Tunnel Doctor

Tunnel Doctor

Diagnose and fix conflicts when Tailscale coexists with proxy/VPN tools on macOS, with specific guidance for SSH access to WSL instances.

Methodology base: the general diagnostic discipline this skill builds on — evidence over assumption, falsification over confirmation, layered isolation, counter-review — lives in the debugging-network-issues skill. This skill is the macOS Tailscale⨯proxy domain layer on top of it; reach for the base skill when the symptom is not a known Tailscale/proxy conflict.

Ownership boundary: this skill owns concrete local network-path operations: Tailscale, route/DNS/proxy interception, VM/WSL propagation, TUN forwarding planes, and proxy node/exit/chain capacity. Use debugging-network-issues as the general method when the boundary is still unknown or the symptom belongs to SSE/CDN/application protocol behavior. This keeps one concrete operator instead of forcing the user to choose between two overlapping proxy doctors.

Conflict Layers

Proxy/VPN tools on macOS create conflicts at several independent layers. Layers 1-3 affect Tailscale connectivity; Layer 4 affects SSH git operations; Layer 5 affects VM/container runtimes. TUN-state failure modes beyond this table — SSH/git connection drops, resolver stall, DIRECT split-brain — are covered in Steps 2H–2J:

Diagnostic Workflow

Step 1: Identify the Symptom

Determine which scenario applies:

• Browser returns HTTP 503, but curl and SSH both work → System proxy bypass conflict (Step 2C)
• local.<domain> fails in browser/default curl, but direct/no-proxy request works → Local vanity domain proxy interception (Step 2C-1)
• Tailscale ping works, SSH works, but curl/HTTP times out → HTTP proxy env var conflict (Step 2A)
• Tailscale ping works, SSH/TCP times out → Route conflict (Step 2B)
• Remote dev server auth redirects to localhost → browser can't follow → SSH tunnel needed (Step 2D)
• make status / scripts curl to localhost fail with proxy → localhost proxy interception (Step 2E)
• git push/pull fails with FATAL: failed to begin relaying via HTTP → SSH double tunnel (Step 2F)
• docker build RUN apk/apt fails with Connection refused instantly → OrbStack transparent proxy + TUN conflict (Step 2G-1, fix: --network host)
• docker pull fails with TLS handshake timeout → VM proxy misconfiguration (Step 2G-2, fix: docker.json with host.internal)
• Container healthcheck (unhealthy) but app runs fine → Lowercase proxy env var leak (Step 2G-4, fix: clear http_proxy+HTTP_PROXY)
• docker build can't fetch base images → VM/container proxy propagation (Step 2G)
• docker pull works but docker build still dies fetching base-image tokens (TLS/i-o timeout) on a WSL2 + Docker Desktop host, or every runbook proxy port on a Windows+v2rayN host silently died after a v2rayN upgrade → buildkit direct-dials past the DD proxy override / v2rayN ≥7.17 LAN-port split (references/windows_host_tun_wsl_cascade.md §2026-08-23)
• git clone fails with Connection closed by 198.18.x.x → two different mechanisms produce this; Step 2H separates them
• Every domestic/DIRECT-rule site fails at once (TLS unexpected EOF mid-handshake, proxy-port CONNECT returns 503, Node CLIs report UNKNOWN_CERTIFICATE_VERIFICATION_ERROR) while proxied overseas sites keep working → TUN DIRECT split-brain (Step 2J)
• SSH connects but operation not permitted → Tailscale SSH config issue (Step 4)
• SSH connects but be-child ssh exits code 1 → WSL snap sandbox issue (Step 5)
• TCP port 22 reachable (nc -z succeeds) but SSH fails with kex_exchange_identification: Connection closed → Tailscale SSH proxy intercept on WSL (Step 5A)
• Same kex_exchange_identification / Connection closed by UNKNOWN port 65535, but there is no Tailscale — you are behind a TUN and the host is a public cloud VM whose HTTPS still works → could be your TUN or the destination's firewall allowlist; they are indistinguishable from the client. Step 2K settles it out-of-band
• tailscale ssh returns "not available on App Store builds" → Wrong Tailscale distribution on macOS (Step 5B)
• Any tool using system DNS (ssh, curl, git) hangs ~60s before resolving, but nslookup returns instantly → Stalled resolver in getaddrinfo chain (Step 2I)
• Windows+WSL host: everything offline at once (domestic AND overseas), WSL dead too, "even Tailscale won't come online" — and/or the user tried several recovery actions (switched NICs, toggled TUN) and can't tell which one fixed it → Windows host TUN cascade + event-log forensics (Step 5C)
• Tailscale, DNS, route ownership, and proxy reachability all pass, but large transfers remain slow and switching the active proxy node changes throughput → proxy node / exit / chain capacity workflow in references/proxy_node_chain_throughput.md

Key distinctions:

