chore(flows): switch default QA LLM from qwen36-fast (4B) to qwen36-deep (27B)

The smaller qwen36-fast was the previous default for OBOL_LLM_MODEL across
release-smoke and flow-{03,04,11,13,14} plus buy-external. It's documented as
flaky on the long single-shot agent-buy prompt at flow-13/14 step 46 (see the
retry-wrapper rationale added in the prior commit, plus
plans/inference-v1337-followup-20260514.md).

Switching the default to qwen36-deep (27B-class, also served by the same
spark1 vLLM endpoint) trades a bit of latency for a much more reliable
tool-call behaviour. Operators can still pin the smaller model explicitly via
OBOL_LLM_MODEL=qwen36-fast for fast iteration on non-agent flows.

Files changed:
- flows/lib.sh, flows/release-smoke.sh, flows/flow-{03,04,11,13,14}*.sh,
  flows/buy-external.sh — default value switch
- flows/lib-dual-stack.sh — WARN box in agent_buy_with_retry now
  recommends checking the model is qwen36-deep first; mentions
  qwen36-35b-heretic as the next escalation
- CLAUDE.md, .agents/skills/obol-stack-dev/{SKILL.md,references/*.md}
  — documentation refreshed

Not changed (intentional):
- internal/{model,hermes}/*_test.go — qwen36-fast is a test fixture for
  the rank parser, not a default; switching would invalidate test
  expectations without changing test intent
- plans/post-490-integration-20260513.md — historical record

Author: bussyjd

.agents/skills/obol-stack-dev/SKILL.md

@@ -47,7 +47,7 @@ OBOL_TOKEN_BASE_SEPOLIA=0x0a09371a8b011d5110656ceBCc70603e53FD2c78
 
 **Payment assertion**: don't bypass the agent buy step with a direct script exec. If the agent times out, diagnose Hermes/LiteLLM/model routing — don't relax the assertion. Required evidence: `PurchaseRequest Ready=True` + paid HTTP 200 + on-chain `Transfer` + exact balance deltas.
 
-**QA LLM**: full seller/buyer QA must route Alice and Bob through `OBOL_LLM_ENDPOINT` (OpenAI-compatible vLLM or llama.cpp on the QA host). Default `OBOL_LLM_MODEL=qwen36-fast`. Sequence: `obol model setup custom` → `obol model prefer` → one `obol model sync`. Local Ollama and cloud-fallback are **not** acceptable green substitutes for full-flow QA.
+**QA LLM**: full seller/buyer QA must route Alice and Bob through `OBOL_LLM_ENDPOINT` (OpenAI-compatible vLLM or llama.cpp on the QA host). Default `OBOL_LLM_MODEL=qwen36-deep` (27B-class). The smaller `qwen36-fast` (~4B) was the previous default but flakes on the long single-shot agent-buy prompt at flow-13/14 step 46 — see the retry-wrapper rationale in `flows/lib-dual-stack.sh::agent_buy_with_retry`. Sequence: `obol model setup custom` → `obol model prefer` → one `obol model sync`. Local Ollama and cloud-fallback are **not** acceptable green substitutes for full-flow QA.
 
 **Public vs private routes**: `/services/*`, `/.well-known/agent-registration.json`, `/skill.md`, and `/` (storefront) are public via the tunnel. **NEVER** remove `hostnames: ["obol.stack"]` from frontend or eRPC HTTPRoutes — exposing them publicly is a critical security flaw.
 

