Act on review findings from SECURITYREVIEW-2026-06-25-1

[mnsmarko]

A new security review has been opened on branch SECURITYREVIEW-2026-06-25-1.

• Report: docs/reviews/2026-06/SECURITYREVIEW-2026-06-25-1.md
• TODO: docs/reviews/2026-06/TODO-SECURITYREVIEW-2026-06-25-1.md

Please review the findings and either fix, defer (ONHOLD), or reject (WONTFIX) each one. Record ONHOLD/WONTFIX in the report file itself so future reviews honor them — a GitHub comment alone is not enough. Only WONTFIX/ONHOLD findings are suppressed later; anything you leave unmarked will be re-reported by the next review if it's still present in the code.

---

Please leave this issue open until you're done

While this issue or its review PR stays open, the scheduler will not open another security review of this repo — so you won't get a second review stacked on top of one you're still working through. To free the repo for its next review you must resolve both:

• Close this issue, and
• Merge or close the review PR — merge it (recommended) to land the report with your WONTFIX/ONHOLD markers on the default branch, the record future runs read to avoid repeating deferred/rejected findings; or close it without merging to abandon this cycle and get a fresh review of newer code (nothing is recorded, so still-present findings come back).

Handover prompt for an AI coding agent (click to expand — paste into Claude Code / Codex in a clone of this repo, with gh authenticated)

You are picking up a completed dual-agent review. There is this open review PR (it contains only the dated report + TODO — no source changes) and a companion issue. Act on the findings, record my defer/reject decisions in the report, then merge the PR and close the issue.

• Report: docs/reviews/2026-06/SECURITYREVIEW-2026-06-25-1.md
• TODO: docs/reviews/2026-06/TODO-SECURITYREVIEW-2026-06-25-1.md
• Review branch: SECURITYREVIEW-2026-06-25-1 · PR base: the default branch

0. Orient. Read the report + TODO on the review branch. Each finding has an ID like 2026-06-03-1-7 and a severity. Check out the base branch and pull latest — the review was a snapshot and code may have moved since.

1. Re-anchor every finding against the CURRENT code (commits may have landed after the review). Report line numbers can be stale — locate each finding now by symbol/content, not by line number. Then: already fixed by a later commit → don't re-fix, leave it unmarked (it just won't be re-reported); no longer applicable → mark WONTFIX with a reason; still live → fix it. Never trust the report's line numbers blindly.

2. Fix the still-live findings via this repo's normal workflow (a fix branch off latest base, or direct commits per convention). Keep code fixes OUT of the review PR branch — it's report-only so it always merges cleanly. Reference the finding ID in commit messages.

3. Record defer/reject decisions IN THE REPORT FILE on the review branch (a GitHub comment is not enough — only the report is read by future runs): WONTFIX to reject, ONHOLD to defer, each with a brief reason. Anything left unmarked and still present will be re-reported next time. Commit + push these edits to the PR branch.

4. Close out with gh — the next review of this type is blocked while either the PR or issue is open, so resolve both:

• gh pr merge <pr> --squash --delete-branch (recommended — lands the report + your markers; it's report-only so it merges into the advanced base. If GitHub says the branch is behind, gh pr update-branch <pr> first.)
• gh issue close <issue> --comment "Fixed: <ids>. Deferred: <ids>. Rejected: <ids>. Already resolved by later commits: <ids>."

Abandoning instead? gh pr close <pr> (no merge) + close the issue — nothing is recorded, markers are lost, still-present findings come back. Only do this deliberately.

Report back: what you fixed, what you marked WONTFIX/ONHOLD (with reasons), what was already resolved by later commits, and the final PR/issue state.

---

Findings summary

11 issues found.

Severity	Count
🟡 Medium	3
🟢 Low	8

🟡 Medium

• 2026-06-25-1-1: Predictable /tmp/.atch-<uid> fallback with symlink-following log/socket creation and unchecked mkdir
• 2026-06-25-1-2: One non-reading client freezes the whole session and stalls the child (head-of-line blocking DoS)
• 2026-06-25-1-11: Unbounded client connections can drive FD_SET past FD_SETSIZE and crash/corrupt the master

🟢 Low

• 2026-06-25-1-3: Control socket has no message origin check — any same-uid process can inject input and signals into a session
• 2026-06-25-1-6: Session log fd leaked to the executed program (missing FD_CLOEXEC)
• 2026-06-25-1-10: Verbatim replay of on-disk logs is a terminal-escape-injection vector (cold attach / tail)
• 2026-06-25-1-7: parse_size integer overflow and unchecked malloc/path-truncation around session paths
• 2026-06-25-1-8: Scrollback ring can be overwritten mid-replay, delivering corrupted/out-of-order history
• 2026-06-25-1-4: Protocol read assumes one packet per read(); a short read disconnects a valid client
• 2026-06-25-1-5: Signal handlers call async-signal-unsafe printf/exit/system
• 2026-06-25-1-9: No security/concurrency tests, no CI or SAST in-repo; security-relevant paths are untested