All plugins

study-tree

Study with the agent without losing the thread: a side question you ask becomes its own branched session, the side threads show as a tree in a Desktop pane, and each one hands one recall-checked takeaway back to the thread it came from. The agent can also draw interactive widgets and one-question quizzes inline, with follow-up buttons. Disclosure: stores its tree in SQLite and agent-written HTML widgets under the active profile's plugin-data/study-tree/ (a widget over 200 KB or with any network reference is refused; Desktop shows it in its sandboxed preview frame); opens Hermes' state.db read-only to follow compression lineage; adds a short section to every session's system prompt, and through pre_llm_call injects a side-thread frame on each side-thread turn and each takeaway once into the parent's next turn; through transform_llm_output rewrites the final reply (study cards moved to the end as their own paragraphs, cards with unknown ids dropped, widget directives put on their own paragraphs, a widget button's prompt quoted at the top); creates a branched session (session.branch or session.branch_stored), renames it with session.title and sends the side question as its first prompt only when the user clicks Open; the Desktop half keeps the last text selected in the window in memory for 60 seconds to prefill the side-question dialog, and writes the glossary to the clipboard only when the user clicks Export glossary. No network calls, no shell commands, no background processes, no stored credentials, no telemetry.

IpastorSanDesktopv0.1.1Updated 3 hours ago

About

Study with the agent without losing the thread: a side question you ask becomes its own branched session, the side threads show as a tree in a Desktop pane, and each one hands one recall-checked takeaway back to the thread it came from. The agent can also draw interactive widgets and one-question quizzes inline, with follow-up buttons. Disclosure: stores its tree in SQLite and agent-written HTML widgets under the active profile's plugin-data/study-tree/ (a widget over 200 KB or with any network reference is refused; Desktop shows it in its sandboxed preview frame); opens Hermes' state.db read-only to follow compression lineage; adds a short section to every session's system prompt, and through pre_llm_call injects a side-thread frame on each side-thread turn and each takeaway once into the parent's next turn; through transform_llm_output rewrites the final reply (study cards moved to the end as their own paragraphs, cards with unknown ids dropped, widget directives put on their own paragraphs, a widget button's prompt quoted at the top); creates a branched session (session.branch or session.branch_stored), renames it with session.title and sends the side question as its first prompt only when the user clicks Open; the Desktop half keeps the last text selected in the window in memory for 60 seconds to prefill the side-question dialog, and writes the glossary to the clipboard only when the user clicks Export glossary. No network calls, no shell commands, no background processes, no stored credentials, no telemetry.

Study Tree

You are halfway through an explanation of normed spaces and want to know what
Cauchy–Schwarz is. Ask it, and the agent finishes the normed spaces first,
then parks your question as its own branched session. You open it when you're
ready, learn it there, and one checked sentence comes back to the main thread.
The side threads, and theirs, show up as a tree beside the chat.

Study Tree is a Hermes plugin with an agent half (four tools, two hooks, a
skill) and a Hermes Desktop half (a pane, a card in the transcript, a
status-bar chip).

Study Tree

The main thread grows; a side question branches off, and only the takeaway flows back

The chat is the trunk. A side question grows its own branch, and only what you
learned there flows back.

Install

hermes plugins install study-tree --enable

Then, in Hermes Desktop, open Capabilities → Plugins and turn on Study
Tree. Its Desktop half is off by default. The status-bar chip turns green
when the Desktop half reaches the plugin's routes; if it says the backend is
not answering, restart the gateway once.

Requires Hermes Agent 0.22.7 or newer.

Use it

Ask a side question

Ask something off the current topic in your own words ("wait, what's
Cauchy–Schwarz? don't derail"). The agent answers the main topic in full, calls
study_branch and ends the reply with a card for the side question. It
branches at most once per reply and never on its own initiative: a concept its
explanation touches is not a side question unless you ask about it.

The card offers Open, Later (hide the card; the thread stays suggested
in the pane) and Drop.

A side-thread card in the main chat

Open it

Open branches the chat into a new session nested under it in the sidebar,
names it after the topic (↳ Cauchy–Schwarz inequality), opens it as a tab
beside the main chat and sends your side question as its first message. The
branch carries the parent's visible user and assistant text, without tool
output, so the side thread knows where you were.

Inside the side thread the agent stays on that topic. A tangent of the tangent
branches again, one level deeper.

Say "got it"

When you say you have it, the agent does not close the thread on that alone.
It asks one short recall question. A right answer closes the thread with your
answer as the takeaway, and the pane keeps your words under it. A wrong one
gets a one-line correction, and the corrected takeaway goes back instead. Say
"just close it" and it closes at once.

The recall question in a side thread

The takeaway reaches the main thread once, on its next turn. The agent brings
it in with a sentence in its own words ("From the Cauchy–Schwarz side thread:
…") and carries on with what you asked. The card in the main chat turns into a
resolved card that shows the takeaway.

