# Structured bug-report Discussion values

[jarekjorlowski-web]

Affected area

Coding agent and CLI

What happened?

What happened

A current prime-agent status --json or prime-agent doctor --json can discover a private control socket owned by a still-running pre-rlm.repl worker during an upgrade. The socket matches prime-agent-forkserver-??????/control.sock, is owned by a process titled prime-agent, but does not speak the daemon protocol. Probing it can terminate the legacy Python forkserver and its active kernel.

Current all-current runtimes do not create this socket. The risk is cross-version coexistence because daemon discovery intentionally scans arbitrary live prime-agent listeners.

Prior art: #1696 identified the collision but was closed unmerged by the contribution gate.

Steps to reproduce

Steps to reproduce

1. On Linux, start a pre-rlm.repl daemon worker that owns an active forkserver-backed Python kernel.
2. Confirm with ss -lxp or lsof that the worker process listens on .../prime-agent-forkserver-??????/control.sock.
3. From a current checkout, run prime-agent status --json or prime-agent doctor --json in the same host/namespace.
4. Observe whether the diagnostic reports the control path and whether the legacy forkserver/kernel exits.
5. Repeat after applying the prepared three-file discovery filter; the control path must be absent while the real daemon/worker endpoints and kernel remain active.

The prepared validation uses disposable PID/mount/network namespaces, two real workers and two real kernels. It never uses doctor --fix, public shutdown/stop, production processes, or external API traffic.

Expected behavior

Expected behavior

Diagnostics must ignore the exact private legacy control-socket shape, still list genuine daemon/worker endpoints, and leave both legacy and current kernels operational with unchanged process/socket identity and state.

Prime Agent version

• current diagnostic baseline: bc0fa7606abb3b7af0f765319518d255e6ae553d; - legacy resident topology fixture: 0940833b771fbaf5144451988604b310a96722fd; - prepared compatibility patch: cb50afed353905330e653247c2601d31b6f0b571.

Environment

Linux; Node.js 22; cross-version daemon/worker coexistence during upgrade. Validation uses isolated user/PID/mount/net/IPC/UTS/cgroup namespaces and real local Python kernels with a faux provider, so no production endpoint or credential is used.

Additional context

Contribution context request

This is a Discussion-first report under CONTRIBUTING.md, not a submitted PR. Please confirm that the report is accepted, explicitly invite implementation/submission, and state whether the GitHub account that will open the PR is maintainer, collaborator, or vouched. Commit authorship is a separate decision and will not be attributed to a person without explicit consent and a GitHub-linked email.

[awhite0030]

Investigated this report and produced a candidate fix, validated on a fork of this repository.

Root cause, fix and validation from the working notes:

Root cause: The daemon discovery process (prime-agent status/doctor) was exclusively ignoring worker sockets but failing to filter out the private control sockets of legacy Python forkserver processes. Probing these sockets caused the forkserver and its active kernels to exit unexpectedly.

Fix: Renamed isWorkerSocketPath to isIgnoredInternalSocketPath in packages/coding-agent/src/cli/daemon-ps.ts and extended its logic to detect and exclude the legacy prime-agent-forkserver-*/control.sock patterns. Updated corresponding call sites and tests.

Validation: Ran npm run check which passed. Executed npx tsx ../../node_modules/vitest/dist/cli.js --run test/suite/regressions/1808-ignore-forkserver-sockets.test.ts to assert correct socket path classifications. All 3 tests passed. Ran existing daemon-ps.test.ts correctly.

Upstream discussion: #1808

Diff: awhite0030/prime-agent@main...fix-ignore-forkserver-sockets-5319261959539241078
Full pull request: awhite0030#25

The change passes npm run check and the focused regression tests against current main. If a fix for this is still wanted, I can open a pull request.