fix: send chat model choices as model ids

e2e/helpers/codex-zero.bash

@@ -85,6 +85,49 @@ disable_codex_beta() {
     return 0
 }
 
+configure_codex_zero_model_policy() {
+    local model="$1"
+    local provider_type="$2"
+    local provider_id="$3"
+    local current payload
+
+    current=$(_codex_zero_curl "/api/zero/model-policies")
+    payload=$(printf '%s' "$current" | jq -c \
+        --arg model "$model" \
+        --arg providerType "$provider_type" \
+        --arg providerId "$provider_id" \
+        '
+        (.policies // [])
+        | map({
+            model,
+            isDefault,
+            defaultProviderType,
+            credentialScope,
+            modelProviderId
+          }) as $existing
+        | ($existing | map(select(.model != $model))) as $others
+        | (
+            ($existing | map(select(.model == $model)) | first)
+            // {model: $model, isDefault: false}
+          ) as $target
+        | {
+            policies: (
+              $others
+              + [
+                  $target
+                  + {
+                      defaultProviderType: $providerType,
+                      credentialScope: "org",
+                      modelProviderId: $providerId
+                    }
+                ]
+            )
+          }
+        ')
+
+    _codex_zero_curl "/api/zero/model-policies" -X PUT -d "$payload" >/dev/null
+}
+
 # Poll /api/zero/chat-threads/:id/messages until the current run has a terminal
 # assistant lifecycle marker (the paged message API no longer exposes agent run
 # status; terminal runs append a null-content lifecycle marker row instead).
@@ -189,17 +232,12 @@ send_chat_run_message() {
     # USE_MOCK_CODEX=true env var causes guest-mock-codex to echo the prompt
     # verbatim — see crates/runner/src/executor.rs (insert_codex_env) and
     # guest_mock_codex::build_events.
-    # Model-first selection can carry either the sentinel provider id for an org
-    # policy route, or a concrete provider id for an explicit model/provider pin.
-    # BYOK smoke tests use a concrete id so they do not mutate shared org policy.
     local selected_model="${CODEX_ZERO_SELECTED_MODEL:-gpt-5.5}"
-    local model_provider_id="${CODEX_ZERO_MODEL_PROVIDER_ID:-00000000-0000-4000-8000-000000000000}"
     payload=$(jq -nc \
         --arg agentId "$agent_id" \
         --arg prompt "$prompt" \
-        --arg modelProviderId "$model_provider_id" \
         --arg selectedModel "$selected_model" \
-        '{agentId: $agentId, prompt: $prompt, modelSelection: {modelProviderId: $modelProviderId, selectedModel: $selectedModel}, hasTextContent: true, realAgentInPreview: true}')
+        '{agentId: $agentId, prompt: $prompt, model: $selectedModel, hasTextContent: true, realAgentInPreview: true}')
     body=$(_codex_zero_curl "/api/zero/chat/messages" \
         -X POST \
         -d "$payload")

e2e/helpers/setup.bash

@@ -235,6 +235,30 @@ zero_model_first_selection_provider_id() {
     printf '%s' "00000000-0000-4000-8000-000000000000"
 }
 
+zero_chat_run_with_model() {
+    local agent_id="$1"
+    local prompt="$2"
+    local selected_model="$3"
+    local real_agent_in_preview="${4:-false}"
+    local payload body
+
+    payload=$(jq -nc \
+        --arg agentId "$agent_id" \
+        --arg prompt "$prompt" \
+        --arg selectedModel "$selected_model" \
+        --argjson realAgentInPreview "$real_agent_in_preview" \
+        '{agentId: $agentId, prompt: $prompt, model: $selectedModel, hasTextContent: true, realAgentInPreview: $realAgentInPreview}')
+
+    body=$(zero_curl "/api/zero/chat/messages" -X POST -d "$payload")
+    LAST_RUN_ID=$(printf '%s' "$body" | jq -r '.runId // ""')
+    LAST_THREAD_ID=$(printf '%s' "$body" | jq -r '.threadId // ""')
+    export LAST_RUN_ID LAST_THREAD_ID
+    [[ -n "$LAST_RUN_ID" && -n "$LAST_THREAD_ID" ]] || {
+        echo "# zero_chat_run_with_model: bad response: $body" >&2
+        return 1
+    }
+}
+
 zero_chat_run_with_model_selection() {
     local agent_id="$1"
     local prompt="$2"

e2e/tests/03-runner/t-codex-zero-byok-smoke.bats

@@ -31,18 +31,20 @@ setup_file() {
     # 1. Feature switch on (also fails the file early if not yet wired)
     enable_codex_beta
 
