feat(cli): add step-into mode for subagents#1372
feat(cli): add step-into mode for subagents#1372Juan Wisznia (juanwisz) wants to merge 12 commits intolangchain-ai:mainfrom
Conversation
Adds step-into subagent data model: ConversationContext dataclass for tracking nested conversations, and extends SessionState with context stack management (push/pop/reset).
Shows a 4th "Step into" option when approving task tool invocations. Adds 's' quick key and dynamic option count.
Adds ExecuteTaskResult return type, _create_branch_context helper, and step_into decision handling in execute_task_textual.
Integrates step-into subagents into the Textual app: - Adds slash commands for context navigation - Handles step-into result in _run_agent_task - Updates /clear to reset context stack - Adds key bindings and autocomplete entries
Adds STEP_INTO.md with full design rationale and usage guide. Updates README to link to the new doc.
Juan Wisznia (juanwisz)
requested a review
from Mason Daugherty (mdrxy)
as a code owner
github-actions
bot
added
external
Remove unused imports (field, Path), fix line lengths, add missing docstring return, suppress magic number warning, and preserve 8-char hex thread IDs in TextualSessionState.
|
cool, thank you Juan Wisznia (@juanwisz) for re-opening on top of the recent changes. would you mind attaching a screen recording of the functionality to the PR description? you should be able to drag and drop a video file and it will upload inline |
step-into-demo.mp4 |
|
Interesting, thank you Juan Wisznia (@juanwisz) for the recording 2 things that stand out immediately
|
|
Mason Daugherty (@mdrxy) I will work on the changes for this. Regarding point 1, you’re right. As for point 2, if you watch the video closely (at 0:08), you’ll see that we indicate when we step into a specific subagent. However, we could make this much clearer by always displaying the current node and the depth level. I think this feature is very strong. Today, it really comes down to two things: good context management and avoiding unnecessary interface changes, even something as small as opening another CLI. This feature lets you micro-manage context very efficiently by only injecting what is actually necessary, and it allows you to do everything from the same terminal. |
# Conflicts: # README.md # libs/cli/deepagents_cli/app.py # libs/cli/deepagents_cli/textual_adapter.py # libs/cli/deepagents_cli/widgets/autocomplete.py
|
scratchpad
|
2 participants
Adds interactive step-into mode for subagent invocations in the CLI. Instead of fire-and-forget delegation, users can enter an interactive session with a subagent while maintaining context isolation.
tasktool call, a new "Step into" option lets users enter the subagent conversation interactively/return(exit subagent, send summary to parent),/summary(view/edit summary file),/context(show conversation stack)This is a re-port of #829 against the current Textual-based CLI architecture (the original PR targeted the old Rich+prompt_toolkit code in
libs/deepagents-cli/which has since been replaced bylibs/cli/).