[shivamhwp]

Refs #30236.

When Clash/Mihomo fake-IP DNS resolves a public CDN hostname to 198.18.0.0/15, HTML previews return only Chromium's generic SOCKS connection error. Add up to three capture-local warnings naming the affected hostname and address, explaining the possible DNS cause, and suggesting inline resources, fake-ip-filter, or redir-host. Each hostname gets its own bounded warning so even a maximum-length hostname retains the explanation and remedies. The user guide explains that DNS settings apply on the machine running the environment.

This implements the report's diagnostic fallback. Synthetic addresses remain blocked, including mixed public/fake DNS answers. Private-network protection and checked-address connection pinning remain intact.

Verification at final head

Candidate 2582ba9a8197e5571b5323202a7378b041fdd41e, incorporating main through 4a23e9e1bf126e858037060996f5278e2ecddaa9. The merge preserves main's Unicode-safe console clipping. The PR changes four files.

• vp run check:changed passed format, lint, both Knip checks, GraphQL validation, affected-package typechecks and focused tests.
• HtmlRender.test.ts and publicProxy.test.ts passed all 16 tests with real Chrome and no skips, including Unicode clipping and private-network/file isolation.
• Two independent Codex review rounds covered the implementation and corrections. A separate review covered the merge resolution. Both CodeRabbit findings are resolved; Macroscope approved this head.
• CI output at this head.

Exact-head command and result output

The attachments render saved output as text; complete transcripts also appear below. All three identify head 2582ba9a8197e5571b5323202a7378b041fdd41e, the environment, command and exit code.

• 30 packaged inline/TUN assertions, original shiv stdout. Retained from completed run r_vy7l2bp6v81srw00; the verifier checks the actual clean source commit and packaged identity against the full SHA before its behavior assertions.
• 16-test real-Chrome run, captured on corex. Reran on 2026-11-05 with Chrome headless shell 156.0.8078.47, Node 26.12.1, and a clean worktree at this exact head. All 16 tests passed, no skips, exit 0.
• Changed-file checks and full original output. Revalidated on 2026-11-05 using vp run check:changed 4a23e9e1bf126e858037060996f5278e2ecddaa9. The content cache confirmed the same four files match the prior passing run. The transcript includes that result and the complete original uncached stdout. Exit 0.

30 focused assertions command and output

Retained command and stdout from shiv run r_vy7l2bp6v81srw00, 2026-11-05.
This is the completed device run, not a rerun after its temporary app was removed.
Source commit: 2582ba9a8197e5571b5323202a7378b041fdd41e
Environment: physical shiv Mac mini; macOS 27.0.1 build 26A434 arm64
Build: unsigned packaged T3 Code (Beta).app 0.1.6; app.isPackaged=true
app.asar SHA-256: 727c6bf629e39a586eea84e45ff79855693c6c1882a56befba6d5953947c73e2
Command: python3 /tmp/t3code-30236-inline/assert-results.py
Exit code: 0
Verifier SHA-256: ab6ed44eb5fbc9cc0538a18741eb859b7ecd0a459051571ba024aa3ed9dc57d9
Verifier source ties the run to this head and checks the real checkout/package:
r=pathlib.Path('/tmp/t3code-30236-inline');head='2582ba9a8197e5571b5323202a7378b041fdd41e';checks=[]
check('exact source commit and clean tracked checkout',subprocess.check_output(['git','-C',str(r/'repo'),'rev-parse','HEAD'],text=True).strip()==head and not subprocess.check_output(['git','-C',str(r/'repo'),'status','--porcelain','--untracked-files=no'],text=True).strip())
 ident=j['identity'];check(label+' genuine package identity',ident['commit']==head and ident['isPackaged'] is True and ident['package']['t3codeCommitHash']==head[:12] and ident['versions']['electron']=='44.10.0' and ident['versions']['node']=='24.22.1')

