[raymondk]

The wait polled the gateway every three seconds, and its limits did not hold while the gateway was out of reach. A single poll had only the HTTP client's limit, six minutes, so a connection that stalled rather than dropped held the live line at "Waiting for payment" that long. The give-up rule counted a hundred consecutive failed polls: five minutes on a network that refuses at once, but about ten hours on one that stalls. And the hour limit was checked only after a successful poll, so it never fired while the gateway was unreachable.

Now a poll gets 10 seconds, and after five minutes without word of the order the wait gives up with ContactLost, which carries the order as last seen over the last poll's failure, so the command can say where the order stands. Both the give-up and the hour are judged on the clock after every poll, answered or not. Still a query every three seconds; nothing else about the order changes. Ctrl-C still lands at any moment, as before.

Code by Josh, cut down from #824 (commit 9cd3e1c, wait part only): WaitOptions { poll_interval, poll_timeout, give_up_after, timeout }, PollTimedOut and ContactLost errors. Not included, after review: the live line saying the gateway has gone quiet and is answering again, and the WaitEvent callback that drove it. A user who sees nothing for a while can Ctrl-C, and #831 then says where the order stands.

Tests: four clock-driven unit tests in cyclepay/tests.rs (a stalled network given up on after five minutes with the poll timeout as the cause, a refusing one with the refusal, an order gone for five minutes not found, a gateway back within five minutes carried on with) and the hour still holding on an unpaid order.

---

Part of restacking #824 into small PRs, one fix each, stacked on main in this order: cycles-buy/06-wait-clock → 07-follow-up → 08-json-fields. Each PR targets the one below it; after the one below merges, retarget to main.

🤖 Generated with Claude Code

https://claude.ai/code/session_015aem4tPs4TYVZb8k4LMbJJ

[Copilot]

🟡 Changes recommended

The overall one-hour timeout remains unenforced during an outage spanning that deadline.

1 open issue

• Enforce overall timeout during failed or null watchdog polls · New

What changed in this PR

Makes icp cycles buy detect gateway outages and bound waits without changing the order flow.

Changes:

• Adds poll timeouts and clock-based outage detection.
• Reports lost and restored contact, preserving the last known order on failure.
• Adds timing tests, an integration test, and documentation.

File	Description
docs/​guides/​tokens-and-cycles.md	Documents outage warnings and wait limits.
crates/​icp-cli/​tests/​cycles_tests.rs	Tests gateway stop/restart reporting.
crates/​icp-cli/​src/​cyclepay/​tests.rs	Tests outage timing, recovery, and timeout behavior.
crates/​icp-cli/​src/​cyclepay/​purchase.rs	Adds timed polling and connectivity events.
crates/​icp-cli/​src/​commands/​cycles/​buy.rs	Displays connectivity changes and failure context.
CHANGELOG.md	Records the wait reliability improvements.

_{Review effort: Balanced}