Repository navigation
Releases
v0.4.3
Full Changelog: v0.4.2...v0.4.3
v0.4.2 - first npm release
First npm release: npm i fast-jev-opencode or opencode plugin add fast-jev-opencode.
Highlights
- v2 adapter reads the payload the host actually sends; pruning verified against a live 2.0.18 session.
- Calibrated thresholds (keepCallThreshold 0.25 / keepResultThreshold 0.15) with one choice question per call, measured on real traffic.
- Credential-shaped tool inputs, and secrets pasted into prose, are redacted before anything is scored.
- Optional per-request reasoning effort, injected only when the host declares support for it.
- logFile for v2, since the host gives plugins no log sink.
- dryRun defaults on: nothing is pruned until you turn it off.
Checks
- Both suites and the engine/adapter benchmark run in CI on Node 22 and 24.
- The published tarball is byte-identical to the artifact both suites passed against.
v0.2.0 - dual host, own engine
Dual-host release, and the point where the engine became ours.
OpenCode v1 and v2 from one package. v2 calls setup(), v1 calls server(), per the official migration guide.
Own engine. src/engine/ is this project's own implementation - no third-party compaction engine is vendored or depended on.
What that bought, over the library everyone else uses:
- Split thresholds - keepCallThreshold 0.5 / keepResultThreshold 0.25. A single 0.5 makes
keepunreachable, because Jev scores a call and its output on different scales; upstream reports the same defect as #56. - Keep-signal guard - a pass that scores >=
minScoredcalls and keeps none is refused. A blanket removal is a bad answer, not a decision. - Credential redaction - tool inputs lose credential-named fields and secret-shaped strings before anything is sent to Jev.
- Result peeks -
peekCharsshows Jev a head/tail excerpt of each output instead of a byte count. - Stub removal by default - a call rated no longer needed keeps its place with the output cut short, so narration never loses its evidence (#65).
- Fail-open everywhere - missing key, timeout, transport error, malformed answer, unfittable history, unreadable config.
Evidence. v1 36 checks, v2 19 checks, engine-to-adapter consistency 24/24, a token-savings sweep, and a size-equalised head+tail baseline (4/5 needed calls kept versus 2/5). CI on Node 22 and 24.
v0.1.0 - OpenCode v1 only
v1-only release. OpenCode 1.x entry (experimental.chat.messages.transform) via the plugin array, before the package became dual-host.
Use this tag only on OpenCode older than 1.18.29, which cannot load an object entrypoint.
"plugin": ["fast-jev-opencode@git+https://github.com/roshan-shaik-ml/fast-jev-opencode.git#v0.1.0"]