diff --git a/plans/phase6-plan.md b/plans/phase6-plan.md
new file mode 100644
index 0000000..0a3b934
--- /dev/null
+++ b/plans/phase6-plan.md
@@ -0,0 +1,11 @@
+# Phase 6 plan
+
+Start from Phase 5 21c1273. Fix cold first-visit delivery only; preserve playback, banner, migration, exact hashes, contracts/pinning, offline and both rollback flags.
+
+1. Extend perf/baseline.mjs/proxy.mjs with page Resource Timing and complete request waterfalls, and a source-worktree option. Create detached historical worktrees beneath perf/.tmp for Phase 2, 3 and 5 tips. Run isolated WebKit/LTE cold measurements, >=3 runs each. Bisect Phase 3 commits with the same instrumented harness. Test head-loader/CSS discovery, fonts, worker start relative to FCP and immutable HTTP-cache behavior with controlled scratch variants. Record actual evidence in plans/phase6-investigation.md.
+2. Add failing tests before each root fix. Likely files: frontend/index.html, lazy.js or a tiny layout bootstrap; frontend/sw.js and asset-sync-core.js only as evidence warrants. Keep the legacy install path and update all-or-nothing behavior intact; cache completed immutable responses safely without any unverifiable body. Preserve selected layout before paint and existing CSS cascade order.
+3. Commit each fix separately with frontend and isolated Bun server suites green. Revert individual fixes for rollback; ASSET_HASHING=0/ASSET_SYNC=0 stay supported. Do not deploy or contact production.
+4. Run final five-run WebKit cold LTE and lossy medians: FCP <= measured Phase 2 (targets ~1,300/3,800 ms); app.js positive-body responses exactly one; cold wire clearly below baseline. Measure Chromium and warm/offline/core-JS/CSS/feature updates against Phase 5, with repeated medians where regression risk exists. Rerun migration both engines with and without stalls, hidden timers, lazy/seam offline checks, rollback paths and all tests.
+5. Self-review master sections 3.2/3.3/6/6b and full diff. Commit <=350-word plans/phase6-report.md with actual before/after table and explicit unmet targets; overwrite DONE06 signal and end.
+
+Risks: CSS cascade/no-flash preservation, duplicate requests crossing browser/worker HTTP-cache partitions, race between page handoff and install, timeout fallbacks delaying cache readiness, CPU contention invalidating timing. Historical worktrees/scratch data stay inside wt-assets; run benchmarks sequentially and preserve the invalid historical media fixture only for comparable home-launch metrics, not playback claims.