Original stdout, unchanged:
PASS exact source commit and clean tracked checkout
PASS control genuine package identity
PASS control live app.asar hash
PASS control real native DNS before/after stays fake
PASS control actual MCP has no protocol/tool error
PASS control unmodified MCP text agrees with structured result
PASS control saved image equals actual returned image
PASS control image and dimensions agree
PASS inline genuine package identity
PASS inline live app.asar hash
PASS inline real native DNS before/after stays fake
PASS inline actual MCP has no protocol/tool error
PASS inline unmodified MCP text agrees with structured result
PASS inline saved image equals actual returned image
PASS inline image and dimensions agree
PASS same desktop package, backend and probe used
PASS external control uses original literal request
PASS external control fails and returns complete warning
PASS actual fetched dependency origin, bytes and hash
PASS inline payload round-trips every original dependency byte
PASS exact inline HTML fits genuine packaged tool cap
PASS inline page has no external script src and safely packs script boundaries
PASS actual library integrity, initialize call and no resource requests returned
PASS inline result has no stale fake-IP warning or SOCKS error
PASS same actual fake-IP config and active TUN before/after
PASS same real Mihomo PID and DNS listeners
PASS real TUN native fetch succeeded and packet proof exists
PASS fresh render browser processes and profiles
PASS inline TUN trace has zero packets and zero drops
PASS normal unmodified pinned render browser
ALL 30 FOCUSED ASSERTIONS PASSED.

16 real-Chrome tests command and output

Captured 2026-11-05T01:06:19.739654+00:00
Environment: corex Linux x86_64; v26.12.1
Source commit before: 2582ba9a8197e5571b5323202a7378b041fdd41e
Tracked and untracked working tree before: clean
Browser for focused Chrome run: Google Chrome for Testing 156.0.8078.47
Command: T3CODE_TEST_HEADLESS_SHELL=/home/shivam/.cache/graft/toolchains/chrome-headless-shell-156.0.8078.47/chrome-headless-shell-linux64/chrome-headless-shell TMPDIR=/tmp/t3code-30236 vp test run apps/server/src/htmlRender/HtmlRender.test.ts apps/server/src/htmlRender/publicProxy.test.ts
note: You are running `vp test` as a Vite+ built-in command. If you meant to run the test npm script, use `vpr test` instead.

 RUN  v5.0.5 /home/shivam/Developer/.state/issue-watch/worktrees/30236


 Test Files  2 passed (2)
      Tests  16 passed (16)
   Start at  01:06:28
   Duration  25.57s (tests 56%, transform 31%, setup 7%, import 5%)

Exit code: 0
Source commit after: 2582ba9a8197e5571b5323202a7378b041fdd41e

Changed-file checks command and output

Captured 2026-11-05T01:06:12.819864+00:00
Environment: corex Linux x86_64; v26.12.1
Source commit before: 2582ba9a8197e5571b5323202a7378b041fdd41e
Tracked and untracked working tree before: clean
Browser for focused Chrome run: Google Chrome for Testing 156.0.8078.47
Command: vp run check:changed 4a23e9e1bf126e858037060996f5278e2ecddaa9
$ node scripts/check-changed.ts 4a23e9e1bf126e858037060996f5278e2ecddaa9 ⊘ cache disabled
check:changed: 4 files since 4a23e9e1bf12 (4a23e9e1bf126e858037060996f5278e2ecddaa9)
Nothing changed since the last passing run.

check:changed summary
  pass  format  0.3s
  pass  lint  0.7s
  pass  knip 1/2  8.4s
  pass  knip 2/2  18.4s
  pass  graphql (apps/server)  0.8s
  pass  typecheck (3 packages)  18.9s
  pass  test (2 files)  17.4s
  CI also runs: full test suites, build
  unused patch check skipped: no dependency files changed

Exit code: 0
Source commit after: 2582ba9a8197e5571b5323202a7378b041fdd41e

Original uncached successful stdout retained with this cached validation:
$ node scripts/check-changed.ts ⊘ cache disabled
check:changed: 4 files since 4a23e9e1bf12 (origin/main)

> vp format  pass  0.3s
Checking formatting...

All matched files use the correct format.
Finished in 177ms on 4 files using 16 threads.