• SSH does NOT use http_proxy/NO_PROXY env vars. If SSH works but HTTP doesn't → Layer 2.
• curl uses http_proxy env var, NOT the system proxy. Browser uses system proxy (set by VPN). If curl works but browser doesn't → Layer 3.
• If tailscale ping works but regular ping doesn't → Layer 1 (route table corrupted).
• If ssh -T git@github.com works but git push fails intermittently → Layer 4 (double tunnel).
• If host curl https://... works but docker pull times out → Layer 5 (VM proxy propagation).
• If docker pull works but docker build RUN apk add fails instantly with Connection refused → OrbStack transparent proxy broken by TUN (Step 2G-1).
• If container healthcheck shows (unhealthy) but app works → lowercase http_proxy leaked into container (Step 2G-4).
• If DNS resolves to 198.18.x.x virtual IPs → a TUN is active, which is not by itself a diagnosis: under TUN every hostname resolves this way, on success as well as failure. Go to Step 2H to find out which mechanism you have.
• If connections to fake-IP-resolved domains die but the same host works via curl --resolve <host>:443:<real-ip> (real IP from dig @<public-resolver>) → the TUN tool's DIRECT forwarding state is broken, not your network and not the destination (Step 2J).
• If nc -z succeeds on port 22 but SSH gets no banner (kex_exchange_identification) → Tailscale SSH proxy intercept (Step 5A). Confirm with tcpdump -i any port 22 on the remote — 0 packets means Tailscale intercepts above the kernel.
• If nc -z also succeeds on a port the destination does not serve (probe 65001) → under a TUN the local stack answers for the destination, so no nc -z result on that host means anything. Never read a port probe you have not calibrated this way. This supersedes the plain nc -z reading in the bullet above whenever a TUN — not just Tailscale — is in the path (Step 2K).
• If the destination's own journalctl -u ssh shows zero entries for your attempts → something before sshd is eating them, but not necessarily the network: a host-local packet filter drops them just as silently. tcpdump on the host separates "never arrived" from "arrived and was dropped locally"; only the first implicates your TUN. Reach the machine through the cloud provider's guest-agent channel to ask, since that path does not use the port you cannot reach. Check the journal, not auth.log: on Ubuntu 24.04 sshd writes nothing to auth.log, so an empty file there is not evidence (Step 2K).
• A cloud host enforces an inbound port twice — provider security group and host firewall (ufw/firewalld/nftables). Adding your IP to one leaves the other blocking, with no signal that you only did half the job (Step 2K).
• If tailscale ssh fails with "not available on App Store builds" → install Standalone Tailscale (Step 5B).
• If nslookup <host> is fast (<0.1s) but dscacheutil -q host -a name <host> takes 60s+ → a supplemental resolver in scutil --dns is dead (Step 2I).
• If ping <resolver-ip> succeeds but dig @<resolver-ip> times out → daemon dead, utun interface zombied. ICMP is answered by the interface; the actual port-53 service is gone (Step 2I).
• If ssh -vvv hangs immediately after debug2: resolving "<host>" port <port> and never reaches debug1: connect to address → DNS resolution stage, not network connect stage. This is Step 2I, not Step 2B/2H.

Diagnosis Discipline (Read Before Committing to a Hypothesis)

When symptoms point at a component (proxy, VPN, route table, DNS), don't commit to a hypothesis from circumstantial evidence — verify with that component's own health endpoint first. Each component has a one-line health check faster and more reliable than ruling out neighbors:

Why this matters: A symptom that matches the description of Step 2X does not, by itself, prove component X is the problem. Multiple layers can produce overlapping symptoms (a 60-second hang during git push could be proxy node death, fakeip route corruption, or DNS resolver stall — all plausible from the user-visible symptom alone). Reaching for the most specific verification first avoids committing to a wrong layer and chasing it down a dead end.

If the failing operation involves DNS at all, run the per-nameserver bisection from Step 2I before suspecting proxy or routing. It rules in/out the largest single class of macOS-on-China-network failures in under 15 seconds.

TUN Measurement Contamination (what your probes lie about while a TUN proxy is up)

When a proxy tool runs in TUN / global mode (Shadowrocket, Clash, Surge), it intercepts traffic at the routing layer and fabricates parts of the network stack locally. Several everyday diagnostic commands then return fabricated or misrouted numbers — trusting them sends the whole investigation the wrong way. Know what each probe actually measures under TUN:

What you can trust under TUN:

• time_appconnect / time_starttransfer from curl (application-layer handshake / TTFB) — these complete only after the tunneled connection actually establishes, so they reflect the real end-to-end path.
• An in-region / domestic IP-geo source for "what is my real local ISP" — an in-region domain hits the proxy's DIRECT rule and exits your real last mile (the foreign source gets tunneled and lies; see table).
• The proxy/TUN config decoded from disk + the tool's own GUI — the authoritative source of which node/route is actually active. Cross-check a file parse against the GUI; do not infer the active node from a network probe.

Counter-move: before citing any latency / reachability number while a TUN is up, ask "would this number be physically possible if the packet really traversed to the destination?" A 0.00s connect or a 0.2ms ping to another continent is the tell that you measured the TUN, not the network. Switch to time_appconnect, or temporarily disable the TUN to get a clean baseline (raw probes become meaningful again once it is off).

Fast Path: Run Automated Checks

For common macOS conflicts (env proxy, system proxy exceptions, direct/proxy path split, local TLS trust), run:

python3 scripts/quick_diagnose.py --host local.example.com --url https://local.example.com/health

Optional route ownership check for a Tailscale destination:

python3 scripts/quick_diagnose.py --host <target-host> --url http://<target-host>:<port>/health --tailscale-ip <100.x.x.x>

Interpretation:

• direct=PASS + forced_proxy=FAIL = host must bypass proxy (skip-proxy + NO_PROXY).
• strict_tls=FAIL + direct=PASS = path is reachable; trust issue only (install/trust local CA).
• host in scutil exceptions: no = browser/system clients still likely proxied.

Step 2A: Fix HTTP Proxy Environment Variables

Check if proxy env vars are intercepting Tailscale HTTP traffic:

env | grep -i proxy

Broken output — proxy is set but NO_PROXY doesn't exclude Tailscale:

http_proxy=http://127.0.0.1:1082
https_proxy=http://127.0.0.1:1082
NO_PROXY=localhost,127.0.0.1          ← Missing Tailscale!

Fix — add Tailscale MagicDNS domain + CIDR to NO_PROXY:

export NO_PROXY=localhost,127.0.0.1,.ts.net,100.64.0.0/10,192.168.*,10.*,172.16.*

Two layers complement each other: .ts.net handles domain-based access, 100.64.0.0/10 handles direct IP access.

NO_PROXY syntax pitfalls — see references/proxy_conflict_reference.md for the compatibility matrix.

Go net/http CIDR caveat: Go's standard net/http does NOT support CIDR notation in NO_PROXY. Setting NO_PROXY=100.64.0.0/10 works for curl and Python, but Go programs (including Tailscale-adjacent tooling) will still send traffic through the proxy. The fix is to use MagicDNS hostnames (e.g., my-wsl-box) instead of raw IPs, or add explicit hostnames to NO_PROXY:

# WRONG for Go programs — CIDR is silently ignored
NO_PROXY=100.64.0.0/10 go-program http://100.101.102.103:8002/health  # → goes through proxy

# CORRECT — use hostname (matched as suffix) or explicit IP
export NO_PROXY=localhost,127.0.0.1,.ts.net,my-wsl-box,100.101.102.103,192.168.*,10.*,172.16.*

This is especially relevant when accessing Tailscale services from Go-based tools (e.g., custom CLIs, Go test suites hitting remote APIs).

