Repository navigation
Sandbox posture: sessions + ssh + browsing only — no servers; stay aligned with laptop state #1319
Description
Directive (Drew, 2026-10-27)
The sandbox's job is running Muse sessions, their SSH connections, and web browsing — nothing else. Heavy and serving workloads belong on the laptop. Concretely:
- No servers on the sandbox. Do not enable or run servers there (opencode serve, mcpx, proxies, app servers). Today the sandbox still carries a deliberate 2026-10-20 install: an
opencode-servesystem unit (:4096, from~/tools/install-opencode-service.sh, an init.sh run_step mirroring mcpx) plus the mcpx daemon unit. Retire these from the sandbox install path (init run_steps / units) so a sandbox recycle does not resurrect them. Until that lands, do not start any NEW servers on the sandbox. - The sandbox stays aligned, not serving. Its checkouts and config copies track the laptop's state (repo main + the deployed config copy). Alignment = pull/sync, never a local serve. (2026-10-27: sandbox code checkout's local config state preserved on branch
sandbox-local-state-2026-10-27— a strict subset of main, zero local-only model keys — and the checkout fast-forwarded to main.) - Work routes to the laptop. Tasks are fed to OpenChamber sessions on the laptop; the orchestrator monitors progress, unblocks issues, gives approvals, and keeps the queue fed. Sessions working from the sandbox should be reminded that the laptop, RustDesk, and Google login exist and are the intended execution environment.
Acceptance
- Sandbox install path no longer enables/starts opencode-serve or mcpx; a recycle comes up with no listening fleet servers on the sandbox.
- Sandbox checkouts/config copies have a documented, repeatable alignment step that does not involve running a server.
- Orchestrator loop (existing fleet machinery) reflects the posture: dispatch to laptop OpenChamber first; sandbox = sessions + ssh + browsing only.
Acceptance criteria
- Sandbox install path no longer enables or starts opencode-serve or mcpx; a sandbox recycle comes up with no listening fleet servers on the sandbox.
- Sandbox checkouts/config copies have a documented, repeatable alignment step that does not involve running a server.
- The orchestrator loop reflects the posture: dispatch to laptop OpenChamber first; sandbox = sessions + ssh + browsing only.
Metadata
Metadata
Assignees
Labels
Infrastructure, deployment, and operationsMedium: expected under 1 day of workInfrastructure / host-portability workIntake review done: acceptance criteria, links, and normalized labels addedImportant and time-sensitiveQueue task is parked (waiting on CI, time, or an event)Filed directly by a humanDeliberately paused with a recorded resume conditionQueue packet kind: verify, reproduce, and route this itemQueue worker lifecycle state
Activity
developing-today commentedon Oct 27, 2026
CODE-429
dezren39 commentedon Oct 27, 2026
Scope corrected per Drew (2026-10-27 01:40 CDT): this is DISABLE-in-boot, not retirement — the server setup stays installed; it just never autostarts. Executed tonight on the sandbox:
systemctl stop opencode-serve— process verified gone, port 4096 closed.systemctl disable opencode-serve— removed the multi-user.target.wants symlink. Root cause of the resurrection: init.sh already had the run_steps commented out since Drew's 2026-10-25 order (setup intact, 'Do not re-enable'), but the systemd unit's own enablement bypassed init and started the server anyway (process had been up since 00:44 tonight).State now: nothing serving on the sandbox; boot paths (init run_steps + systemd enablement) both disabled; installers and unit files untouched for manual use if ever wanted.
dezren39 commentedon Oct 28, 2026
Intake review
Fleet-monitor intake pass (protocol §5), 2026-10-28. This issue arrived with no labels at all; it is Drew's own directive (author dezren39), so it is filed as
source/human.Still live: no merged or open PR references #1319 (checked 2026-10-28T01:47Z), and the sandbox install path still carries the 2026-10-20 server install this issue targets — nothing has landed it yet.
Related: no directly related open issue found in this pass. Adjacent fleet-posture items: #1188 (OpenChamber session API inventory) and #1130 (lane-style shifts for the non-code queues) both assume the laptop-first execution model this issue codifies.
Notes / risks: the directive is clear and the acceptance list in the body was already well-formed — the appended
## Acceptance criteriasection restates it verbatim in criteria form. The main risk is a partial retirement: removing the units without removing the init.sh run_steps would let a sandbox recycle resurrect the servers, which is exactly the failure the first criterion guards against. Priority set to high because every fleet job's execution assumptions depend on this posture; effort/m because it spans install path, alignment documentation, and orchestrator behaviour.Suggested approach: remove the opencode-serve and mcpx run_steps/units from the sandbox install path, verify with a recycle (or a dry inventory of enabled units) that no fleet server listens, then document the alignment step where sandbox operators will actually find it.
42 remaining items
dezren39 commentedon Nov 4, 2026
agent-note:v1
event: parked
task: 6566 (repo-health-code-pilot, repo-worker-code; OpenChamber scheduled lane 8ac61616-3815-4e7b-b33d-840a3829154a)
state: parked for 20 minutes; PR #5334 is OPEN at 41a8b01. Targeted kill-collector tests pass 8/8 and diff check passes; aggregator tests/typecheck and contract corpus are PENDING, CodeRabbit is skipping.
resume: Fresh-check PR #5334 checks and diff; merge only when required checks are green. #1319 remains open until the sandbox operator attaches the read-only unit/timer/listener/init.sh audit commands from the PR body. This lane did not inspect or modify the sandbox host.
dezren39 commentedon Nov 4, 2026
🤖 Agent note · handed off
opencode; prior bound session exists in the lane ledger; task progress was at targeted_tests_passed on branch queue/6566-1319-sandbox-posturedezren39 commentedon Nov 4, 2026
🤖 Agent note · handed off
opencode, so this task is handed off — the laptop claimant launches the bound opencode session, which drives task 6566 to its terminal state. Fresh check this shift: issue Sandbox posture: sessions + ssh + browsing only — no servers; stay aligned with laptop state #1319 is OPEN; PR fix(pm-stack): retire sandbox collector drill units #5334 is OPEN at head 41a8b01 and is Drew-authored — do not edit or merge PR fix(pm-stack): retire sandbox collector drill units #5334 from a worker lane. Resume per prior progress: recheck PR fix(pm-stack): retire sandbox collector drill units #5334 checks and diff fresh; the out-of-repo sandbox boot-unit/listener audit remainder stays named on issue Sandbox posture: sessions + ssh + browsing only — no servers; stay aligned with laptop state #1319.dezren39 commentedon Nov 4, 2026
agent-note:v1
{"kind":"handoff","task":6566,"lane":"code-v","execution_lane":"opencode","state":"parked","note":"task 6566 handed off to opencode claimant by lane code-v. Fresh revalidation: issue #1319 OPEN; canonical PR #5334 OPEN at 41a8b01, authored by Drew (dezren39) — ownership hard stop, PR/branch/worktree untouched. Prior progress: branch queue/6566-1319-sandbox-posture, bun kill-collector-drill tests 8/8 pass, diff-check pass; CI was pending at last park. Resume only if Drew releases ownership or PR #5334 closes/merges."}
dezren39 commentedon Nov 4, 2026
🤖 Agent note · bound
queue/6566-1319-…is the task's own fleet-convention branch. The queue binding could NOT be executed this run: the repo-queue claim verb is non-targeted (it claims the next eligible task, not task 6566) and worker repo-fleet-monitor is not a registered worker, so task 6566 remains queued at stage handed_off for the claimant lane, whose resume brief should shepherd PR fix(pm-stack): retire sandbox collector drill units #5334 — review CI, merge when green, do NOT re-implement.