Evenfire is an intelligence layer platform for running LLM agents that take real actions, on your infrastructure
TypeScript
fix(minikube): refuse a Host model name without a provider and check it
Alfredo Lopez216c67709e71
diff --git a/deploy/overlays/minikube/instances/host.yaml b/deploy/overlays/minikube/instances/host.yamlindex 37da38839..ffa86f5b3 100644--- a/deploy/overlays/minikube/instances/host.yaml+++ b/deploy/overlays/minikube/instances/host.yaml@@ -8,10 +8,11 @@ spec:contextRef: context1secretRef: chatllm-api-keysmodel:- # Overwritten by full-setup.sh step 6f, which auto-detects the provider- # from .env. This literal is only the apply-time placeholder.- provider: zai- name: glm-5.1+ # Overwritten by full-setup.sh step 6f, which resolves the pair from .env+ # (scripts/minikube/host-model.sh) and checks it against llm_allowed_models.+ # This literal is only the apply-time placeholder: the no-key default.+ provider: openai+ name: gpt-5.4-miniworkflowControl:scopes:- workflow:listdiff --git a/scripts/bootstrap-cluster.sh b/scripts/bootstrap-cluster.shindex 24b21cc3c..c575c5178 100755--- a/scripts/bootstrap-cluster.sh+++ b/scripts/bootstrap-cluster.sh@@ -37,10 +37,21 @@ BOLD='\033[1m'log() { echo -e "${CYAN}[BOOTSTRAP]${NC} $*"; }ok() { echo -e "${GREEN} OK${NC} — $*"; }+warn() { echo -e "${YELLOW} WARN${NC} — $*"; }err() { echo -e "${RED} ERROR${NC} — $*"; }KC="kubectl --context=${PROFILE}"+# Resolve the Host model up front, by the same rule as full-setup.sh+# (scripts/minikube/host-model.sh), so a CLERUM_MODEL_NAME without+# CLERUM_MODEL_PROVIDER stops here instead of after the cluster is built.+# This script runs no control-api migrations, so there is no+# llm_allowed_models table to check the pair against.+# shellcheck source=scripts/minikube/host-model.sh+source "${SCRIPT_DIR}/minikube/host-model.sh"+resolve_host_model || exit 1+log "Host model: ${RESOLVED_PROVIDER}/${RESOLVED_MODEL}"+# ─── Step 1: Verify cluster ───────────────────────────────────────echo -e "\n${BOLD}═══ Step 1: Verify Cluster ═══${NC}"if ! $KC cluster-info &>/dev/null; then@@ -221,8 +232,8 @@ spec:contextRef: context1secretRef: chatllm-api-keysmodel:- provider: ${CLERUM_MODEL_PROVIDER:-zai}- name: ${CLERUM_MODEL_NAME:-glm-4.7}+ provider: ${RESOLVED_PROVIDER}+ name: ${RESOLVED_MODEL}workflowControl:scopes:- workflow:listdiff --git a/scripts/check-prereqs.sh b/scripts/check-prereqs.shindex 20b028cc9..cfef988c5 100755--- a/scripts/check-prereqs.sh+++ b/scripts/check-prereqs.sh@@ -295,11 +295,20 @@ azure:AZURE_OPENAI_API_KEY"if [ -n "$LLM_KEYS_SET" ]; thenok ".env LLM key set —$LLM_KEYS_SET"else- # Non-fatal: setup boots with test placeholders (default zai). The agent- # just can't reach a model until a real key is added — matches the- # "optional" behavior documented in docs/deploy/minikube.md.- warn ".env no LLM key set — setup will boot with placeholders (default zai); the agent can't call a model until you add one"- echo " Set one of 22 providers in .env (OPENAI_API_KEY / CLAUDE_API_KEY / GEMINI_API_KEY / GROQ_API_KEY / MISTRAL_API_KEY …); setup infers the provider. Full list: docs/deploy/llm-providers.md"+ # Non-fatal: setup boots with test placeholders (default+ # openai/gpt-5.4-mini). The agent just can't reach a model until a real+ # key is added — matches the "optional" behavior documented in+ # docs/deploy/minikube.md.