Verify the fix:

# Both must return HTTP 200:
NO_PROXY="...(new value)..." curl -s --connect-timeout 5 http://<host>.ts.net:<port>/health -w "HTTP %{http_code}\n"
NO_PROXY="...(new value)..." curl -s --connect-timeout 5 http://<tailscale-ip>:<port>/health -w "HTTP %{http_code}\n"

Then persist in shell config (~/.zshrc or ~/.bashrc).

Step 2B: Detect Route Conflicts

Check if a proxy tool hijacked the Tailscale CGNAT range:

route -n get <tailscale-ip>

Healthy output — traffic goes through Tailscale interface:

destination: 100.64.0.0
interface: utun7    # Tailscale interface (utunN varies)

Broken output — proxy hijacked the route:

destination: 100.64.0.0
gateway: 192.168.x.1    # Default gateway
interface: en0           # Physical interface, NOT Tailscale

Important: Not all utun interfaces are Tailscale's. Verify which utun belongs to Tailscale before concluding the route is correct:

# Find Tailscale's utun interface (has a 100.x.x.x IP)
ifconfig | grep -A2 'inet 100\.'

Quick indicators by MTU:

• MTU 1280 → typically Tailscale
• MTU 4064 → typically Shadowrocket TUN

If route -n get shows traffic going to a utun with MTU 4064, it is hitting Shadowrocket's TUN, not Tailscale — this is still a route conflict even though the interface name starts with utun.

Confirm with full route table:

netstat -rn | grep 100.64

Two competing routes indicate a conflict:

100.64/10  192.168.x.1   UGSc  en0       ← Proxy added this (wins)
100.64/10  link#N        UCSI  utun7     ← Tailscale route (loses)

Root cause: On macOS, UGSc (Static Gateway) takes priority over UCSI (Cloned Static Interface) for the same prefix length.

Step 2C: Fix System Proxy Bypass (Browser 503)

Symptom: Browser shows HTTP 503 for http://<tailscale-ip>:<port>, but both curl --noproxy '*' and curl (with proxy env var) return 200. SSH also works.

Root cause: The browser uses the system proxy configured by the VPN profile (Shadowrocket/Clash/Surge). The proxy matches IP-CIDR,100.64.0.0/10,DIRECT and tries to connect directly — but "directly" means via the Wi-Fi interface (en0), NOT through Tailscale's utun interface. The proxy process itself doesn't have a route to Tailscale IPs, so the connection fails with 503.

Diagnosis:

# curl with proxy env var works (curl connects to proxy port, but traffic flows differently)
curl -s -o /dev/null -w "%{http_code}" http://<tailscale-ip>:<port>/
# → 200

# Browser gets 503 because it goes through the VPN system proxy, not http_proxy env var

Fix — add Tailscale CGNAT range to skip-proxy in the proxy tool config:

For Shadowrocket, in [General]:

skip-proxy = 192.168.0.0/16, 10.0.0.0/8, 172.16.0.0/12, 100.64.0.0/10, localhost, *.local, captive.apple.com

skip-proxy tells the system "bypass the proxy entirely for these addresses." The browser then connects directly through the OS network stack, where Tailscale's routing table correctly handles the traffic.

Why skip-proxy works but tun-excluded-routes doesn't:

• skip-proxy: Bypasses the HTTP proxy layer only. Traffic still flows through the TUN interface and Tailscale utun handles it. Safe.
• tun-excluded-routes: Removes the CIDR from the TUN routing entirely. This creates a competing en0 route that overrides Tailscale. Breaks everything.

Step 2C-1: Fix Local Vanity Domain Interception (local.<domain>)

Symptom: https://local.<domain> fails in browser or default curl, but succeeds with direct/no-proxy command:

env -u http_proxy -u https_proxy curl -k -I https://local.<domain>/health
# -> 200
curl -I https://local.<domain>/health
# -> proxy CONNECT then TLS reset/failure

Root cause: The domain is routed through system/shell proxy instead of local direct path.

Fix:

1. Add domain to proxy app bypass list (skip-proxy for Shadowrocket).
2. Add domain to shell bypass list (NO_PROXY/no_proxy).
3. If local TLS uses internal CA, trust the local root certificate.

# ~/.zshrc
export NO_PROXY=localhost,127.0.0.1,.ts.net,100.64.0.0/10,192.168.*,10.*,172.16.*,local.<domain>,www.local.<domain>
export no_proxy="$NO_PROXY"

Verification:

python3 scripts/quick_diagnose.py --host local.<domain> --url https://local.<domain>/health

Expected:

• host in NO_PROXY: yes
• host in scutil exceptions: yes
• ambient=PASS and direct=PASS

Step 2D: Fix Auth Redirect for Remote Dev (SSH Tunnel)

Symptom: Dev server runs on a remote machine (e.g., Mac Mini via Tailscale). You access http://<tailscale-ip>:3010 in the browser. Login/signup works, but after auth, the app redirects to http://localhost:3010/ which fails — localhost on your machine isn't running the dev server.

Root cause: The app's APP_URL (or equivalent) is set to http://localhost:3010. Auth libraries (Better-Auth, NextAuth, etc.) use this URL for callback redirects. Changing APP_URL to the Tailscale IP introduces Shadowrocket proxy conflicts and breaks local development on the remote machine.

Fix — SSH local port forwarding. This avoids all three conflict layers entirely:

# Forward local port 3010 to remote machine's localhost:3010
ssh -NL 3010:localhost:3010 <tailscale-ip>

# Or with autossh for auto-reconnect (recommended for long sessions)
autossh -M 0 -f -N -L 3010:localhost:3010 \
    -o "ServerAliveInterval=30" \
    -o "ServerAliveCountMax=3" \
    -o "ExitOnForwardFailure=yes" \
    <tailscale-ip>

Now access http://localhost:3010 in the browser. Auth redirects to localhost:3010 → tunnel → remote dev server → works correctly.

Why this is the best approach:

• No .env changes needed — APP_URL=http://localhost:3010 works everywhere
• No Shadowrocket conflicts — localhost is always in skip-proxy
• No code changes — same behavior as local development
• Industry standard — VS Code Remote SSH, GitHub Codespaces use the same pattern

Install autossh: brew install autossh (macOS) or apt install autossh (Linux)

Kill background tunnel: pkill -f 'autossh.*<tailscale-ip>'

Step 2E: Fix localhost Proxy Interception in Scripts

Symptom: Makefile targets or scripts that curl localhost (health checks, warmup routes) fail or timeout when http_proxy is set globally in the shell.

