capture-cc-traffic

name

capture-cc-traffic

description

Capture the exact HTTP request that real Claude Code CLI sends to the upstream API — headers, body, system blocks — for debugging stealth proxies, comparing cloak output against real CLI, or reverse-engineering CLI behavior. Use this skill whenever the user mentions mimicking Claude Code, debugging stealth headers, comparing real CLI traffic to a proxy/relay, the Claude Cloak project, diagnosing why a relay (中转站) rejects a cloak request, comparing cloak body shape against relay policy, investigating 中转站政策与真 CLI 流量的差异, or asks "what does the real CLI actually send", even if they don't explicitly say "capture traffic".

Capturing Real Claude Code Outbound Traffic

The goal is to get a complete recording of one real Claude Code CLI request — every header, the full JSON body including all system blocks — and use it as ground truth when building a stealth proxy or relay that needs to look like Claude Code.

This skill exists because three subtle traps make naive capture attempts silently fail.

The Three Traps

Trap 1 — ~/.claude/settings.json has an env block that overrides shell env. If settings.json contains env: { ANTHROPIC_BASE_URL: "..." }, that wins over any ANTHROPIC_BASE_URL=... you set in your shell. Even env -i ANTHROPIC_BASE_URL=... does not help — the CLI re-injects the settings.json env after spawn. The capture run will appear to succeed (exit 0) but route to the wrong host and our echo server stays empty.

Trap 2 — --settings <file> merges, it does not replace. The help text says "additional settings", which is literal. Passing --settings /tmp/empty-env.json cannot null out the global env block; an empty {} merge leaves the existing keys untouched. The only reliable override is to temporarily rewrite ~/.claude/settings.json itself.

Trap 3 — --bare and --print have non-obvious input requirements. --bare disables OAuth and requires ANTHROPIC_API_KEY (not ANTHROPIC_AUTH_TOKEN), and it reads the prompt from stdin — it will not consume the argv prompt. --print accepts argv but blocks waiting on stdin unless you redirect < /dev/null. Forget either and the run hangs into a timeout with no captured request.

What to Do

Run scripts/capture.sh from this skill directory. It handles the whole flow: starts a localhost echo server, backs up and neutralizes ~/.claude/settings.json, runs claude --print with ANTHROPIC_BASE_URL=http://localhost:7700, then restores settings even if the run fails.

bash <skill_dir>/scripts/capture.sh "say OK"

Outputs:

• /tmp/cc-capture/echo.log — raw ::REQ:: chunks for every HTTP request the CLI made
• /tmp/cc-capture/parsed.json — structured summary (model, system block sizes, headers, metadata, thinking, tools count)
• /tmp/cc-capture/system-blocks/block{0,1,2,...}.txt — each system block dumped as its own file so they can be diffed or copied into the cloak's template assets

Then read parsed.json and the system-blocks/*.txt files to compare against whatever the cloak is currently sending.

Two Modes That Look Identical But Are Not

claude --print "..." runs in entrypoint sdk-cli, which produces the Agent SDK identity block:

You are a Claude agent, built on Anthropic's Claude Agent SDK.

True interactive claude (no --print) uses entrypoint cli and produces the main CLI identity block:

You are Claude Code, Anthropic's official CLI for Claude.

The large second system block (the multi-thousand-token instructions body) is shared between both modes, so --print is good enough for capturing it. But if a middleman strictly matches against the first identity block, the cloak should emit the cli variant (Tb1 in cli.js), not the sdk-cli variant (Fq4).

To capture the interactive-mode identity instead, run scripts/capture.sh --interactive "say OK". This pipes the prompt via stdin in a way that mimics an interactive turn rather than --print.

What a Healthy Capture Looks Like

After a successful run, expect roughly:

• 1 HEAD request to / (CLI's reachability probe)
• 1 or more POST requests to /v1/messages?beta=true
• The POST body's system field is an array of 3 blocks: billing header → identity → main instruction body (~26 KB)
• metadata.user_id is a JSON-stringified {device_id, account_uuid, session_id} object
• The billing header text begins with x-anthropic-billing-header: cc_version=X.Y.Z.<3hex>; cc_entrypoint=cli; cch=00000; — note the literal 00000, not a real hash

If only the HEAD request shows up and no POST follows, the CLI bailed out before sending the message — usually because Trap 3 (stdin) was hit. If the POST goes somewhere other than localhost:7700, Trap 1 (settings.json env) was not neutralized.

Restoring State

capture.sh always restores ~/.claude/settings.json from the backup, including on Ctrl-C, but if something goes very wrong (the script is killed -9, the system reboots), the backup will sit at ~/.claude/settings.json.cc-capture.bak. Restoring it is a plain mv. Until restored, the user's normal Claude Code CLI will route to http://localhost:7700 and fail to reach Anthropic.

Anchoring Findings to cli.js Source

When a capture surfaces a field whose origin is unclear, the fastest way to identify the source code is anchor-search against the bundled CLI implementation. Use scripts/find-cli.sh to locate it regardless of install method:

CLI="$(bash <skill_dir>/scripts/find-cli.sh)"
node -e 'const s=require("fs").readFileSync(process.argv[1],"utf8"); console.log(s.indexOf(process.argv[2]))' \
  "$CLI" 'x-anthropic-billing-header: cc_version='

For CLI versions <~2.1.166, find-cli.sh returns cli.js (a JS bundle, directly readable). For 2.1.167+, it returns the platform-specific native binary that ships under @anthropic-ai/claude-code-<plat>-<arch>/bin/claude. Anchor literals still work against the native binary via strings -n 30 "$CLI" | grep -F <anchor> or grep -aoE because Bun's SEA embeds the JS source verbatim.

Useful anchors and what they locate in cli.js 2.1.112 (most still work in 2.1.167):

Anchor literal	Function it points to	What it builds
x-anthropic-billing-header: cc_version=	dk8(q)	The billing block text, including the negated bedrock/aws/mantle ternary that controls cch=00000;
claude-code-20250219	ZR1(model) in 2.1.112, N4_ registry in 2.1.167	Runtime anthropic-beta assembly
claude-cli/${	OI()	User-Agent string with optional agent-sdk / client-app / workload suffixes
You are Claude Code, Anthropic's	Tb1/pq4/Fq4	The three identity-line constants
function I8(	I8()	Returns B8.sessionId for the X-Claude-Code-Session-Id header

The 2.1.167 refactor moved per-feature betas into a frozen registry N4_ whose elements are { header, ... } objects, with order mapping to feature names like claude_code, oauth_auth, interleaved_thinking, effort, task_budgets, mid_conversation_system, etc. The WPK map (N4_.map(h => [h.header, h])) is the reverse lookup from header token to feature object.