Merge pull request #2645 from Molecule-AI/staging

staging → main: auto-promote f689c81
2026-05-03 22:54:00 +00:00 · 2026-05-03 22:54:00 +00:00 · 1141a42910
commit 1141a42910
parent bae2727074 84448d452b
5 changed files with 283 additions and 60 deletions
--- a/.github/workflows/continuous-synth-e2e.yml
+++ b/.github/workflows/continuous-synth-e2e.yml
@ -39,9 +39,14 @@ on:
  workflow_dispatch:
    inputs:
      runtime:
-        description: "Runtime to provision (langgraph = fastest, default; hermes = slower but covers SDK-native path; claude-code = needs OAUTH token in tenant env)"
+        description: "Runtime to provision (claude-code = default + cheapest via MiniMax; langgraph = OpenAI-only; hermes = SDK-native path, slower)"
        required: false
-        default: "langgraph"
+        default: "claude-code"
+        type: string
+      model_slug:
+        description: "Model id to provision the workspace with (default MiniMax-M2.7-highspeed; e.g. 'sonnet' to test direct Anthropic, 'openai/gpt-4o' for hermes)"
+        required: false
+        default: "MiniMax-M2.7-highspeed"
        type: string
      keep_org:
        description: "Skip teardown for post-mortem debugging (only manual dispatch — never set this for cron runs)"
@ -70,13 +75,23 @@ jobs:
    runs-on: ubuntu-latest
    timeout-minutes: 12
    env:
-      # langgraph default keeps cold-start under 5 min on staging EC2.
-      # hermes is slower (~7-10 min) and isn't needed for the
-      # regression class this gate exists to catch (deployment-pipeline
-      # + schema-drift + integration). Operators can pick hermes via
-      # workflow_dispatch when they need to exercise the SDK-native
-      # session path.
-      E2E_RUNTIME: ${{ github.event.inputs.runtime || 'langgraph' }}
+      # claude-code default: cold-start ~5 min (comparable to langgraph),
+      # but uses MiniMax-M2.7-highspeed via the template's third-party-
+      # Anthropic-compat path (workspace-configs-templates/claude-code-
+      # default/config.yaml:64-69). MiniMax is ~5-10x cheaper than
+      # gpt-4.1-mini per token AND avoids the recurring OpenAI quota-
+      # exhaustion class that took the canary down 2026-05-03 (#265).
+      # Operators can pick langgraph / hermes via workflow_dispatch
+      # when they specifically need to exercise the OpenAI or SDK-
+      # native paths.
+      E2E_RUNTIME: ${{ github.event.inputs.runtime || 'claude-code' }}
+      # Pin the canary to a specific MiniMax model rather than relying
+      # on the per-runtime default ("sonnet" → routes to direct
+      # Anthropic, defeats the cost saving). Operators can override
+      # via workflow_dispatch by setting a different E2E_MODEL_SLUG
+      # input if they need to exercise a specific model. M2.7-highspeed
+      # is "Token Plan only" but cheap-per-token and fast.
+      E2E_MODEL_SLUG: ${{ github.event.inputs.model_slug || 'MiniMax-M2.7-highspeed' }}
      # Bound to 10 min so a stuck provision fails the run instead of
      # holding up the next cron firing. 15-min default in the script
      # is for the on-PR full lifecycle where we have more headroom.
@ -88,19 +103,26 @@ jobs:
      E2E_KEEP_ORG: ${{ github.event.inputs.keep_org == 'true' && '1' || '' }}
      MOLECULE_CP_URL: ${{ vars.STAGING_CP_URL || 'https://staging-api.moleculesai.app' }}
      MOLECULE_ADMIN_TOKEN: ${{ secrets.CP_STAGING_ADMIN_API_TOKEN }}
-      # Provisioned tenant's default model (langgraph: openai:gpt-4.1-mini)
-      # needs OPENAI_API_KEY at first call. Sibling workflows
-      # e2e-staging-saas.yml + canary-staging.yml use the same secret;
-      # without this wire-up the tenant boots, accepts a2a messages,
-      # then returns "Could not resolve authentication method" — masked
-      # earlier by the a2a-sdk task-mode contract bugs PR #2558+#2563
-      # fixed. tests/e2e/test_staging_full_saas.sh:325 reads this and
-      # persists it as a workspace_secret on tenant create.
+      # MiniMax key is the canary's PRIMARY auth path. claude-code
+      # template's `minimax` provider routes ANTHROPIC_BASE_URL to
+      # api.minimax.io/anthropic and reads MINIMAX_API_KEY at boot.
+      # tests/e2e/test_staging_full_saas.sh branches SECRETS_JSON on
+      # which key is present — MiniMax wins when set.
+      E2E_MINIMAX_API_KEY: ${{ secrets.MOLECULE_STAGING_MINIMAX_API_KEY }}
+      # OpenAI fallback — kept wired so operators can dispatch with
+      # E2E_RUNTIME=langgraph or =hermes and still have a working
+      # canary path. The script picks the right blob shape based on
+      # which key is non-empty.
      E2E_OPENAI_API_KEY: ${{ secrets.MOLECULE_STAGING_OPENAI_KEY }}
    steps:
      - uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd # v6.0.2

