Repository navigation
bug(ghosts): Codex text ghosts are never superseded (response id vs rollout turn id) #1231
Description
Found in the #1227 review (reviewer A, finding 1).
Codex ghosts are minted with turnId = the proxy response id (resp_…). Committed entries carry codexTurnId = the rollout's turn_context.turn_id (a UUID, providers/codex/renderer/transcript/rollout.ts:242). The two never match, and text blocks have no tool id to fall back on, so reconcileUpstream cannot supersede any Codex text ghost. Every one orphans 30 s later.
- Live panes: these ghosts stay hidden only by render rule 3 (the turn is in semantic history) or rule 4.
- Restored panes: fix(workspace): restored panes read their persisted ghost log #1227 hides them by re-dating each to its
createdAt, which treats the symptom rather than the identity. - Scale: on the owner's machine, 818 logs contain Codex ghosts, and 388 visible text rows would have painted on restore before fix(workspace): restored panes read their persisted ghost log #1227's re-dating.
- Fix direction: the rollout carries the
resp_…id. Every one of these turn ids greps in its rollout. Stamp it on the mapped entry and match on it.
Activity
fix(workspace): restore holds only ghosts the committed tail has not …
Juliusolsson05 commentedon Sep 25, 2026
Scope note after #1227 closed (the on-disk ghost log is being removed): this issue still matters for LIVE Codex ghosts, which reconcile can never supersede. Only the restore half is moot.
Juliusolsson05 commentedon Sep 25, 2026
Re-audit after #1235 (quality loop, 2026-09-25):
Two identity routes, both verified on a real rollout (
rollout-2026-09-25T05-55-26-…):token_usage_recordcarries{turn_id, response_id}(33 in that rollout). A response's output items (reasoning, message, tool call) precede its usage record, and the tool output follows it. The mapper can't stamp items at ingest (no lookahead), but on a usage record the ingest sites (useIpcSubscriptions,initialHistory,reconstructSlices) can supersede ghosts whoseturnId === response_id. This needsMappedTranscriptEntryto carry that fact.response_itemcarries its id (msg_,rs_,ctc_), and the proxy SSE sees the same ids. Semantic blocks don't record them yet (a codex-headless package change), but they would give an exact per-block match.Route 2 is the more precise one.
Juliusolsson05 commentedon Sep 25, 2026
C8 hunt (2026-09-25): still reachable. Codex text ghosts carry the live response id (
resp_…) as turnId, while committed entries carry the rollout turn UUID incodexTurnId, and nothing stamps the response id onto mapped entries. So Codex ghosts are never superseded; they orphan after 30 s and are hidden only by selectMergedEntries rules 3-4 (ghosts.ts:405-436).docs(ghosts): correct three comments that still described the turn-id…
test(ghosts): drive the production mapper; say reasoning ghosts are n…