ODS V3 Pre-Release: Public testing and refinement ahead of the official V3 launch. Turn your PC, Mac, or Linux box into a private AI server.
Python ai-agentsamdcomfyuidockerllama-cppllmlocal-ain8nnvidiaopen-webuiragself-hostedspeech-to-textstrix-halotext-to-speechworkflow-automation
#!/usr/bin/env bash
# External Ollama / LM Studio discovery and installer-selection contracts.
set -euo pipefail
ROOT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")/.." && pwd)"
# shellcheck source=../installers/lib/external-services.sh
source "$ROOT_DIR/installers/lib/external-services.sh"
PASSED=0
FAILED=0
pass() {
printf '[PASS] %s\n' "$1"
PASSED=$((PASSED + 1))
}
fail() {
printf '[FAIL] %s\n' "$1" >&2
FAILED=$((FAILED + 1))
}
assert_eq() {
local actual="$1" expected="$2" label="$3"
if [[ "$actual" == "$expected" ]]; then
pass "$label"
else
fail "$label (expected '$expected', got '$actual')"
fi
}
assert_true() {
local label="$1"
shift
if "$@"; then
pass "$label"
else
fail "$label"
fi
}
assert_eq "$(external_llm_normalize_model_name 'qwen3.5:9b')" \
"qwen3.5-9b" "normalizes Ollama tags"
assert_eq "$(external_llm_normalize_model_name 'Qwen3.5-9B-Q4_K_M.gguf')" \
"qwen3.5-9b" "removes GGUF quantization suffixes"
assert_eq "$(external_llm_container_url 'http://localhost:11434/v1')" \
"http://host.docker.internal:11434" "normalizes localhost for containers"
assert_eq "$(external_llm_container_url 'http://[::1]:1234/api/v1')" \
"http://host.docker.internal:1234" "normalizes IPv6 loopback for containers"
assert_eq "$(external_llm_host_url 'http://host.docker.internal:11434/v1')" \
"http://127.0.0.1:11434" "normalizes the Docker host alias for host probes"
assert_true "accepts a supported external base URL" \
external_llm_validate_url "http://127.0.0.1:11434/v1"
if external_llm_validate_url "file:///tmp/models"; then
fail "rejects non-HTTP external URLs"
else
pass "rejects non-HTTP external URLs"
fi
if external_llm_validate_url "http://user:[email protected]:11434"; then
fail "rejects credentials embedded in external URLs"
else
pass "rejects credentials embedded in external URLs"
fi
if external_llm_validate_url "http://127.0.0.1:11434?token=secret"; then
fail "rejects query data embedded in external URLs"
else
pass "rejects query data embedded in external URLs"
fi
assert_true "matches equivalent Ollama and GGUF model names" \
external_llm_model_matches "Qwen3.5-9B-Q4_K_M.gguf" "qwen3.5:9b"
if external_llm_model_matches "qwen3.5:9b" "llama3.2:3b"; then
fail "rejects unrelated model families"
else
pass "rejects unrelated model families"
fi
run_phase_case() {
set -euo pipefail
local case_name="$1"
local install_dir="$2"
ods_progress() { :; }
log() { :; }
ai() { :; }
ai_ok() { :; }
ai_bad() { :; }
resolve_compose_config() { :; }
curl() {
local url="${*: -1}"
case "$url" in
*/api/tags)
[[ "${MOCK_OLLAMA:-down}" == "up" ]] || return 22
printf '{"models":[{"name":"qwen3.5:9b"},{"name":"llama3.2:3b"}]}'
;;
*/v1/models)
[[ "${MOCK_LMSTUDIO:-down}" == "up" ]] || return 22
printf '{"data":[{"id":"qwen3.5-9b"},{"id":"local-model"}]}'
;;
*/v1/chat/completions)
printf '{"choices":[{"message":{"content":"OK"}}]}'
;;
*)
return 22
;;
esac
}
INTERACTIVE=false