-      - name: Verify required secret present
+      - name: Verify required secrets present
+        env:
+          # Re-bind so the per-runtime LLM key check below sees the right
+          # secret. The job-level env block already reads both; this just
+          # makes them visible inside the conditional shell.
+          IS_DISPATCH: ${{ github.event_name == 'workflow_dispatch' }}
        run: |
          # Schedule-vs-dispatch hardening (mirrors the sweep-cf-* and
          # redeploy-tenants-on-* workflows): hard-fail on missing secret
@ -109,7 +131,7 @@ jobs:
          # dispatch — operators can dispatch ad-hoc to verify a fix
          # without setting up the secret first.
          if [ -z "${MOLECULE_ADMIN_TOKEN:-}" ]; then
-            if [ "${{ github.event_name }}" = "workflow_dispatch" ]; then
+            if [ "$IS_DISPATCH" = "true" ]; then
              echo "::warning::CP_STAGING_ADMIN_API_TOKEN not set — synth E2E cannot run"
              echo "::warning::Set it at Settings → Secrets and Variables → Actions"
              exit 0
@ -119,6 +141,36 @@ jobs:
            exit 1
          fi

+          # LLM-key requirement is per-runtime: claude-code uses MiniMax
+          # (MOLECULE_STAGING_MINIMAX_API_KEY), langgraph + hermes use
+          # OpenAI (MOLECULE_STAGING_OPENAI_KEY). Cron firing must have
+          # the right key for the active runtime; dispatch can soft-skip.
+          case "${E2E_RUNTIME}" in
+            claude-code)
+              required_secret_name="MOLECULE_STAGING_MINIMAX_API_KEY"
+              required_secret_value="${E2E_MINIMAX_API_KEY:-}"
+              ;;
+            langgraph|hermes)
+              required_secret_name="MOLECULE_STAGING_OPENAI_KEY"
+              required_secret_value="${E2E_OPENAI_API_KEY:-}"
+              ;;
+            *)
+              echo "::warning::Unknown E2E_RUNTIME='${E2E_RUNTIME}' — skipping LLM-key check"
+              required_secret_name=""
+              required_secret_value="present"
+              ;;
+          esac
+          if [ -n "$required_secret_name" ] && [ -z "$required_secret_value" ]; then
+            if [ "$IS_DISPATCH" = "true" ]; then
+              echo "::warning::${required_secret_name} not set — synth E2E with runtime=${E2E_RUNTIME} cannot reach an LLM"
+              echo "::warning::Set it at Settings → Secrets and Variables → Actions, OR dispatch with a different runtime"
+              exit 0
+            fi
+            echo "::error::${required_secret_name} secret missing — runtime=${E2E_RUNTIME} cannot authenticate against its LLM provider"
+            echo "::error::Set it at Settings → Secrets and Variables → Actions"
+            exit 1
+          fi
+
      - name: Install required tools
        run: |
          # The script depends on jq + curl (already on ubuntu-latest)
--- a/canvas/src/components/CookieConsent.tsx
+++ b/canvas/src/components/CookieConsent.tsx
@ -98,9 +98,17 @@ export function CookieConsent() {
  };

