test(markdown-format): SessionStart probe case failed once on Windows with only the jq notice

[kyle-sexton]

Problem

The markdown-format suite's "SessionStart probe with the kill switch unset" case (plugins/markdown-format/hooks/markdown-format.test.sh:1979-1990) failed once on test-windows on a PR that did not touch markdown-format, then passed on the next head of the same branch. It expects the probe's notice to name markdownlint-cli2; the failing output named only jq.

Evidence

• Failure: run 37094085446, job 111120317567, head 12532bb89 of feat(effort): phase effort advice, event-log content opt-in and typed OTEL columns, with the upstream record #5992, step "Test the markdown-format hook on Windows": PASS=166 FAIL=1 SKIPPED=4. Output: {"hookSpecificOutput":{... "additionalContext":"markdown-format: jq was not found on PATH. Hooks that need it will skip until it is installed. ..."}}, with no markdownlint-cli2 line. The line just before reads bash.exe: warning: could not find /tmp, please create!.
• The next head, a65512646, passed test-windows. git diff --stat 12532bb89 a65512646 -- plugins/markdown-format is empty.
• Flake rate: over the 44 most recent completed, non-cancelled test-windows runs (PR and push, through 2026-10-03T04:46Z), the markdown-format step failed in 1 and passed in 43. No other failure in that window.

Likely cause (not confirmed)

run_probe_launcher runs the probe under env -i PATH="$PROBE_BIN", which removes TMP/TEMP/TMPDIR; hence the /tmp warning. probe-prerequisite.sh reports each missing tool through hook::notice_once "$plugin-$name", keyed on session s1, which every probe case reuses. If notice-once state lands in a location shared across cases when the temp variables are gone, an earlier case can mark markdownlint-cli2 as already noticed, leaving only jq (which is never in PROBE_BIN on Windows because the test links only /usr/local/bin, /usr/bin and /bin). Where that state is written was not checked; #4612 (notice_once keys on agent) is in the same area.

Impact

Low: 1 in 44 runs, and test-windows is not a required check. It still turns a PR's checks red and sends an agent to debug a plugin the PR did not change.

Proposed fix options

1. Give each probe case its own notice-once state (pass a fresh TMPDIR and whatever state root hook::notice_once reads into the env -i call) and a unique session_id.
2. Assert on what the case is about (the markdownlint line is present when it is missing) and stop depending on which other tools happen to be absent from PROBE_BIN.

Recommendation: confirm where hook::notice_once writes under env -i on Git Bash, then apply option 1. Basis: the failing output and the /tmp warning above; the cause is a hypothesis until that check is done.

Acceptance criteria

• The probe cases pass TMPDIR (or the notice state root) and a per-case session id inside env -i.
• The /tmp warning no longer appears in the step log.
• No markdown-format failure in the next 50 test-windows runs.

[kyle-sexton]

This was generated by AI during triage.

Routing: human-gated, proposed close. Priority: low. Work class: scoped (C3) if it stays open.

Question: Close #6011 as not reproducible on current code? Recommended: yes.

Evidence:

• The failing head 12532bb89 (2026-10-02, a feat(effort): phase effort advice, event-log content opt-in and typed OTEL columns, with the upstream record #5992 branch commit, not on main) ran the probe through hooks/exec-bash.mjs and hooks/probe-prerequisite.sh. refactor(prerequisites): convert the existing prerequisites.json files and move the fleet reader to the node checker #5980 (merged 2026-10-03 05:33Z) replaced that path. The test's run_probe_launcher now runs lib/prerequisites.mjs probe (plugins/markdown-format/hooks/markdown-format.test.sh:1968-1973). That probe keeps its once-per-session notice state under $CLAUDE_PLUGIN_DATA/skip-notices (plugins/markdown-format/lib/prerequisites.mjs:479-481), and each case gets a fresh CLAUDE_PLUGIN_DATA from mktemp -d.
• The issue's hypothesis (notice state shared across cases) was already weak at the failing head, which also gave each case its own CLAUDE_PLUGIN_DATA.
• Since refactor(prerequisites): convert the existing prerequisites.json files and move the fleet reader to the node checker #5980 merged, the step "Test the markdown-format hook on Windows" ran in 185 test-windows runs (2026-10-03T05:40Z to now), and every one succeeded. At the issue's 1-in-44 rate, about 4 failures would be expected. The issue's own acceptance bar, no failure in the next 50 runs, is met.

Caveat: the root cause was never confirmed. This is "not reproducible after the launcher path was replaced", not "fixed". The bash.exe: warning: could not find /tmp line under env -i was not re-checked in current logs.

Alternatives:

1. Recommended: close as not reproducible.
2. Keep it open and delegate a hardening change: per-case TMPDIR and a unique session_id inside env -i for every probe case, plus an assertion that does not depend on which other tools are missing. That is a C3 test-only change, worth it only if the flake recurs.