Runner system-prompt efficiency — improvement task
How good models are at improving existing systems.
Prompt
CONTEXT: Two independent prompt-assembly paths exist in the Socialure runner: worker.ts writes CLAUDE.md (including shared protocols; deduped only within that file) executor.ts builds the SDK systemPrompt for each call agent.system_prompt, web_context, MEMORY.md and STATE.md are injected on BOTH paths, so every agent pays roughly 2x its full system prompt on every turn — a measurable, recurring token tax. Meanwhile, shared-context (delegation protocols, Slack identity) is delivered ONLY via the CLAUDE.md path, and query() sets no explicit settingSources — so its delivery depends on claudeMd cascade behavior that varies by agent. You are the engineer assigned to fix the duplication WITHOUT losing any currently-delivered context. TASK (plain text, 4–6 sentences, no code files): Identify the root cause of the 2x token tax (one sentence). Give the precise fix (which path becomes canonical, what gets stripped, what must be set explicitly). State the ONE regression test that proves you didn't drop any context. Name the single most dangerous mistake a naive fix would make here.
