stet v1.2.0
Features
- Cursor rules reach the review prompt as compact review digests (heading,
❌ Bad: and **Detect:** lines per section), cached in <state_dir>/rules-digest-cache.json. Rules matching the file come before alwaysApply rules, and the budget scales with the context limit (15%, clamped to 1000–8192 tokens) or is set with rules_max_tokens / STET_RULES_MAX_TOKENS / --rules-max-tokens. Dropped sections produce one warning per run (#11, partial)
- Findings must sit on or within 3 lines of a line the diff added; issues in unchanged context lines are no longer reported (#21)
info/nitpick findings that only praise the change, with no suggestion, are dropped, and the prompt tells the reviewer not to emit them (#25, partial)stet run prints a warning to stderr (even with --quiet) when review flags such as --strictness or --verify had no effect because there were no new hunks (#23)
Fixes
stet benchmark and the run summary show throughput for OpenAI-compatible providers such as LM Studio, labelled end-to-end (#4)- One JS/TS extension list (
.js .jsx .mjs .cjs .ts .tsx) drives both function-context expansion and symbol lookup: .jsx files now get symbol definitions, and .mjs/.cjs files get enclosing-function context (#16)
- A JS/TS function whose body contains an arrow callback keeps its full enclosing scope (#24)
- JS symbol lookup picks up functions passed as arguments and names bound by
import/require, and skips built-in method names (#19, partial)
- The spine contract script's recovery
stet start uses the requested strictness (#22)
Documentation
- The configuration contract and README document
rules_max_tokens and rule digests
docs/review-process-internals.md describes the current rule ordering, JS/TS extension list, praise and added-line filters, and truncated-response handling
Install
curl -fsSL https://raw.githubusercontent.com/beettlle/stet/main/install.sh | STET_RELEASE_TAG=v1.2.0 bash
