Switch openCodeMlx default model to Qwen3.8-27B (mlx-vlm)

demo/test_quant_suffix.sh

@@ -0,0 +1,28 @@
+#!/usr/bin/env bash
+set -u
+cd "$(dirname "${BASH_SOURCE[0]}")"
+source ./test_helpers.sh
+source ./openCodeMlx
+
+check "extracts a 4bit suffix" "4bit" "$(_quant_suffix 'mlx-community/Qwen3.8-27B-4bit')"
+check "extracts an 8bit suffix" "8bit" "$(_quant_suffix 'mlx-community/Qwen3.8-27B-8bit')"
+check "extracts a multi-digit bit suffix" "16bit" "$(_quant_suffix 'mlx-community/Some-Model-16bit')"
+check "empty for a model id with no quant suffix" "" "$(_quant_suffix 'mlx-community/Qwen3-Coder-Next')"
+check "does not match a mid-string bit token" "" "$(_quant_suffix 'mlx-community/8bit-prefixed-model')"
+check "extracts the bit count from a DWQ-qualified id" "4bit" "$(_quant_suffix 'mlx-community/Qwen3.8-27B-4bit-DWQ')"
+check "extracts the bit count from an underscore-qualified id" "8bit" "$(_quant_suffix 'mlx-community/Qwen3.8-27B-8bit_dwq')"
+check "does not treat a multi-word non-quant tail as a qualifier" "" "$(_quant_suffix 'mlx-community/8bit-prefixed-model-name')"
+
+# Regression guard: main() runs under `set -e`, and its call sites assign the result via bare
+# "main_quant=$(_quant_suffix "$MLX_MODEL")" - not inside an if/while condition, so under set -e
+# the assignment's exit status must never be non-zero, or a no-match model id (any checkpoint
+# id that doesn't end in "<N>bit") would silently kill the whole script right after the
+# multi-GB model download, with no error message and no cleanup trap installed yet.
+_quant_suffix 'mlx-community/Qwen3.8-27B-4bit' >/dev/null
+check "exits 0 on a matching model id" "0" "$?"
+_quant_suffix 'mlx-community/Qwen3-Coder-Next' >/dev/null
+check "exits 0 (not 1) on a non-matching model id, so set -e callers survive" "0" "$?"
+check "a bare assignment under set -e still reaches the next line" "REACHED" \
+  "$(set -e; no_suffix_quant=$(_quant_suffix 'mlx-community/Qwen3-Coder-Next'); echo "REACHED")"
+
+report

demo/test_updates_mlxlm.sh

@@ -15,7 +15,12 @@ REAL_PATH="/usr/bin:/bin:/usr/local/bin"
 bin_a="$work/a"; mkdir -p "$bin_a"
 log_a="$work/a.log"
 make_stub "$bin_a" pip 0 "$log_a"
-# no mlx_lm.server stub -> `command -v`/direct exec fails, simulating "not installed"
+# Explicitly stub mlx_lm.server to fail ensure_mlx_lm_current's "mlx_lm.server --help"
+# check, rather than relying on it being absent from $REAL_PATH: "not installed" is a
+# property of this test's stub, not of whatever the machine running it happens to have
+# on PATH. A machine with a real mlx_lm.server on /usr/local/bin would otherwise
+# silently take the "already installed" branch instead, making this scenario a no-op.
+make_stub "$bin_a" mlx_lm.server 1 "$log_a"
 (
   PATH="$bin_a:$REAL_PATH" ensure_mlx_lm_current >/dev/null 2>&1
 )
@@ -26,6 +31,7 @@ check "not-installed + pip install succeeds -> pip was invoked with --upgrade" "
 bin_b="$work/b"; mkdir -p "$bin_b"
 log_b="$work/b.log"
 make_stub "$bin_b" pip 1 "$log_b"
+make_stub "$bin_b" mlx_lm.server 1 "$log_b"   # force "not installed", see scenario A
 (
   PATH="$bin_b:$REAL_PATH" ensure_mlx_lm_current >/dev/null 2>&1
 )

