herdr and tuios

Run mendcode in a herdr pane, or in a tuios pane, and the multiplexer's sidebar shows what the session is doing, as it does for its built-in agents: nothing to install, no hook, no plugin.

• Working while a turn runs (answering, running a tool, compacting).
• Blocked while a tool call waits for approval, with the request as its message (approve shell: git push). tuios lists it as needs_input in its Inbox.
• Idle when nothing runs. herdr and tuios show a turn that just ended as done until someone looks at the pane.
• The session id goes with each report, with the command that resumes it (mendcode --session <id>), so the multiplexer can bring the session back after a restart.
• Quitting MendCode releases the pane. A MendCode that is killed leaves it to the multiplexer, which clears it once the pane's shell prompt is back.

MendCode reports when it finds herdr's pane variables, HERDR_SOCKET_PATH and HERDR_PANE_ID (herdr sets them in its panes, and tuios sets them in every pane, since it answers herdr's socket API). Each change of state is one pane.report_agent request on that socket, with source mendcode, and the same state again sends nothing. A multiplexer that is gone or slow costs the session nothing: a report gives up after two seconds. MENDCODE_HERDR=0 turns it off.

Inside tuios, tuios get-agent-state and tuios list-agents name the harness mendcode; ask-agent, wait-for and fan can then tell when the pane is busy. The state is the terminal's own session, the one on screen: the side chat (/btw) does not change it. A sub-agent waiting on an approval or a question holds the pane as blocked too, with the sub-agent and what it waits on as the message (Fix the flaky test: Allow shell): herdr has no sub-agent rows, so that is how a waiting child shows there.

In a tuios pane (TUIOS_ENV=1) the rail also shows three facts under MendCode's row, as tuios set-agent-meta would write them with source mendcode: model (the session's model), context (how full the context is, as the prompt's status shows it) and cost (what the session has cost, when it has a price). They are sent on tuios's own socket (TUIOS_SOCKET, for the pane TUIOS_PANE_ID) only when one changes; a fact that goes away is removed, and tuios drops them all when MendCode leaves the pane. tuios get-agent-state --json lists them under meta. MENDCODE_HERDR=0 turns this off too.

tuios also lists the session's sub-agents under MendCode's row, the same ones the Agents strip shows: the daemon's child tasks (task and delegate_task, on any engine) and the sub-agents an engine runs inside the session (Claude Code's Agent calls, in the foreground or the background). Each is sent with report-agent-activity on tuios's socket: it starts with its type (agent, model and engine, as explore · Sonnet 5.5 · claude), says what it was asked and the tool it runs, and ends as it ended, done with the first line of its answer, failed with its error, or stopped. tuios list-agents --json shows them in subagent_list and counts the running ones in subagents. Only changes are sent. Another session opened in the pane, and the first report of a new MendCode process, start with session_start, so tuios forgets the children it had and none is left from a session switch, a crash or a restart; a child that leaves the list while running is stopped, and quitting clears them. The reports carry the harness mendcode and the session's id, so tuios never puts them on another agent. A sub-agent Codex or pi runs inside its own turn is not reported: those engines do not tell MendCode about them.