Root cause: http_proxy=http://127.0.0.1:1082 is set in ~/.zshrc but no_proxy doesn't include localhost. All curl commands send localhost requests through the proxy.

Fix — add --noproxy localhost to all localhost curl commands in scripts:

# WRONG — fails when http_proxy is set
@curl -sf http://localhost:9000/minio/health/live && echo "OK"

# CORRECT — always bypasses proxy for localhost
@curl --noproxy localhost -sf http://localhost:9000/minio/health/live && echo "OK"

Alternatively, set no_proxy globally in ~/.zshrc:

export no_proxy=localhost,127.0.0.1

Step 2F: Fix SSH ProxyCommand Double Tunnel (git push/pull failures)

Symptom: ssh -T git@github.com succeeds consistently, but git push or git pull fails intermittently with:

FATAL: failed to begin relaying via HTTP.
Connection closed by UNKNOWN port 65535

Small operations (auth, fetch metadata) work; large data transfers fail.

Root cause: When Shadowrocket TUN is active, it already routes all TCP traffic through its VPN tunnel. If SSH config also uses ProxyCommand connect -H, data flows through two proxy layers — the landing proxy drops large/long-lived HTTP CONNECT connections.

Diagnosis:

# 1. Confirm Shadowrocket TUN is active
ifconfig | grep '^utun'

# 2. Check SSH config for ProxyCommand
grep -A5 'Host github.com' ~/.ssh/config

# 3. Confirm: removing ProxyCommand fixes push
GIT_SSH_COMMAND="ssh -o ProxyCommand=none" git push origin main

Fix — remove ProxyCommand and switch to ssh.github.com:443. See references/proxy_conflict_reference.md § SSH ProxyCommand and Git Operations for the full SSH config, why port 443 helps, and fallback options when VPN is off.

Step 2G: Fix VM/Container Runtime Proxy Propagation (Docker pull/build failures)

Symptom: docker pull or docker build fails with net/http: TLS handshake timeout, Connection refused from Alpine/Debian repos, or Internal Server Error from auth.docker.io, while host curl to the same URLs works fine.

Applies to: OrbStack, Docker Desktop, or any VM-based Docker runtime on macOS with Shadowrocket/Clash TUN active.

Root cause: VM-based Docker runtimes (OrbStack, Docker Desktop) run the Docker daemon inside a lightweight VM. The VM's outbound traffic takes a different network path than host processes:

Host process (curl):   Process → TUN (Shadowrocket) → landing proxy → internet ✅
VM process (Docker):   Docker daemon → VM bridge → host network → TUN → ??? ❌

The TUN handles host-originated traffic correctly but may drop or delay VM-bridged traffic (different TCP stack, MTU, keepalive behavior).

Critical distinction: docker pull vs docker build use different proxy paths:

Fixing docker.json alone will NOT fix docker build — the RUN commands inside the build container don't inherit daemon proxy settings.

Diagnosis — identify which sub-problem:

# 1. Can the Docker daemon pull images?
docker pull --quiet alpine:latest 2>&1

# 2. Can a RUN command inside a build reach the internet?
docker build --no-cache - <<'EOF' 2>&1
FROM alpine:latest
RUN apk update && echo "APK OK"
EOF

# 3. Can a running container reach the internet?
docker run --rm alpine:latest sh -c "apk update 2>&1 | head -3"

Four sub-problems and their fixes:

2G-1: docker build fails but host works (most common with OrbStack + Shadowrocket)

Symptom: RUN apk add or RUN apt-get install inside docker build fails with Connection refused instantly (< 0.2s), even though host curl to the same URL works.

Root cause: OrbStack's network_proxy: auto creates a transparent proxy inside the VM that intercepts all HTTPS traffic. When Shadowrocket TUN is also active, the transparent proxy's upstream connection breaks — it redirects HTTPS to 127.0.0.1 inside the VM, which has nothing listening.

Diagnosis:

# Verify: inside the container, HTTPS goes to 127.0.0.1 (broken transparent proxy)
docker run --rm alpine:latest sh -c "wget -q --timeout=5 -O /dev/null https://dl-cdn.alpinelinux.org/ 2>&1"
# → "wget: can't connect to remote host (127.0.0.1): Connection refused"
#                                        ^^^^^^^^^^^^ This is the smoking gun

# Verify: --network host bypasses the VM bridge and works
docker run --rm --network host alpine:latest sh -c "apk update 2>&1 | head -3"
# → "v3.23.x ... OK: 27431 distinct packages available"  ← Works!

Fix — use --network host for docker build:

docker build --network host -f Dockerfile -t myimage .

This bypasses OrbStack's VM network bridge entirely. The build container uses the host's network stack directly, where Shadowrocket TUN correctly handles traffic.

Trade-off: --network host disables build-time network isolation. For CI/CD, prefer fixing the proxy config (2G-2). For local development, --network host is the pragmatic fix.

Permanent fix — if all your builds need this, add to ~/.docker/daemon.json or use a shell alias:

# Shell alias (add to ~/.zshrc)
alias docker-build='docker build --network host'

2G-2: OrbStack auto-detects and caches proxy config

OrbStack's network_proxy: auto reads http_proxy from the shell environment and configures the Docker daemon. The config is stored in ~/.orbstack/config/docker.json.

Key behaviors:

• network_proxy: auto — OrbStack reads host env, creates transparent proxy in VM
• network_proxy: none — Disables transparent proxy, but VM bridge traffic still routes through TUN (may timeout)
• docker.json — Controls docker pull proxy, NOT docker build RUN commands

Diagnosis:

# Check all three layers
echo "=== OrbStack config ==="
orbctl config get network_proxy

echo "=== docker.json (daemon proxy) ==="
cat ~/.orbstack/config/docker.json

echo "=== Docker info (effective proxy) ==="
docker info | grep -iE "proxy|No Proxy"

Fix — configure docker.json with host.internal (OrbStack resolves this to the host IP):

python3 -c "
import json, os
config = {
    'proxies': {
        'http-proxy': 'http://host.internal:1082',
        'https-proxy': 'http://host.internal:1082',
        'no-proxy': 'localhost,127.0.0.1,::1,192.168.128.0/24,100.64.0.0/10,host.internal,*.local'
    }
}
path = os.path.expanduser('~/.orbstack/config/docker.json')
json.dump(config, open(path, 'w'), indent=2)
print('Written:', path)
"

# Full restart required
orbctl stop && sleep 3 && orbctl start

Important: Use host.internal (OrbStack-specific), NOT 127.0.0.1 (points to VM loopback) and NOT host.docker.internal (may not resolve in all contexts).