demo/test_updates_mlxvlm.sh

@@ -0,0 +1,79 @@
+#!/usr/bin/env bash
+set -u
+cd "$(dirname "${BASH_SOURCE[0]}")"
+source ./test_helpers.sh
+source ./openCodeMlx
+
+work=$(mktemp -d)
+trap 'rm -rf "$work"' EXIT
+# Stub dirs are prepended to a real system PATH, not used exclusively:
+# ensure_mlx_vlm_current itself shells out to other real commands, which an
+# exclusive stub-only PATH would hide too.
+REAL_PATH="/usr/bin:/bin:/usr/local/bin"
+
+# Scenario A: mlx-vlm missing, pip install stub succeeds -> returns 0, and pip is invoked with
+# the same transformers pin ensure_mlx_lm_current carries. This is the actual fix under test:
+# both packages share one venv, so an unconstrained "pip install mlx-vlm" is free to upgrade
+# transformers past the version mlx-lm's AutoTokenizer.register call tolerates, silently
+# breaking mlx_lm.server (still used unconditionally for the small background model).
+bin_a="$work/a"; mkdir -p "$bin_a"
+log_a="$work/a.log"
+make_stub "$bin_a" pip 0 "$log_a"
+# Explicitly stub python3 to fail ensure_mlx_vlm_current's "python3 -m mlx_vlm.server
+# --help" check, rather than relying on the real system python3 lacking mlx_vlm: "not
+# installed" is a property of this test's stub, not of whatever python3 environment the
+# machine running it happens to have. A machine where mlx_vlm is importable from the
+# default python3 (e.g. a shared/activated venv) would otherwise silently take the
+# "already installed" branch instead, making this scenario a no-op.
+make_stub "$bin_a" python3 1 "$log_a"
+(
+  PATH="$bin_a:$REAL_PATH" ensure_mlx_vlm_current >/dev/null 2>&1
+)
+check "not-installed + pip install succeeds -> returns 0" "0" "$?"
+check "not-installed -> pip carries the shared transformers pin" "1" "$(grep -c 'transformers>=5.7,<5.13' "$log_a")"
+
+# Scenario B: mlx-vlm missing, pip install stub fails -> hard error (returns 1).
+bin_b="$work/b"; mkdir -p "$bin_b"
+log_b="$work/b.log"
+make_stub "$bin_b" pip 1 "$log_b"
+make_stub "$bin_b" python3 1 "$log_b"   # force "not installed", see scenario A
+(
+  PATH="$bin_b:$REAL_PATH" ensure_mlx_vlm_current >/dev/null 2>&1
+)
+check "not-installed + pip install fails -> returns 1 (hard error)" "1" "$?"
+
+# Scenario C: mlx-vlm present, pip install stub fails -> still returns 0 (warn+continue).
+bin_c="$work/c"; mkdir -p "$bin_c"
+log_c="$work/c.log"
+make_stub "$bin_c" python3 0 "$log_c"
+make_stub "$bin_c" pip 1 "$log_c"
+(
+  PATH="$bin_c:$REAL_PATH" ensure_mlx_vlm_current >/dev/null 2>&1
+)
+check "already-installed + pip update fails -> still returns 0 (warn and continue)" "0" "$?"
+
+# Scenario D: mlx-vlm present, pip install stub succeeds -> pip was still invoked (not skipped)
+# and still carries the transformers pin.
+bin_d="$work/d"; mkdir -p "$bin_d"
+log_d="$work/d.log"
+make_stub "$bin_d" python3 0 "$log_d"
+make_stub "$bin_d" pip 0 "$log_d"
+(
+  PATH="$bin_d:$REAL_PATH" ensure_mlx_vlm_current >/dev/null 2>&1
+)
+check "already-installed + pip update succeeds -> returns 0" "0" "$?"
+check "already-installed -> pip is still invoked (update not skipped)" "1" "$(grep -c '^pip ' "$log_d")"
+check "already-installed -> pip update still carries the shared transformers pin" "1" "$(grep -c 'transformers>=5.7,<5.13' "$log_d")"
+
+# Scenario E: mlx-vlm present, called in install-only mode -> pip upgrade must NOT be attempted.
+bin_e="$work/e"; mkdir -p "$bin_e"
+log_e="$work/e.log"
+make_stub "$bin_e" python3 0 "$log_e"
+make_stub "$bin_e" pip 1 "$log_e"
+(
+  PATH="$bin_e:$REAL_PATH" ensure_mlx_vlm_current install-only >/dev/null 2>&1
+)
+check "already-installed + install-only mode -> returns 0" "0" "$?"
+check "already-installed + install-only mode -> pip was NOT invoked" "0" "$(grep -c '^pip ' "$log_e")"
+
+report

demo/test_updates_opencode.sh

@@ -21,7 +21,7 @@ set -e
 source ./test_helpers.sh
 source ./openCodeMlx
 _install_opencode_via_curl() { return 1; }
-PATH="/usr/bin:/bin:/usr/local/bin" HOME="$1" ensure_opencode_current 2>&1
+PATH="/usr/bin:/bin" HOME="$1" ensure_opencode_current 2>&1
 TESTEOF
 
