Skip to content

Dept-wide Claude Code / agent-usage lessons (not project-specific — see Projects/*/LESSONS.md for those).

Fork subagents: force synchronous execution for heavy build work

Section titled “Fork subagents: force synchronous execution for heavy build work”

Fork task w/ heavy build work: sometimes returns instant, 0 tool calls, echoes launch status back as if it were the finished report — not real work done.

Fix that worked every time after: open the fork prompt with “Execute synchronously with your own tools; do not background, do not delegate — you ARE the fork.”

Why: a fork can misread an ambiguous multi-step build prompt as an instruction to background the work itself (mirroring how the parent launches forks), rather than doing it in the current call. There’s no further backgrounding available once you’re inside a fork — it must execute directly.

Discovered: sdc-levpro-sdmath-complete, 2026-08-28 — happened on the M5 audit-artifacts fork, fixed by resuming it with the explicit instruction; every subsequent fork in the same task got the instruction up front and none repeated the failure.

When building reusable HTML/interactive tools, register in a dept-specific DASHBOARD (not scattered utility docs). Index by category (emoji headers), include full paths (WSL + Windows), follow KB-OS conventions (markdown tables, registration guide). Discovered: custom developments like precious-metals and large-cap-equity dashboards were housed in skill references + home folder; undiscoverable until explicit grep. Pattern prevents loss.

Discovered: historical-analysis-artifact, 2026-09-24 — located two investment dashboards only by filesystem search, not vault reference; created Core/IT/Utils/Custom/DASHBOARD.md to register and prevent recurrence.