  return (
-    <div
-      role="dialog"
-      aria-modal="true"
+    // role="region" + aria-label, NOT role="dialog" + aria-modal. The
+    // banner is informational — it never blocks the page, never traps
+    // focus, and the user can keep using the canvas while it's up.
+    // Claiming aria-modal="true" without a focus trap is genuinely
+    // harmful for screen-reader users: they get told the rest of the
+    // page is inert, jump into the banner, and then can't escape.
+    // Region semantics let assistive tech navigate around it normally.
+    // (Also: forcing a modal cookie banner would be a dark pattern —
+    // GDPR explicitly discourages it.)
+    <section
+      role="region"
      aria-labelledby="cookie-consent-title"
      aria-describedby="cookie-consent-body"
      className="fixed bottom-0 left-0 right-0 z-[9999] border-t border-line bg-surface/95 backdrop-blur-sm p-4 shadow-[0_-4px_12px_rgba(0,0,0,0.4)]"
@ -117,7 +125,7 @@ export function CookieConsent() {
            workspaces). See our{" "}
            <a
              href="https://moleculesai.app/legal/privacy"
-              className="text-accent underline hover:text-accent"
+              className="text-accent underline underline-offset-2 hover:text-accent-strong focus:outline-none focus-visible:ring-2 focus-visible:ring-accent/60 rounded-sm"
              target="_blank"
              rel="noreferrer"
            >
@ -130,20 +138,20 @@ export function CookieConsent() {
          <button
            type="button"
            onClick={() => decide("rejected")}
-            className="rounded border border-line bg-surface-sunken px-4 py-2 text-sm text-ink hover:bg-surface-card"
+            className="rounded border border-line bg-surface-sunken px-4 py-2 text-sm text-ink hover:bg-surface-card focus:outline-none focus-visible:ring-2 focus-visible:ring-accent/60 focus-visible:ring-offset-2 focus-visible:ring-offset-surface"
          >
            Necessary only
          </button>
          <button
            type="button"
            onClick={() => decide("accepted")}
-            className="rounded border border-accent bg-accent-strong px-4 py-2 text-sm font-medium text-white hover:bg-accent"
+            className="rounded border border-accent bg-accent-strong px-4 py-2 text-sm font-medium text-white hover:bg-accent focus:outline-none focus-visible:ring-2 focus-visible:ring-accent/60 focus-visible:ring-offset-2 focus-visible:ring-offset-surface"
          >
            Accept all
          </button>
        </div>
      </div>
-    </div>
+    </section>
  );
 }

--- a/canvas/src/components/tests/CookieConsent.test.tsx
+++ b/canvas/src/components/tests/CookieConsent.test.tsx
@ -40,7 +40,7 @@ afterEach(() => {
 describe("CookieConsent", () => {
  it("renders the banner when no decision is stored", () => {
    render(<CookieConsent />);
-    expect(screen.getByRole("dialog")).toBeTruthy();
+    expect(screen.getByRole("region")).toBeTruthy();
    expect(screen.getByRole("button", { name: "Accept all" })).toBeTruthy();
    expect(screen.getByRole("button", { name: "Necessary only" })).toBeTruthy();
  });
@ -48,7 +48,7 @@ describe("CookieConsent", () => {
  it("stores 'accepted' and hides the banner when user clicks Accept all", () => {
    render(<CookieConsent />);
    fireEvent.click(screen.getByRole("button", { name: "Accept all" }));
-    expect(screen.queryByRole("dialog")).toBeNull();
+    expect(screen.queryByRole("region")).toBeNull();

    const raw = window.localStorage.getItem(STORAGE_KEY);
    expect(raw).not.toBeNull();
@ -61,7 +61,7 @@ describe("CookieConsent", () => {
  it("stores 'rejected' and hides the banner when user clicks Necessary only", () => {
    render(<CookieConsent />);
    fireEvent.click(screen.getByRole("button", { name: "Necessary only" }));
-    expect(screen.queryByRole("dialog")).toBeNull();
+    expect(screen.queryByRole("region")).toBeNull();

    const parsed = JSON.parse(window.localStorage.getItem(STORAGE_KEY)!);
    expect(parsed.decision).toBe("rejected");
@ -73,7 +73,7 @@ describe("CookieConsent", () => {
      JSON.stringify({ decision: "accepted", decidedAt: new Date().toISOString(), version: 1 }),
    );
    render(<CookieConsent />);
-    expect(screen.queryByRole("dialog")).toBeNull();
+    expect(screen.queryByRole("region")).toBeNull();
  });

  it("re-prompts when the stored decision is on an older policy version", () => {
@ -82,13 +82,13 @@ describe("CookieConsent", () => {
      JSON.stringify({ decision: "accepted", decidedAt: new Date().toISOString(), version: 0 }),
    );
    render(<CookieConsent />);
-    expect(screen.getByRole("dialog")).toBeTruthy();
+    expect(screen.getByRole("region")).toBeTruthy();
  });

  it("re-prompts when localStorage contains invalid JSON", () => {
    window.localStorage.setItem(STORAGE_KEY, "{not json");
    render(<CookieConsent />);
-    expect(screen.getByRole("dialog")).toBeTruthy();
+    expect(screen.getByRole("region")).toBeTruthy();
  });

  it("exposes a privacy-policy link with target='_blank'", () => {
@ -99,11 +99,19 @@ describe("CookieConsent", () => {
    expect(link.getAttribute("rel")).toContain("noreferrer");
  });

