@vladchatware/dsh-subprocess-docker

Docker implementation of the subprocess seam (ctx.subprocess). Load @vladchatware/dsh-docker first, then this service in place of dsh-subprocess-local; existing Bash, PTY, and LSP consumers then execute in the shared container.

Substrate facts (verified live): each exec is its own process-group leader (the owner's wrapper writes pid == pgid), so tree-scoped termination is one in-guest group kill; tty(true) execs give programmatic PTY text I/O plus resize. Output frames are Docker's multiplexed 8-byte-headered stream, which the owner's parseFrames demultiplexes.

Known substrate limits:

• No signal fact — Docker reports a signal death as exit code 128+N; the provider maps 128+N to the signal and attributes SIGTERM/SIGKILL when it performed the kill itself.
• Foreground-group inspection is best-effort — dockerode exposes no foreground-group primitive; the terminal reports its session leader.
• Collect/spill keeps a bounded host tail and, when configured, a host spill file with offset reads (output.ts).
• Ambient credentials are scrubbed — host DSH_* and credential-shaped names are not inherited; every spec.env entry is an explicit opt-in.

Model Experience

Indirectly, through the Bash executor behind dsh-tool-bash, which renders remote output, exit facts, and spill paths.

KV Cache effect

No direct invalidation; the named consumers own any request-prefix changes.

Known Limitations and Deferred Work

• 128+N exit coding is a convention — unrequested signal deaths are reported only when no wrapper-published code wins; the provider cannot prove a signal it did not send.
• Synchronous-PID consumers are unsupported — pid resolves shortly after start (from the pid file), not synchronously at spawn.
• Private state lives for the container lifetime — process pid files remain under .dsh-dkr until the owner removes the container.