> vp lint  pass  0.7s

> vp knip 1/2  pass  8.4s

> vp knip 2/2  pass  18.4s

> vp graphql (apps/server)  pass  0.8s
~/apps/server$ graphqlgen --check ⊘ cache disabled

> vp typecheck (3 packages)  pass  18.9s
~/scripts$ tsc --noEmit ○ cache miss: 'packages/contracts/src/settings.ts' modified, executing

~/evals$ tsc --noEmit ○ cache miss: 'packages/provider-core/node_modules/@t3tools/shared/src/semver.ts' modified, executing

~/apps/server$ tsc --noEmit ○ cache miss: 'packages/provider-cursor/node_modules/@t3tools/provider-core/src/server/subagentProjection.ts' modified, executing

---
vp run: 0/3 cache hit (0%), 3/3 successful. @t3tools/scripts#typecheck not cached because it modified its inputs. (Run `vp run --last-details` for full details)

> vp test (2 files)  pass  17.4s

 RUN  v5.0.5 /home/shivam/Developer/.state/issue-watch/worktrees/30236


 Test Files  2 passed (2)
      Tests  5 passed (5)
   Start at  23:52:06
   Duration  13.46s (transform 64%, import 18%, setup 16%, tests 1%)

check:changed summary
  pass  format  0.3s
  pass  lint  0.7s
  pass  knip 1/2  8.4s
  pass  knip 2/2  18.4s
  pass  graphql (apps/server)  0.8s
  pass  typecheck (3 packages)  18.9s
  pass  test (2 files)  17.4s
  CI also runs: full test suites, build
  unused patch check skipped: no dependency files changed

Packaged desktop, real Mihomo TUN and both workarounds

Recorded on physical shiv on 2026-11-04, macOS 27.0.1 / 26A434 arm64, run r_0hunblvijey3q000. Both unsigned T3 Code (Beta).app builds used the repository's dist:desktop:artifact --platform mac --target dir --arch arm64 workflow. Runtime inspection confirmed app.isPackaged=true, genuine app.asar paths and embedded commits.

Build	Full source commit	app.asar SHA-256
Main baseline	dd69b316656119b4613b13f3c076f45d4dbda2ce	86a1303d11ec5f97e8c7a7bb2e04d8a5715625042efb97c962bfc61fa53a5023
Final PR head	2582ba9a8197e5571b5323202a7378b041fdd41e	7dbed9060a0afe937ec5cd5197610ee0476201cb86571b2949c5aa95977c885a

Both report version 0.1.6, embedded Node 24.22.1 and UI Chromium 152.0.7977.169. Main uses Electron 44.10.1 and candidate uses 44.10.0, as selected by their revisions. Both HTML renderers used Chrome headless shell 156.0.8078.80.

Mihomo Meta v1.19.33 ran with TUN enabled, gVisor and a real UP/RUNNING utun99. Only 198.18.0.0/16 was routed through it. A temporary macOS scoped resolver sent cdn.jsdelivr.net queries to Mihomo's TCP/UDP DNS listener at 127.0.0.1:15353. Runtime API, interface, route, native DNS and packet evidence confirm the setup. Native curl reached the same CDN URL through 198.18.0.4:443 with HTTP 200, and tcpdump observed that traffic on the TUN. Global networking and existing VPN configuration were preserved.

Every case sent the identical issue HTML, including https://cdn.jsdelivr.net/npm/mermaid@11/dist/mermaid.min.js, to the packaged server's actual HTTP MCP html_preview. A free task-local ACP probe received the session-provided connection and displayed the genuine returned image and consoleMessages in the native thread. No DNS stub, replay, source-function call or classifier patch was used. This exercises HTTP MCP and agent-message display; it does not exercise a paid model, the stdio bridge or native tool-result presentation.

Case	Condition, with TUN active	Actual result
A, main	fake-IP, no exclusion; OS DNS 198.18.0.4	Generic SOCKS error and external-script-failed; no warning.
B, final head	Same runtime and fake-IP condition	Same refusal plus the complete hostname/address/cause/remedies warning.
C, same final package	fake-IP with only the CDN hostname in fake-ip-filter	Public CDN DNS, control hostname still fake. Image says Script loaded; sole console message is external-script-loaded.
D, same final package	Separate clean redir-host runtime, empty filter	CDN and control DNS public. Image says Script loaded; sole console message is external-script-loaded.