-  it("uses role=dialog with aria-labelledby and aria-describedby for screen readers", () => {
+  it("uses role=region (NOT dialog) with aria-labelledby/describedby — banner is informational, not modal", () => {
+    // Regression guard: an earlier version claimed role="dialog"
+    // aria-modal="true" without a focus trap. That falsely told screen
+    // readers the rest of the page was inert, trapping AT users in a
+    // banner they couldn't escape. role="region" lets assistive tech
+    // navigate around it normally; the banner stays informational.
    render(<CookieConsent />);
-    const dialog = screen.getByRole("dialog");
-    expect(dialog.getAttribute("aria-labelledby")).toBe("cookie-consent-title");
-    expect(dialog.getAttribute("aria-describedby")).toBe("cookie-consent-body");
+    const banner = screen.getByRole("region");
+    expect(banner.getAttribute("aria-labelledby")).toBe("cookie-consent-title");
+    expect(banner.getAttribute("aria-describedby")).toBe("cookie-consent-body");
+    // No aria-modal claim — explicit guard against regression.
+    expect(banner.getAttribute("aria-modal")).toBeNull();
+    expect(screen.queryByRole("dialog")).toBeNull();
  });

  it("does NOT render on local dev (non-SaaS hostname)", () => {
@ -116,7 +124,7 @@ describe("CookieConsent", () => {
      value: { ...window.location, hostname: "localhost" },
    });
    render(<CookieConsent />);
-    expect(screen.queryByRole("dialog")).toBeNull();
+    expect(screen.queryByRole("region")).toBeNull();
  });

  it("does NOT render on a LAN hostname (192.168.*, *.local)", () => {
@ -125,7 +133,7 @@ describe("CookieConsent", () => {
      value: { ...window.location, hostname: "192.168.1.74" },
    });
    render(<CookieConsent />);
-    expect(screen.queryByRole("dialog")).toBeNull();
+    expect(screen.queryByRole("region")).toBeNull();
  });
 });

