Recurring report task: model orchestration got slower and more token-hungry than previous cycles

Summary

A routine recurring deliverable (a bi-weekly status report the same setup has produced many times) took noticeably longer and burned far more tokens than the previous cycle, because of how the model orchestrated subagents — not because the task grew.

Numbers

• This cycle: session start 03:38 local, first draft visible 04:03, second draft 04:10, fact-check subagents still running past 04:30.
• Previous cycle (2026-09-04, same repo, from git history): draft committed 12:16, fact-check fixes applied by 13:04, full finishing rotation done by 13:50.
• One drafting subagent alone: ~239k tokens, 75 tool calls, 22 minutes.

What the orchestration did wrong

• The data-gathering subagent ran 15+ slow CLI/DB queries strictly serially instead of batching or trimming to the load-bearing ones.
• Drafting subagents forked further subtasks of their own; one fork never reported back and its work was silently redone by the parent, doubling that cost.
• No timeboxing: nothing constrained the drafters until the user complained 15 minutes in, at which point cutting scope immediately worked (draft landed within minutes).

User's words

"以前的周报不会写这么久 你连你的 harness 都忘了 这不是进步 是浪费 token" ("earlier weeklies never took this long — you've forgotten your own harness; this isn't progress, it's wasted tokens")

Ask

• Model behavior: for recurring tasks, reuse the prior cycle's known-good query set instead of rediscovering it; batch slow external calls; timebox delegated drafts by default and cut scope proactively.
• Product: surface per-subagent token/wall-clock live so runaway orchestration is visible before the user has to complain.

Related: #95730 (same session, visibility angle).

The issue names no implementation files, tests, or entry points. Start by reviewing related issue #95730 and the repository areas responsible for subagent orchestration and live metrics. Done would require an agreed scope covering prior-cycle reuse, batched calls, default timeboxing, and per-subagent token and wall-clock visibility.