-    # 2. Org-level openai-api-key provider. The selected model is explicitly
-    # pinned to this provider on the chat message, without provider defaults or
-    # shared org policy mutation.
+    # 2. Org-level openai-api-key provider. Chat writes only carry the selected
+    # model; provider resolution comes from org model policy.
     $ZERO_CLI org model-provider setup --type "openai-api-key" --secret "$OPENAI_API_KEY" >/dev/null
     export OPENAI_PROVIDER_ID
     OPENAI_PROVIDER_ID=$(zero_model_provider_id_by_type "openai-api-key")
     export CODEX_ZERO_SELECTED_MODEL="gpt-5.4-mini"
-    export CODEX_ZERO_MODEL_PROVIDER_ID="$OPENAI_PROVIDER_ID"
+    configure_codex_zero_model_policy \
+        "$CODEX_ZERO_SELECTED_MODEL" \
+        "openai-api-key" \
+        "$OPENAI_PROVIDER_ID"
 
     # 3. Compose declares framework: codex explicitly. The framework is
-    # resolved from the explicit model-first provider pin. At secret resolution
-    # the provider's declared framework wins (Epic #11520) and is propagated
+    # resolved from the org model policy route. At secret resolution the
+    # provider's declared framework wins (Epic #11520) and is propagated
     # downstream via build-zero-context.ts's resolvedFramework, with no
     # compose-vs-provider equality check.
     cat > "$TEST_DIR/vm0-basic.yaml" <<EOF
@@ -103,10 +105,10 @@ teardown_file() {
     # survive the subshell boundary.
     wait_for_chat_assistant_done "$THREAD_ID"
 
-    # Assert: real codex produced the expected sentinel. The org's only
-    # provider is openai-api-key, so a real (non-mock) codex completion
-    # already proves the BYOK provider routing chain resolved end-to-end —
-    # without the key the run cannot produce the sentinel.
+    # Assert: real codex produced the expected sentinel. The selected model's
+    # policy routes to openai-api-key, so a real (non-mock) codex completion
+    # proves the BYOK provider routing chain resolved end-to-end — without the
+    # key the run cannot produce the sentinel.
     [[ "$LAST_MSG_CONTENT" == *"RESULT=579"* ]] \
         || fail "Expected 'RESULT=579' in assistant content, got: $LAST_MSG_CONTENT"
 }

e2e/tests/03-runner/t30-vm0-model-provider-inject.bats

@@ -2,41 +2,43 @@
 
 # Test model-provider credential injection into container environment
 #
-# Verifies that an explicit model-first provider pin is honored by zero run and
+# Verifies that an explicit runner provider type is honored by vm0 run and
 # injected into the container as CLAUDE_CODE_OAUTH_TOKEN.
 
 load '../../helpers/setup'
 
 setup() {
     export UNIQUE_ID="$(date +%s%3N)-$RANDOM"
-    export AGENT_ID=""
-    export THREAD_ID=""
+    export TEST_DIR="$(mktemp -d)"
+    export AGENT_NAME="e2e-mp-inject-${UNIQUE_ID}"
+
+    cat > "$TEST_DIR/vm0.yaml" <<EOF
+version: "1.0"
+agents:
+  ${AGENT_NAME}:
+    description: "Model provider credential injection test"
+    framework: claude-code
+EOF
+
+    $VM0_CLI compose "$TEST_DIR/vm0.yaml" >/dev/null
 }
 
 teardown() {
-    [ -n "$THREAD_ID" ] && zero_curl "/api/zero/chat-threads/$THREAD_ID" -X DELETE >/dev/null 2>&1 || true
-    [ -n "$AGENT_ID" ] && $ZERO_CLI agent delete "$AGENT_ID" -y 2>/dev/null || true
+    if [ -n "$TEST_DIR" ] && [ -d "$TEST_DIR" ]; then
+        rm -rf "$TEST_DIR"
+    fi
 }
 
 @test "model-provider credential is injected into container" {
-    local provider_id
-    provider_id=$(zero_model_provider_id_by_type "claude-code-oauth-token")
-
-    AGENT_ID=$(create_private_zero_agent "e2e-mp-inject-${UNIQUE_ID}") || {
-        echo "# Failed to create private zero agent" >&2
-        return 1
-    }
-
-    zero_chat_run_with_model_selection \
-        "$AGENT_ID" \
-        "case \"\$CLAUDE_CODE_OAUTH_TOKEN\" in \"\"|\"***\") marker=MISMATCH ;; *) marker=OK ;; esac; printf 'INJECTED_%s\n' \"\$marker\"" \
-        "$provider_id" \
-        "claude-sonnet-4-6" \
-        false
-    THREAD_ID="$LAST_THREAD_ID"
+    zero_model_provider_id_by_type "claude-code-oauth-token" >/dev/null
 