DRY_RUN=false
ODS_MODE=local
INSTALL_DIR="$install_dir"
GGUF_FILE="Qwen3.5-9B-Q4_K_M.gguf"
LLM_MODEL="qwen3.5-9b"
unset EXTERNAL_LLM_URL EXTERNAL_LLM_CONTAINER_URL EXTERNAL_LLM_PROVIDER
unset EXTERNAL_LLM_MODEL EXTERNAL_LLM_AUTO_REUSE EXTERNAL_LLM_DISABLE
unset EXTERNAL_LLM_RESET SKIP_MODEL_DOWNLOAD NATIVE_LLM_BASE_URL
unset EXTERNAL_LLM_API_KEY_FILE EXTERNAL_LLM_API_KEY_RESET EXTERNAL_LLM_API_KEY_DISABLE
case "$case_name" in
default)
MOCK_OLLAMA=up
;;
auto)
MOCK_OLLAMA=up
EXTERNAL_LLM_AUTO_REUSE=true
;;
persisted)
MOCK_OLLAMA=up
;;
disabled)
MOCK_OLLAMA=up
EXTERNAL_LLM_DISABLE=true
;;
disabled-cloud)
MOCK_OLLAMA=up
ODS_MODE=cloud
EXTERNAL_LLM_DISABLE=true
;;
explicit-offline)
EXTERNAL_LLM_URL="http://127.0.0.1:11434"
EXTERNAL_LLM_PROVIDER="ollama"
EXTERNAL_LLM_MODEL="qwen3.5:9b"
;;
explicit-openai|detect-openai|explicit-switch|explicit-same|explicit-same-no-key)
MOCK_LMSTUDIO=up
MOCK_OLLAMA=down
EXTERNAL_LLM_URL="http://10.0.2.2:18080"
EXTERNAL_LLM_PROVIDER="openai-compatible"
[[ "$case_name" != detect-openai ]] || EXTERNAL_LLM_PROVIDER=auto
EXTERNAL_LLM_MODEL="local-model"
if [[ "$case_name" == explicit-same || "$case_name" == explicit-same-no-key ]]; then
EXTERNAL_LLM_URL="http://127.0.0.1:18080"
fi
[[ "$case_name" != explicit-same-no-key ]] || EXTERNAL_LLM_API_KEY_DISABLE=true
;;
explicit-cloud)
MOCK_OLLAMA=up
ODS_MODE=cloud
EXTERNAL_LLM_URL="http://127.0.0.1:11434"
EXTERNAL_LLM_PROVIDER="ollama"
EXTERNAL_LLM_MODEL="qwen3.5:9b"
;;
explicit-hybrid)
MOCK_OLLAMA=up
ODS_MODE=hybrid
EXTERNAL_LLM_URL="http://127.0.0.1:11434"
EXTERNAL_LLM_PROVIDER="ollama"
EXTERNAL_LLM_MODEL="qwen3.5:9b"
;;
explicit-native)
MOCK_OLLAMA=up
NATIVE_LLM_BASE_URL=http://localhost:8080
EXTERNAL_LLM_URL="http://127.0.0.1:11434"
EXTERNAL_LLM_PROVIDER="ollama"
EXTERNAL_LLM_MODEL="qwen3.5:9b"
;;
*)
return 99
;;
esac
# shellcheck source=../installers/phases/02b-external-services.sh
source "$ROOT_DIR/installers/phases/02b-external-services.sh"
}
TEMP_DIR="$(mktemp -d)"
trap 'rm -rf "$TEMP_DIR"' EXIT
cat > "$TEMP_DIR/commented.env" <<'EOF'
EXTERNAL_LLM_URL=http://127.0.0.1:11434 # retained endpoint
ODS_GATEWAY_ONLY=true # preserve API-only selection
ENABLE_DEVTOOLS=false # preserve lean host tools
EXTERNAL_LLM_MODEL="model # literal" # trailing comment
EOF
assert_eq "$(external_llm_env_value "$TEMP_DIR/commented.env" EXTERNAL_LLM_URL)" \
"http://127.0.0.1:11434" "retained route follows Compose inline-comment grammar"
assert_eq "$(external_llm_env_value "$TEMP_DIR/commented.env" ODS_GATEWAY_ONLY)" \
"true" "gateway-only selection survives an inline comment"
assert_eq "$(external_llm_env_value "$TEMP_DIR/commented.env" ENABLE_DEVTOOLS)" \
"false" "developer-tool selection survives an inline comment"
assert_eq "$(external_llm_env_value "$TEMP_DIR/commented.env" EXTERNAL_LLM_MODEL)" \
"model # literal" "quoted hash stays inside a model name"
for external_case in explicit-openai detect-openai; do
if output="$(run_phase_case "$external_case" "$TEMP_DIR/$external_case"; printf '%s|%s\n' \
"${EXTERNAL_LLM_PROVIDER:-}" "${EXTERNAL_LLM_MODEL:-}")"; then
assert_eq "$output" "openai-compatible|local-model" "$external_case preserves protocol and exact model identity"