--- a/tests/e2e/test_secrets_dispatch.sh
+++ b/tests/e2e/test_secrets_dispatch.sh
@ -0,0 +1,145 @@
+#!/usr/bin/env bash
+# Regression test for the SECRETS_JSON branching in
+# tests/e2e/test_staging_full_saas.sh (lines ~322-368).
+#
+# The synth-E2E canary picks one of two LLM auth paths based on which
+# E2E_*_API_KEY is set. The branch order is load-bearing:
+#
+#   E2E_MINIMAX_API_KEY first   → claude-code MiniMax path (cheap canary
+#                                  default since 2026-05-03; routes via
+#                                  workspace-configs-templates/claude-
+#                                  code-default/config.yaml's `minimax`
+#                                  provider entry).
+#
+#   E2E_OPENAI_API_KEY second  → langgraph + hermes legacy path (kept
+#                                  as fallback for operator dispatches
+#                                  that need the OpenAI-shaped
+#                                  HERMES_CUSTOM_* env block).
+#
+# Without this gate, a future "tidy up the if/elif" refactor could
+# silently flip the precedence (OpenAI wins when both are set →
+# claude-code workspace boots without MINIMAX_API_KEY → 401 at first
+# turn → canary red without any signal that the wrong key shape was
+# selected). The 2026-05-03 OpenAI-quota incident took ~16h to
+# diagnose for exactly this class of "looks like an LLM problem,
+# was actually a wiring problem" failure.
+
+set -uo pipefail
+
+SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
+SAAS_SCRIPT="$SCRIPT_DIR/test_staging_full_saas.sh"
+
+if [ ! -f "$SAAS_SCRIPT" ]; then
+  echo "FATAL: cannot locate test_staging_full_saas.sh at $SAAS_SCRIPT" >&2
+  exit 2
+fi
+
+PASS=0
+FAIL=0
+
+assert_eq() {
+  local label="$1" got="$2" want="$3"
+  if [ "$got" = "$want" ]; then
+    echo "  ✓ $label"
+    PASS=$((PASS+1))
+  else
+    echo "  ✗ $label" >&2
+    echo "      got:  $got" >&2
+    echo "      want: $want" >&2
+    FAIL=$((FAIL+1))
+  fi
+}
+
+# Extract just the SECRETS_JSON block from the saas script and source
+# it into a sub-shell so we can run the branching logic in isolation.
+# Anchor on the comment header so a structural refactor that moves the
+# block fails this test loudly rather than silently sourcing nothing.
+extract_block() {
+  awk '
+    /^# ─── 5\. Provision parent workspace/ {capture=1; next}
+    capture && /^MODEL_SLUG=/ {exit}
+    capture {print}
+  ' "$SAAS_SCRIPT"
+}
+
+BLOCK=$(extract_block)
+if [ -z "$BLOCK" ]; then
+  echo "FATAL: SECRETS_JSON block not found in $SAAS_SCRIPT — refactor anchor changed?" >&2
+  exit 2
+fi
+
+# Run the extracted block in a clean env, capturing SECRETS_JSON.
+run_block() {
+  # Caller passes vars on the command line, e.g.
+  #   run_block E2E_MINIMAX_API_KEY=mx-test
+  env -i PATH="$PATH" "$@" bash -c "
+    set -uo pipefail
+    $BLOCK
+    echo \"\$SECRETS_JSON\"
+  " 2>/dev/null | tail -1
+}
+
+# Resolve a JSON key from the captured payload using python3 (already
+# a hard dep of the saas script). Returns empty string on missing key.
+get_json_key() {
+  local payload="$1" key="$2"
+  python3 -c "
+import json, sys
+p = json.loads(sys.argv[1])
+print(p.get(sys.argv[2], ''))
+" "$payload" "$key"
+}
+
+list_json_keys() {
+  python3 -c "
+import json, sys
+p = json.loads(sys.argv[1])
+print(','.join(sorted(p.keys())))
+" "$1"
+}
+
+echo "Test: SECRETS_JSON branching in test_staging_full_saas.sh"
+echo
+
+# ── Branch 1: MiniMax wins when set ──
+SECRETS_JSON=$(run_block E2E_MINIMAX_API_KEY=mx-test)
+assert_eq "MiniMax key set → MINIMAX_API_KEY in payload" \
+  "$(get_json_key "$SECRETS_JSON" MINIMAX_API_KEY)" "mx-test"
+assert_eq "MiniMax-only payload contains exactly MINIMAX_API_KEY" \
+  "$(list_json_keys "$SECRETS_JSON")" "MINIMAX_API_KEY"
+
+# ── Branch 1 precedence: MiniMax beats OpenAI when both set ──
+# Critical: the 2026-05-03 incident shape was "two paths exist, wrong
+# one wins". The bash if/elif must keep MiniMax above OpenAI so the
+# claude-code default canary doesn't accidentally use the (more
+# expensive, quota-burnt) OpenAI key.
+SECRETS_JSON=$(run_block E2E_MINIMAX_API_KEY=mx-priority E2E_OPENAI_API_KEY=oai-loser)
+assert_eq "Both keys set → MiniMax wins" \
+  "$(get_json_key "$SECRETS_JSON" MINIMAX_API_KEY)" "mx-priority"
+assert_eq "Both keys set → OpenAI block NOT emitted" \
+  "$(get_json_key "$SECRETS_JSON" OPENAI_API_KEY)" ""
+assert_eq "Both keys set → no HERMES_* leakage from OpenAI branch" \
+  "$(get_json_key "$SECRETS_JSON" HERMES_INFERENCE_PROVIDER)" ""
+
+# ── Branch 2: OpenAI used when MiniMax absent ──
+SECRETS_JSON=$(run_block E2E_OPENAI_API_KEY=oai-test)
+assert_eq "Only OpenAI set → OPENAI_API_KEY in payload" \
+  "$(get_json_key "$SECRETS_JSON" OPENAI_API_KEY)" "oai-test"
+assert_eq "Only OpenAI set → HERMES_CUSTOM_API_KEY mirrors OpenAI key" \
+  "$(get_json_key "$SECRETS_JSON" HERMES_CUSTOM_API_KEY)" "oai-test"
+assert_eq "Only OpenAI set → MODEL_PROVIDER pinned to colon-form" \
+  "$(get_json_key "$SECRETS_JSON" MODEL_PROVIDER)" "openai:gpt-4o"
+assert_eq "Only OpenAI set → MINIMAX_API_KEY NOT emitted" \
+  "$(get_json_key "$SECRETS_JSON" MINIMAX_API_KEY)" ""
+
+# ── No keys: empty payload ──
+SECRETS_JSON=$(run_block)
+assert_eq "No keys set → SECRETS_JSON is empty object" \
+  "$SECRETS_JSON" "{}"
+
+echo
+echo "─────────────────────────────────────────────────"
+echo "PASSED: $PASS"
+echo "FAILED: $FAIL"
+echo "─────────────────────────────────────────────────"
+[ "$FAIL" -eq 0 ]
--- a/tests/e2e/test_staging_full_saas.sh
+++ b/tests/e2e/test_staging_full_saas.sh
@ -320,29 +320,39 @@ tenant_call() {
 }

