Skip to content

ADR-v2-082 — Spoken Shell Pipeline Builder (dry-run first)

Status: Accepted (2026-07-02) · Wave J Context links: [[adr-v2-065-terminal-command-safety-gate]] (classifies formed cmd vs builds pipeline), [[adr-v2-076-voice-git-choreographer]], [[adr-011]]

Context

Wave J research (#10) — speak pipeline stages ("list files, pipe to grep error, pipe to word count") → the daemon renders ls | grep error | wc -l into the terminal as text without executing; "dry run" wraps a mutating tool with its native --dry-run/-n; nothing runs until "run it". Distinct from the Terminal Command Safety Gate, which classifies an already-formed command — this constructs the pipeline from spoken stages and defaults to preview. Impossible in cloud dictation. Anchors: NaSh (arXiv 2506.13028), NL2SH (Westenfelder et al., NAACL 2025, arXiv 2502.06858).

Decision

Add an opt-in Spoken Shell Pipeline Builder: [shellpipe] enabled=false. Pure cores: parse_stages(text) splits on "pipe to"/"then"/","; render_pipeline(stages) maps each stage to a command (a curated read-only vocab + grep/filter <quoted arg> via shlex.quote) and joins with |, returning None if any stage is unknown; dryrun_wrap(pipeline) swaps a mutating leading tool (rm/rsync/mv/git clean) for its dry-run form. Dependency-free — it only ever renders text. The NL2Bash-tuned SLM for free-form fallback is deferred. OFF by default.

Consequences

  • Constructs pipelines from speech and defaults to preview — nothing runs without explicit "run it".
  • shlex.quote on interpolated args removes injection risk.
  • Distinct from the Terminal Safety Gate (builds vs classifies).
  • Privacy (ADR-011): pure string assembly; nothing executes here.
  • Caveat: curated stage vocab → free-form stages need the deferred SLM; unknown stages abort to None. Off by default.