else
fail "$external_case accepts a generic OpenAI-compatible endpoint"
fi
done
if output="$(run_phase_case default "$TEMP_DIR/default"; printf '%s|%s\n' "${EXTERNAL_LLM_URL:-}" "${SKIP_MODEL_DOWNLOAD:-}")"; then
assert_eq "$output" "|false" "non-interactive ambient discovery is inert by default"
else
fail "default non-interactive phase completes"
fi
if output="$(run_phase_case auto "$TEMP_DIR/auto"; printf '%s|%s|%s|%s\n' \
"${EXTERNAL_LLM_PROVIDER:-}" "${EXTERNAL_LLM_MODEL:-}" \
"${EXTERNAL_LLM_CONTAINER_URL:-}" "${SKIP_MODEL_DOWNLOAD:-}")"; then
assert_eq "$output" \
"ollama|qwen3.5:9b|http://host.docker.internal:11434|true" \
"explicit auto-reuse validates and persists an exact external route"
else
fail "opt-in auto-reuse phase completes"
fi
mkdir -p "$TEMP_DIR/persisted"
cat > "$TEMP_DIR/persisted/.env" <<'EOF'
EXTERNAL_LLM_URL=http://127.0.0.1:11434
EXTERNAL_LLM_CONTAINER_URL=http://host.docker.internal:11434
EXTERNAL_LLM_PROVIDER=ollama
EXTERNAL_LLM_MODEL=qwen3.5:9b
SKIP_MODEL_DOWNLOAD=true
EOF
if output="$(run_phase_case persisted "$TEMP_DIR/persisted"; printf '%s|%s\n' \
"${EXTERNAL_LLM_MODEL:-}" "${SKIP_MODEL_DOWNLOAD:-}")"; then
assert_eq "$output" "qwen3.5:9b|true" \
"rerun revalidates and preserves a reachable external selection"
else
fail "persisted external selection rerun completes"
fi
for route_case in explicit-switch explicit-same explicit-same-no-key; do
mkdir -p "$TEMP_DIR/$route_case/config/litellm"
printf 'EXTERNAL_LLM_URL=http://127.0.0.1:18080\n' >"$TEMP_DIR/$route_case/.env"
printf 'test-secret-123\n' >"$TEMP_DIR/$route_case/config/litellm/external-upstream.key"
chmod 600 "$TEMP_DIR/$route_case/config/litellm/external-upstream.key"
done
if output="$(run_phase_case explicit-switch "$TEMP_DIR/explicit-switch"; printf '%s|%s\n' \
"${EXTERNAL_LLM_API_KEY_FILE:-}" "${EXTERNAL_LLM_API_KEY_RESET:-}")"; then
assert_eq "$output" '|true' 'new external endpoint never inherits the previous key'
else
fail 'new external endpoint validation completes without old key'
fi
if output="$(run_phase_case explicit-same-no-key "$TEMP_DIR/explicit-same-no-key"; printf '%s|%s\n' \
"${EXTERNAL_LLM_API_KEY_FILE:-}" "${EXTERNAL_LLM_API_KEY_RESET:-}")"; then
assert_eq "$output" '|true' '--no-external-llm-key stops saved key reuse'
else
fail '--no-external-llm-key route validation completes'
fi
if output="$(run_phase_case explicit-same "$TEMP_DIR/explicit-same"; printf '%s|%s\n' \
"${EXTERNAL_LLM_API_KEY_FILE:-}" "${EXTERNAL_LLM_API_KEY_RESET:-}")"; then
assert_eq "$output" "$TEMP_DIR/explicit-same/config/litellm/external-upstream.key|false" \
'same external endpoint reuses its installed private key'
else
fail 'same external endpoint key reuse completes'
fi
mkdir -p "$TEMP_DIR/disabled"
cp "$TEMP_DIR/persisted/.env" "$TEMP_DIR/disabled/.env"
if output="$(run_phase_case disabled "$TEMP_DIR/disabled"; printf '%s|%s|%s\n' \
"${EXTERNAL_LLM_URL:-}" "${SKIP_MODEL_DOWNLOAD:-}" "${EXTERNAL_LLM_RESET:-}")"; then
assert_eq "$output" "|false|true" \
"--no-external-llm clears the persisted topology for managed inference"
else
fail "external reset phase completes"
fi
if output="$(run_phase_case disabled-cloud "$TEMP_DIR/disabled"; printf '%s|%s|%s\n' \
