@rahularya01/pi-cursor 1.4.18 → 1.4.20

package/CHANGELOG.md

@@ -1,5 +1,26 @@
 # Changelog
 
+## [1.4.20] - 2026-08-18
+
+### Fixed
+
+- **An interrupted assistant turn replayed as one that simply trailed off.** Cursor's turn structure carries only the text a turn produced, so a turn that was aborted, errored, or truncated came back on the next request indistinguishable from a model that chose to stop — pi's `stopReason` and `errorMessage` were dropped entirely. Resuming a session after an interrupted turn therefore looked to the model like missing context rather than incomplete work: in the session that surfaced this, "continue please" sent the model listing the workspace and reading unrelated agent transcripts, from which it confabulated a prior discussion that never happened. Completed turns are untouched; a turn that did not finish now carries an explicit trailing note (`[pi-cursor: this assistant turn was interrupted before it finished; …]`) placed after any tool calls it managed to emit. Error detail rides along, redacted through `redactSecrets()` and capped at 200 characters. A trailing interrupted turn — the one being retried, not history behind us — is deliberately left unannotated so the live user text is not stranded.
+- **Conversation journals silently dropped all but the newest 64 blobs.** The live blob store holds up to `MAX_ACTIVE_BLOB_ENTRIES` (512) entries, but the on-disk journal that survives a restart persisted only 64. A checkpoint addresses its history by blob id, so a restored checkpoint referencing an evicted blob asked for content that was gone — and `getBlobArgs` answers a miss with an empty result, which is indistinguishable from an empty blob. The conversation came back structurally intact with its older turns blank, with no error at any layer. Journals now persist the whole store under a byte budget derived from the record's actual checkpoint size, and a record that still could not fit everything is marked so the reader drops its checkpoint and rebuilds from pi's transcript instead of resuming with holes. Roughly 15–20 turns was enough to cross the old cap.
+- **A blob miss is no longer invisible.** `getBlobArgs` for a blob we do not hold now records `kv_blob_miss` in the lifecycle log and in `/cursor.doctor`'s `lastStreamEvent`, so this class of silent history loss is diagnosable from the sanitized log alone.
+
+## [1.4.19] - 2026-08-17
+
+### Performance
+
+- **Stopped reading whole journals to check one timestamp.** The run-journal TTL sweep runs once per user turn, on the main thread, and read plus `JSON.parse`d every journal file — megabytes of base64 blobs each — only to look at the `savedAt` field near the front of the record. It now reads a 512-byte head through a single file descriptor, falling back to a full parse only when the head does not carry the field; stale journals are rejected from the same head instead of decoding their blobs first. ~138x faster over a 25-journal cache directory (896ms → 6.5ms).
+- **Image dedup no longer hashes every payload.** `mergeImages()` sha256'd every image on every history replay. Duplicates must agree on MIME type and byte length, so shapes are bucketed first and only images colliding on shape are digested, with digests memoized by buffer identity. ~700x faster replaying a 12-image transcript, and ~5x on cold buffers with distinct sizes.
+- **Removed a `process.env` read from the per-token path.** `debugLog()` re-resolved `PI_CURSOR_PROVIDER_DEBUG` on every call, including once per streamed token. `process.env` is a native-backed proxy costing roughly 100x an ordinary property read, which made it the single largest cost in `processServerMessage()`. The flag is now resolved once, lazily. ~8x faster per server message.
+- **Fixed O(n²) model catalog building on the activation path.** `hasVariantParameterSet()` re-normalized all of a model's variants for each of that model's variants, so catalog building scaled quadratically in string sorts. The advertised parameter sets are now derived once per model. ~5x faster, taking the whole activation model chain from 3.75ms to 1.0ms. The canonical key sorts also moved off `localeCompare`, which was needlessly slow and could rank distinct strings as equal — leaving the key dependent on input order.
+- **Dropped dead work from side-channel splitting.** `splitUserTextAndSideChannel()` evaluated seven regexes and seven full-string trims per user message inside an `if` block with an empty body, then discarded the result.
+- **Memoized journal blob encoding.** A conversation's blobs are carried over unchanged across the journal writes triggered by each tool-call pause, so their base64 encoding is now cached by buffer identity rather than recomputed per write.
+
+No behavior change; startup module load was measured and deliberately left alone (the extension's own bundle imports in ~13ms — the rest is pi's peer dependencies).
+
 ## [1.4.18] - 2026-08-16
 
 ### Added