fix object issue#32
Merged
Merged
Conversation
Kitenite
force-pushed
the
object-issue-m78z
branch
from
8f856a4 to
0ca58fb
Compare
Kitenite
force-pushed
the
object-issue-m78z
branch
from
0ca58fb to
91d9d1b
Compare
Kitenite
added a commit
that referenced
this pull request
Feb 11, 2026
Every error response now includes a machine-readable `code` field (SESSION_NOT_FOUND, WRITE_FAILED, FINISH_FAILED, INVALID_BODY, etc.) for deterministic client error handling.
Kitenite
added a commit
that referenced
this pull request
Feb 11, 2026
Kitenite
added a commit
that referenced
this pull request
Feb 11, 2026
…rhaul (#1391) * Reduce stream latency * Smooth stream * Smooth stream * fix(streams): fail /generations/finish when producer had background errors Track producer background errors per session. finishGeneration now flushes, clears per-message seq state, and throws if any producer errors occurred during the run. The finish route returns structured error response (code: FINISH_FAILED) instead of silent success. * refactor(streams): extract producer error helpers for clarity Extract recordProducerError and drainProducerErrors as private helpers. Simplifies the onError callback and finishGeneration method, making the error lifecycle (record → drain → throw) explicit. * fix(desktop): check res.ok for /generations/finish and send messageId Non-2xx responses from finish are now logged with the response body. The finish request now sends the messageId so the server can clear per-message seq state. * fix(streams): await producer flush and detach in deleteSession deleteSession is now async and awaits producer.flush() then producer.detach() before cleaning up session state. Prevents returning 204 while queued chunks are still in flight. * fix(streams): flush producer before reset control event Ensures all queued chunks are durably written before the reset event is appended, preventing reset from racing ahead of buffered data. Also clears producer errors on reset. * fix(streams): route all writes through producer for global ordering Extract appendToStream helper that prefers the producer when available, falling back to direct stream.append. writeChunk, writeUserMessage, and writePresence all use this single write path now. User messages and presence flush immediately for durability while streaming chunks remain buffered. * fix(desktop): add abort signal to chunk POSTs for fast interrupt Pass the agent abort signal to streaming chunk fetch calls so that interrupting an agent cancels in-flight chunk sends immediately. AbortError is silently swallowed since it's the expected outcome. * fix(desktop): emit error event when generation finish fails If /generations/finish returns non-2xx or the network request fails, emit an explicit error event so the UI shows a visible failure instead of silently appearing done. * fix(streams): guard session delete/reset with per-session mutex Add a promise-chain based per-session lock so concurrent delete, reset, and close operations serialize rather than race. Prevents interleaved lifecycle transitions from corrupting session state. * fix(streams): write user messages directly to stream for txid immediacy Flush the producer first to preserve global ordering, then append the user message directly to the stream. This avoids producer queue latency that was causing txid timeout errors on the client side (5s default timeout in stream-db awaitTxId). * perf(desktop): remove /generations/start round trip Generate messageId client-side with crypto.randomUUID() instead of blocking on POST /generations/start. Eliminates a full HTTP round trip before the first token can stream. * perf: batch chunk sends to reduce per-chunk HTTP overhead Add writeChunks method to protocol and POST /chunks/batch endpoint that accepts an array of chunks in a single HTTP request. On the desktop side, replace the sequential per-chunk POST chain with a ChunkBatcher that coalesces chunks within a 5ms window (or 50-chunk max) before sending as a batch. This reduces HTTP round trips from N to ~N/batch_size during active streaming. * Update docs * perf(streams): tune producer lingerMs and add flush timeout Reduce producer lingerMs from 5ms to 1ms since the desktop ChunkBatcher already coalesces at 5ms — avoids double-buffering latency. Add 10s timeout to flushSession so flush/finish cannot hang indefinitely on a stuck producer. * perf: skip Zod on batch endpoint, add bounded queue to ChunkBatcher The /chunks/batch endpoint now does a lightweight array check instead of full Zod schema validation on every chunk — this is an authenticated internal path from the desktop client. ChunkBatcher now has a maxBufferSize (default 2000) that drops oldest chunks when the buffer exceeds the cap, preventing OOM when the network or proxy is slower than the agent. * Add retry with exponential backoff for batch sends (#21) ChunkBatcher now retries failed sendBatch calls up to 3 times with exponential backoff (50ms base). sendBatch callback throws on non-ok responses so the retry logic can catch transient failures. AbortError is rethrown immediately to respect cancellation. * Add producer health tracking with sync fallback (#23, #24) Track per-session producer health. When a producer fires onError, mark it unhealthy and route subsequent writes through direct stream.append instead. A successful flush restores healthy status. Prevents cascading failures when the producer is in a bad state. * Track active generation per session for single-writer enforcement (#25, #33) Add startGeneration/getActiveGeneration/finishGeneration lifecycle to protocol. Chunk routes auto-register the generation from the first chunk if none is active. finishGeneration clears the active generation. Reset and delete also clean up generation state. * Add sessionId and messageId to all route responses (#34) Include sessionId (and messageId where applicable) in both success and error responses from chunk and session routes for tracing. * Add structured error codes to all route responses (#32) Every error response now includes a machine-readable `code` field (SESSION_NOT_FOUND, WRITE_FAILED, FINISH_FAILED, INVALID_BODY, etc.) for deterministic client error handling. * Add structured error codes and sessionId to auth routes (#32, #34) * Format: lint fixes across streams and desktop * Update recommendations doc: mark completed items 30 of 51 items now done. Remaining items are larger architectural changes (14, 29, 31), operational (16-18), and observability/test/ rollout work (35-50). * Add perf * Chunks * Remove /generations/start endpoint (#29) Generation is now auto-registered from the first chunk written. Desktop generates messageId client-side, so this endpoint was dead code. Replaced with chunksBatch in the discovery listing. * Document terminal semantics convention (#31) message-end chunk = UI signal (isLoading → false) /generations/finish = server lifecycle cleanup (flush, seq clear, error drain) Both are required and always sent in that order. * Fix restore * Refactor * More splitting * More split * Fixed feedback * Chunk * Fixed comments * Update CI env
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Description
Related Issues
Type of Change
Testing
Screenshots (if applicable)
Additional Notes