"${EXTERNAL_LLM_URL:-}" "${SKIP_MODEL_DOWNLOAD:-}" "${EXTERNAL_LLM_RESET:-}")"; then
assert_eq "$output" "|false|true" \
"--no-external-llm permits an intentional transition to cloud mode"
else
fail "external reset followed by cloud mode completes"
fi
if run_phase_case explicit-offline "$TEMP_DIR/offline"; then
fail "explicit unavailable provider must fail closed"
else
pass "explicit unavailable provider fails closed"
fi
if run_phase_case explicit-cloud "$TEMP_DIR/cloud"; then
fail "external reuse must reject cloud mode"
else
pass "external reuse rejects cloud mode without mutating the topology"
fi
if run_phase_case explicit-hybrid "$TEMP_DIR/hybrid"; then
fail "external reuse must reject hybrid mode"
else
pass "external reuse rejects hybrid mode without bypassing LiteLLM"
fi
if run_phase_case explicit-native "$TEMP_DIR/native"; then
fail "external reuse must reject a second host-managed backend"
else
pass "external reuse rejects a simultaneous host-native llama-server"
fi
run_phase06_env_cycle() (
set -euo pipefail
local install_dir="$TEMP_DIR/phase06-install"
mkdir -p "$install_dir"
export HOME="$install_dir/test-home"
mkdir -p "$HOME"
tar -C "$ROOT_DIR" \
--exclude='./.env' \
--exclude='./extensions/services/dashboard/node_modules' \
--exclude='./extensions/services/dashboard/dist' \
-cf - . | tar -C "$install_dir" -xf -
export INSTALL_DIR="$install_dir"
export SCRIPT_DIR="$install_dir"
export LOG_FILE="$install_dir/phase06.log"
export DRY_RUN=false
export INTERACTIVE=false
export ODS_MODE=local
export GPU_BACKEND=cpu
export TIER=T1
export TIER_NAME=Entry
export LLM_MODEL=qwen3-1.7b
export GGUF_FILE=Qwen3-1.7B-Q4_K_M.gguf
export MAX_CONTEXT=4096
export ODS_VERSION=2.1.0
export ENABLE_VOICE=false
export ENABLE_WORKFLOWS=false
export ENABLE_RAG=false
export ENABLE_HERMES=false
export EXTERNAL_LLM_URL=http://127.0.0.1:11434
export EXTERNAL_LLM_CONTAINER_URL=http://host.docker.internal:11434
export EXTERNAL_LLM_PROVIDER=ollama
export EXTERNAL_LLM_MODEL=qwen3.5:9b
export NATIVE_LLM_BASE_URL=
# This fixture exercises ONLY external routing/.env generation. The copied
# tree still contains the real APE compose with its ./data/ape:/data/ape:z
# bind, so phase 06's private-state preparation would run here and (on the
# non-1000 CI runner) correctly fail ownership verification for a directory
# this fixture never provisions or asserts. Drop only the unrelated APE
# service declaration so the fixture stays isolated to its stated scope;
# the dedicated real-image APE ownership probe covers that contract.
printf 'services:\n ape-not-a-bind-fixture:\n image: scratch\n' \
>"$install_dir/extensions/services/ape/compose.yaml"
# shellcheck source=../installers/lib/constants.sh
source "$install_dir/installers/lib/constants.sh"
# shellcheck source=../installers/lib/logging.sh
source "$install_dir/installers/lib/logging.sh"
# shellcheck source=../installers/lib/ui.sh
source "$install_dir/installers/lib/ui.sh"
# shellcheck source=../installers/lib/detection.sh
source "$install_dir/installers/lib/detection.sh"
# shellcheck source=../installers/lib/progress.sh
source "$install_dir/installers/lib/progress.sh"
