@miller-tech/uap 1.172.9 → 1.172.10

package/config/llama-profiles/gemma4-26b-a4b-mtp.env

@@ -142,4 +142,11 @@ LLAMA_CHAT_TEMPLATE_FILE=embedded
 # on this binary and makes the server exit 1 at startup. Do not copy EXTRA_ARGS
 # across from the Qwen profile (which runs on mainline) without re-checking
 # each flag against `llama-server --help` for THIS binary.
-LLAMA_EXTRA_ARGS=--metrics --jinja --mlock --prio 2 --temp 1.0 --top-k 64 --top-p 0.95 --min-p 0.0 --ctx-checkpoints 32 --checkpoint-min-step 2048 --cache-ram 8192
+# --cache-ram: 8192 was the llama.cpp DEFAULT, i.e. never actually chosen. The
+# host-RAM prompt cache holds KV prefixes so a follow-up turn re-uses its prompt
+# instead of re-prefilling it; undersize it and every eviction turns the next
+# turn on that prefix into a full reprocess. Measured on this box under the Qwen
+# profile at 16384: 87% of all prefill time went on re-doing evicted work. This
+# model is smaller, but the working set is the SESSION's, not the model's, so
+# the same 32 GiB applies. See the Qwen profile for the measurement method.
+LLAMA_EXTRA_ARGS=--metrics --jinja --mlock --prio 2 --temp 1.0 --top-k 64 --top-p 0.95 --min-p 0.0 --ctx-checkpoints 32 --checkpoint-min-step 2048 --cache-ram 32768

package/config/llama-profiles/qwen36-35b-a3b.env

@@ -61,9 +61,28 @@ LLAMA_N_PREDICT=16384
 # are built around exactly this template. Do NOT switch it to "embedded".
 LLAMA_CHAT_TEMPLATE_FILE=/home/cogtek/dev/miller-tech/universal-agent-protocol/tools/agents/config/qwen3.5-enhanced.jinja
 
+# --cache-ram sizing — measured, not guessed
+# The host-RAM prompt cache holds KV prefixes so a follow-up turn re-uses its
+# prompt instead of re-prefilling it. Undersize it and it thrashes: every
+# eviction turns the next turn on that prefix into a full reprocess.
+#
+# Measured on this box at 16384 (2026-07-29), one hour of a live agent session:
+#   27 evictions, ~23.5 GB churned  (avg entry 869 MiB)
+#   30 large reprocesses >5k tokens, 829,587 tokens re-prefilled
+#   418s of the 482s total prefill time — 87% — spent re-doing work
+#   worst single stall 44.4s
+# The eviction count and the large-reprocess count track ~1:1, which is the
+# causal link: the cache was holding a working set 1.5x its size.
+#
+# 32768 (32 GiB) clears that ~23.5 GB working set with margin. Host has 121 GB
+# with ~50 GB available and 0 swap in use, so this leaves ~18 GB free — sizing
+# to the measured working set rather than to all free RAM, because this cache
+# competes with the page cache and with whatever else the operator runs.
+# Re-measure with the eviction grep above before raising it further.
+#
 # FLAG DRIFT: upstream renamed --checkpoint-every-n-tokens (old b8780 spelling,
 # which this profile used until 2026-07-28) to --checkpoint-min-step. An unknown
 # flag makes llama-server exit 1, and systemd then burns its restart budget
 # looping on it. Re-check every flag here against `llama-server --help` whenever
 # LLAMA_BIN moves.
-LLAMA_EXTRA_ARGS=--metrics --jinja --mlock --prio 2 --reasoning auto --reasoning-format deepseek --reasoning-budget 8000 --ctx-checkpoints 64 --checkpoint-min-step 2048 --cache-ram 16384
+LLAMA_EXTRA_ARGS=--metrics --jinja --mlock --prio 2 --reasoning auto --reasoning-format deepseek --reasoning-budget 8000 --ctx-checkpoints 64 --checkpoint-min-step 2048 --cache-ram 32768

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@miller-tech/uap",
-  "version": "1.172.9",
+  "version": "1.172.10",
   "description": "Autonomous AI agent memory system with CLAUDE.md protocol enforcement",
   "type": "module",
   "main": "dist/index.js",