+ warn ".env no LLM key set — setup will boot with placeholders (default openai/gpt-5.4-mini); the agent can't call a model until you add one"+ echo " Set one of 22 providers in .env (OPENAI_API_KEY / CLAUDE_API_KEY / GEMINI_API_KEY / GROQ_API_KEY / MISTRAL_API_KEY …). Setup infers the provider from OPENAI_API_KEY / CLAUDE_API_KEY / ZAI_API_KEY / BAILIAN_API_KEY; any other provider needs CLERUM_MODEL_PROVIDER and CLERUM_MODEL_NAME. Full list: docs/deploy/llm-providers.md"+ fi++ # CLERUM_MODEL_NAME names one provider's model; without CLERUM_MODEL_PROVIDER+ # full-setup.sh refuses it (scripts/minikube/host-model.sh) rather than pair+ # it with whichever key is present.+ if [ -n "$(env_value CLERUM_MODEL_NAME "$ENV_FILE")" ] &&+ [ -z "$(env_value CLERUM_MODEL_PROVIDER "$ENV_FILE")" ]; then+ warn ".env CLERUM_MODEL_NAME is set without CLERUM_MODEL_PROVIDER — full-setup will refuse it; set both, or remove CLERUM_MODEL_NAME"fi# Provider/key consistency — non-fatal, catches a common mismatch.diff --git a/scripts/minikube/apply-llm-secret.sh b/scripts/minikube/apply-llm-secret.shindex 7bdd07065..f7c133dec 100755--- a/scripts/minikube/apply-llm-secret.sh+++ b/scripts/minikube/apply-llm-secret.sh@@ -13,8 +13,11 @@## Backward compatible: the four original providers (openai/claude/zai/# bailian) keep their test-placeholder fallbacks so an empty .env (CI, first-# boot) still stands up the default `zai` agent. Real values override the+# boot) still stands up the default `openai` agent (openai/gpt-5.4-mini, which+# does not reply until a real key is added). Real values override the# placeholders; the other 18 providers appear only when their key is set.+# Which provider the Host uses is decided by scripts/minikube/host-model.sh,+# the same rule full-setup.sh applies; the summary line below reports it.## The Secret stores keys in the registry `dataKey` form (lowercase-hyphen,# e.g. `openai-api-key`), which is what mcp-host's LLM-secret watch expects.@@ -113,8 +116,16 @@ kubectl --context="$CONTEXT" create secret generic chatllm-api-keys \"${ARGS[@]}" \--dry-run=client -o yaml | kubectl --context="$CONTEXT" apply -f -+# The provider full-setup.sh will pick for the Host, by the shared rule: an+# explicit CLERUM_MODEL_PROVIDER, else the first of the four original keys+# present, else openai.+# shellcheck source=scripts/minikube/host-model.sh+source "${SCRIPT_DIR}/host-model.sh"+HOST_PROVIDER="$(resolve_model_provider)"+HOST_PROVIDER="${HOST_PROVIDER:-openai}"+if [ "${#REAL_KEYS[@]}" -gt 0 ]; then- echo " LLM API keys applied — real keys from .env: ${REAL_KEYS[*]} (provider: ${CLERUM_MODEL_PROVIDER:-zai})"+ echo " LLM API keys applied — real keys from .env: ${REAL_KEYS[*]} (provider: ${HOST_PROVIDER})"else- echo " LLM API keys applied — no keys in .env; using test placeholders (provider: ${CLERUM_MODEL_PROVIDER:-zai})"+ echo " LLM API keys applied — no keys in .env; using test placeholders (provider: ${HOST_PROVIDER})"fidiff --git a/scripts/minikube/full-setup.sh b/scripts/minikube/full-setup.shindex 211a8237c..18c4c5058 100755--- a/scripts/minikube/full-setup.sh+++ b/scripts/minikube/full-setup.sh@@ -310,6 +310,9 @@ fi# `make minikube-deploy-all` and `build-images.sh --verify-only` already use.