Feedback overnight workflow model tier scope fence

name: Overnight multi-agent runs need Fable/Opus tiering and explicit scope fences vs parallel sessions description: Justin corrected an overnight fanout run: use Fable for cheap bulk work, Opus only for verify/judge/gate, and explicitly fence off zones already owned by his other concurrent chats. type: feedback

When launching unattended multi-agent workflows (fanout/Workflow tool) for CreatorTrack or similar, two corrections apply:

Model tiering: bulk finder/builder/researcher/designer agents should run on Fable (cheap), not Sonnet. Opus is reserved for adversarial verification, judge/synthesis, and gate steps (typecheck+lint+commit). Sonnet is acceptable only for easy read-only mapping tasks.
Scope fencing: before launching an overnight run, check whether Justin has other Claude Code chats open on the same codebase and explicitly exclude their zones from the workflow's scope (e.g. one chat owned content-ideas/Link-in-Bio/analytics, another owned CreatorTrack perf, so the audit run was fenced to exclude links, ideas/planner code, analytics, and performance changes).

Why: Justin runs multiple concurrent Claude Code sessions on the same repo (CreatorTrack) working different feature areas; an overnight fanout that doesn't tier models burns budget faster (session limit hit in ~75 min for two overlapping Sonnet-heavy runs) and risks duplicate/conflicting work across sessions.

How to apply: before launching any overnight/token-max fanout or Workflow run, (a) ask or infer what other sessions are active and what they own, and write explicit exclusions into the workflow prompt; (b) default finder/builder/researcher agents to Fable, verify/gate/judge agents to Opus, reserving Sonnet for simple read-only mapping only.

[auto-memory session 8eb446b3-1279-4576-82eb-374262ab5375, confidence 0.85, mode direct]