Main and final head with fake-IP DNS and TUN enabled. Candidate starts at about 1:11. Each segment shows the native request, build/DNS identity and returned diagnostics.

before-after.mp4

Exact additional warning returned at final head:

DNS returned blocked 198.18.0.0/15 addresses, which may come from Clash or Mihomo fake-IP DNS. Inline the resource, add its host to fake-ip-filter, or switch DNS to redir-host. Affected host: cdn.jsdelivr.net (198.18.0.4)

Case C, CDN-only fake-ip-filter, with successful script loading and returned console messages:

C-filter.mp4

Case D, independent redir-host runtime with an empty filter, with successful loading and returned console messages:

D-redir.mp4

C and D used fresh render-browser processes. Packet captures observed new CDN TLS connections. Public requests followed the normal network path while Mihomo TUN stayed enabled.

The evidence assertions checked source/package identity, live bundle hashes, identical requests, agreement between unmodified MCP text and structured responses, returned image bytes, native DNS, actual TUN state, independent workaround configurations and fresh browser processes. The command exited 0:

PASS A-main commit dd69b316656119b4613b13f3c076f45d4dbda2ce
PASS B-head commit 2582ba9a8197e5571b5323202a7378b041fdd41e
PASS C-filter commit 2582ba9a8197e5571b5323202a7378b041fdd41e
PASS D-redir commit 2582ba9a8197e5571b5323202a7378b041fdd41e
ALL FOUR CASES PASSED.

The recordings are contiguous cuts of native window captures, with no overlays, speed changes or generated frames. Their decoded frames were inspected directly. All tracked task processes exited; the owned resolver, route and TUN were removed; original DNS/proxy/interface state and all 2,438 original profile file hashes were restored. Temporary builds and browser profiles were removed.

Inline dependency with fake-IP DNS and TUN unchanged

A focused follow-up on physical shiv on 2026-11-05, run r_vy7l2bp6v81srw00, rebuilt the same exact head 2582ba9a8197e5571b5323202a7378b041fdd41e as an unsigned packaged 0.1.6 app. Runtime identity confirmed app.isPackaged=true; this build's app.asar SHA-256 is 727c6bf629e39a586eea84e45ff79855693c6c1882a56befba6d5953947c73e2. OS, Electron, embedded Node and render-browser versions match the candidate above.

The original external CDN control still failed with the complete diagnostic. In the same Mihomo runtime, fake-IP mode and TUN stayed enabled, fake-ip-filter stayed empty, and native DNS remained 198.18.0.4 before and after the inline request. Native curl fetched the actual Mermaid 11.17.2 bundle through the synthetic IP and active TUN, with HTTP 200 and matching TUN packets.

The complete dependency is 3,572,661 bytes, SHA-256 581ed7d74bd9048d0e3a91363927d72ef22942d7722546b27f7cc29e35390eb8. It exceeds MCP's 512,000-character HTML cap as a plain inline script. This test therefore used lossless gzip plus Unicode packing, fitting the complete bundle into a 492,366-character self-contained HTML document. The page decompresses the original bytes, verifies their length and SHA-256 with Web Crypto, executes the actual script and calls mermaid.initialize({startOnLoad:false}). It has no external script source or substitute library.

The returned image says Inline Mermaid initialized successfully. Returned consoleMessages are exactly:

[
  {"level":"log","text":"inline-dependency-integrity: 3572661 bytes; SHA-256 581ed7d74bd9048d0e3a91363927d72ef22942d7722546b27f7cc29e35390eb8"},
  {"level":"log","text":"mermaid.initialize type: function"},
  {"level":"log","text":"inline-dependency-success: actual Mermaid initialized with startOnLoad=false"},
  {"level":"log","text":"inline-resource-requests: []"}
]

Native recording of the final inline request submission and response, including setup attempts. This uses the same real packaged HTTP MCP and ACP message presentation described above:

inline-request.mp4

Short native review of the actual control and inline responses, showing complete returned image/console output and native DNS. This reviews already returned messages:

inline-result-review.mp4

Decoded frame from that recording, showing the successful inline result and unchanged fake-IP DNS:

