Skip to content

Sandbox gate: direct-kill fallback for MCP fixture reap - #141

Merged
w4ffl35 merged 1 commit into
masterfrom
fix/sandbox-mcp-reap
Oct 13, 2026
Merged

w4ffl35 merged 1 commit into
masterfrom
fix/sandbox-mcp-reap

Conversation

@w4ffl35

@w4ffl35 w4ffl35 commented Oct 13, 2026

Copy link
Copy Markdown
Member

RSI baseline eval (rsi-20261013060952) failed at testMcpCloseRetainsFailedTransportForDispose:369 — 'test reaps the MCP server after restoring real signals', false !== true. Full 677KB trial output recovered via the output artifact (the 2KB preview hides it).

The test reaped its fixture server with kill(-serverPid) only. In sealed eval guests that group kill demonstrably misses (target alive 2s later) while direct pid kill — the descendant reap two lines below — works everywhere. Root cause of the group miss is unestablished (same-user SIGKILL to a live group should land; user/mount/worktree/network/uid all eliminated as factors by direct reproduction); the fallback closes it regardless.

Verified: suite green on host and in headlesscode-openshell-rsi:local; tsc clean; check-standards clean. Re-running improve after merge.

…dbox gate)

The mcp-close test reaped its fixture server with a group kill only.
In sealed RSI eval guests the group kill demonstrably misses (server
alive 2s later, reap assert fails, baseline red) while direct pid kill
— used by the descendant reap right below — works. Belt and suspenders:
group kill first, then the server pid itself. No behavior change when
the group kill lands.
@w4ffl35
w4ffl35 merged commit 00a7059 into master Oct 13, 2026
2 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant