From e942ecc34aedde8d280164bd5f97fda26cb141a8 Mon Sep 17 00:00:00 2001
From: Jeff Smith <jeff@unbiasedgeek.com>
Date: Mon, 6 Apr 2026 21:59:01 -0600
Subject: [PATCH] docs(plan): add Phase 2.5 context budget reliability (#44)

#5 smoke test showed the dir loop exhausts the 126k context budget on
a 13-file Python lib. Sequencing #44 between Phase 2 and Phase 3 so
the foundation is solid before planning + external tools add more
prompt and tool weight.
---
 PLAN.md | 9 +++++++++
 1 file changed, 9 insertions(+)

diff --git a/PLAN.md b/PLAN.md
index a74c83c..295aa90 100644
--- a/PLAN.md
+++ b/PLAN.md
@@ -541,6 +541,15 @@ without 9 phases of rework.
   Until then, the survey prompt carries a Band-Aid warning that the
   histogram is biased toward source code.
 
+### Phase 2.5 — Context budget reliability (#44)
+- Dir loop exhausts the 126k context budget on a 13-file Python lib
+  (observed during #5 smoke test). Must be addressed before Phase 3
+  adds longer prompts and more tools, or every later phase will
+  inherit a broken foundation.
+- Investigate root cause (tool result accumulation, parse_structure
+  output size, redundant reads) before picking a fix.
+- Add token-usage instrumentation so regressions are visible.
+
 ### Phase 3 — Investigation planning
 - Planning pass after survey, before dir loops
 - `submit_plan` tool