mirror of
https://github.com/stablyai/orca.git
synced 2026-11-19 04:23:57 +00:00
* fix(native-chat): terminal restore no longer waits for chats an older Orca saved The open that terminal restore and SSH reconnect wait for now only reads and writes per-chat records. A converted chat an older Orca continued gets its copy record at once; the copy's history converts from the original like any unconverted chat, paged on a reader. A converted chat with no saved revision is compared in the background, one chat per idle step. A chat that fails is recorded with its error and absorbed, so it no longer reruns the pass at every launch; storage failures pause the pass for the next launch instead. Tracked in #42783. * fix(native-chat): keep copies an older build readied early, and pending work across pauses A copy's mapping now clears when its history converts here, so a copy another build readied first is found and kept again from the original. Copy-due markers clear per chat, so a paused pass loses none. A pending copy is found by its original's id, however many times the original moved. Forced copies use a set lookup. * fix(native-chat): bound and absorb the remaining recovery work A copy whose history fails to convert is recorded and left to an open instead of every idle pass. The background compare restores at most 20,000 original rows per chat; a longer chat keeps its turns as a copy. Corruption fails only its own chat. Compare markers are written with the pass, and a later absorbed revision clears an earlier failure. * fix(native-chat): never mark a copy failed for a quit or a lost reader Only a failure of the original's own content records a copy as failed, and never while maintenance is stopping. Corruption after the original's revision read pauses the pass instead of absorbing chats, since it may be this file's. * fix(native-chat): compare a whole chat per idle step and absorb one that crashes its reader Per the council's ruling, the row budget is removed: it forced duplicate copies of long chats whose conversion then did the same full restore. A compare that takes its reader down is recorded and absorbed instead of retried every launch; stopping and contention still leave it for the next launch. * test: give the forgotten-chats maintenance fake the legacy compare step * fix(agent-session-journal): forgotten chats absorb their pending compare; this file's damage leaves a compare pending A chat forgotten while its older-Orca compare or copy waited takes that revision, so the next open never brings it back as a copy. Compare reads this file before the original, and only the original's failures settle as the chat's answer; this file's damage, contention and stopping stay pending. A reader exit gets one retry before it is absorbed, and a compare error no longer skips that launch's idle conversion. * fix(agent-session-journal): a damaged original settles its compare instead of staying pending * fix(agent-session-journal): a message an older Orca queued makes its compare keep a copy * chore: count the compare's card read and the stranded-copy point read in the sqlite baseline * test(agent-session-journal): the copy keeps the older Orca's queued message beside the chat's own * perf(agent-session-journal): a copy due on an unchanged original no longer reads or replays every chat * chore: prune the blocking baselines main's worker-only rules cleared; count the reconcile's split-out files * perf(agent-session-journal): keep the older-Orca reconcile on workers and off whole-profile reads The chat-store path moves beside the journal's own, so the main thread no longer reaches the reconcile chain. A pass over copies due reads only those chats' claims and looks tabs up one at a time, and a compare with no original reads nothing from this file. * perf(agent-session-journal): recording the tab index keeps this file's tabs as a set * perf(agent-session-journal): a missing original ends the compare pass instead of asking about every chat * docs(agent-session-journal): say what a pass over copies due still scans * fix(agent-session-journal): damage to one chat here leaves only its compare pending * docs(agent-session-journal): a pass over copies due scans tabs once per due chat * fix(agent-session-journal): a copy too large for one reader reply is recorded, not retried every launch * chore: drop the tab lookup's cast; prune the sync-fs rows the worker-only move cleared * fix(agent-session-journal): an original that won't open ends the compare pass; a stranded rollback stops the reconcile Drops the statement-count test: the repo doesn't test work counts. * fix(agent-session-journal): a forgotten copy drops its mapping and failure keys * test(e2e): take the chat-store path from journal-host-database * fix(agent-session-journal): name the older-Orca compare in reader crash diagnostics