6 · Pipelines that last

A pipeline that worked once wants to run every week. This chapter turns one-liners into artifacts.

Save it: .sem files

A .sem file pins one verb invocation (or a multi-stage pipeline) as TOML:

#!/usr/bin/env -S smartpipe run
verb = "map"
prompt = "Extract {vendor, total}"
schema-from = "vendor string; total number >= 0"

smartpipe run extract.sem executes it; unknown keys are loud errors, so a typo can't silently change an unattended run. Drop it in ~/.config/smartpipe/verbs/ and it becomes a verb: smartpipe extract. Details: .sem files · custom verbs.

Land the results: write and readable

# translate a folder, mirror it back under fr/
smartpipe 'notes/*.txt' --as lines \
| smartpipe map "translate to French" \
| smartpipe write 'fr/{name}'
# {name} = the source file's name, carried by __source - notes/a.txt becomes fr/a.txt

# a human-readable report of a JSONL run
smartpipe readable < results.jsonl > report.txt

write routes items to files by template. {name}, {stem}, {ext}, {path}, and {index} fill from the item's provenance (the __source spine that remembers which file each item was cut from - {name} is NOT a record field); any other {field} fills from the record's own data, for fan-out like by-lang/{lang}.jsonl. Line-cut items reassemble in order. readable renders the same blocks you see at the terminal, anywhere.

Trust it: doctor and the receipts

smartpipe doctor          # config, models, keys, completions - with fixes
smartpipe doctor --probe  # what your models can actually see and hear
smartpipe usage           # the run ledger: tokens and media over time

Every run already ends with a receipt on stderr (tokens, media, cache hits) - the numbers that go in the report.

That's the track. From here: the cookbook for task-shaped recipes, concepts for how it works, and the CLI reference for every flag.