using the proprietary Claude Agent SDK interface, you should try using mngr instead-- its API is considerably cleaner, and it abstracts over a variety of different coding agents already.

Running the live SDK tests

This project contains an opt-in, live integration suite (imbue/mngr_robinhood/test_sdk_*.py) that verifies the documented query() and ClaudeSDKClient interfaces of the Claude Agent SDK end-to-end against the real API. The tests are clean-room: they import only documented public names from claude_agent_sdk and assert the documented behavior.

The suite is parametrized by the sdk fixture over two targets: real_sdk (the real claude_agent_sdk package, run via claude --print) and mngr_sdk (this module, which drives an interactive claude agent in tmux). Every test runs against both unless it exercises a surface the mngr transport cannot provide (currently only fork_session), which is gated to real_sdk via the requires_native_sdk fixture. The mngr target therefore needs a local claude CLI and tmux on PATH. Use -k mngr_sdk / -k real_sdk to run a single target.

These tests make real, paid API calls, so they are excluded from every CI run (via the sdk_live marker, which is filtered out in offload-modal.toml) and only run when explicitly opted in. Each test runs the agent in an isolated temp directory.

To run them, export a first-party ANTHROPIC_API_KEY and set RUN_SDK_LIVE_TESTS=1:

set -a; source .env; set +a   # exports ANTHROPIC_API_KEY from a local .env
just test-sdk-live

If RUN_SDK_LIVE_TESTS=1 and ANTHROPIC_API_KEY are not both set, the suite is skipped.