From 7f6be8a56a3e0a958234937adfc2c4822856660a Mon Sep 17 00:00:00 2001
From: Vuong Hoang <vh@phasefinal.com>
Date: Tue, 8 Sep 2026 23:01:26 -0700
Subject: [PATCH] =?UTF-8?q?fix(erp-seat):=20tool=5Fchoice=3Dnone=20returne?=
 =?UTF-8?q?d=20an=20empty=20turn=20=E2=80=94=20add=20--exclude-tools-when-?=
 =?UTF-8?q?tool-choice-none=20(vLLM=20kept=20the=20tools=20in=20the=20prom?=
 =?UTF-8?q?pt,=20the=20model=20called=20one,=20parsing=20was=20off);=2012-?=
 =?UTF-8?q?shape=20tool=20matrix=20green=20before/after?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

---
 persistent-memory.md         | 6 ++++++
 stacks/erp-seat/README.md    | 7 +++++++
 stacks/erp-seat/compose.yaml | 5 +++++
 3 files changed, 18 insertions(+)

diff --git a/persistent-memory.md b/persistent-memory.md
index 31c4ae6..fca8e69 100644
--- a/persistent-memory.md
+++ b/persistent-memory.md
@@ -187,6 +187,12 @@ COMPLETE and gated RESCUED (02:13 PDT). Live open items:_
   60 router + 191 vision) from `/tank/aimodels/erp-tune-v6-bf16` (merged-run06 relayed gx10→nh3-dev→ana-ml2 in 17 min,
   no key path gx10→ana-ml2). Quant = `services/erp-seat-quant/` — DATA-FREE (~90 s; playbook §3.16), tokenizer cap
   reset, encode-equal to source. Smoke: prose in `content`, ~200 tok/s single-stream (n=3, spread <1%).
+  **Tool calling FIXED 2026-09-08 23:00 PT (operator: "fix toolcalling with the trial seat"):** the one reproducible
+  defect in a 12-shape matrix was `tool_choice:"none"` → empty turn (content AND tool_calls null, 3/3) — vLLM kept the
+  tools in the prompt, the model called one, parsing was off. Fix = `--exclude-tools-when-tool-choice-none` (seat
+  recreated; none→prose 3/3; auto/required/named/parallel/nested/empty-list/streaming/round-trip all green).
+  ⚠ `stacks/gemma4-charrp` (char-rp seat) has the same exposure and no flag yet — bouncing it is a consumer-visible
+  restart, so left for the operator's word.
   ⚠ NOT gate-parity (gated artifact = bf16 arm on the GX10; this seat has a smoke test + 3-run decode probe only) —
   **operator ruled "no gate"**; the config block states it as unrated on every safety axis.
   ⚠ **ana-ml2 mesh return routes are NON-PERSISTENT** (`ip route replace 10.100/16, 10.0/16, 10.6.110/24, 100.64/10
diff --git a/stacks/erp-seat/README.md b/stacks/erp-seat/README.md
index 52912f0..8e6ddd7 100644
--- a/stacks/erp-seat/README.md
+++ b/stacks/erp-seat/README.md
@@ -12,6 +12,13 @@ checkpoint so the GX10 is free to train the next run. First occupant: **run 6**
   and reasoning parsers, `enable_thinking` pinned false, the model's own stock template
   (`ae53464b…`, the one it trained through). Without the reasoning parser the post-tool turn leaks
   `<|channel>` markers; without the kwargs pin all prose lands in `reasoning_content`.
+- **`tool_choice: "none"` trap (measured 2026-09-08, fixed with `--exclude-tools-when-tool-choice-none`).**
+  Without the flag vLLM still renders the tools into the prompt, the model emits a tool call
+  anyway, and because parsing is off for `none` the reply is `content: null, tool_calls: null`
+  — an empty turn, 3/3 reproductions. With the flag the tools are dropped from the prompt and
+  the model answers in prose (3/3). The rest of the matrix (auto / required / named / parallel /
+  nested schema / empty `tools: []` / streaming / tool-result round trip) was green before and
+  after. `stacks/gemma4-charrp` has the same exposure and does NOT carry the flag yet.
 - **GPU1 is shared** — check real usage (`nvidia-smi --query-compute-apps=pid,used_memory`) before
   raising `ERP_GPU_MEM_UTIL`; the flag sizes KV, not CUDA context.
 - **Rollback / next run:** point `ERP_MODEL` + `ERP_SERVED_NAME` at the next quant dir, keep the
diff --git a/stacks/erp-seat/compose.yaml b/stacks/erp-seat/compose.yaml
index 0d1fa61..0f2fb40 100644
--- a/stacks/erp-seat/compose.yaml
+++ b/stacks/erp-seat/compose.yaml
@@ -35,6 +35,11 @@ services:
       - gemma4
       - --default-chat-template-kwargs
       - '{"enable_thinking": false}'
+      # tool_choice:"none" trap (measured 2026-09-08): without this flag vLLM still renders the
+      # tools into the prompt, the model emits a tool call anyway, and because tool parsing is
+      # off for tool_choice=none the reply comes back with content=null AND tool_calls=null —
+      # an empty turn. The flag drops the tools from the prompt so the model answers in prose.
+      - --exclude-tools-when-tool-choice-none
       - --chat-template
       - ${ERP_CHAT_TEMPLATE:-/tank/aimodels/erp-tune-v6-nvfp4a16/chat_template.jinja}
       - --max-model-len