     # GitHub Actions masks sk-ant-like values in logs, so emit a non-secret marker
     # only after verifying the injected token is present inside the container.
-    WAIT_FOR_LOG_TIMEOUT=60 wait_for_log "$LAST_RUN_ID" -- "INJECTED_OK"
+    run $VM0_CLI run "$AGENT_NAME" \
+        --model-provider-type "claude-code-oauth-token" \
+        "case \"\$CLAUDE_CODE_OAUTH_TOKEN\" in \"\"|\"***\") marker=MISMATCH ;; *) marker=OK ;; esac; printf 'INJECTED_%s\n' \"\$marker\""
+
+    echo "$output"
+    assert_success
     assert_output --partial "INJECTED_OK"
 }

e2e/tests/03-runner/t54-vm0-billable-firewall.bats

@@ -2,11 +2,12 @@
 
 # Verify firewall_billable propagation through the full stack.
 #
-# Uses explicit model-first pins for provider selection. The test never changes
-# the shared e2e org's workspace default, so no race with other chunks.
+# BYOK selection uses the runner provider-type path so chat writes do not carry
+# provider pins. The vm0 case exercises the web chat model-only path and relies
+# on org model policy for provider resolution.
 #
-# t54-0: run pinned to org BYOK anthropic-api-key; "$" marker absent.
-# t54-1: run uses the model policy's vm0 route; billableFirewalls covers
+# t54-0: direct run with org BYOK anthropic-api-key; "$" marker absent.
+# t54-1: chat run uses the model policy's vm0 route; billableFirewalls covers
 #   the concrete anthropic firewall → "$" marker present.
 
 load '../../helpers/setup'
@@ -17,13 +18,26 @@ setup_file() {
     fi
 
     export UNIQUE_ID="$(date +%s%3N)-$RANDOM"
+    export TEST_DIR="$(mktemp -d)"
+    export RUN_AGENT_NAME="e2e-billable-runner-${UNIQUE_ID}"
     export THREAD_IDS=""
 
     $ZERO_CLI org model-provider setup \
         --type anthropic-api-key \
         --secret "$ANTHROPIC_API_KEY" >/dev/null
-    export ANTHROPIC_PROVIDER_ID
-    ANTHROPIC_PROVIDER_ID=$(zero_model_provider_id_by_type "anthropic-api-key")
+    zero_model_provider_id_by_type "anthropic-api-key" >/dev/null
+
+    cat > "$TEST_DIR/vm0.yaml" <<EOF
+version: "1.0"
+agents:
+  ${RUN_AGENT_NAME}:
+    description: "Billable firewall BYOK e2e"
+    framework: claude-code
+    environment:
+      ANTHROPIC_MODEL: "claude-sonnet-4-6"
+EOF
+
+    $VM0_CLI compose "$TEST_DIR/vm0.yaml" >/dev/null
 
     # Create a private zero agent for this file so CI does not consume the
     # shared org's limited public-agent slots.
@@ -46,25 +60,26 @@ teardown_file() {
         zero_curl "/api/zero/chat-threads/$thread_id" -X DELETE >/dev/null 2>&1 || true
     done
     [ -n "$AGENT_ID" ] && $ZERO_CLI agent delete "$AGENT_ID" -y 2>/dev/null || true
+    if [ -n "$TEST_DIR" ] && [ -d "$TEST_DIR" ]; then
+        rm -rf "$TEST_DIR"
+    fi
 }
 
 @test "t54-0: BYOK provider — firewall not billable" {
-    zero_chat_run_with_model_selection \
-        "$AGENT_ID" \
-        "Reply with exactly: DONE" \
-        "$ANTHROPIC_PROVIDER_ID" \
-        "claude-sonnet-4-6" \
-        true
-    THREAD_IDS="$THREAD_IDS $LAST_THREAD_ID"
-    export THREAD_IDS
+    run $VM0_CLI run "$RUN_AGENT_NAME" \
+        --model-provider-type "anthropic-api-key" \
+        --real-agent-in-preview \
+        "Reply with exactly: DONE"
 
