ral does not suspend

A stopped child is resumed, at once, by whoever waits on it: Stopped → SIGCONT, one rule, every role, every platform that has stops. There is no park, no job table, no fg/bg/jobs/disown, no Ctrl-Z gate for stage threads, no stop witnessed by the anchor. Background work is spawn, which already returns a value where bg returned a line in jobs.

Why

Stop/continue job control is the 1980 BSD answer to “I started something long in the foreground and want my prompt back” on a terminal you could not open another of. None of ral’s three programs need that: exarch never suspends anything it runs, synod runs the shell in a VM, and a REPL user has spawn, a second window, and Ctrl-C. It was also, empirically, the single largest source of complexity and defects in ral’s process layer — every review finding logged against the pipeline collector this year was a stop-tracking seam: a level with one edge, a stale edge after settlement, a parked latch that never cleared (see the “risks accepted” and “an early draft got exactly this line wrong” entries in event-driven-pipeline-collector).

The record elsewhere agrees. fish implements job control in full, on one reaper and one job table, and it has been a perpetual defect source for expert maintainers (fish issue #3952: SIGTTOU/foreground-restore races), to the point of disabling job control outside interactive mode entirely (fish PR #8021). nushell shipped for years with none of it, and the failure mode of doing nothing — Ctrl-Z stops the foreground program with no way to resume it (elvish #988 is the identical bug) — eventually forced a minimal freeze/unfreeze in nushell 0.103 (PR #14883), which promptly grew demand for more (#15196, #15202: resume without refocusing, a jobs list). Nobody in this small sample has found a middle point that stays small.

The design space this leaves is binary, not a dial. Either stops are tracked — which costs the gate, the levels, the park, and the table whether or not bg/fg exist as commands, because the expense is tracking a stop, not resuming one — or stops do not exist here at all. The second, done right, is one row of the collector’s fold: an ownerless stop answered with SIGCONT and nothing else.

The mechanism, and why it is the one chosen

Clearing VSUSP on the tty is not enough by itself: vim reads a literal ^Z keystroke and calls raise(SIGTSTP) on itself regardless of the tty’s special-character table, so a “stop cancels the pipeline” rule would kill the editor on a habitual keystroke — elvish’s bug, with worse consequences, and exactly what drove nushell toward freeze/unfreeze instead.

Answering every stop with SIGCONT is total, and that totality is the point: vim stops, is resumed within the same wait loop, and redraws — Ctrl-Z visibly does nothing, which is the honest message for a shell that does not park. kill -STOP sent from outside is undone the same way, correctly, for the same reason. SIGTTIN (a background process touching the controlling tty) would in principle cycle stop → CONT → re-read → stop again, but with no bg there are no backgrounded, tty-attached pipelines to begin with: a spawn worker never receives the tty (route_parent_stdin gives it Null/Empty instead), so the cycle’s precondition cannot arise. The invariant lives where the routing decision is made, not where the signal would land.

The anchor goes one step further and ignores SIGTSTP/SIGTTIN/SIGTTOU outright (core/src/runtime/pipeline/group.rs), so it essentially never stops and never needs resuming — one fewer party the one rule has to reach.

Correction, 2026-09-04: what remains of “stop” in the code is one arm, once, not two — reaper::scan_one (core/src/process/reaper/unix.rs) answers a stop with SIGCONT and posts nothing for it, so no subscriber, pipeline or standalone, ever sees one.

Consequences

Deferral goes with the park it existed for. A write could formerly outlive the frame that staged it only because a stopped pipeline could sit parked, unresolved, for an arbitrary human-scale interval; once no wait can stall on a stop, no write can either. WriteFate::Defer, WriteOutcome::Deferred (and its wire word "deferred"), and StagedWrite are gone. PendingWrite (core/src/runtime/command/redirect.rs:63) now carries its own abandoning Drop — commit-by-value on an Option<Staged>, so the ordinary drop path is a no-op once commit/abandon has already emptied it, and the only remaining path is “nobody committed, so clean up the temp file.”

GroupRole collapses to a bool. Foreground versus Background was a distinction job control drew — a background job doesn’t own the terminal, so fg exists to promote it. With no promotion, the only question left is owned: bool, and terminal ownership itself moves to whichever site holds the TerminalPlan that decides it, rather than being read off the role.

What stays, because it was never about stops. ForegroundGuard / tcsetpgrp — vim still needs the terminal, whether or not it can ever be stopped and resumed by a job table. TerminalLease. PipelineRelay (Ctrl-C). The anchor as pgid-keeper and Ctrl-C witness. PgidPolicy. TtyInputPermit. Cancellation with grace. Thread stages. The collector’s fold itself, which shrank by exactly the rows this decision removed rather than being replaced.

Risks accepted

• A user who genuinely wants fg. The answer today is spawn for work meant to run in the background from the start, and Ctrl-C followed by a re-run under spawn for work that was not. If this proves wrong in practice, nushell’s minimal freeze/unfreeze-to-foreground is the fallback shape — but it buys back the gate and the levels, so it should be resisted until real use, not speculation, demands it.
• A program that must be stopped — a debugger target, kill -STOP for inspection from outside. Not a shell’s job under this decision: spawn it, and inspect it from outside ral with the operating system’s own tools.
• SIGTTIN cycling is impossible only as long as nothing ral launches reads the controlling tty from a non-foreground process group. route_parent_stdin and ForegroundDecision enforce that today; the invariant is named here because it is the one precondition the whole mechanism leans on, and it is enforced at the routing site, not the signal site.
• Windows has none of this — no SIGTSTP, no process groups in the Unix sense — and loses nothing by the decision; several cfg(unix) branches simply disappear rather than needing a Windows equivalent.

Non-goals

• No per-command process groups, subreapers, or cgroups for an unconfined run; a killed external’s descendants remain the group owner’s to end, as in every POSIX shell.
• No change to terminal ownership (tcsetpgrp), the terminal lease, or Ctrl-C handling — those are the parts of “job control” that were never about stop/continue.

Superseded and unamended

Supersedes, in part, a-stage-ral-stopped-has-no-failure: its forgiven-death rule for a reader-gone kill is unchanged, but its two clauses describing a parked pipeline no longer hold — there is no park to abandon held read ends, and a stop no longer wedges a collector until fg.

Amended by one-delivery-path: PipelineRelay is gone from “What stays” — Ctrl-C reaches a pipeline’s processes through the collector alone. The anchor stays, as pgid-keeper and as the witness of what the kernel already delivered to the group.

See stages-are-threads, event-driven-pipeline-collector, terminal-foreground-ownership, terminal-lease, pipeline-execution, cancellation, io-process, repl; docs/SPEC.md §11.6, §7.6.