Summary

Fixes #79

Three bugs fixed, all affecting the same user flow.

Bug 1 — Health check fails when Docker proxy is configured

When users have http_proxy/https_proxy set in the Docker daemon config (e.g. /etc/systemd/system/docker.service.d/ or ~/.config/docker/config.json), Docker injects those proxy env vars into every container. The health check curl commands then attempt to route internal Docker network requests (e.g. http://elasticsearch:9200, http://kibana:5601) through the proxy, which either can't reach those hostnames or rejects them — causing the health check to time out after all retries (~584s) even though Elasticsearch is running fine (as confirmed by the reporter: manual curl http://localhost:9200 from the host returns 200).

Fix: Add --noproxy '*' to all curl commands that contact internal Docker network hostnames:

• Elasticsearch container health check
• Kibana container health check

Bug 1b — kibana_settings password setup silently broken inside bash -c '...'

The kibana_settings service command is bash -c '...' (outer single quotes). Adding --noproxy '*' inside that context breaks the shell quoting: the ' in '*' closes the outer single-quoted string, causing bash to receive a truncated script and exit immediately with an error — so the Kibana system password is never set and the stack fails to start.

Fix: Use --noproxy "*" (double quotes) for the kibana_settings curl command. Inside a single-quoted bash -c '...' string, double quotes are literal characters passed through to bash, which then interprets "*" as a double-quoted glob-safe * — the correct value for curl's --noproxy argument.

Bug 2 — VERSION: unbound variable crash in error log generation

The script uses set -eu. When a failure occurs, generate_error_log() calls get_os_info(), which sources /etc/os-release and then prints $VERSION. On rolling-release distributions like Arch Linux, /etc/os-release does not define VERSION (only VERSION_ID). With set -eu, this causes an unbound variable crash — so the error log is never written and the user sees a confusing VERSION: unbound variable message instead of the actual failure details.

Fix: Use ${VERSION:-} to safely default to an empty string when VERSION is not defined.

Tests added

tests/get_os_info_test.sh (runs in all CI jobs)

• test_get_os_info_succeeds_when_VERSION_is_not_defined: creates a mock /etc/os-release without a VERSION field (simulating Arch Linux) and verifies the fixed code path (${VERSION:-}) exits 0 under set -eu
• test_bare_VERSION_crashes_under_set_eu_when_not_defined: confirms the unfixed pattern ($VERSION) would have crashed, making the above test meaningful

tests/docker/proxy.sh (runs in Docker CI jobs)

Configures ~/.docker/config.json to inject a non-existent proxy (http://127.0.0.1:19999) into all containers — simulating a user with a corporate proxy in their Docker daemon config. Runs start-local.sh and asserts both services respond:

• test_elasticsearch_accessible_with_proxy_configured: HTTP 200 from localhost:9200
• test_kibana_accessible_with_proxy_configured: HTTP 200 from localhost:5601

All 4 CI jobs (ubuntu-22.04 Docker, ubuntu-24.04 Docker, ubuntu-22.04 Podman, ubuntu-24.04 Podman) pass.

Test plan

• Run start-local with http_proxy set in Docker daemon config — health checks pass (verified in CI via tests/docker/proxy.sh)
• Run start-local on Arch Linux — error log generates without crashing (verified in CI via tests/get_os_info_test.sh)
• Run start-local on a standard distro (Ubuntu, Debian) — no regression (verified across all 4 CI matrix jobs)

🤖 Generated with Claude Code