# shellcheck source=scripts/minikube/image-mode.shsource "${SCRIPT_DIR}/image-mode.sh"+# Host model resolution and its llm_allowed_models check (step 6f).+# shellcheck source=scripts/minikube/host-model.sh+source "${SCRIPT_DIR}/host-model.sh"if [ "$SKIP_BUILD" = true ]; thenRECORDED_IMAGE_SOURCE="$(image_mode_source "$PROJECT_DIR")" || exit 1if [ "$RECORDED_IMAGE_SOURCE" != "$IMAGE_SOURCE" ]; then@@ -1919,55 +1922,18 @@ elselog "Skipping demo MCP servers (SEED_PROFILE=minimal) — context1 starts empty."fi-# Resolve the model provider from whichever key the user actually supplied.-# Priority order matches docs/architecture/overview.md:724 and-# .env.quickstart.example. An explicit CLERUM_MODEL_PROVIDER always wins.-# Without this, the Host pins zai while zai-api-key is a placeholder, so the-# agent never replies for anyone who set a different provider's key.-resolve_model_provider() {- if [ -n "${CLERUM_MODEL_PROVIDER:-}" ]; then- printf '%s' "${CLERUM_MODEL_PROVIDER}"- return 0- fi- if [ -n "${OPENAI_API_KEY:-}" ]; then printf 'openai'; return 0; fi- if [ -n "${CLAUDE_API_KEY:-}" ]; then printf 'claude'; return 0; fi- if [ -n "${ZAI_API_KEY:-}" ]; then printf 'zai'; return 0; fi- if [ -n "${BAILIAN_API_KEY:-}" ]; then printf 'bailian'; return 0; fi- printf ''-}--# These MUST stay in sync with the canonical registry at-# mcp-host/src/llm/registryCore.ts:50,57,64,72 — that is the single source of-# truth (config.ts:301 reads it via descriptorFor(provider).defaultModel).-# Verify against that file before committing; a drifted model id fails at the-# first message, which is the exact bug this task fixes.-default_model_for_provider() {- case "$1" in- openai) printf 'gpt-5.4-mini' ;;- claude) printf 'claude-sonnet-4-6' ;;- zai) printf 'glm-5.1' ;;- bailian) printf 'qwen3-coder-plus' ;;- *) printf '' ;;- esac-}--RESOLVED_PROVIDER="$(resolve_model_provider)"-if [ -z "$RESOLVED_PROVIDER" ]; then- # No key at all. Warn, do not abort — the platform still comes up fully and- # quickstart.md:39 already documents that the agent will not reply. This is- # deliberately softer than the ADMIN_PASSWORD precondition (Task 4).- warn "No LLM API key found in .env (OPENAI_API_KEY / CLAUDE_API_KEY / ZAI_API_KEY / BAILIAN_API_KEY)."- warn "Defaulting Host to zai with a placeholder key — the chatllm agent will NOT reply."- warn "Set a key in .env and re-run to fix."- RESOLVED_PROVIDER="zai"-fi-RESOLVED_MODEL="${CLERUM_MODEL_NAME:-$(default_model_for_provider "$RESOLVED_PROVIDER")}"-if [ -z "$RESOLVED_MODEL" ]; then- err "Unknown provider '${RESOLVED_PROVIDER}' — set CLERUM_MODEL_NAME explicitly in .env."- exit 1-fi-# 6f. Apply Host model from .env/default E2E model+# The rule lives in scripts/minikube/host-model.sh: an explicit+# CLERUM_MODEL_PROVIDER wins, CLERUM_MODEL_NAME without it is refused, and+# otherwise the provider follows the first key present (openai/gpt-5.4-mini,+# with a warning, when there is none). The pair must then be an enabled+# llm_allowed_models row in this profile's control-postgres, which the+# migrations above created; a refusal stops setup before the Host is applied.+# A no-key run still comes up fully: docs/get-started/quickstart.md documents+# that the agent does not reply until a key is added.