.agents/skills/obol-stack-dev/references/llm-routing.md

@@ -47,15 +47,15 @@ Canonical user flow for vLLM / sglang / mlx-lm / a remote GPU box. **No ConfigMa
 obol stack up
 
 # Drop auto-detected Ollama entries — they will out-rank the new custom entry.
-# Internal/model/rank.go parses ":9b" as 90 deci-billions; "qwen36-fast" (no
+# Internal/model/rank.go parses ":9b" as 90 deci-billions; "qwen36-deep" (no
 # ":Nb" tag) ranks 0. Without removing them, the agent stays on slow host Ollama.
 obol model remove qwen3.5:9b
 obol model remove qwen3.5:4b
 
 obol model setup custom \
     --name spark1-vllm \
     --endpoint http://192.168.18.23:8000/v1 \
-    --model qwen36-fast
+    --model qwen36-deep
 # `setup custom` validates the endpoint, patches LiteLLM, and internally calls
 # syncAgentModels → hermes.Sync → rewrites the default agent's deployment files
 # with the new primary model. No manual restart needed.

.agents/skills/obol-stack-dev/references/paid-flows.md

@@ -64,7 +64,7 @@ The runner has a `warn_unpaid_base_sepolia_rpc` preflight. The CLI scrubs paid-R
 - Alice ServiceOffer reaches `Ready=True`.
 - ERC-8004 registration tx published to Base Sepolia (`/.well-known/agent-registration.json` reachable via tunnel for live flows).
 - Bob `PurchaseRequest` reaches `Ready=True`.
-- LiteLLM exposes `paid/<OBOL_LLM_MODEL>` (default `qwen36-fast`).
+- LiteLLM exposes `paid/<OBOL_LLM_MODEL>` (default `qwen36-deep`).
 - Paid inference returns HTTP 200 and **final-answer** content (not reasoning metadata or tool-catalogue text).
 - On-chain `Transfer(Bob signer → Alice, <PAID_AMOUNT>)` receipt is archived.
 - Alice balance increases and Bob signer balance decreases by exactly `PAID_AMOUNT` wei (USDC for flow-11, OBOL for flow-13/14).

.agents/skills/obol-stack-dev/references/remote-qa.md

@@ -60,7 +60,7 @@ Set `OBOL_LLM_MODEL` to an id returned by `/models`.
 cd "$QA"
 export PATH="$QA/.workspace/bin:$FOUNDRY_BIN:$TOOL_ROOT:$PATH"
 export OBOL_LLM_ENDPOINT=${OBOL_LLM_ENDPOINT:-http://127.0.0.1:8000/v1}
-export OBOL_LLM_MODEL=${OBOL_LLM_MODEL:-qwen36-fast}
+export OBOL_LLM_MODEL=${OBOL_LLM_MODEL:-qwen36-deep}
 ts=$(date +%Y%m%d-%H%M%S)
 log="$QA/.tmp/flow-14-$ts.log"
 art="$QA/.tmp/flow-14-$ts-artifacts"

CLAUDE.md

@@ -37,7 +37,7 @@ go test -tags integration -v -run TestIntegration_Tunnel_SellDiscoverBuySidecar_
 
 # Release-gate seller/buyer smoke (requires OBOL_LLM_ENDPOINT pointing at OpenAI-compatible vLLM/llama.cpp)
 RELEASE_SMOKE_INCLUDE_OBOL=true RELEASE_SMOKE_INCLUDE_OBOL_FORK=true \
-  OBOL_LLM_ENDPOINT=http://127.0.0.1:8000/v1 OBOL_LLM_MODEL=qwen36-fast \
+  OBOL_LLM_ENDPOINT=http://127.0.0.1:8000/v1 OBOL_LLM_MODEL=qwen36-deep \
   bash flows/release-smoke.sh
 
 just up    # obol stack init + up
@@ -246,13 +246,13 @@ obol model remove qwen3.5:4b
 obol model setup custom \
     --name spark1-vllm \
     --endpoint http://192.168.18.23:8000/v1 \
-    --model qwen36-fast
+    --model qwen36-deep
 # `setup custom` validates the endpoint, patches LiteLLM, and internally calls
 # syncAgentModels → hermes.Sync → rewrites the default agent's deployment files
 # with the new primary model. No manual restart needed.
 