All 30 focused assertions passed with exit 0. They checked source/package identity, dependency byte-for-byte round-trip and browser integrity verification, actual MCP text/structured/image agreement, unchanged fake-IP/TUN state, fresh render-browser processes and profiles, successful initialization and no stale warning. The inline page recorded no resource requests; its scoped TUN trace had zero packets and zero drops. A broader process-filtered trace had drops and is not used as exhaustive evidence. All 34 captured task PIDs, including ACP probes, exited; 2,625 original profile hashes matched after restoration; task DNS/proxy changes, resolver, route, TUN, build and browser scratch were removed. Natural system AWDL address rotation was left intact.

Not shown in these native runs: full-device or public IPv6 TUN routing, other clients/providers, the stdio bridge and native tool-result presentation. The inline case verifies full-bundle integrity and initialization, not Mermaid diagram rendering or lazy-loaded modules. A plain uncompressed full-bundle inline request exceeds the tool input cap. Earlier source-engine runtime captures also verified capture isolation and complete warnings for three 253-character hostnames. Maria's independent real-model Linux web verification supplements the native HTTP MCP evidence.

Implemented with Claude Opus 5.5 through Claude Code; investigated, verified and reviewed with GPT-6.1 Sol through Codex in T3 Code.

[macroscopeapp]

Approvability

Verdict: Approved at 2582ba9

Macroscope's review found this PR approvable — HTML previews now explain blocked fake-IP DNS results and suggest remediation. Existing network restrictions remain unchanged, and diagnostics are bounded and isolated to each capture.

^{You can add or adjust custom eligibility rules. Learn more.}

[coderabbitai]

No actionable comments were generated in the recent review.

Walkthrough

The proxy tracks hostnames that resolve only to fake-IP addresses. Screenshot results include clipped console messages and warnings for recorded hosts. Documentation describes the address restrictions and possible workarounds.

Changes: Fake-IP preview diagnostics

Layer / File(s)	Summary
Track fake-IP hostnames apps/server/src/htmlRender/publicProxy.ts, apps/server/src/htmlRender/publicProxy.test.ts	The proxy records up to three hostnames that resolve only to the 198.18.0.0/15 range. Its result includes the listening port and recorded hostnames. Tests check that mixed public and fake-IP results are not recorded.
Add fake-IP warnings to screenshot results apps/server/src/htmlRender/headlessChrome.ts, docs/user/html-renders.md	Browser startup uses the proxy result object. Screenshot results include clipped page console messages and warnings for recorded fake-IP hosts. The documentation describes the restriction and listed workarounds.

Estimated code review effort: 3 (Moderate) | ~20 minutes

Change: Bug fix

Sequence Diagram(s)

sequenceDiagram
  participant publicProxy
  participant launchBrowser
  participant captureHtmlScreenshot
  publicProxy-->>launchBrowser: return port and fakeIpHosts
  launchBrowser->>launchBrowser: create fake-IP warning messages
  launchBrowser-->>captureHtmlScreenshot: return page console messages and warnings
  captureHtmlScreenshot->>captureHtmlScreenshot: combine console messages and warnings

Loading

Possibly related issues

• Issue 30236: The changes add diagnostics and workarounds for fake-IP DNS conflicts. The proxy continues to refuse connections to those synthetic addresses.

Suggested reviewers: juliusmarminge, saphid

Merge Risk: Minimal · up to 2582b

The change preserves blocking of unsafe destinations while adding bounded fake-IP diagnostics; no unresolved merge-blocking risk was identified.

Pre-merge checks: 3 · 1

Failed checks: 1 warning

Check name	Status	Explanation	Resolution
Description check		The description is detailed and covers the problem, implementation, scope, and verification. However, it does not use the required section headings and does not provide explicit maintainer approval ev…	Reorganize the description under ## Problem, ## Change, ## Scope and approval, and ## Verification. In the scope section, link the triaged issue or approval discussion and include the maintainer approval comment, or explain why this focused…

Passed checks: 3 passed

Check name	Status	Explanation
Title check		The title clearly identifies the server-side fix that explains fake-IP DNS failures in HTML previews.
Linked Issues check		Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check		Check skipped because no linked issues were found for this pull request.