+resolve_host_model || exit 1+log "Checking Host model ${RESOLVED_PROVIDER}/${RESOLVED_MODEL} against llm_allowed_models..."+assert_host_model_allowed "$RESOLVED_PROVIDER" "$RESOLVED_MODEL" || exit 1log "Applying Host model ${RESOLVED_PROVIDER}/${RESOLVED_MODEL} (source: ${CLERUM_MODEL_PROVIDER:+CLERUM_MODEL_PROVIDER}${CLERUM_MODEL_PROVIDER:-auto-detected from .env})..."cat <<HOSTEOF | $KC apply -f -apiVersion: clerum.io/v1alpha1diff --git a/scripts/minikube/host-model.sh b/scripts/minikube/host-model.shnew file mode 100644index 000000000..89ef40310--- /dev/null+++ b/scripts/minikube/host-model.sh@@ -0,0 +1,138 @@+#!/usr/bin/env bash+# ======================================================================+# Which model the minikube chatllm Host runs, and proof that it may+# ======================================================================+#+# Sourced library, shared by scripts/minikube/full-setup.sh (step 6f) and+# scripts/bootstrap-cluster.sh. It defines functions only: it sets no shell+# options and reports through the caller's err/warn functions.+#+# The rule (docs/llm-providers/README.md):+# - CLERUM_MODEL_PROVIDER and CLERUM_MODEL_NAME set -> used as given.+# - CLERUM_MODEL_PROVIDER only -> that provider's+# default model.+# - CLERUM_MODEL_NAME without CLERUM_MODEL_PROVIDER -> refused. The name+# belongs to one provider, and pairing it with whichever key happens to be+# present is how openai/glm-4.7 reached a Host and failed on its first+# message.+# - neither -> the provider of the first key present, in the order+# OPENAI_API_KEY, CLAUDE_API_KEY, ZAI_API_KEY, BAILIAN_API_KEY, with its+# default model. With no key at all: openai/gpt-5.4-mini, with a warning+# that the agent will not reply until a key is added.+#+# Usage (library):+# source scripts/minikube/host-model.sh+# resolve_host_model || exit 1 # sets RESOLVED_PROVIDER/MODEL+# assert_host_model_allowed "$RESOLVED_PROVIDER" "$RESOLVED_MODEL" || exit 1+# ======================================================================++# The provider named by CLERUM_MODEL_PROVIDER, else the provider of the first+# key present. Prints nothing when neither exists.+resolve_model_provider() {+ if [ -n "${CLERUM_MODEL_PROVIDER:-}" ]; then+ printf '%s' "${CLERUM_MODEL_PROVIDER}"+ return 0+ fi+ if [ -n "${OPENAI_API_KEY:-}" ]; then printf 'openai'; return 0; fi+ if [ -n "${CLAUDE_API_KEY:-}" ]; then printf 'claude'; return 0; fi+ if [ -n "${ZAI_API_KEY:-}" ]; then printf 'zai'; return 0; fi+ if [ -n "${BAILIAN_API_KEY:-}" ]; then printf 'bailian'; return 0; fi+ printf ''+}++# These MUST equal the defaultModel values in mcp-host/src/llm/registryCore.ts,+# the single source of truth mcp-host reads through+# descriptorFor(provider).defaultModel. scripts/tests/test-minikube-host-model.sh+# (H11) reads that file and fails on any drift.+default_model_for_provider() {+ case "$1" in+ openai) printf 'gpt-5.4-mini' ;;+ claude) printf 'claude-sonnet-4-6' ;;+ zai) printf 'glm-5.1' ;;+ bailian) printf 'qwen3-coder-plus' ;;+ *) printf '' ;;+ esac+}++# Sets RESOLVED_PROVIDER and RESOLVED_MODEL from the environment by the rule in+# the header. Returns 1, after err, when the pair cannot be resolved.