# Match install-core's privilege helpers; sudo itself is stubbed below.
# shellcheck source=../installers/lib/sudo.sh
source "$install_dir/installers/lib/sudo.sh"
ods_progress() { :; }
ai() { :; }
ai_ok() { :; }
ai_warn() { :; }
ai_bad() { :; }
chapter() { :; }
signal() { :; }
show_phase() { :; }
sudo() { return 0; }
ods_sudo() {
# Permission setup now verifies its result. chmod is owner-safe;
# CI's non-1000 runner needs real sudo for the fixture's group change.
case "$1" in
chmod) "$@" ;;
chgrp) command sudo -n "$@" ;;
*) return 0 ;;
esac
}
docker() {
if [[ "${1:-}" == "info" && "${2:-}" == "--format" ]]; then
printf '4\n'
return 0
fi
command docker "$@"
}
# shellcheck source=../installers/phases/06-directories.sh
source "$install_dir/installers/phases/06-directories.sh"
grep -qx 'LLM_BACKEND=external' "$install_dir/.env"
grep -qx 'LLM_MODEL=qwen3.5:9b' "$install_dir/.env"
grep -qx 'LLM_API_URL=http://litellm:4000' "$install_dir/.env"
grep -qx 'OPEN_WEBUI_LLM_BASE_URL=http://litellm:4000/v1' "$install_dir/.env"
grep -qx 'HERMES_LLM_BASE_URL=http://litellm:4000/v1' "$install_dir/.env"
grep -q 'model_name: "ods/current"' "$install_dir/config/litellm/local.yaml"
grep -q 'model: "openai/qwen3.5:9b"' "$install_dir/config/litellm/local.yaml"
grep -q 'api_base: "http://host.docker.internal:11434/v1"' "$install_dir/config/litellm/local.yaml"
grep -q 'master_key: os.environ/LITELLM_MASTER_KEY' "$install_dir/config/litellm/local.yaml"
[[ -f "$install_dir/config/litellm/external-upstream.key" ]]
[[ ! -s "$install_dir/config/litellm/external-upstream.key" ]]
[[ "$(stat -c '%a' "$install_dir/config/litellm/external-upstream.key")" == 600 ]]
grep -q 'api_key: not-needed' "$install_dir/config/litellm/local.yaml"
grep -qx 'EXTERNAL_LLM_PROVIDER=ollama' "$install_dir/.env"
grep -qx 'SKIP_MODEL_DOWNLOAD=true' "$install_dir/.env"
grep -qx 'MODEL_RECOMMENDED_MODEL=qwen3-1.7b' "$install_dir/.env"
printf 'test-secret-123\n' >"$TEMP_DIR/operator-key"
chmod 600 "$TEMP_DIR/operator-key"
export EXTERNAL_LLM_API_KEY_FILE="$TEMP_DIR/operator-key"
source "$install_dir/installers/phases/06-directories.sh"
[[ "$(cat "$install_dir/config/litellm/external-upstream.key")" == test-secret-123 ]]
grep -q 'api_key: os.environ/EXTERNAL_LLM_API_KEY' "$install_dir/config/litellm/local.yaml"
unset EXTERNAL_LLM_API_KEY_FILE
export EXTERNAL_LLM_API_KEY_RESET=true
export EXTERNAL_LLM_URL=http://127.0.0.1:18080
export EXTERNAL_LLM_CONTAINER_URL=http://host.docker.internal:18080
source "$install_dir/installers/phases/06-directories.sh"
[[ ! -s "$install_dir/config/litellm/external-upstream.key" ]]
grep -q 'api_key: not-needed' "$install_dir/config/litellm/local.yaml"
