Feedback wda model tiering
Justin (2026-07-02, CreatorTrack round 4/5 session): "I feel like we have used more Fable 5 usage than we need, can we have a more flushed-out process that is less context and token intensive."
Why: A full WDA round ran 8-10 background agents, all inheriting Fable 5 (~100-270k tokens each), plus a heavy orchestrator loop. Most worker tasks (build a component from a written spec, mechanical refactors, verification passes) don't need frontier reasoning.
How to apply:
- Agent tool spawns default to model: "sonnet" for build-from-spec workers; "haiku" for mechanical sweeps (renames, config edits, alias tables); reserve Fable 5 (inherit) for merge reconciliation, root-cause debugging, and architecture calls. Matches [[feedback_parallel_subagents_haiku]] + subscription-over-API cost preference.
- Slim briefs: point agents at the ledger file + INTENT conventions instead of restating specs; include a short repo-map (key dirs/files for their seam) so each worker doesn't re-explore the codebase from scratch.
- Worker self-QA = typecheck + one smoke check; full visual QA happens ONCE post-merge on :3060, not per-slot Playwright suites.
- Orchestrator keeps merge conflict resolution surgical (grep hunks, don't read whole files) and checkpoints via the ledger, not long chat recaps.
[Claude Code]