+resolve_host_model() {+ RESOLVED_PROVIDER=""+ RESOLVED_MODEL=""+ if [ -n "${CLERUM_MODEL_NAME:-}" ] && [ -z "${CLERUM_MODEL_PROVIDER:-}" ]; then+ err "CLERUM_MODEL_NAME='${CLERUM_MODEL_NAME}' is set without CLERUM_MODEL_PROVIDER. A model name belongs to one provider: set CLERUM_MODEL_PROVIDER in .env, or remove CLERUM_MODEL_NAME to use the detected provider's default model."+ return 1+ fi++ RESOLVED_PROVIDER="$(resolve_model_provider)"+ if [ -z "$RESOLVED_PROVIDER" ]; then+ warn "No LLM API key found in .env (OPENAI_API_KEY / CLAUDE_API_KEY / ZAI_API_KEY / BAILIAN_API_KEY)."+ warn "Defaulting Host to openai with a placeholder key — the chatllm agent will NOT reply."+ warn "Set a key in .env and re-run to fix."+ RESOLVED_PROVIDER="openai"+ fi++ if [ -n "${CLERUM_MODEL_NAME:-}" ]; then+ RESOLVED_MODEL="${CLERUM_MODEL_NAME}"+ else+ RESOLVED_MODEL="$(default_model_for_provider "$RESOLVED_PROVIDER")"+ fi+ if [ -z "$RESOLVED_MODEL" ]; then+ err "Provider '${RESOLVED_PROVIDER}' has no default model here — set CLERUM_MODEL_NAME explicitly in .env."+ return 1+ fi+ return 0+}++# Proves (provider, model) is an enabled row of llm_allowed_models in the+# profile's control-postgres before the Host is applied. Needs KC (the kubectl+# command with the profile's --context). Only `enabled` is read: `stale` is a+# soft quarantine that leaves a model usable (control-api llmAllowedModels.ts).+#+# The values travel as psql variables and the SQL reads them as :'provider'+# and :'model', so neither is ever interpolated into SQL text.+#+# Refusals (each an err line and a non-zero return):+# HOST_MODEL_UNKNOWN no row for the pair+# HOST_MODEL_DISABLED the row exists with enabled = false+# HOST_MODEL_CHECK_FAILED the query could not run or answered something else+assert_host_model_allowed() {+ local provider="${1:-}" model="${2:-}" answer+ if [ -z "$provider" ] || [ -z "$model" ]; then+ err "HOST_MODEL_CHECK_FAILED: assert_host_model_allowed needs a provider and a model (got '${provider}' / '${model}')."+ return 1+ fi+ if [ -z "${KC:-}" ]; then+ err "HOST_MODEL_CHECK_FAILED: KC is not set, so the profile's control-postgres cannot be queried for ${provider}/${model}."+ return 1+ fi+ # $KC is a command plus its --context flag and must split into words.+ # shellcheck disable=SC2086+ if ! answer="$($KC exec -i -n control-plane deployment/control-postgres -- \+ psql -U postgres -d profiles -v ON_ERROR_STOP=1 \+ -v "provider=${provider}" -v "model=${model}" -Atq -f - <<'SQL'+SELECT enabled FROM llm_allowed_models+ WHERE provider = :'provider' AND model = :'model';+SQL+ )"; then+ err "HOST_MODEL_CHECK_FAILED: the llm_allowed_models query for ${provider}/${model} failed in control-postgres (see the kubectl/psql error above)."+ return 1+ fi+ case "$answer" in+ t)+ return 0+ ;;+ f)+ err "HOST_MODEL_DISABLED: ${provider}/${model} is in llm_allowed_models but disabled. Enable it in the Control UI, or set CLERUM_MODEL_PROVIDER and CLERUM_MODEL_NAME to an enabled model."+ return 1+ ;;+ '')+ err "HOST_MODEL_UNKNOWN: ${provider}/${model} is not in llm_allowed_models. Set CLERUM_MODEL_PROVIDER and CLERUM_MODEL_NAME in .env to a model that catalog enables for that provider."+ return 1+ ;;+ *)+ err "HOST_MODEL_CHECK_FAILED: unexpected answer from llm_allowed_models for ${provider}/${model}: '$(printf '%s' "$answer" | tr '\n' ' ')'."