 # ─── 5. Provision parent workspace ─────────────────────────────────────
-# Runtimes like hermes crash at boot with "No provider API key found"
-# if nothing in the standard env-var list is set. Inject the API key
-# from E2E_OPENAI_API_KEY so the runtime can actually start — it's
-# per-workspace secret, so it's persisted as a workspace_secret and
-# materialized into the container env. Missing key falls through to
-# an empty secrets map; workspace will still fail but the error is
-# expected and actionable.
+# Inject the LLM provider key so the runtime can authenticate at boot.
+# Branch by which secret is set so the script supports both paths
+# without forcing every dispatch to ship both keys:
+#
+#   E2E_MINIMAX_API_KEY → claude-code MiniMax path. Cheapest, default
+#     for the cron canary post-2026-05-03. Routes via the claude-code
+#     template's `minimax` provider (workspace-configs-templates/
+#     claude-code-default/config.yaml:64-69) which sets
+#     ANTHROPIC_BASE_URL=https://api.minimax.io/anthropic at boot.
+#     MINIMAX_API_KEY is the vendor-specific env name the adapter
+#     reads (PR #244 — per-vendor envs prevent ANTHROPIC_AUTH_TOKEN
+#     collisions when a user runs MiniMax + Z.ai workspaces side-by-
+#     side).
+#
+#   E2E_OPENAI_API_KEY → langgraph + hermes paths. Kept as fallback
+#     for operator dispatches that explicitly want to exercise the
+#     OpenAI path. The HERMES_* fields pin hermes-agent's bridge to
+#     api.openai.com (template-hermes' derive-provider.sh otherwise
+#     resolves openai/* → openrouter.ai and 401s). MODEL_PROVIDER
+#     follows workspace/config.py:258's 'provider:model' format.
+#
+# Both empty → '{}' (workspace will fail at first turn with an
+# expected, actionable auth error rather than masking the test).
 SECRETS_JSON='{}'
-if [ -n "${E2E_OPENAI_API_KEY:-}" ]; then
-  # MODEL_PROVIDER is a full model slug in 'provider:model' format per
-  # workspace/config.py:258. Using just "openai" gets parsed as the
-  # model name → 404 model_not_found. Also set OPENAI_BASE_URL to
-  # OpenAI's own endpoint — default is openrouter.ai which would need
-  # a different key format.
-  #
-  # The HERMES_* fields below bypass template-hermes/scripts/derive-provider.sh
-  # — verified 2026-04-24 that even with template-hermes#19's fix in main,
-  # staging tenants sometimes resolve openai/* to PROVIDER=openrouter and
-  # emit {'message':'Missing Authentication header','code':401} (OpenRouter's
-  # shape) in the A2A reply. Setting HERMES_INFERENCE_PROVIDER=custom +
-  # HERMES_CUSTOM_{BASE_URL,API_KEY,API_MODE} pins the bridge deterministically
-  # so the test doesn't depend on every tenant EC2 having a freshly-cloned
-  # template-hermes.
+if [ -n "${E2E_MINIMAX_API_KEY:-}" ]; then
+  SECRETS_JSON=$(python3 -c "
+import json, os
+k = os.environ['E2E_MINIMAX_API_KEY']
+print(json.dumps({
+    'MINIMAX_API_KEY': k,
+}))
+")
+elif [ -n "${E2E_OPENAI_API_KEY:-}" ]; then
  SECRETS_JSON=$(python3 -c "
 import json, os
 k = os.environ['E2E_OPENAI_API_KEY']