Add cluster.disable_console_workers (reverse gh-602's console-routing half; preserve controller-exclusion)

[lukstafi]

Goal

Add a cluster.disable_console_workers boolean config option (default false, full back-compat). When true, machines whose role is console are never selected to run worker slots, and the keepalive never auto-starts slots on them — freeing the console machine (currently macbook-pro) for interactive/SSH use.

Policy: a faithful 180° of gh-ludics-602's console-routing half

gh-ludics-602 ("self-orchestration breaks the controller") has two halves for ludics self-modifying tasks:

1. Keep them OFF the controller/leader — self-orchestration there repoints the global ludics symlink and boots the dashboard. This MUST stay.
2. Prefer/route them to the console — the escape hatch.

This flag reverses only half (2): the console flips from preferred to forbidden for all worker slots. Half (1) stays.

Net effect with the flag on:

• ludics self-modifying tasks (Orchestrating a ludics self-task on the controller node takes down the live controller (dashboard launchd bootout + global symlink repoint via worktree ludics init) #602's exact predicate, project == "ludics" case-insensitive) are excluded from both the console and the controller/leader. If no other eligible worker exists, selectMachineForSlot returns null — it blocks (awaits manual/SSH handling) rather than ever falling back to the controller.
• non-ludics tasks only lose the console; the controller stays available to them.
• flag off (default): Orchestrating a ludics self-task on the controller node takes down the live controller (dashboard launchd bootout + global symlink repoint via worktree ludics init) #602 behavior is unchanged — console-preferred, controller-protected.

What changed

• Config schema/types + parsing (src/config.ts, templates/config.reference.yaml): added disable_console_workers (YAML snake_case) to the cluster reference schema and the LudicsFullConfig.cluster interface. Parsed in clusterConfig() (src/cluster.ts) into ClusterConfig.disableConsoleWorkers (camelCase TS), default false. New disableConsoleWorkers() helper.
• Machine selection (selectMachineForSlot, src/cluster.ts): hoisted Orchestrating a ludics self-task on the controller node takes down the live controller (dashboard launchd bootout + global symlink repoint via worktree ludics init) #602's isLudicsSelfTask predicate; when the flag is on, console-role machines are removed from the candidate set for every task, and for ludics self-tasks the leader/controller is additionally removed. Empties-to-null (blocks) rather than falling back onto the console or the controller. Applied after the requirement filter.
• Keepalive auto-start (maybeAutoStartSlots, src/mag.ts): skips auto-starting any slot whose assigned machine has role console when the flag is on.
• Tests (src/cluster.test.ts): parse of true/false/absent; non-ludics task keeps the controller but loses the console; ludics self-task with flag ON and only leader+console available returns null (asserts it does not pick the controller), incl. a console-down variant; ludics self-task still routes to a real non-console/non-leader worker; flag-off Orchestrating a ludics self-task on the controller node takes down the live controller (dashboard launchd bootout + global symlink repoint via worktree ludics init) #602 console routing stays green; the no-eligible-machine path returns null without falling back to the console.

Activating config

The federation controller harness already sets cluster.disable_console_workers: true in its config.yaml (currently inert until this merges + deploys).

Operational consequence

With the flag on and the console as the only otherwise-eligible online machine, assignments block rather than running on the console. For ludics self-modifying tasks specifically, when neither a non-console/non-leader worker nor the (forbidden) console/controller is available, the task blocks and awaits manual/SSH handling — by design, this is preferable to re-triggering the gh-602 controller-disruption.

Validation

• bun run build — ✅
• bun run typecheck — ✅
• bun run lint — ✅
• bun run lint:config-reference — ✅
• bun test src/cluster.test.ts — ✅ 41 pass / 0 fail
• bun test src/mag-remote-dispatch-path.test.ts — ✅ 3 pass / 0 fail

🤖 Generated with Claude Code

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 221ca129ad

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

Filter consoles out of health-routing selectors

With disable_console_workers enabled, this new filter only affects slot assignment; routeOrSkipRequirementsUnmet in src/health.ts still calls selectOnlineCapableMachine, which builds its candidate list from all clusterMachines() and returns the first fresh capability match. In a setup where the console is the only online machine matching a project's os/gpu requirements, nightly test-health will still SSH to the console and run the suite there, defeating the config's purpose of reserving the console for interactive use. Apply the same console-role exclusion to the capability selector or share the eligibility filter.

Useful? React with 👍 / 👎.