+ return 1+ ;;+ esac+}diff --git a/scripts/tests/test-minikube-host-model.sh b/scripts/tests/test-minikube-host-model.shindex c36fc1e1e..0ddb979e7 100755--- a/scripts/tests/test-minikube-host-model.sh+++ b/scripts/tests/test-minikube-host-model.sh@@ -23,8 +23,11 @@ pass() { echo "PASS: $1"; PASSED=$((PASSED + 1)); }fail() { echo "FAIL: $1"; FAIL=1; FAILED=$((FAILED + 1)); }# The library reports through the caller's err/warn, as full-setup.sh defines-# them. These print plain prefixes the cases can match.+# them. These print plain prefixes the cases can match. Only the sourced+# library calls them, which shellcheck cannot see.+# shellcheck disable=SC2329err() { printf 'ERROR -- %s\n' "$*"; }+# shellcheck disable=SC2329warn() { printf 'WARN -- %s\n' "$*"; }WORK="$(mktemp -d)"@@ -68,6 +71,8 @@ run_case() {CLERUM_MODEL_PROVIDER CLERUM_MODEL_NAME RESOLVED_PROVIDER RESOLVED_MODELexport PATH="$WORK/bin:$PATH"export KUBECTL_STUB_LOG="$stub_log" KUBECTL_STUB_MODE="$stub_mode"+ # Scoped to this case's subshell on purpose.+ # shellcheck disable=SC2030export KC="kubectl --context=fixture"while [ "$#" -gt 0 ] && [ "$1" != "--" ]; doexport "${1?}"@@ -84,7 +89,9 @@ run_case() {) </dev/null 2>&1}-# What full-setup.sh step 6f does before the Host heredoc.+# What full-setup.sh step 6f does before the Host heredoc. run_case invokes it+# inside the case's subshell, where resolve_host_model sets the two variables.+# shellcheck disable=SC2329,SC2031resolve_then_check() {resolve_host_model || return $?assert_host_model_allowed "$RESOLVED_PROVIDER" "$RESOLVED_MODEL"@@ -253,6 +260,7 @@ assert_h9_full_setup_sources_the_library_and_checks_before_the_host() {# shellcheck disable=SC2016source_line="$(grep -nF 'source "${SCRIPT_DIR}/host-model.sh"' "$FULL_SETUP" | head -n 1 | cut -d: -f1)"resolve_line="$(grep -nE '^resolve_host_model \|\| exit 1$' "$FULL_SETUP" | head -n 1 | cut -d: -f1)"+ # shellcheck disable=SC2016assert_line="$(grep -nE '^assert_host_model_allowed "\$RESOLVED_PROVIDER" "\$RESOLVED_MODEL" \|\| exit 1$' "$FULL_SETUP" | head -n 1 | cut -d: -f1)"heredoc_line="$(grep -nF 'cat <<HOSTEOF' "$FULL_SETUP" | head -n 1 | cut -d: -f1)"if [ -n "$source_line" ] && [ -n "$resolve_line" ] && [ -n "$assert_line" ] && [ -n "$heredoc_line" ] &&@@ -291,15 +299,15 @@ registry_default_model() {}assert_h11_default_models_match_the_mcp_host_registry() {- local provider registry_value library_value- for provider in openai claude zai bailian; do- registry_value="$(registry_default_model "$provider")"+ local registry_provider registry_value library_value+ for registry_provider in openai claude zai bailian; do+ registry_value="$(registry_default_model "$registry_provider")"# shellcheck source=scripts/minikube/host-model.sh- library_value="$( (source "$LIB" && default_model_for_provider "$provider") 2>/dev/null)"+ library_value="$( (source "$LIB" && default_model_for_provider "$registry_provider") 2>/dev/null)"if [ -n "$registry_value" ] && [ "$library_value" = "$registry_value" ]; then- pass "H11 default_model_for_provider $provider = registryCore.ts ($registry_value)"+ pass "H11 default_model_for_provider $registry_provider = registryCore.ts ($registry_value)"else- fail "H11 $provider: library '${library_value}' vs registryCore.ts '${registry_value}'"+ fail "H11 $registry_provider: library '${library_value}' vs registryCore.ts '${registry_value}'"fidone}