 # (b) OR keep Ollama and force-promote the custom entry to the head:
-obol model prefer qwen36-fast
+obol model prefer qwen36-deep
 obol model sync                                                # propagate to Hermes
 
 obol model list                                                # confirm head of model_list

flows/buy-external.sh

@@ -60,7 +60,7 @@
 #   EXTERNAL_PR_TIMEOUT_S       default: 300 (5 min)
 #   EXTERNAL_LOG_BLOCKS_BACK    default: 30 (~6 min on Base Sepolia at 2s/blk)
 #   OBOL_LLM_ENDPOINT           default: http://127.0.0.1:8000/v1
-#   OBOL_LLM_MODEL              default: qwen36-fast
+#   OBOL_LLM_MODEL              default: qwen36-deep (27B-class)
 #   OBOL_LLM_NAME               default: external-llm
 #
 # Exit code: 0 on PASS (every step pass), 1 on any FAIL.
@@ -106,7 +106,7 @@ EXTERNAL_PR_TIMEOUT_S="${EXTERNAL_PR_TIMEOUT_S:-300}"
 EXTERNAL_LOG_BLOCKS_BACK="${EXTERNAL_LOG_BLOCKS_BACK:-30}"
 
 OBOL_LLM_ENDPOINT="${OBOL_LLM_ENDPOINT:-http://127.0.0.1:8000/v1}"
-OBOL_LLM_MODEL="${OBOL_LLM_MODEL:-qwen36-fast}"
+OBOL_LLM_MODEL="${OBOL_LLM_MODEL:-qwen36-deep}"
 OBOL_LLM_NAME="${OBOL_LLM_NAME:-external-llm}"
 
 # Resolve OBOL_ROOT before sourcing helpers — lib.sh re-derives it but

flows/flow-03-inference.sh

@@ -5,7 +5,7 @@ source "$(dirname "$0")/lib.sh"
 
 if [ -n "${OBOL_LLM_ENDPOINT:-}" ]; then
     run_step "Route LiteLLM through QA LLM endpoint" route_llm_via_obol_cli "$OBOL"
-    LITELLM_MODEL="${OBOL_LLM_MODEL:-qwen36-fast}"
+    LITELLM_MODEL="${OBOL_LLM_MODEL:-qwen36-deep}"
 else
     LITELLM_MODEL="$FLOW_MODEL"
 

flows/flow-04-agent.sh

@@ -110,8 +110,8 @@ fi
 
 model_name=$("$OBOL" kubectl get cm hermes-config -n "$NS" -o jsonpath='{.data.config\.yaml}' 2>/dev/null | sed -n 's/^[[:space:]]*default: //p' | tr -d '"' | head -1)
 [ -n "$model_name" ] || model_name="qwen3.5:35b"
-if [ -n "${OBOL_LLM_ENDPOINT:-}" ] && [ "$model_name" != "${OBOL_LLM_MODEL:-qwen36-fast}" ]; then
-    fail "Hermes default model $model_name does not match QA LLM model ${OBOL_LLM_MODEL:-qwen36-fast}"
+if [ -n "${OBOL_LLM_ENDPOINT:-}" ] && [ "$model_name" != "${OBOL_LLM_MODEL:-qwen36-deep}" ]; then
+    fail "Hermes default model $model_name does not match QA LLM model ${OBOL_LLM_MODEL:-qwen36-deep}"
     cleanup_pid "$PF_PID"
     emit_metrics
     exit 0

flows/flow-11-dual-stack.sh

@@ -36,7 +36,7 @@
 #   FLOW11_BOB_HTTP_PORT   FLOW11_BOB_HTTP_ALT_PORT
 #   FLOW11_BOB_HTTPS_PORT  FLOW11_BOB_HTTPS_ALT_PORT
 #   OBOL_LLM_ENDPOINT      required vLLM/llama.cpp/OpenAI-compatible endpoint
-#   OBOL_LLM_MODEL         endpoint model name (default: qwen36-fast)
+#   OBOL_LLM_MODEL         endpoint model name (default: qwen36-deep)
 source "$(dirname "$0")/lib.sh"
 
