self.md radar — 2026-11-05
Conductor gives its cloud a deletion date
Conductor’s announcement leaves its next product unnamed
it gives the cloud service a deletion-start date: November 20
1. export Conductor Cloud before November 20
Conductor is joining Meta
new sign-ups and app downloads stopped November 4
cloud workspaces become read-only on November 13; users can still read chats and export data
on November 20, the service shuts down and deletion of cloud user data begins
to move a cloud workspace and its chats to a Mac, right-click it and choose Transfer to localSettings → Account → Export data downloads an archive of all your Conductor data
the local app also starts winding down in early 2027
it will receive new models and critical bug fixes until then
active subscribers are promised refunds
reading: Conductor’s shutdown timeline and export FAQ
2. HACS moves into Home Assistant; its plugins remain unreviewed
Home Assistant 2026.11 brings HACS inside the application as the Marketplace
existing installations migrate on startup, carrying installed projects, custom repositories and update entities with them
separate HACS installation ends on this version
the catalog can be browsed, installed from and updated without a GitHub account
custom repositories still need a GitHub connection
the official Settings entry does not mean the plugins have been checked
the documentation says community content is neither tested, audited nor supported by the project, and can access the home, its data and the host system
installs and updates, including automated ones, are blocked until the warning is accepted
a community integration with the same internal identifier as a built-in one can replace it after a restart
Home Assistant’s subsequent fixes to that built-in integration then stop reaching the installation
reading: Marketplace installation, replacement behavior and HACS migration
3. SlackBench’s agents do better when handed the files
same model, same harness, different way into the Slack history
in SlackBench’s paired experiment, Codex CLI with GPT-5.5 scores 45.52 through MCP tools and 59.90 with direct terminal access to exported workspace files
every tested pair improves overall
the authors observe that message-search responses carry metadata that swells the prompt
file access lets agents read whole passages directly
the workspace contains about 3,000 messages across 15 channels, tied to 14 GitHub repositories and 9 Google Docs
it was built by humans and agents performing research projects, then manually cleaned up; it is not an untouched company’s Slack archive
the comparison does not establish a win on every task or test preservation of live Slack permissions
reading: SlackBench’s data construction and access-method comparison
Opus 5.5 in Claude Code gains much less: 72.65 with MCP, 73.68 with files
those overall scores give equal weight to four task categories, rather than pooling all 140 tasks