Too much token consumed by sub agents

[manisharanixyz-ctrl]

i saw my sub agent consume around 40M tokens i did not even gave him long task.
expected tokens where around 2M.
model used qwen 3.8 flash.
even it did not did much task

[shubham-kr2002]

There can be something like subagent inside subagent or reading a long file of test case i encountered it few day ago.

[PerryLink]

Your 40M-vs-2M gap can't be explained by the GUI counter alone — subagent sessions are separate sessions with their own logs, so their tokens don't roll into the parent's in-view total. Facts from master (c291e79), then a checklist:

• spawn vs fork: a spawned child starts with NO parent context (packages/subagent/subagent-spawn-in-process/src/index.ts:49-59), while a forked child is seeded with the parent's completed-turn prefix (subagent-fork-in-process/src/index.ts:71-81). If your task used fork (or an out-of-process provider that inherits context), a long parent conversation is the #1 multiplier.
• Each child is its own session (subagent-in-process-driver/src/index.ts:113 mints a fresh id), and retries multiply too: the deepseek provider documents retry omission as normal mode with five retries (packages/llm/llm-deepseek/src/index.ts:171-172), capped per step (llm-retry/src/index.ts:223).
• Tool results are pruned to 8192 chars only in the standard preset (packages/preset/agent-presets/presets/standard/agent.cordis.yml:151-156); if your deployment doesn't mount that pruner, big tool outputs re-enter every request.
• Recursion: maxDepth defaults to 3 (tool-subagent README), and every goal round appends a fresh followup prompt into history (packages/goal/goal-round-driver/src/prompt.ts:12-18), so goal loops grow linearly per round.

Checklist to find the 40M: 1) was it fork or spawn? 2) look at the child session's log (not the parent's) and find the largest request step — its usage fields are what the meter anchors on (packages/llm/token-meter/src/index.ts:145-186); 3) check for large file/image content or unpruned tool outputs; 4) check the retry count in that step; 5) check whether the agent kept re-reading the same big workspace files.

If you share the child session log, I can point at the exact step. (unverified: qwen adapter behavior — that model isn't in this repo's catalog.)