export LLM_MODEL=qwen3-1.7b
export GGUF_FILE=Qwen3-1.7B-Q4_K_M.gguf
export MAX_CONTEXT=4096
export EXTERNAL_LLM_URL=
export EXTERNAL_LLM_CONTAINER_URL=
export EXTERNAL_LLM_PROVIDER=
export EXTERNAL_LLM_MODEL=
export EXTERNAL_LLM_RESET=true
# Turning API mode off forgets the stored key (fleet row 25).
printf 'stored-secret\n' >"$install_dir/config/litellm/external-upstream.key"
source "$install_dir/installers/phases/06-directories.sh"
[[ ! -e "$install_dir/config/litellm/external-upstream.key" ]]
grep -qx 'LLM_BACKEND=llama-server' "$install_dir/.env"
grep -qx 'LLM_MODEL=qwen3-1.7b' "$install_dir/.env"
grep -qx 'LLM_API_URL=http://llama-server:8080' "$install_dir/.env"
grep -qx 'OPEN_WEBUI_LLM_BASE_URL=' "$install_dir/.env"
grep -qx 'HERMES_LLM_BASE_URL=http://llama-server:8080/v1' "$install_dir/.env"
grep -qx 'EXTERNAL_LLM_URL=' "$install_dir/.env"
grep -qx 'EXTERNAL_LLM_PROVIDER=' "$install_dir/.env"
grep -qx 'SKIP_MODEL_DOWNLOAD=false' "$install_dir/.env"
grep -q 'api_base: http://llama-server:8080/v1' "$install_dir/config/litellm/local.yaml"
! grep -q 'host.docker.internal:11434\|openai/qwen3.5:9b' "$install_dir/config/litellm/local.yaml"
# A Windows Portal reinstall (host-native llama-server) can inherit the
# local route from an earlier CPU fallback. Recompute only that obsolete
# route: the native server is reached through LiteLLM, which holds its key.
export EXTERNAL_LLM_RESET=false
export NATIVE_LLM_BASE_URL=http://localhost:13305
export ODS_MODE=local
source "$install_dir/installers/phases/06-directories.sh"
grep -qx 'LLM_API_URL=http://litellm:4000' "$install_dir/.env"
grep -qx 'NATIVE_LLM_CONTAINER_BASE_URL=http://host.docker.internal:13305' "$install_dir/.env"
sed -i 's#^LLM_API_URL=.*#LLM_API_URL=http://llama-server:8080/v1#' "$install_dir/.env"
source "$install_dir/installers/phases/06-directories.sh"
grep -qx 'LLM_API_URL=http://litellm:4000' "$install_dir/.env"
sed -i 's#^LLM_API_URL=.*#LLM_API_URL=http://custom-litellm:4000#' "$install_dir/.env"
source "$install_dir/installers/phases/06-directories.sh"
grep -qx 'LLM_API_URL=http://custom-litellm:4000' "$install_dir/.env"
# A rerun that pauses a model API (Settings > Remote model) restores the
# LLM_API_URL the API replaced (fleet, laptop: http://litellm:4000 stayed
# and Portal chat reached LiteLLM without a key), or the mode's default.
export NATIVE_LLM_BASE_URL=
sed -i 's#^LLM_API_URL=.*#LLM_API_URL=http://litellm:4000#' "$install_dir/.env"
export ODS_REMOTE_ROUTE_PAUSED=true ODS_REMOTE_ROUTE_PREVIOUS_API_URL=http://llama-server:8080/v1
source "$install_dir/installers/phases/06-directories.sh"
grep -qx 'LLM_API_URL=http://llama-server:8080/v1' "$install_dir/.env"
sed -i 's#^LLM_API_URL=.*#LLM_API_URL=http://litellm:4000#' "$install_dir/.env"
export ODS_REMOTE_ROUTE_PREVIOUS_API_URL=
source "$install_dir/installers/phases/06-directories.sh"
grep -qx 'LLM_API_URL=http://llama-server:8080' "$install_dir/.env"
export ODS_REMOTE_ROUTE_PAUSED=false
)
if run_phase06_env_cycle; then
pass "phase 06 preserves external routing, restores managed inference, and repairs stale host-native routes"
else
fail "phase 06 external routing/reset/host-native cycle"
fi
run_phase06_amd_external() (
set -euo pipefail
local install_dir="$TEMP_DIR/phase06-amd"
mkdir -p "$install_dir"
export HOME="$install_dir/test-home"
mkdir -p "$HOME"
tar -C "$ROOT_DIR" \
--exclude='./.env' \
--exclude='./extensions/services/dashboard/node_modules' \
--exclude='./extensions/services/dashboard/dist' \
-cf - . | tar -C "$install_dir" -xf -
export INSTALL_DIR="$install_dir"
export SCRIPT_DIR="$install_dir"
export LOG_FILE="$install_dir/phase06.log"
