feat(account): claim links expire after 30 days and accounts get claimed in the browser, so refresh the stored link and stop offering it once claimed

[AnderRV]

Why

Claim links expire 30 days after they are issued, and the account is now claimed in the browser by proving an email (Google, GitHub or a 6-digit code). The CLI kept showing the link stored at signup, so after 30 days zenrows account claim and the quota-error hint pointed at a dead page, and both kept offering the link after the account was claimed.

Client attribution was also too broad: any exported OPENAI_API_KEY labelled the shell as Codex, and any CLAUDE_* variable as Claude Code.

What

• zenrows account claim. Asks the account endpoint whether the account is already claimed, and says so if it is. Otherwise it requests a fresh link (POST /api/agent/claim) when the stored one is missing, untrusted or 25+ days old, and saves it with its issue time.
• Quota/402 hints. Same logic, checked at most every 6 hours, with 5-second timeouts. If the backend can't be reached, the stored link is kept only while it is under 30 days old; otherwise the error says to run zenrows account claim.
• Parallel commands. A lock file with an owner token ensures only one process talks to the claim endpoints at a time; a stale lock is taken over safely. The 6-hour slot is claimed before any network call, and the stored state is re-read before minting and before writing. In a test, eight parallel zenrows fetch processes hitting a 402 made exactly one claim request.
• The key only goes to Zenrows. Claim and account requests are sent only to https zenrows.com hosts (no userinfo), or to loopback for local testing. A signup URL from the environment, a project config, or a discovery document that fails this check falls back to the built-in endpoint, so a cloned project's config can't redirect the key.
• Files. secrets.json and account.json are created 0600 from the start (temp file plus rename), in a 0700 directory.
• Other fixes. A 409 (already_claimed or not_agent_account) is treated as final. A non-JSON 200 gets an accurate error. Claim links with userinfo are rejected.
• Client detection. openai-codex only for CODEX_SANDBOX, CODEX_SANDBOX_NETWORK_DISABLED or CODEX_THREAD_ID. claude-code only for CLAUDECODE, CLAUDE_CODE_ENTRYPOINT or CLAUDE_CODE_CHILD_SESSION. Anything else is unknown.
• Both SKILL.md copies and the README describe the current claim flow.

Verification

• npm run typecheck clean. npm test: 409 pass, 0 fail.
• New tests cover refresh thresholds, the throttle, claimed detection, network fallbacks, the lock (including a live lock held by another process), override validation, discovery validation, the 409s, a non-JSON 200, file permissions and detection. Each behaviour was broken once to confirm its test fails.
• End to end, the built CLI ran against a local stub with an isolated home directory. It covered: signup writing 0600 files; claim at under 25 days (no request) and at 26 days (refresh); eight parallel 402s (one claim request, no leftover lock); backend down with a 31-day link; and a claimed account (hints stop, "already claimed").