fix(sandbox): wrap fetch() to route HTTPS through EnvHttpProxyAgent

[lcsmontiel]

Summary

• Microsoft Teams channel @mentions silently fail in NemoClaw sandboxes — the bot never replies. DMs work fine. Symptom in the gateway log is msteams debounce flush failed: fetch failed followed by ECONNREFUSED against graph.microsoft.com with agent=custom.
• Root cause: OpenClaw's Teams adapter calls the Microsoft Graph API via fetch() with a custom undici dispatcher. The custom dispatcher creates its own connection pool, bypassing Node's default EnvHttpProxyAgent — which is the only thing routing HTTPS through the OpenShell L7 proxy. Direct egress is blocked by the sandbox netns.
• Follow-up to PR fix(sandbox): rewrite #2109 proxy fix as http.request wrapper (signed) NVIDIA/NemoClaw#2344 (http.request rewrite for axios requests fail with ERR_BAD_RESPONSE inside NemoClaw sandbox — double proxy conflict with NODE_USE_ENV_PROXY NVIDIA/NemoClaw#2109). fix(sandbox): rewrite #2109 proxy fix as http.request wrapper (signed) NVIDIA/NemoClaw#2344 covers axios / follow-redirects / proxy-from-env; this covers fetch() + custom dispatcher. Same preload file, same delivery path (via /usr/local/lib/nemoclaw/preloads/http-proxy-fix.js after the refactor(runtime): extract entrypoint preload modules NVIDIA/NemoClaw#3109 refactor).

Reproduction

1. Deploy NemoClaw with the Teams adapter (proxy-enabled sandbox).
2. Invite the bot to a team channel (not a DM).
3. @mention the bot.
4. Gateway log:

   msteams debounce flush failed: fetch failed
   [FETCH] https://graph.microsoft.com/v1.0/teams/.../messages/... agent=custom
   [FETCH FAIL] ... msg=fetch failed cause=none code=ECONNREFUSED
   

DMs continue to work because the webhook payload carries the message body — no Graph API call is needed.

The fix

Add a globalThis.fetch wrapper alongside the existing http.request wrapper in nemoclaw-blueprint/scripts/http-proxy-fix.js. When fetch() is called with a custom dispatcher and an HTTPS URL, strip the dispatcher and let the default EnvHttpProxyAgent handle the request through the proxy.

var origFetch = globalThis.fetch;
if (typeof origFetch === 'function') {
  var dispatcherStripWarned = false;
  globalThis.fetch = function (url, opts) {
    if (opts && opts.dispatcher) {
      var urlStr = '';
      if (typeof url === 'string') urlStr = url;
      else if (url && typeof url.href === 'string') urlStr = url.href;   // URL
      else if (url && typeof url.url === 'string')  urlStr = url.url;    // Request
      if (urlStr.startsWith('https://')) {
        if (!dispatcherStripWarned) {
          dispatcherStripWarned = true;
          console.warn('[nemoclaw http-proxy-fix] stripping custom fetch() dispatcher ...');
        }
        var newOpts = Object.assign({}, opts);
        delete newOpts.dispatcher;
        return origFetch.call(this, url, newOpts);
      }
    }
    return origFetch.apply(this, arguments);
  };
}

Non-HTTPS URLs and dispatcher-free fetch() calls pass through unchanged — direct intra-sandbox HTTP traffic and any non-proxy use of the dispatcher option keep working.

Design notes

• URL extraction handles all three fetch() input forms — string, URL object (.href), Request object (.url) — in that order. A naive url?.url would miss URL instances (which have .href, not .url) and silently leave the dispatcher attached, defeating the fix.
• One-shot console.warn so the strip is auditable in logs without spamming on every call. The Teams adapter polls Graph frequently under load — per-call logs would flood. The warned flag is closure-scoped so a process restart re-arms it.
• Defensive typeof origFetch === 'function' guard so a Node runtime that ever ships without globalThis.fetch (or a future embedding context that strips it) silently no-ops instead of throwing at preload time.
• Delivery: after refactor(runtime): extract entrypoint preload modules NVIDIA/NemoClaw#3109 the preload is shipped as a standalone module under /usr/local/lib/nemoclaw/preloads/. This PR only touches the canonical http-proxy-fix.js — no heredoc to keep in sync. The existing http-proxy-fix-sync.test.ts end-to-end test (which extracts the entrypoint block, runs it, and reads the generated /tmp/... file) is extended with explicit assertions that both the http.request wrapper and the globalThis.fetch wrapper are present in the generated preload, so a future accidental deletion of either trips CI.

Why not patch undici / replace the dispatcher with a proxy-aware one?

• Patching undici reaches into vendored bundled code; same fragility as the original Module._load hook in fix(proxy): resolve axios + NODE_USE_ENV_PROXY double-proxy conflict NVIDIA/NemoClaw#2110 that fix(sandbox): rewrite #2109 proxy fix as http.request wrapper (signed) NVIDIA/NemoClaw#2344 superseded.
• Constructing a proxy-aware replacement dispatcher is more code and risks losing caller-intent behavior (custom TLS, retry, telemetry interceptors). Stripping cleanly and letting EnvHttpProxyAgent route is the smallest correct change.
• The strip is observable via console.warn, so if a caller's dispatcher carried meaningful behavior other than direct routing, that's discoverable in logs.

Relationship to NVIDIA#2344 / NVIDIA#2296 / NVIDIA#2109

Path	Bug class	Fixed by
http.request() (axios, follow-redirects, proxy-from-env)	Library configures FORWARD-mode proxy request	NVIDIA#2344
https.request() Upgrade: websocket (Discord)	EnvHttpProxyAgent picks FORWARD instead of CONNECT for WS	NVIDIA#2296 (ws-proxy-fix.js)
fetch() + custom dispatcher (Teams Graph)	Custom dispatcher bypasses EnvHttpProxyAgent entirely	This PR

Same root cause family (HTTP code path that doesn't respect NODE_USE_ENV_PROXY), three different surface code paths, three orthogonal fixes.

Test plan

• npx vitest run --project cli test/http-proxy-fix-sync.test.ts — 1/1 pass (extracts entrypoint block, runs it, reads the generated preload; now asserts both http.request and globalThis.fetch wrappers + delete newOpts.dispatcher)
• npx vitest run --project cli test/service-env.test.ts — passes (no regression in the persist block tests)
• npm run typecheck:cli — clean
• shellcheck scripts/nemoclaw-start.sh — clean (unchanged in this PR)
• End-to-end on real proxy-enabled sandbox: Teams channel @mentions now succeed; DM regression check passes; non-Graph fetch() calls and fetch() calls without a custom dispatcher confirmed unaffected.

Closes the channel-message gap left by NVIDIA#2344.

[github-actions]

🚀 Docs preview ready!

https://NVIDIA.github.io/NemoClaw/pr-preview/pr-2/