Full details: Description check

Explanation

The description is detailed and covers the problem, implementation, scope, and verification. However, it does not use the required section headings and does not provide explicit maintainer approval evidence or explain an approval exemption under “Scope and approval.”

Resolution

Reorganize the description under ## Problem, ## Change, ## Scope and approval, and ## Verification. In the scope section, link the triaged issue or approval discussion and include the maintainer approval comment, or explain why this focused bug fix qualifies without prior approval.

Finishing Touches

Generate unit tests (beta)

• Create a new PR

• Autofix · Keep fixing CodeRabbit findings and required CI, and resolving merge conflicts

Recent review info

Run configuration

• Source mode: Auto
• Repository: pingdotgg/t3code
• Branch: main
• File: .coderabbit.config.ts
• Commit: 4a23e9e
• Layers: Repository TS config
• Review profile: CHILL
• Plan: Advanced
• Run ID: efa5d15d-1438-4050-acb2-1346d4c22afd

Commits

Reviewing files that changed from the base of the PR and between 41c67f7 and 2582ba9.

Files selected for processing: 1

• apps/server/src/htmlRender/headlessChrome.ts

Included review availability: This review used your included allowance. Your plan provides up to 10 included reviews per hour; 2 remain after this review.

[coderabbitai]

• Fix comments: Fix CodeRabbit comments on this PR

Dismissing prior approval to re-evaluate 41c67f7

[maria-rcks]

Note

Written by gpt-6.1-sol on behalf of Maria

Verified in a real dev environment

Verified unchanged head 41c67f7cca47f313f829cd65c8d9b9faf7fd3f8b against github/main at 7c22e24e3fa1a3bb3a95f27aea000067f198fe7f in separate checkouts, with one dev server running at a time. Installed dependencies with vp i. Used Playwright from the repo dependencies and /usr/bin/google-chrome to create a project pointing at a real Git repository and send real Codex gpt-6.1-sol turns through the web composer. Codex called the dev server's real html_preview MCP tool. No mocks, fixtures, replay provider, DNS function overrides, or proxy patches.

Mihomo v1.19.33 supplied fake-IP DNS on loopback port 15353. A systemd-resolved routing domain scoped it to cdn.jsdelivr.net; the OS and Node lookup returned 198.18.0.4. Both runs used identical HTML loading the Mermaid CDN script. Main returned only ERR_SOCKS_CONNECTION_FAILED and external-script-failed. The PR returned those errors plus the complete warning naming the hostname/address, possible DNS cause, and inline / fake-ip-filter / redir-host remedies. The resource stayed blocked; both returned screenshot PNGs were byte-identical.

Neighboring behavior passed through the same web/composer/provider/MCP path:

• A later inline script executed and returned only inline-script-loaded, with no stale warning.
• After restoring public DNS, the identical CDN page displayed external script loaded and returned an empty consoleMessages array.

Blacksmith checks passed: publicProxy.test.ts (4 tests) and HtmlRender.test.ts with real Chrome headless shell (11 tests). The existing CI reports 32 successful checks. No code changes or pushes were needed.

Linux runtime: Node 24.22.1, UI Chrome 151.0.7922.108, T3-managed capture Chrome 156.0.8078.80 with its sandbox enabled. Both dev servers used DBUS_SESSION_BUS_ADDRESS=unix:path=/tmp/pr30318-no-session-bus to avoid this VM's desktop keyring startup wait; without that environment setting, main's inline and external previews both timed out. No keyring configuration was changed. Scoped DNS configuration was restored after verification.

Runtime transcript

Not verified

Native desktop, iOS and Android clients; macOS/Windows; remote/relay/tunnel connections; other providers; Mihomo TUN routing; applying the specific fake-ip-filter or redir-host settings; maximum-length hostname diagnostics. Verification covers the shared server behavior through the Linux web client and real Codex path.

Dismissing prior approval to re-evaluate 2582ba9