Why NOT remove the proxy: When TUN is active, removing the Docker proxy means VM traffic goes directly through the bridge → TUN path, which causes TLS handshake timeouts. The proxy provides a working outbound channel.

2G-3: Removing proxy makes Docker worse (counter-intuitive)

Decision tree:

• docker pull broken → Fix docker.json with host.internal proxy (2G-2)
• docker build broken → Use --network host (2G-1) OR pass --build-arg http_proxy=http://host.internal:1082
• Both broken → Fix both: docker.json + --network host

2G-4: Deploy scripts and container healthchecks probe localhost through proxy

Deploy scripts that curl localhost inside containers or Docker healthchecks that use wget http://localhost will route through the proxy if env vars leak into the container.

Common symptoms:

• Container healthcheck shows (unhealthy) but the app inside is running fine
• wget: can't connect to remote host (127.0.0.1): Connection refused in healthcheck logs (proxy port, not app port)

Root cause: Docker inherits uppercase AND lowercase proxy env vars from the host. Many tools only clear uppercase (HTTP_PROXY=) but forget lowercase (http_proxy=http://127.0.0.1:1082). The healthcheck wget uses lowercase.

Fix in docker-compose.yml — clear BOTH cases:

environment:
  # Must clear both uppercase and lowercase — wget/curl check different vars
  - HTTP_PROXY=
  - HTTPS_PROXY=
  - http_proxy=
  - https_proxy=
  - NO_PROXY=*
  - no_proxy=*

Fix in deploy scripts:

_local_bypass="localhost,127.0.0.1,::1"
export NO_PROXY="${_local_bypass}${NO_PROXY:+,${NO_PROXY}}"
export no_proxy="$NO_PROXY"

# Use 127.0.0.1 instead of localhost in probe URLs (some proxy implementations
# only match exact string "localhost" in no-proxy, not the resolved IP)
curl http://127.0.0.1:3001/health   # ✅ bypasses proxy
curl http://localhost:3001/health    # ❌ may still go through proxy

Verify the fix:

# Docker proxy check (should show proxy + no-proxy)
docker info | grep -iE "proxy|No Proxy"

# Pull test
docker pull --quiet hello-world

# Build test (the real verification)
docker build --network host --no-cache - <<'EOF'
FROM alpine:latest
RUN apk update && echo "BUILD OK"
EOF

# Container env check (no proxy leak)
docker exec <container> env | grep -i proxy
# Expected: all empty or not set

Step 2H: SSH/Git Failing Through a TUN (Connection closed by 198.18.x.x)

Symptom: git clone/fetch/push fails with Connection closed by 198.18.0.x port 443. ssh -T git@github.com may also fail. DNS resolution returns 198.18.x.x addresses instead of real IPs.

First, a warning about this symptom: the 198.18.x.x in the error message is not evidence that fake-IP caused the failure. Under TUN, every hostname resolves to a fake IP, so that address appears in the error whether the connection failed for this reason or any other. It tells you a TUN is active — nothing more. Two very different mechanisms produce an identical error line, and they need opposite fixes:

Tell them apart before doing anything — run the same proxy path with plain HTTPS to an unrelated host, enough times to see a rate rather than an anecdote:

# Same proxy, plain HTTPS, unrelated destination — does it also fail?
for i in $(seq 1 40); do
  curl -s -o /dev/null -x http://127.0.0.1:<proxy-port> \
       -w '%{http_code}\n' -m 10 https://www.google.com/generate_204
done | sort | uniq -c

• Some HTTPS requests fail too (000) at a rate comparable to your SSH failures → (B). The failure is not protocol- or port-specific. Rerouting SSH will not help, because the thing dropping connections is upstream of the protocol.
• HTTPS is 100% clean across many trials while SSH reliably fails → (A), and the DIRECT-rule fix below is warranted.

One measured data point for (B), so the rate is concrete rather than hand-waved: on a macOS host running Shadowrocket in TUN mode, ssh -T git@github.com succeeded 20/20 in a good window and failed roughly 15–20% of attempts in a bad one — while plain HTTPS through the explicit proxy to an unrelated host failed at a comparable rate in the same window. Same-day git push failed twice in a row and then succeeded on the third try, unchanged. That machine was (B), so the port-443 story did not survive contact with the measurement. This is one host, not a universal refutation — run the test above and find out which one you have.

Fix for (B) — retry, and make it automatic. The failure is episodic, so the cheapest correct response is to try again. Make it transparent at git's transport layer rather than remembering to re-run commands:

git config --global core.sshCommand /path/to/ssh-retry-wrapper.sh

Two things that wrapper must get right:

• Only retry failures that happen before key exchange completes. Match on stderr signatures such as kex_exchange_identification, Connection closed|reset by <host> port <n>, and Connection timed out during banner exchange. If key exchange never completed, ssh never ran the remote command, so git received zero protocol bytes and reopening the connection is invisible to it. A mid-transfer drop carries none of those signatures and must pass through untouched — retrying there is what corrupts things.
• Don't reach for ConnectionAttempts instead; it does not work here. It retries only the TCP connect phase, and under TUN the handshake completes locally no matter what the destination does — so connect always "succeeds" and the failure lands afterwards, in the banner exchange. Measured with a listener that accepts and immediately closes: ConnectionAttempts=1, 3, and 5 all produced exactly one accept.

⚠️ Beware stacking retries. If a wrapper script or cron job already retries pushes, adding a transport-layer retry underneath multiplies them (5 × 3 = 15 connections), and the outer layer's logs will undercount actual connection attempts.

Fix for (A) — a DIRECT rule (requires proxy tool config access), so the TUN passes this traffic through without protocol inspection:

IP-CIDR,140.82.112.0/24,DIRECT
IP-CIDR,192.30.252.0/22,DIRECT

This degrades softly: if it doesn't help, you are back where you started.