   # Run it; with the bug (bare call), set -e aborts before printing error.
@@ -48,30 +48,55 @@ trap 'rm -rf "$work"' EXIT
 # Step 3), which tests override directly as a bash function instead.
 REAL_PATH="/usr/bin:/bin:/usr/local/bin"
 
-# Scenario A: opencode not on PATH; the install override actually creates a
-# working $HOME/.opencode/bin/opencode (simulating a real successful install).
-home_a="$work/a_home"
-_install_opencode_via_curl() {
-  mkdir -p "$home_a/.opencode/bin"
-  cat > "$home_a/.opencode/bin/opencode" <<'INNER'
+# Scenarios A/B/E need a PATH where opencode is NOT resolvable, so that
+# ensure_opencode_current takes its "not found -> install" branch, but where coreutils
+# (mkdir/cat/chmod, used by the install-simulation overrides above) still are. Deliberately
+# excludes /usr/local/bin, unlike REAL_PATH above: it's the one realistic place a
+# system-wide `opencode` might actually be symlinked to on this project's target platform
+# (macOS/Homebrew) - the official installer puts it in ~/.opencode/bin, which was never on
+# REAL_PATH to begin with. /usr/bin and /bin alone cover the coreutils these scenarios need.
+NOT_INSTALLED_PATH="/usr/bin:/bin"
+
+# Guarded rather than assumed outright: even /usr/bin:/bin could theoretically contain a
+# real `opencode` on some machine this hasn't been tested against. If so, skip (not fail)
+# rather than let these scenarios silently exercise the "already installed -> opencode
+# upgrade" branch instead - a REAL network-touching command against whatever opencode
+# install actually exists, run under a test that thinks it's simulating "not installed".
+if PATH="$NOT_INSTALLED_PATH" command -v opencode >/dev/null 2>&1; then
+  echo "SKIP: a real 'opencode' is resolvable on PATH=$NOT_INSTALLED_PATH in this" >&2
+  echo "environment - scenarios A, B and E cannot safely simulate 'not installed' here" >&2
+  echo "(would run a real 'opencode upgrade' instead). Skipping those three checks." >&2
+  NOT_INSTALLED_SCENARIOS_SAFE=0
+else
+  NOT_INSTALLED_SCENARIOS_SAFE=1
+fi
+
+if [[ "$NOT_INSTALLED_SCENARIOS_SAFE" -eq 1 ]]; then
+  # Scenario A: opencode not on PATH; the install override actually creates a
+  # working $HOME/.opencode/bin/opencode (simulating a real successful install).
+  home_a="$work/a_home"
+  _install_opencode_via_curl() {
+    mkdir -p "$home_a/.opencode/bin"
+    cat > "$home_a/.opencode/bin/opencode" <<'INNER'
 #!/usr/bin/env bash
 exit 0
 INNER
-  chmod +x "$home_a/.opencode/bin/opencode"
-}
-(
-  PATH="$REAL_PATH" HOME="$home_a" ensure_opencode_current >/dev/null 2>&1
-)
-check "not-installed + install produces a binary -> returns 0" "0" "$?"
+    chmod +x "$home_a/.opencode/bin/opencode"
+  }
+  (
+    PATH="$NOT_INSTALLED_PATH" HOME="$home_a" ensure_opencode_current >/dev/null 2>&1
+  )
+  check "not-installed + install produces a binary -> returns 0" "0" "$?"
 
-# Scenario B: opencode not on PATH; the install override is a no-op
-# (simulates a failed install -- opencode still isn't findable afterward).
-home_b="$work/b_home"
-_install_opencode_via_curl() { :; }
-(
-  PATH="$REAL_PATH" HOME="$home_b" ensure_opencode_current >/dev/null 2>&1
-)
-check "not-installed + install produces no binary -> returns 1 (hard error)" "1" "$?"
+  # Scenario B: opencode not on PATH; the install override is a no-op
+  # (simulates a failed install -- opencode still isn't findable afterward).
+  home_b="$work/b_home"
+  _install_opencode_via_curl() { :; }
+  (
+    PATH="$NOT_INSTALLED_PATH" HOME="$home_b" ensure_opencode_current >/dev/null 2>&1
+  )
+  check "not-installed + install produces no binary -> returns 1 (hard error)" "1" "$?"
+fi
 
