Two commands that ask a model you name.

What leaves the machine

unswell check, the Go library, the GitHub Action and the MCP server send nothing anywhere. unswell review and unswell trim send the complete text of every selected document to the endpoint or the local program you name, and to nothing else. An API key goes to that endpoint in the Authorization header. It is not written to a report or to an error message.

Without a named model both commands fail before they read a document.

Run them only where your policy lets that text reach that model.

Name the model

An endpoint that implements the OpenAI chat completions API:

export REVIEW_API_KEY=...
unswell review README.md docs \
  --endpoint https://api.example.com/v1/chat/completions \
  --model example-model \
  --api-key-env REVIEW_API_KEY

--api-key-env names the environment variable that holds the key, so the key never appears in a command line. Plain http is accepted for this machine only, and a model served here needs no key.

A local program can be the model. It reads the prompt on standard input and prints the answer on standard output. It runs by its arguments, without a shell, in an empty temporary directory:

unswell trim docs/guide.md --command assistant --command-arg=--print

Paths, --stdin, --filename, --format, --config and --profile select and extract documents as they do for unswell check. The model is asked about the same prose that the rules see. Code blocks and protected identifiers are context.

Review

A rule names a construction. It cannot tell that a sentence repeats an earlier one or that an opening adds nothing. unswell review asks the model to read the whole document and flag items under a fixed rubric of seven categories. Unswell binds every flag to source bytes and drops a flag that does not quote its item.

docs/guide.md:3:1: review [empty_framing] The opening only announces the statement.
    "It is important to note that"
Reviewed 1 document: 1 finding.
Reviewer: example-model at api.example.com. No second model confirmed these findings.
Findings are a model's judgment, bound to source bytes. They are advisory and never gate.

--confirm-model names a second model that answers yes, no or uncertain for every bound flag without seeing the first model’s reasons. The text report then lists the flags it accepted. That gives a shorter list.

Trim

unswell trim asks the model for sentences with empty words removed. The model proposes, and a deterministic check keeps a proposal only if all of this holds:

• Every word of the trimmed sentence occurs in the original, in the same order. Nothing is added, replaced or moved.
• At least one word is removed and at least half of the words remain.
• No negation, number, modal or hedging word is removed.
• No code or link is removed, and a kept word keeps its quotation marks and its emphasis markers.

Only sentences that lie in one piece of source are offered. A sentence with inline code, a link or emphasis inside is left alone, as are headings, tables and code.

docs/guide.md:3:1: trim
    - These features significantly improve the ability to type-check your code.
    + These features improve the ability to type-check your code.
1 trim in 1 document.
Model: example-model at api.example.com.
Nothing was written. Use --write to apply these trims.
A trim only deletes words. A deleted word can still carry meaning that the check cannot see.

Files change only with --write. The command reads each file again first and leaves it alone if it changed during the run or is read-only. Review the change in your version control before you keep it.

The check is exact about words and knows nothing about their meaning. A removed adjective can be the point of a sentence, so read each trim.

Reports and exit codes

--report json:PATH writes the complete result, with schema unswell-review-v1 or unswell-trim-v1: byte spans, the accepted items, and the rejected ones with the reason. --report text:PATH writes the text shown above. SARIF and HTML are not produced.

Both commands exit with 0 when every selected document was processed, with or without findings, 2 when a document could not be processed or the options were invalid, and 130 when the run was interrupted. There is no exit code for findings: a model’s answer differs between runs, so it cannot fail a build.

How they were measured

Both commands were compared with what people edited and judged by a consensus of models. The numbers and the judges are in the documentation of review and of trim.