[harness] Lead session has no live view of subagent cost, and subagents have no default channel back to the lead
Context. Solo-operator project using Claude Code (desktop app, Windows 11) with a lead session on Claude Fable 5.1 orchestrating Opus 5.5 / Sonnet subagents (in-process, worktree-isolated) under a six-seat ceiling. Measured 2026-09-28; numbers are from tool output. No secrets or private paths.
Finding. The lead session cannot read a running subagent's token or tool-call count, while the desktop UI shows both per subagent. The operator caught a STARVED subagent from that UI (a third of its siblings' token and tool-call growth over the same two hours) before the lead did; the lead's own probe (the transcript file's mtime) said "alive" because a hook kept writing. Separately, the default subagent tool set carries no SendMessage, so a mid-run progress request by message is one the subagent cannot answer; its only outbound channels are its final handback and files it writes.
Evidence (measured). Six subagents running; one made 85 tool calls / 1.58 MB of transcript in 2 h while a sibling launched two minutes later made 326 calls / 3.8 MB. Six progress requests by message went unanswered until re-issued as "write a progress file"; all six then wrote it within minutes. The lead built a prototype rate instrument from the task output files (bytes and tool calls per minute per subagent; under 0.4x the median = flag).
Minimal reproduction. Dispatch two subagents; slow one down with a heavy PostToolUse hook on its files; from the lead session try to (a) read either subagent's token or tool-call count, (b) receive a mid-run message from a subagent without adding a tool to its definition by name.
Ask. (1) Expose per-subagent tokens, tool-call count and last-tool-time to the lead session (a query tool, or fields on the task notification stream). (2) A lead-directed message tool in the default subagent tool set, or a harness-native progress field a subagent can set. (3) A per-call grant/deny of SendMessage on the Agent tool.
This issue has 1 comment on GitHub. Read the full discussion on GitHub ↗