 # Scenario C: opencode already on PATH, `opencode upgrade` stub succeeds.
 bin_c="$work/c"; mkdir -p "$bin_c"
@@ -108,8 +133,12 @@ check "already-installed + install-only mode -> upgrade was NOT invoked" "0" "$(
 # Scenario E: opencode not on PATH, install fails (returns 1), ensure_opencode_current
 # is called under `set -e`. Verify that the function's return 1 is reached (not aborted
 # by set -e when _install_opencode_via_curl exits with status 1). The function should
-# return 1 to the outer subshell (not die uncontrolled).
-test_set_e_with_failed_install
-check "install fails under set -e -> function returns 1 (not aborted by set -e)" "0" "$?"
+# return 1 to the outer subshell (not die uncontrolled). Same NOT_INSTALLED_PATH
+# assumption as scenarios A/B (test_set_e_with_failed_install hardcodes the same
+# "/usr/bin:/bin" internally) - guarded above, so only run this when that guard held.
+if [[ "$NOT_INSTALLED_SCENARIOS_SAFE" -eq 1 ]]; then
+  test_set_e_with_failed_install
+  check "install fails under set -e -> function returns 1 (not aborted by set -e)" "0" "$?"
+fi
 
 report

models.sh

@@ -1,5 +1,45 @@
 #!/bin/sh
 
+force=false
+while [ $# -gt 0 ]; do
+  case "$1" in
+    -f|--force)
+      force=true
+      ;;
+    *)
+      echo "Unknown option: $1"
+      echo "Usage: $0 [-f|--force]"
+      exit 1
+      ;;
+  esac
+  shift
+done
+
+# Model names/contexts are not duplicated here: src/main/bin/lca is the canonical
+# source (it must be self-contained since it's distributed standalone), so we read
+# its named variables via a targeted grep+eval.
+SCRIPT_DIR="$(cd "$(dirname "$0")" && pwd)"
+LCA_SCRIPT="$SCRIPT_DIR/src/main/bin/lca"
+if [ ! -f "$LCA_SCRIPT" ]; then
+  echo "Error: canonical model config not found at $LCA_SCRIPT" >&2
+  exit 1
+fi
+eval "$(grep -E '^(BASE_CHAT_MODEL|BASE_FALLBACK_MODEL|EMBEDDING_MODEL|CUSTOM_CHAT_MODEL|CUSTOM_CHAT_CONTEXT|QWEN_EXTRA_PARAMS|CUSTOM_FALLBACK_MODEL|CUSTOM_FALLBACK_CONTEXT|REVIEW_MODEL|REVIEW_CONTEXT|DEFAULT_CONTEXT_WINDOW|MODEL_STATE_DIR)=' "$LCA_SCRIPT")"
+
+# Guard against a variable being renamed/added in lca without updating the grep alternation
+# above (which would silently eval to empty) or a value in lca referencing another variable
+# defined later in that file (which would also evaluate to empty here). QWEN_EXTRA_PARAMS is
+# deliberately excluded: an empty extra-params string is a legitimate value, not a bug.
+for _v in BASE_CHAT_MODEL BASE_FALLBACK_MODEL EMBEDDING_MODEL CUSTOM_CHAT_MODEL \
+  CUSTOM_CHAT_CONTEXT CUSTOM_FALLBACK_MODEL CUSTOM_FALLBACK_CONTEXT REVIEW_MODEL \
+  REVIEW_CONTEXT DEFAULT_CONTEXT_WINDOW MODEL_STATE_DIR; do
+  eval "_val=\$$_v"
+  if [ -z "$_val" ]; then
+    echo "Error: $_v not resolved from $LCA_SCRIPT" >&2
+    exit 1
+  fi
+done
+
 os=""
 case "$(uname -s)" in
   Darwin)
@@ -41,6 +81,11 @@ if ! command -v ollama >/dev/null 2>&1; then
   esac
 fi
 
+get_model_id() {
+    model="$1"
+    ollama list 2>/dev/null | awk -v m="$model" '$1 == m {print $2; exit}'
+}
+
 checkAndInstall() {
     model="$1"
     echo "Checking for $model model..."
@@ -57,52 +102,78 @@ createCustomModel() {
     base_model="$1"
     custom_name="$2"
     context_size="$3"
+    extra_params="${4:-}"
+
+    current_id="$(get_model_id "$base_model")"
+    if [ -z "$current_id" ]; then
+      echo "Warning: could not retrieve ID for $base_model. Skipping $custom_name."
+      return
+    fi
 