Not recommended: hardcoding GitHub's IP in ~/.ssh/config. Earlier versions of this skill prescribed HostName 140.82.112.35. Three reasons to avoid it:

1. It goes stale, and fails hard when it does. GitHub rotates these addresses; a dead hardcoded IP surfaces as connection refused, which reads like an outage. Prescribing a weekly cron job to babysit an IP is a sign the fix is in the wrong place.
2. It may not bypass what it claims to. Proxy tools route by rule, not by whether the destination happens to be a fake IP — so unless a matching DIRECT rule exists, a literal IP takes the same forwarding path as the hostname did, and inherits the same instability.
3. It introduces a new known_hosts lookup key. known_hosts is keyed by connection string, not by host key, so [140.82.112.35]:443 is unseen and triggers a TOFU prompt — which either blocks unattended automation or pressures you into loosening StrictHostKeyChecking. That is a real cost paid for an uncertain benefit.

If you already applied the hardcoded-IP fix and things improved, run the HTTPS test above before concluding it was the IP: episodic failures also "improve" on their own when the bad window ends.

Step 2I: Fix Stalled DNS Resolver in getaddrinfo Chain

Symptom: ssh, curl (no -x), git, and any other tool using system DNS hangs ~60 seconds before resolving. ssh -vvv freezes immediately after:

debug2: resolving "<host>" port <port>
debug3: resolve_host: lookup <host>:<port>

…and never reaches debug1: connect to address. After the wait it eventually succeeds — but every new connection pays the same penalty. nslookup <host> returns instantly (~10ms) but dscacheutil -q host -a name <host> takes 60s+.

Root cause: macOS getaddrinfo consults every entry in scutil --dns whose domain filter matches (or has no filter at all). If one resolver's nameserver is unreachable but its interface is still in the routing table, getaddrinfo waits the full UDP retry timeout (typically 30-60s) before falling through to the next resolver. The most common real-world trigger is a tunneling daemon (Tailscale, Cisco AnyConnect, Pulse Secure) that crashed without unwinding its utun and DNS injection.

Why nslookup lies: nslookup reads only /etc/resolv.conf (one nameserver). dscacheutil and getaddrinfo go through DirectoryService, which queries the whole resolver chain in scutil --dns. A divergence between these two is the smoking gun.

The "ping ok but DNS dead" trap: ping <resolver-ip> may answer in <1ms even when port 53 is dead, because the utun interface still claims the IP and replies to ICMP locally. Don't infer resolver health from ping. Test the actual service: dig @<ip> +tries=1 +timeout=3 example.com.

Diagnosis: Bisect by Nameserver

Find the dead resolver in under 15 seconds:

# 1. Read every resolver's nameserver, interface, and matching scope
scutil --dns | grep -E "^resolver|nameserver|domain :|search domain|if_index"

# 2. Time each nameserver in isolation (3-second cap)
for ns in <each_unique_nameserver_from_step_1>; do
  printf "  %s: " "$ns"
  /usr/bin/time -p dig @$ns +tries=1 +timeout=3 +short example.com 2>&1 | tr '\n' ' '
  echo
done

Healthy nameservers respond in <0.1s. The dead one returns connection timed out; no servers could be reached after exactly 3.01s.

For IPv6 resolvers, run the same dig @<ipv6> test — Tailscale and several VPNs inject both v4 and v6 addresses, and either side dying produces the same symptom.

Read Resolver Attributes — Determines Blast Radius

Each scutil --dns resolver has attributes that decide which queries it participates in:

A dead resolver with a domain filter is annoying but localized. A dead resolver with no domain filter (very common with VPN-injected DNS like Tailscale's 100.100.100.100) tanks every system lookup until you fix it.

Confirm the Suspect Component

Once the bisection identifies the dead nameserver, identify which app injected it (interface name in if_index is the strongest hint — utun* interfaces usually trace back to a VPN daemon).

For Tailscale specifically:

tailscale status
# Healthy: lists peers
# Dead:    failed to connect to local Tailscale service; is Tailscale running?

The "failed to connect" error means the daemon process is gone but the network configuration it injected (utun interface + DNS resolver entry) hasn't been cleaned up. The same pattern applies to any VPN/tunneling tool.

Fix

Restart the responsible app at the application level so its cleanup hooks run and remove the stale interface:

Tailscale (App Store and Standalone macOS builds):

osascript -e 'quit app "Tailscale"' && sleep 3 && open -a Tailscale

For other VPN/tunneling tools, prefer a clean app-level quit (menu bar → Quit, or osascript -e 'quit app "<name>"') over kill -9. Forced kill skips cleanup and can leave the same dead-interface state. Only escalate to pkill -9 <name> if the app refuses to exit normally.

Why "restart the app" beats "flush DNS cache": sudo dscacheutil -flushcache; sudo killall -HUP mDNSResponder flushes cached results, but the resolver chain in scutil --dns is rebuilt from network configuration, not from the cache. The dead resolver is still there after a flush. The fix has to come from the app that registered the resolver in the first place.

Verify End-to-End (4 Dimensions)

A DNS-resolver fix is easy to half-verify. All four must pass before declaring the system path healed:

# 1. The owning daemon is back (not just its UI)
tailscale status | head -3

# 2. The previously-dead nameserver responds fast
dig @<previously-dead-ns> +tries=1 +timeout=3 +short example.com
# Expected: <0.1s, returns IP

# 3. macOS system path is unblocked (proves getaddrinfo recovered)
/usr/bin/time -p dscacheutil -q host -a name example.com
# Expected: <0.1s, returns IP

# 4. The original failing command works WITHOUT any workaround
ssh -o "ProxyCommand=none" -T git@github.com
# Expected: "Hi <user>! You've successfully authenticated..."

The fourth dimension is the one that matters most. If you applied a workaround during diagnosis (a ProxyCommand that delegates DNS to a SOCKS5 proxy, a /etc/hosts entry, a hardcoded IP), running the original command with the workaround disabled (ProxyCommand=none) is the only way to know you actually healed the system DNS path rather than just routed around it.

See references/dns_resolver_chain_stall.md for the full mental model of macOS resolver ordering, the IPv4-vs-IPv6 split, and a worked example walking through every diagnostic command and its real output.

