Repository navigation
One malformed session_start recording bricks all --continue resume for a project (SessionDiscovery.listContinueTargets throws on any unreadable session file) #3732
Description
Activity
LLxprt Issue Planner — infrastructure failure
The planner did not produce output. Please inspect the workflow logs and re-run once resolved.
Still reproduces on
0.12.0-nightly.261009.6445094cf(installed build). After upgrading,--continuecrashed in several projects with:Error: Cannot discover continue targets: .../chats/session-2026-10-08T21-18-08-6a0e7af5-520.jsonl: Invalid session_start: missing or malformed required fields at listContinueTargets (.../bundle/llxprt.js:219853:22)New detail: these files are not damaged by crashes. Our own writer produces them. The bad headers are well-formed JSON with
"provider":"unknown","model":"", followed by a normalprovider_switchand a full conversation.buildNewRecordingServiceincliSessionBootstrap.tsrecordsconfig.getProvider() ?? 'unknown'andconfig.getModel()before the profile's provider/model is applied, andhandleSessionStartinReplayEngine.tsthen rejects the empty model.Count on one machine: about 20 such files across 10 project dirs, written between 2026-09-14 and 2026-10-08. Each one makes bare
--continueand--continue <id>fatal for its whole project. Projects without one (for example uqm, branch-2) resume fine, including codex/gpt → claudecode/opus switches.Two fixes are needed: discovery should skip unreadable candidates instead of throwing (the original report), and the recorder should not write a
session_startwith an empty model.- added a commit that references this issue
on Oct 10, 2026 - added a commit that references this issue
on Oct 11, 2026
Summary
--continue(bare) and--continue <explicit-session-id>both crash withCannot discover continue targets ... Invalid session_start: missing or malformed required fieldswhen ANY single session recording in the project's chats dir has a malformedsession_startrecord. One corrupt file (in this case asession-2026-09-17T20-08-46-*.jsonlleft behind by a crashed/killed session) blocks resuming EVERY session in that project, including by explicit id where discovery of other candidates is unnecessary.Repro (observed 2026-09-18, branch-5 working tree at ~ea297be0c + issue3663 changes)
session_start(observed after a hard-killed session from 2026-09-17).bun scripts/start.ts --continue <valid-session-id> --prompt ....Impact
--continue <id>: the explicit id makes scanning other candidates unnecessary, but the scan still throws, so explicit resume is impossible until the corrupt file is manually removed.--continue: same total failure; a single unreadable file should not hide all healthy sessions.Expected behavior
--continue <id>should resolve the explicit id without requiring every other file to parse (or at worst skip unreadable candidates).Environment
~/Library/Logs/llxprt-code/tmp/<hash>/chats/.Workaround used
Moved the corrupt file out of the chats dir into a quarantine folder; resume works again.