 # ═════════════════════════════════════════════════════════════════
@@ -60,7 +60,7 @@ BOB_HTTP_ALT_PORT="${FLOW11_BOB_HTTP_ALT_PORT:-$(pick_free_port)}"
 BOB_HTTPS_PORT="${FLOW11_BOB_HTTPS_PORT:-$(pick_free_port)}"
 BOB_HTTPS_ALT_PORT="${FLOW11_BOB_HTTPS_ALT_PORT:-$(pick_free_port)}"
 FACILITATOR_URL="${FLOW11_FACILITATOR_URL:-https://x402.gcp.obol.tech}"
-OBOL_LLM_MODEL="${OBOL_LLM_MODEL:-qwen36-fast}"
+OBOL_LLM_MODEL="${OBOL_LLM_MODEL:-qwen36-deep}"
 export OBOL_LLM_MODEL
 FLOW11_ARTIFACT_DIR="${FLOW11_ARTIFACT_DIR:-$OBOL_ROOT/.tmp/flow-11-$(date +%Y%m%d-%H%M%S)}"
 if ! BASE_SEPOLIA_RPC="$(resolve_base_sepolia_rpc "${FLOW11_BASE_SEPOLIA_RPC:-${BASE_SEPOLIA_RPC:-}}")"; then

flows/flow-13-dual-stack-obol.sh

@@ -36,7 +36,7 @@
 #   FLOW13_BOB_HTTP_PORT,   _ALT, _HTTPS_PORT, _HTTPS_ALT_PORT
 #   FLOW13_ARTIFACT_DIR           where receipts + logs land
 #   OBOL_LLM_ENDPOINT             required vLLM/llama.cpp/OpenAI-compatible endpoint
-#   OBOL_LLM_MODEL                endpoint model name (default: qwen36-fast)
+#   OBOL_LLM_MODEL                endpoint model name (default: qwen36-deep, 27B-class)
 #
 source "$(dirname "$0")/lib.sh"
 DUAL_STACK_FLOW_PREFIX="FLOW13"
@@ -61,7 +61,7 @@ BOB_HTTP_ALT_PORT="$(dual_stack_env_or_free_port BOB_HTTP_ALT_PORT)"
 BOB_HTTPS_PORT="$(dual_stack_env_or_free_port BOB_HTTPS_PORT)"
 BOB_HTTPS_ALT_PORT="$(dual_stack_env_or_free_port BOB_HTTPS_ALT_PORT)"
 
-OBOL_LLM_MODEL="${OBOL_LLM_MODEL:-qwen36-fast}"
+OBOL_LLM_MODEL="${OBOL_LLM_MODEL:-qwen36-deep}"
 export OBOL_LLM_MODEL
 
 ANVIL_PORT="${FLOW13_ANVIL_PORT:-$(pick_free_port)}"
@@ -923,9 +923,9 @@ buyer_status=$(buyer_sidecar_status)
 # Mirror flow-14's relaxed assertion. Two reasons to allow remaining>=5
 # rather than exact-5: (a) controller may merge into an existing auth
 # pool on rerun (remaining=10 etc.); (b) the agent prompt asks for
-# --count 5, but qwen36-fast occasionally hallucinates --count 1, which
-# is an LLM-stochasticity issue not a buy-flow correctness issue. We
-# only care that the buy step actually provisioned at least the
+# --count 5, but the LLM occasionally hallucinates a different count,
+# which is an LLM-stochasticity issue not a buy-flow correctness issue.
+# We only care that the buy step actually provisioned at least the
 # requested count.
 remaining_n=$(echo "$buyer_status" | grep -oE 'remaining=[0-9]+' | head -1 | cut -d= -f2)
 if [ -n "$remaining_n" ] && [ "$remaining_n" -ge 5 ] 2>/dev/null; then