Step 2J: Fix TUN DIRECT Split-Brain (domain connections die, real-IP connections work)

Symptom: every site that the TUN tool routes DIRECT (domestic sites, cloud-provider APIs, your own API domain) fails at the same time, while proxied overseas sites keep working. The failures wear several disguises depending on the client:

• curl direct: SSL routines::unexpected eof while reading mid-handshake
• curl -x http://127.0.0.1:<proxy-port>: CONNECT tunnel failed, response 503
• Node-based CLIs (Claude Code, npm, etc.): UNKNOWN_CERTIFICATE_VERIFICATION_ERROR with endless retries — the fake-IP↔domain table is mismapped, so the TUN forwards your SNI to the wrong backend, which presents another site's certificate
• A proxy health watchdog, if you run one, keeps reporting healthy — see the warning below

Root cause: the TUN tool (Shadowrocket / Clash / Surge in fake-IP mode) has two independent forwarding planes: proxied traffic (overseas, via the node) and DIRECT traffic (matched by domain rules, forwarded from the tool's own network stack). The DIRECT plane's state can break alone — DNS-over-tunnel failure, stale fake-IP table after a network change, or internal state corruption — while the proxied plane stays perfectly healthy. Everything resolved to a fake IP whose domain matches a DIRECT rule then dies inside the tool.

Diagnosis — three commands separate this from a real network outage:

# 1. Confirm fake-IP DNS is in effect (domain resolves into 198.18.0.0/15)
dig +short <any-direct-rule-domain>
# → 198.18.x.x

# 2. Get the real IP from a public resolver (UDP/53 usually bypasses the broken path)
dig +short @223.5.5.5 <domain>   # or @1.1.1.1 / @8.8.8.8
# → real address

# 3. Connect by real IP with correct SNI — bypasses the fake-IP table
curl -sS -o /dev/null -w '%{http_code}\n' --resolve <domain>:443:<real-ip> https://<domain>/
# → normal HTTP status  ← physical network is FINE; the TUN's DIRECT state is broken

If step 3 also fails, this is not split-brain — treat it as a real local-network outage.

Fix — restart the tunnel, then flush the OS DNS cache (stale fake-IP entries survive the reconnect):

# Shadowrocket (URL scheme; Clash/Surge: use their API or GUI toggle)
open "shadowrocket://disconnect" && sleep 3 && open "shadowrocket://connect" && sleep 6

sudo killall -HUP mDNSResponder

Verify all four planes — a fix that restores one plane can leave (or put) another down:

curl -sS -o /dev/null -w 'domestic direct: %{http_code}\n' --max-time 10 https://<domestic-site>/
curl -sS -o /dev/null -w 'own DIRECT-rule domain: %{http_code}\n' --max-time 12 https://<your-api-domain>/health
curl -sS -o /dev/null -w 'cloud API: %{http_code}\n' --max-time 10 https://<cloud-provider-endpoint>/
curl -x http://127.0.0.1:<proxy-port> -sS -o /dev/null -w 'overseas via proxy: %{http_code}\n' --max-time 10 https://www.google.com/generate_204

Watchdog blind spot: if you run an automated proxy health checker, check what it actually probes. A watchdog that only tests an overseas endpoint through the proxy certifies the proxied plane and nothing else — in the incident that produced this section, the watchdog reported "healthy" every 5 minutes for 2+ hours while the DIRECT plane was completely down. A green health check is evidence only for the path it probes; probe each plane the tool forwards.

Step 2K: Decide Whether the Blocker Is Your Tunnel or the Destination's Firewall

Symptom: SSH to a public host fails with kex_exchange_identification: Connection closed by <ip> — or Connection closed by UNKNOWN port 65535 — while HTTPS to the same host keeps working. nc -z <ip> 22 reports the port open. The method below is written for SSH; for another blocked service substitute its own port, log path and service unit throughout.

Why this needs its own step: it looks exactly like Step 2H, and Step 2H's fix (restart the TUN, flush DNS) will not help, because the cause may not be local at all. It also looks exactly like Step 5A — but 5A is specifically Tailscale-on-WSL, while this step is for a TUN with no Tailscale in the path. The third mechanism neither of those covers is the destination's own firewall silently dropping you — at either of the two layers a cloud host enforces (the provider's security group outside the machine, and ufw/firewalld/nftables inside it). From the client, a dropped-by-allowlist connection and a TUN forwarding failure are indistinguishable: both show "the connection opens, then dies before the protocol banner." Under a TUN, every additional local probe produces more misleading evidence rather than less.

First: stop trusting nc -z. Under a TUN the local stack completes the TCP handshake on the destination's behalf, so nc -z succeeds for every port, including ports the destination has closed. Calibrate before reading any port probe:

# Probe a port the destination certainly does NOT serve
nc -z -G 5 -w 5 <destination-ip> 65001 && echo "nc is lying: the TUN answered for a closed port"
# Control: an address that cannot answer at all (TEST-NET-3, RFC 5737)
nc -z -G 5 -w 5 203.0.113.199 443 || echo "control OK: this probe can still report failure"

• First line prints → on this machine nc -z <that host> <any port> carries zero information about the destination. Continue with this step and ignore every port probe from here on.
• First line does not print (and the control line does) → nc -z is trustworthy on this path, so this step's premise does not hold. Re-check whether a TUN is actually in play before continuing; if it is not, an open/closed port reading is real evidence and Step 2B/5A may fit better.

The same trap applies to hostnames: fake-IP DNS answers for names that do not exist, so an HTTP 200 does not identify which machine served it. Calibrate that too — dig +short @1.1.1.1 <a-name-you-just-invented> should come back empty. If it returns an address (typically 198.18.x.x), DNS answers are synthetic and no name-based result on this machine identifies a host.

Then ask the destination out-of-band. The only way to settle "did my packets arrive?" is a channel that does not traverse the failing data path. On a cloud VM that is the provider's guest-agent control plane: it reaches the instance over the provider's own network, so it works even when every port you can dial is blocked.