See it, play with it

Ask the agent to show, plot or visualise an idea and it draws an interactive
widget inline with study_visual: an ℓᵖ unit ball with a slider for p, a
sequence converging. Under the widget sit up to three buttons with the next
questions it invites. A button sends its question to the agent, and the reply
quotes it at the top (↪ Why is ℓ² the only one from an inner product?),
since a widget's prompt shows no message bubble of its own.

An ℓᵖ unit-ball widget with follow-up buttons

Ask for a quiz and you get one multiple-choice question of the agent's own on
the current topic. Pick an option and the reply opens with Your answer: …,
then says whether it is right and why, in two lines.

The pane

The study tree pane on the right shows every thread of the current tree,
with its status: suggested, open, or resolved with its takeaway. Click a
thread to open it. A thread with a widget can show it in the pane. The pane
refreshes when the agent branches, resolves or draws.

The Study Tree pane with a multi-level tree and a takeaway

Side question… in the pane parks a side question, or opens it at once,
without asking the agent. The same dialog opens from the palette command
Study Tree: Side question…, prefilled with text you selected in the last
minute, and from Ask as a side question when you right-click selected text in
the composer. Ctrl+Shift+Y (Cmd+Shift+Y on macOS) shows or hides the pane.

A chat that Hermes compacts continues under a new session id. The plugin
follows Hermes' compression lineage, so a compacted thread stays in its tree.

Export the glossary

Once a thread is resolved, Export glossary in the pane copies the tree as
Markdown: one line per resolved thread with its takeaway, and your own answer
to the recall question under it when you got it right. A side thread's own side
threads go under a heading named after it. Suggested, open and dropped threads
are left out. Paste it into a note, in Obsidian or anywhere else.

Limits

  • The cards, the pane and the widgets live in Hermes Desktop. On other surfaces
    (CLI, TUI, Telegram and other messaging platforms) the tools still work, but a
    card shows as a plain ::study-tree-node{id="…"} line and a widget as a
    plain ::preview{file="…"} line.
  • Widgets in the pane are view-only. Their buttons work in the chat, where
    Desktop wires them to the agent.
  • While a reply streams, a card shows as its raw directive line. Desktop draws
    the card when the turn ends.

What it does to your Hermes

  • Storage. The tree lives in a SQLite file at
    $HERMES_HOME/plugin-data/study-tree/data.db. Widgets are written to
    $HERMES_HOME/plugin-data/study-tree/visuals/<node id>/. Nothing goes in the
    install directory.
  • Widgets are agent-written HTML. study_visual refuses a widget over
    200 KB or one that references the network (<script src>, external URLs in
    src, href or CSS url(), @import, fetch, XMLHttpRequest,
    WebSocket, EventSource). Desktop renders it in its sandboxed preview
    frame, with scripts allowed and no network.
  • Reads Hermes' session database, read-only. To follow compression
    lineage it opens Hermes' state.db read-only through SessionDB. It reads no
    other files outside its own data directory.
  • Adds to the model's context. A short Study Tree section goes into every
    session's system prompt. Through pre_llm_call it adds a frame on each turn of
    an open side thread (its topic, what it branched from, the recall-check rule)
    and each takeaway once, on the parent's next turn.
  • Rewrites the final reply. Through transform_llm_output it moves each
    card to the end of the reply as its own paragraph, drops cards whose id is not
    in this session's tree, puts each widget directive on its own paragraph, and
    quotes a widget button's question or quiz answer at the top. Earlier turns are
    never touched.
  • Creates sessions only when you click. Open (on a card, in the pane or
    in the side-question dialog) creates a branched session (session.branch or session.branch_stored),
    renames it with session.title, and sends the side question as its first
    prompt. Nothing else creates, renames or sends.
  • Remembers your last selection, in memory. The Desktop half keeps the
    last text you selected in the window for 60 seconds, to prefill the
    side-question dialog. It is never written to disk, and it leaves the window
    only if you send it as a side question.
  • No network calls, no shell commands, no background processes, no stored
    credentials, no telemetry.     The Desktop half talks only to the plugin's own
    routes on your gateway (/api/plugins/study-tree/).

Development

git clone <repo-url> study-tree
cd study-tree
scripts/install.sh                 # links the checkout into $HERMES_HOME/plugins/study-tree
hermes plugins enable study-tree   # lets Hermes import the agent half
scripts/check.sh                   # ruff, pytest, node tests, hermes plugins doctor and validate

scripts/install.sh creates one symlink and refuses to replace anything that
is not a symlink. After it, restart the gateway and turn on Study Tree in
Desktop under Capabilities → Plugins.

| Path | What it is |
|---|---|
| plugin.yaml, __init__.py | Agent half: manifest and register(ctx) |
| study_tree/ | Tree store, tools, hooks, widget checks, compression lineage |
| skills/study-tree/ | The skill the agent loads, with a unit-ball and a quiz template |
| dashboard/ | Routes under /api/plugins/study-tree/ for the Desktop half |
| desktop/plugin.js | Desktop half: pane, card, chip, dialog (plain ESM, no build step) |

License

MIT

README from the reviewed commit aae3bf5; it updates when the author re-pins.