Repository navigation
Immutable
release. Only release title and notes can be modified.
Patch release: a manager restart no longer closes healthy long workflow runs,
subagents cancel stops a run's whole process tree, sleep cycles stop reading
other agents' transcripts, and refused Slack channel reads now fail visibly.
Fixed
- Manager restarts closed healthy workflow runs (#1151, #1124, MOD-434). The
restart reconciler closed any live run paststarted_at + timeout + 300s. For
a workflow run that timeout is a flat 3600s default the runner never enforces,
so each restart failed long runs that were still within their step budgets. The
runner now arms a deadline from the current step's timeout before each piece of
step work that can block, and the reconciler closes a run only after the later
of that deadline andstarted_at + timeout, plus the grace. subagents cancelleft a run's processes running (#1148, #1122). Cancel
signalled only the runner, so the claude CLI and its tool commands kept writing
afterCancelled. It now sends SIGTERM to the recorded process and every live
descendant, sends SIGKILL to what remains, and blocks until they are gone. Exit
1 means something survived and is named, or that the cancel was refused with
nothing signalled.SessionEntrygains an additivepid_startfield in
state.json, so a recycled pid is never signalled. The reconciler's deadline
branch uses the same kill.subagents liststops listing the manager, and the
runs view stops counting a finished run as running.- Sleep cycles read other agents' transcripts (#1144, #1142). Agents sharing
an OS user share one Claude projects tree, so a sleep cycle could copy another
agent's transcripts into long-term memory. The Claude brain now records each
session id it runs, and history indexes and returns only those sessions.
Existing cursors stay valid, and an agent upgrading with an existing
history.dbclaims that database's sessions by exact working directory. - Refused Slack channel reads looked empty (#1149, #1140).
bobi read-conversationon a ref with no thread printed[]and exited 0 when Slack
refused the read. It now fails with HTTP 502 and exit 1. Login bootstrap's
ask-id recovery fails the run with its marker intact instead of posting a fresh
ask. Channel reads with no thread also follow Slack's cursor past 200 messages
and keep the newestlimit. These changes are in the event server. - Tests overwrote the developer's global instructions (#1145, #1143, #1150,
#1147). Unit and stub-brain integration tests now run under a per-testHOME
withCLAUDE_CONFIG_DIRandCODEX_HOMEunset. Tests markedclaude,liveor
docker, plus e2e, keep the real home because they need real credentials.