lsm / HyperNeo / 31982860859 / 32

79%

dev: 79%

push

github

fix(daemon): stop SDK transcript purges + un-wedge sdk_resume_choice sessions (#2547)

* fix(daemon): pin SDK transcript retention for every spawned session

The SDK/Claude CLI purges chat transcripts idle > cleanupPeriodDays
(default 30) at startup, scanning all of ~/.claude/projects. HyperNeo
keeps sessions active/resumable indefinitely in its DB, so a swept
transcript wedges the session's delivery queue on sdk_resume_choice
with no path to recovery.

Pin cleanupPeriodDays (3650) via the inline `settings` flag-settings
layer for every session the daemon spawns, across all providers. That
layer overrides on-disk user/project settings, so retention is owned
solely by HyperNeo (session archive moves transcripts to
claude-session-archives); the SDK documents a large value as the
sanctioned way to opt out of cleanup.

* fix(daemon): treat bun:sqlite null row as no unresolved action card

bun:sqlite `.get()` returns null (not undefined) for no rows, so
hasUnresolvedHyperNeoAction's `row !== undefined` check inverted its
polarity: no card read as "an unresolved card exists". The
emitSdkResumeChoiceMessage dedupe then suppressed the resume-choice
card entirely, so a session whose SDK transcript was purged parked on
sdk_resume_choice forever with no card to answer (and isWaitingForInput
kept steers from ever burning their park budget).

Use the null-safe comparison already used by isHandleTaken/isNameTaken,
with regression tests against a real bun:sqlite table (verified to fail
on the old comparison).

* fix(daemon): apply transcript retention to every direct SDK launch

Codex review (PR #2547): the retention pin only covered queries built
through QueryOptionsBuilder, but the daemon also spawns SDK subprocesses
directly — title generation, workflow selection, model discovery,
GitHub router/security agents, and the evolution services. Cleanup runs
at subprocess startup and scans all of ~/.claude/projects, so any of
those routine paths could still purge a long... (continued)

6822 of 24936 branches covered (27.36%)

Branch coverage included in aggregate %.

10622 of 30618 relevant lines covered (34.69%)

773.89 hits per line