-    RUN_ID="$LAST_RUN_ID"
+    echo "$output"
+    assert_success
+
+    RUN_ID=$(echo "$output" | grep -oP 'Run ID:\s+\K[a-f0-9-]{36}' | head -1)
     [ -n "$RUN_ID" ] || {
         echo "# Failed to extract Run ID"
         return 1
     }
 
-    wait_for_zero_run_completed "$RUN_ID"
     WAIT_FOR_LOG_TIMEOUT=60 wait_for_log "$RUN_ID" --network -- "[model-provider:anthropic-api-key]"
     refute_output --partial '[model-provider:anthropic-api-key $]'
 
@@ -73,10 +88,9 @@ teardown_file() {
 }
 
 @test "t54-1: vm0 meta-provider — firewall billable" {
-    zero_chat_run_with_model_selection \
+    zero_chat_run_with_model \
         "$AGENT_ID" \
         "Reply with exactly: DONE" \
-        "$(zero_model_first_selection_provider_id)" \
         "claude-sonnet-4-6" \
         true
     THREAD_IDS="$THREAD_IDS $LAST_THREAD_ID"

turbo/apps/api/src/signals/routes/__tests__/chat-callbacks.bdd.test.ts

@@ -48,8 +48,6 @@ function goalsClient() {
   return setupApp({ context })(zeroGoalsContract);
 }
 
-const MODEL_FIRST_SELECTION_PROVIDER_ID =
-  "00000000-0000-4000-8000-000000000000";
 const USER_ARTIFACTS_BUCKET = "test-user-artifacts";
 const CHAT_CALLBACK_PRE_CREATE_TIMING_PREFIX =
   "api_dispatch_pre_create_zero_chat_callback_";
@@ -163,13 +161,10 @@ async function startChatRun(
         ? {}
         : { attachFiles: body.attachFiles }),
       ...(body.selectedModel === undefined
-        ? { modelProvider: "anthropic-api-key" }
-        : {
-            modelSelection: {
-              modelProviderId: MODEL_FIRST_SELECTION_PROVIDER_ID,
-              selectedModel: body.selectedModel,
-            },
-          }),
+        ? body.threadId === undefined
+          ? { model: "claude-sonnet-4-6" }
+          : {}
+        : { model: body.selectedModel }),
     },
     [201],
   );
@@ -2362,10 +2357,11 @@ describe("CHAT-02: auto-send across a model switch", () => {
       prompt: "And stringify?",
     });
     const secondHeaders = await claimChatRun(runnerGroup, second.runId);
-    await chat.updateThreadModelSelection(actor, first.threadId, {
-      modelProviderId: MODEL_FIRST_SELECTION_PROVIDER_ID,
-      selectedModel: "claude-sonnet-4-6",
-    });
+    await chat.updateThreadModelSelection(
+      actor,
+      first.threadId,
+      "claude-sonnet-4-6",
+    );
     await queueChatMessage(actor, {
       agentId,
       threadId: first.threadId,

turbo/apps/api/src/signals/routes/__tests__/chat-files.bdd.test.ts

@@ -33,8 +33,6 @@ const context = testContext();
 const bdd = createBddApi(context);
 const api = createChatFilesBddApi(context);
 const authOrg = createAuthOrgAgentsBddApi(context);
-const MODEL_FIRST_SELECTION_PROVIDER_ID =
-  "00000000-0000-4000-8000-000000000000";
 
 describe("CHAT-01 chat thread lifecycle", () => {
   it("creates, mutates, searches, and deletes a thread through visible APIs", async () => {
@@ -173,10 +171,7 @@ describe("CHAT-01 chat thread lifecycle", () => {
     });
 
     await api.renameThread(owner, thread.id, "Pinned launch plan");
-    await api.updateThreadModelSelection(owner, thread.id, {
-      modelProviderId: MODEL_FIRST_SELECTION_PROVIDER_ID,
-      selectedModel: "gpt-5.4-mini",
-    });
+    await api.updateThreadModelSelection(owner, thread.id, "gpt-5.4-mini");
     await api.pinThread(owner, thread.id);
     const readEmpty = await api.markThreadRead(owner, thread.id);
 
@@ -558,10 +553,7 @@ describe("CHAT-02 chat messages and visible validation", () => {
       {
         agentId: agent.agentId,
         prompt: "Persist the model selected at send time",
-        modelSelection: {
-          modelProviderId: MODEL_FIRST_SELECTION_PROVIDER_ID,
-          selectedModel: "gpt-5.4-mini",
-        },
+        model: "gpt-5.4-mini",
       },
       [201],
     );