export DRY_RUN=false
export INTERACTIVE=false
export ODS_MODE=local
export GPU_BACKEND=amd
export TIER=T1
export TIER_NAME=Entry
export LLM_MODEL=qwen3-1.7b
export GGUF_FILE=Qwen3-1.7B-Q4_K_M.gguf
export MAX_CONTEXT=4096
export ODS_VERSION=2.1.0
export ENABLE_VOICE=false
export ENABLE_WORKFLOWS=false
export ENABLE_RAG=false
export ENABLE_HERMES=false
export EXTERNAL_LLM_URL=http://127.0.0.1:11434
export EXTERNAL_LLM_CONTAINER_URL=http://host.docker.internal:11434
export EXTERNAL_LLM_PROVIDER=ollama
export EXTERNAL_LLM_MODEL=qwen3.5:9b
export NATIVE_LLM_BASE_URL=
# Same isolation as run_phase06_env_cycle: this AMD external-reuse fixture
# asserts only the .env routing contract, so remove the unrelated APE bind
# service that the fixture neither provisions nor verifies.
printf 'services:\n ape-not-a-bind-fixture:\n image: scratch\n' \
>"$install_dir/extensions/services/ape/compose.yaml"
source "$install_dir/installers/lib/constants.sh"
source "$install_dir/installers/lib/logging.sh"
source "$install_dir/installers/lib/ui.sh"
source "$install_dir/installers/lib/detection.sh"
source "$install_dir/installers/lib/progress.sh"
source "$install_dir/installers/lib/sudo.sh"
ods_progress() { :; }
ai() { :; }
ai_ok() { :; }
ai_warn() { :; }
ai_bad() { :; }
chapter() { :; }
signal() { :; }
show_phase() { :; }
sudo() { return 0; }
ods_sudo() {
case "$1" in
chmod) "$@" ;;
chgrp) command sudo -n "$@" ;;
*) return 0 ;;
esac
}
docker() {
if [[ "${1:-}" == "info" && "${2:-}" == "--format" ]]; then
printf '4\n'
return 0
fi
command docker "$@"
}
source "$install_dir/installers/phases/06-directories.sh"
grep -qx 'ODS_MODE=local' "$install_dir/.env"
grep -qx 'LLM_BACKEND=external' "$install_dir/.env"
grep -qx 'LLM_API_BASE_PATH=/v1' "$install_dir/.env"
grep -qx 'AMD_INFERENCE_RUNTIME=' "$install_dir/.env"
grep -qx 'AMD_INFERENCE_BACKEND=' "$install_dir/.env"
grep -qx 'AMD_INFERENCE_LOCATION=' "$install_dir/.env"
grep -qx 'AMD_INFERENCE_PORT=' "$install_dir/.env"
grep -qx 'AMD_INFERENCE_MANAGED=' "$install_dir/.env"
)
if run_phase06_amd_external; then
pass "AMD external reuse writes one coherent external backend contract"
else
fail "AMD external reuse .env contract"
fi
probe_retries_after_one_transport_failure() (
local curl_failure="$1" calls=0
curl() {
calls=$((calls + 1))
[[ "${*: -1}" == "http://127.0.0.1:18080/v1/chat/completions" ]] || return 99
[[ "$calls" -eq 2 ]] || return "$curl_failure"
}
sleep() { [[ "$1" == 2 ]]; }
external_llm_probe_completion 'http://127.0.0.1:18080' 'test-model' >/dev/null 2>&1 &&
[[ "$calls" -eq 2 ]]
)
assert_true "external completion probe retries one transient timeout" probe_retries_after_one_transport_failure 28
assert_true "external completion probe retries one connection failure" probe_retries_after_one_transport_failure 7
probe_fails_after_two_timeouts() (
local calls=0
curl() {
calls=$((calls + 1))
return 28
}
sleep() { [[ "$1" == 2 ]]; }
if external_llm_probe_completion 'http://127.0.0.1:18080' 'test-model' >/dev/null 2>&1; then
return 1
fi
[[ "$calls" -eq 2 ]]
)
assert_true "external completion probe stays red after bounded retries" probe_fails_after_two_timeouts
probe_does_not_retry_http_error() (
local calls=0
curl() {
calls=$((calls + 1))
return 22
}
sleep() { return 99; }
if external_llm_probe_completion 'http://127.0.0.1:18080' 'test-model' >/dev/null 2>&1; then
return 1
fi
[[ "$calls" -eq 1 ]]
)
assert_true "external completion probe does not retry a hard HTTP error" probe_does_not_retry_http_error
if python3 "$ROOT_DIR/tests/test-external-completion-probe.py"; then
pass "completion probe distinguishes reasoning budget exhaustion from invalid responses"
else
fail "external completion response validation"
fi
printf '\nResult: %d passed, %d failed\n' "$PASSED" "$FAILED"
[[ "$FAILED" -eq 0 ]]