The idea: the cook-off's cost gap was context per call, not number of calls (29.5 vs 28.7 calls per row). Two causes:
- The lead ran with autocompact accidentally set to 1M.
- Sub-agents load the whole always-on set plus the full build plan and BRD (~145–215k), then never compact. 34 of 38 sub-agents peaked at 250–580k.
Proposals:
- Role-specific stubs that name 2–4 skills.
- Give each agent its per-row spec slice, not the whole spec.
- A 250k autocompact default, or a preflight warning, written by
init-project.sh.
- A cap on how many rows are bundled per agent.
- Per-role mean context per call recorded in the time log.
Field evidence: contrib/inbox/2026-09-29-subagent-context-cost.md (#165). Re-estimated at 250k autocompact, the build cost drops from $264 to ~$205.
Done = on the next comparable run, calls per row are unchanged, mean context per call is about halved, and cost per row is at or below the MCP arm's.
Rough size: a cluster (agent-roles.md, agents/ stubs, init-project.sh).
The idea: the cook-off's cost gap was context per call, not number of calls (29.5 vs 28.7 calls per row). Two causes:
Proposals:
init-project.sh.Field evidence:
contrib/inbox/2026-09-29-subagent-context-cost.md(#165). Re-estimated at 250k autocompact, the build cost drops from $264 to ~$205.Done = on the next comparable run, calls per row are unchanged, mean context per call is about halved, and cost per row is at or below the MCP arm's.
Rough size: a cluster (
agent-roles.md,agents/stubs,init-project.sh).