[mvalancy]

Fixes the red in the Comprehensive suite's "Installer & Deploy Config" job on #281 (job 130418062294). The "stranded" installer case ended with:

⚠ Network connectivity test failed
ℹ This may cause download failures. Continue? [y/N] ✗ Installation cancelled - network required

Cause: the pre-flight check made one 5-second request to github.com. With no terminal attached, the Continue? prompt reads "no", so one dropped request cancelled the whole install. Anyone running the one-line installer on a shaky connection would see the same thing.

Fix: the new network_reachable asks up to 3 times, 2 seconds apart, before calling the network unreachable. GRAPHDONE_NETWORK_TRIES and GRAPHDONE_NETWORK_PAUSE override the count and the pause. Nothing else in the check changes:

• wget is still used when there is no curl;
• with neither tool, the check passes;
• the prompt is the same.

Test: installer-contract.mjs now runs the real network_reachable and check_network against a fake curl:

• two failures and then a success: the install continues, with no warning;
• three failures: it cancels;
• TRIES=1: it asks once.

The new test fails on the previous install.sh. node --test tests/toolchain/installer-contract.mjs passes 6/6.

🤖 Generated with Claude Code

[github-actions]

🧪 Comprehensive Test Suite

• Unit suites (Node 22.18.0 & 24.x) — core, web, server, mcp-server: ✅ passed
• Installer & deploy config: ✅ passed

Full-stack smoke gate runs in the CI workflow.

[mvalancy]

Independent review (72c44db)

node --test tests/toolchain/installer-contract.mjs passes 6/6. I also ran both functions under dash, the /bin/sh that the shebang and curl … | sh use, with set -e. It tried 3 times then cancelled, and GRAPHDONE_NETWORK_TRIES=0 asked once. set -e is safe here: the function only runs under if !, and the && return lists are exempt anyway. install.sh does not use set -u.

The test harness is sound. The fake curl comes first on PATH, so a real /usr/bin/curl is never reached. On a noexec tmpdir the test fails loudly (no calls file) instead of passing by mistake.

Findings

1. should-fix: the wget call (install.sh:453) retries inside every attempt. GNU wget defaults to --tries=20 and retries connect timeouts with a 1–10 s backoff. wget 1.21 made 20 connects to a black-holed address. That is about 4 min per attempt (20×5 s + 145 s). Failure scenario: a machine with only wget, behind a firewall that drops traffic to github.com, now shows a frozen "Checking network ● ● ● ●" for about 12 min before cancelling (about 4 min before this PR). Fix: add --tries=1, now that the loop does the retrying.
2. should-fix (older than this PR; a follow-up is fine): the [y/N] prompt (install.sh:468) can never be answered. check_network & runs as an async job, and a shell without job control gives such a job /dev/null as stdin. In dash, echo y | sh -c 'f(){ read -r r || r=n; echo $r; }; f & wait' prints n. So after 3 failures every install cancels, with or without a terminal, and wait $network_pid then exits under set -e. The test comment's "on a run with no terminal" understates this.
3. nit: install.sh:457. If GRAPHDONE_NETWORK_TRIES is not an integer (abc, 2.5), [ -ge ] errors, so return 1 never runs and the loop never ends. Under dash with pause 0 it made about 1,200 curl calls in 3 s. A fix: case $tries in ''|*[!0-9]*) tries=3;; esac.
4. nit: the curl worst case is 3×5 s + 2×2 s = 19 s, but the dots stop after 3 s (6 × 0.5 s), so the line sits still for about 16 s. The animation loop could run for as long as the job does.
5. nit: the test runs the functions under bash (installer-contract.mjs:175). Users run them under sh, so running the test under sh would catch dash-only breakage. They do pass under dash today. The wget branch has no test.

🤖 Generated with Claude Code

[mvalancy]

Thanks. All five findings are addressed in d4235e8:

1. wget retries: wget now gets --tries=1. A test checks the flag and that the wget-only path retries.
2. Unanswerable prompt: only network_reachable runs in the background now. The prompt runs after wait, in the foreground.
   • It reads the terminal (ask_continue, which reads /dev/tty), not stdin, because under curl … | sh stdin is the script itself. With no terminal the answer is no.
   • The disk-space and macOS prompts use the same helper, since they had the same stdin problem.
   • Tests cover three cases: a terminal "y" continues, a terminal "n" cancels, and a "y" on stdin with no terminal cancels.
3. Bad retry values: a non-number GRAPHDONE_NETWORK_TRIES or GRAPHDONE_NETWORK_PAUSE falls back to the default, with a test.
4. Animation: the dots keep moving until the check ends. A continued install now says GitHub wasn't reachable instead of "Connected".
5. Shell: the contract runs the functions under /bin/sh, with only fakes on PATH.

installer-contract.mjs passes 8/8. 3 of those 8 fail on the previous commit and on dev.

🤖 Generated with Claude Code

[github-actions]

🧪 Comprehensive Test Suite

• Unit suites (Node 22.18.0 & 24.x) — core, web, server, mcp-server: ✅ passed
• Installer & deploy config: ✅ passed

Full-stack smoke gate runs in the CI workflow.

[mvalancy]

Merge verification for d4235e8:

• CI passed:
  • CI run 43027322476.
  • Comprehensive Test Suite run 43027322669, whose "Installer & Deploy Config" job runs the real installer in a container, fresh, update and stranded.
  • Prod Stack Smoke is skipped on PRs.
• Merge ref: the parents of refs/pull/287/merge are af3f55d (dev's tip) and d4235e8 (the head). Its tree 6b596d8 is identical to git merge-tree --write-tree origin/dev d4235e8.
• Review: both rounds' findings are fixed (see the reply above). installer-contract.mjs passes 8/8, and 3 of those tests fail on dev.

Merging to dev.