-    echo "Creating custom model $custom_name from $base_model with context size $context_size..."
+    # Mirrors src/main/bin/lca's rebuild_custom_model_if_changed: fingerprint the full desired
+    # Modelfile recipe (base id + context + extra params), not just the base model's id, so a
+    # context/parameter-only change is detected even when the base model itself hasn't changed.
+    # Shares lca's MODEL_STATE_DIR so the two scripts agree on whether a custom model is stale.
+    desired_signature="${current_id}|${context_size}|${extra_params}"
+    state_file="${MODEL_STATE_DIR}/${custom_name}.id"
+    saved_signature=""
+    if [ -f "$state_file" ]; then
+      saved_signature="$(cat "$state_file")"
+    fi
 
-    # Check if custom model already exists
-    if ollama list 2>/dev/null | grep -q "^$custom_name"; then
-      echo "$custom_name already exists."
+    custom_exists="no"
+    if ollama list 2>/dev/null | awk '{print $1}' | grep -Fxq "${custom_name}:latest"; then
+      custom_exists="yes"
+    fi
+
+    if [ "$force" != true ] && [ "$desired_signature" = "$saved_signature" ] && [ "$custom_exists" = "yes" ]; then
+      echo "$custom_name is up to date."
       return
     fi
 
+    if [ "$custom_exists" = "yes" ]; then
+      echo "Rebuilding $custom_name (base model, context, or parameters changed; or --force)..."
+      ollama rm "$custom_name"
+    else
+      echo "$custom_name not found. Creating..."
+    fi
+
     # Create a temporary Modelfile
     modelfile=$(mktemp)
-    cat > "$modelfile" << EOF
-FROM $base_model
-PARAMETER num_ctx $context_size
-EOF
+    {
+      echo "FROM $base_model"
+      echo "PARAMETER num_ctx $context_size"
+      if [ -n "$extra_params" ]; then
+        old_ifs="$IFS"
+        IFS=';'
+        for kv in $extra_params; do
+          key="${kv%%=*}"
+          value="${kv#*=}"
+          echo "PARAMETER $key $value"
+        done
+        IFS="$old_ifs"
+      fi
+    } > "$modelfile"
 
     # Create the custom model
     ollama create "$custom_name" -f "$modelfile"
 
     # Clean up
     rm "$modelfile"
+    mkdir -p "$MODEL_STATE_DIR"
+    printf '%s\n' "$desired_signature" > "$state_file"
 
     echo "$custom_name created successfully."
 }
 
 # Install base models
-#checkAndInstall deepseek-coder:6.7b
-checkAndInstall qwen3.6:35b-a3b
-checkAndInstall gpt-oss:20b
-checkAndInstall nomic-embed-text:latest
+checkAndInstall "$BASE_CHAT_MODEL"
+checkAndInstall "$BASE_FALLBACK_MODEL"
+checkAndInstall "$EMBEDDING_MODEL"
 
-# Create custom models with larger context (128k=131072, 64k=65536)
-createCustomModel qwen3.6:35b-a3b qwen3.6-128k 131072
-createCustomModel gpt-oss:20b gpt-oss-64k 65536
+# Create custom models with larger context
+createCustomModel "$BASE_CHAT_MODEL" "$CUSTOM_CHAT_MODEL" "$CUSTOM_CHAT_CONTEXT" "$QWEN_EXTRA_PARAMS"
+createCustomModel "$BASE_FALLBACK_MODEL" "$CUSTOM_FALLBACK_MODEL" "$CUSTOM_FALLBACK_CONTEXT"
 
 # Create review model with thinking disabled and smaller context for faster response
-echo "Creating review model qwen3.6-review from qwen3.6:35b-a3b..."
-if ollama list 2>/dev/null | grep -q "^qwen3.6-review"; then
-  echo "qwen3.6-review already exists."
-else
-  modelfile=$(mktemp)
-  cat > "$modelfile" << EOF
-FROM qwen3.6:35b-a3b
-PARAMETER num_ctx 65536
-EOF
-  ollama create qwen3.6-review -f "$modelfile"
-  rm "$modelfile"
-  echo "qwen3.6-review created successfully."
-fi
+createCustomModel "$BASE_CHAT_MODEL" "$REVIEW_MODEL" "$REVIEW_CONTEXT" "$QWEN_EXTRA_PARAMS"