diff --git a/_kos/ideas/subagent-session-display.md b/_kos/ideas/subagent-session-display.md
new file mode 100644
index 0000000..fb6b30b
--- /dev/null
+++ b/_kos/ideas/subagent-session-display.md
@@ -0,0 +1,117 @@
+# Subagent display in the sessions table
+
+- **Status:** idea (pre-hypothesis, no commitment)
+- **Date:** 2026-08-05
+- **Origin:** operator question during the act-2 demo review, same session
+  that shipped the LLM/RUNTIME column shuffle and the statusline context
+  feed (finding-011, PRs #110/#112)
+- **Related:** aae-orc-7hzb (statusline probe, carries the "no daemon
+  surface for per-subagent context" gap), finding-011 (the
+  subagentStatusLine payload shape)
+
+## The question
+
+Some harnesses share live subagent information. Claude Code's
+subagentStatusLine feed delivers, per refresh tick, a tasks array under
+the parent session id: task id, type, status, description, startTime,
+model, contextWindowSize, tokenCount, and a tokenSamples time series.
+`marvel ctx-forward` already parses this and throws it away, because the
+daemon has nowhere to put it.
+
+Should `marvel get sessions` display subagents at all, and if so, how?
+
+## Display sketches (from the operator, verbatim intent)
+
+- A fold indicator on the parent row: an icon signalling "there are more
+  sessions here" that expands to child rows. Collapsed by default so the
+  fleet view stays one row per session.
+- Tree glyphs on expanded children, in the AGENT NAME or RUNTIME column:
+  `↳`, `├──`, `├`, `└─` or similar, marking the row as a subprocess of
+  the row above.
+- Runtime-aware rendering: since marvel knows the parent runtime is
+  claude (or whatever), the child row's RUNTIME cell could do something
+  interesting rather than repeat the parent, e.g. carry the glyph plus
+  the harness name (`└ claude`), or the task type (`local_agent`).
+
+## Column applicability (the gaps problem)
+
+Subagent rows are not Sessions. They have no pane, no lifecycle managed
+by marvel, no restart policy, no generation. A child row would populate:
+
+| Column | Child value |
+|---|---|
+| WORKSPACE / TEAM / ROLE | inherit from parent (or blank to de-emphasize) |
+| GEN | n/a |
+| AGENT NAME | task description or label ("Count files in directory") |
+| STATE | task status (running / completed) |
+| HEALTH | n/a (no healthcheck applies) |
+| CTX% | tokenCount / contextWindowSize |
+| CPU% / RSS | n/a (subagents share the parent process; procstat cannot split them) |
+| DESK | n/a (same pane as parent) |
+| RUNTIME | glyph + inherited harness, or task type |
+| LLM | task model |
+
+Roughly half the columns are inapplicable. Options: render `·` in n/a
+cells (visually quieter than `-`, which already means "not measured"),
+or give expanded children a compressed layout that does not pretend to
+be full rows.
+
+## Sorting (the hard interaction)
+
+Sorting a flat table with attached children breaks naively: children
+must stick to their parent or the tree reads as garbage.
+
+Prior art worth stealing:
+
+- **htop tree view:** a mode toggle. In tree mode, sorting applies among
+  siblings within each subtree; the hierarchy is never broken. Leaving
+  tree mode restores flat global sort.
+- **ps f / pstree:** ASCII art forest, no interactive sorting at all.
+- **k9s:** hierarchical resources get their own views rather than
+  inlined rows.
+
+Candidate rule: children always travel with (immediately below) their
+parent; sort keys order parents globally and children within a parent.
+A sort on a column that is n/a for children (CPU%, DESK) simply leaves
+child order untouched. Expansion state is a watch-mode toggle (say
+`x` or Space on the parent row); one-shot `get sessions` prints either
+collapsed-with-indicator or a `--subagents` flag expands all.
+
+## What has to exist first (the data plane)
+
+Display is the second half. The first half is a daemon surface:
+
+1. Extend the heartbeat RPC (or add a sibling RPC) so ctx-forward can
+   deliver the tasks array it already parses.
+2. Store shape: ephemeral child readings hanging off the Session
+   (`[]SubagentReading` with a TTL or last-seen timestamp), never
+   persisted to bolt (they are gone when the parent's statusline stops
+   reporting them).
+3. Staleness: subagent rows vanish from the feed when tasks complete;
+   the table needs a decay rule (drop after N seconds unseen) so
+   finished subagents do not linger as ghosts.
+
+Only claude exposes this today (finding-011). If codex/opencode never
+grow an equivalent, the fold indicator is claude-only, which argues for
+the indicator being data-driven (rows appear when readings exist)
+rather than runtime-driven.
+
+## Open questions
+
+- Is the sessions table even the right surface, vs `marvel describe
+  session` (already per-session, no sorting problem) or a dedicated
+  `marvel top`-style view?
+- Does a subagent row ever warrant an event (`agent.subagent.started`?)
+  or is that ring noise?
+- tokenSamples is a small time series; is a sparkline in the CTX% cell
+  worth the render complexity, or strictly a describe-level detail?
+- Does the B14 taxonomy have anything to say about naming these rows
+  (they are agents in the B14 sense, running inside a session that is
+  itself one agent's pane)?
+
+## Crystallization signal
+
+If a second harness exposes subagent data, or the first fleet workload
+routinely runs foreground subagents worth watching, extract a frontier
+question and a probe brief (display prototype behind a watch-mode
+toggle, fed by a heartbeat extension).