• Alibaba Cloud — aliyun ecs DescribeCloudAssistantStatus --RegionId <region> --InstanceId.1 <instance-id> (a LastHeartbeatTime within the last minute proves the machine is alive and its agent is healthy), then aliyun ecs RunCommand --RegionId <region> --InstanceId.1 <instance-id> --Type RunShellScript --ContentEncoding Base64 --CommandContent <base64> --Timeout 60, and aliyun ecs DescribeInvocationResults --RegionId <region> --InvokeId <id> to collect output (the Output field is base64). Commands executed and verified 2026-08-30.
• AWS / GCP / Azure have equivalents (SSM Run Command, guest-agent based access, the Run Command extension). Syntax not verified here — check the provider's docs.
• No guest agent, or the host is not on a provider that offers one → this step cannot settle the question; you need someone with physical or console access to read the host's logs. Do not substitute another client-side probe.
• Heartbeat stale or absent → the agent or the host is down. That is a fact about the host, not evidence for either side of the tunnel-vs-firewall question; fix or investigate that first.

Ask the one question that settles ownership:

# run these INSIDE the destination, through the out-of-band channel
journalctl -u ssh -n 20 --no-pager || journalctl -u sshd -n 20 --no-pager   # unit is `ssh` on Debian/Ubuntu, `sshd` on RHEL
grep -a sshd /var/log/auth.log 2>/dev/null | tail -20    # Debian/Ubuntu, only if rsyslog is installed
grep -a sshd /var/log/secure   2>/dev/null | tail -20    # RHEL family

⚠️ Check the journal first, and do not read an empty auth.log as an answer. On Ubuntu 24.04 the file still exists and still receives CRON and sudo lines, but rsyslog is not installed by default and sshd writes nothing to it — measured on a live 24.04 host: 90 lines in auth.log, 0 of them from sshd, while journalctl -u ssh had the full record. Reading auth.log alone there manufactures exactly the false "no attempts arrived" this step exists to prevent.

To see an attempt land in real time, run this on the host while you retry the connection from the client — it shows only sockets open at that instant, so it says nothing about an attempt that already failed and was torn down:

ss -tn | grep ':22 '

• Entries present → packets arrive and sshd is closing you. That is a server-side problem (MaxStartups exhaustion, AllowUsers/DenyUsers, hosts.deny, full disk, fork failure) — not a tunnel problem, and outside this skill.
• No entries for your attempts → do not conclude "the packets never arrived." Zero journal entries only proves nothing reached sshd, which a host-local packet filter produces just as reliably as a network that never delivered anything. The two have opposite fixes and are separated by one command:

# on the host, while the client retries; needs root
timeout 20 tcpdump -nni any "tcp port 22 and host <your-egress-ip>" -c 10
ss -tan '( sport = :22 )'          # look for SYN-RECV, not just LISTEN

• SYNs appear in tcpdump, but ss shows only LISTEN and no SYN-RECV → the packets did arrive and a host-local firewall dropped them before the socket layer. The network path is fine; stop looking at your tunnel. Go to the host-firewall check below.
• No SYNs in tcpdump at all → now you have earned "never arrived." The blocker sits before the host: its cloud firewall / network ACL, or your own TUN. Continue below.

⚠️ Retransmitted SYNs (the same sequence number every second) are the giveaway for a silent drop; a rejected connection produces an RST instead, and a closed port produces one immediately.

If nothing arrived, check the allowlist against your real egress IP. Under a TUN, curl ifconfig.me returns the proxy exit, not the address the destination's firewall sees. The authoritative value is the source IP the provider recorded for your own API calls:

# Alibaba Cloud example. Do NOT grep for '"sourceIpAddress":"' — the CLI pretty-prints with a
# space after the colon, so that pattern silently matches nothing. Parse the JSON instead.
aliyun actiontrail LookupEvents --RegionId <region> --MaxResults 20   | python3 -c 'import json,sys,re; d=json.load(sys.stdin);       print("\n".join(sorted({e.get("sourceIpAddress","") for e in d.get("Events",[])
            if re.fullmatch(r"[0-9.]+", e.get("sourceIpAddress",""))})))'

The IP filter is load-bearing: internal service-to-service events carry a hostname in that field (e.g. rmc.resourcemanager.aliyuncs.com), not your address.

⚠️ This is authoritative only if your provider-API traffic and your blocked connection take the same forwarding plane (both DIRECT, or both proxied). Step 2J is the reason: a TUN routes by rule, so the provider's API domain and your destination host can sit on different planes with different exits — in which case the recorded IP is not the IP the destination's firewall sees. Confirm the two match a common rule before trusting the comparison.

There are usually two allowlists, and fixing one is not fixing the problem. A cloud host typically enforces the port twice: at the provider's security group (outside the machine) and again at the host's own firewall (ufw, firewalld, raw iptables/nftables). They are edited by different tools, drift independently, and a rule added to one is invisible in the other. Check both, always:

# outside the machine — provider security group (Alibaba Cloud shown)
aliyun ecs DescribeSecurityGroupAttribute --RegionId <region> --SecurityGroupId <sg-id>

# inside the machine, via the out-of-band channel — host firewall
ufw status verbose 2>/dev/null            # Debian/Ubuntu, if ufw is in use
firewall-cmd --list-all 2>/dev/null       # RHEL family
iptables -S 2>/dev/null | head -30        # raw rules, and the INPUT policy
nft list ruleset 2>/dev/null | head -60

• Your current egress IP is missing from either list → that is the fix. Add it to that layer (security group: aliyun ecs AuthorizeSecurityGroup, reverse RevokeSecurityGroup; ufw allow from <ip> to any port <n> proto tcp, reverse ufw delete allow ...). Allowlists pinned to a dynamic residential or mobile address rot silently: the rule still reads as correct, and nothing logs the day your address changes.
• It is present in the security group but packets still never reach sshd → this is the case the previous bullet's older wording got wrong. It does not mean the blocker is local. Run the tcpdump + ss pair above first: arriving SYNs with no SYN-RECV mean the host firewall is dropping you, and no amount of work on your own machine will change it. Only after tcpdump shows nothing arriving is the blocker local; continue below.

⚠️ iptables -S first, not just the service-level tools. A default -P INPUT DROP with explicit accepts for 80/443 looks like a healthy machine from every angle this skill checks elsewhere — sshd active and listening on 0.0.0.0:22, config valid, no fail2ban, empty hosts.deny, disk and memory fine — and still drops every SSH packet. Checking fail2ban and hosts.deny while skipping the packet filter is the single easiest way to certify a blocked host as "completely healthy."

View the full SKILL.md on GitHub