Every version, and what it changed for you.

Four things Archis was reporting about itself that were not true.

Before you upgrade — Run chi doctor after upgrading. If it reports the session briefing as registered but running a binary that is gone, chi hook install --global rewrites it — and Claude Code reads hooks at startup, so restart it afterwards.

• FixedA hook that can no longer run says so, and chi hook install repairs it. The entry Archis registers carries the full path of the binary that wrote it, so renaming your home directory or moving the install left a hook that failed on every single tool call — with a shell error that never mentioned Archis. Every question you could ask about it answered "installed", because all three matched the entry by its text: reinstalling was a no-op, the setup wizard called the step done, and chi doctor reported it healthy. Registered and runnable are now separate questions. chi doctor fails the check and names the binary that is gone, chi hook status marks the entry broken and prints the command that fixes it, and chi hook install writes the dead entry again in place — never a second copy, and never an entry whose binary is still there.
• FixedA repository you are working in reaches the dashboard even when your agent never asks Archis anything. A project was recorded on its first Archis tool call, which is the moment that exists for the server and the wrong one for everybody else: an agent that spends the afternoon reading and editing with its own tools asks Archis nothing, so a repository Archis had indexed — and whose activity it was recording the whole time — stayed off the projects list for as long as the work lasted.
• Fixedget_symbols asked for source now returns source. It answered with the compact projection instead, while charging the request against the session's exact-source budget and recording it as a source-level read, so the only way to get what you asked for was to ask again through get_source — a second turn spent on content the session had already paid for. A batch that runs out of budget part-way now says which symbols it did not serve and why, rather than serving all of them at the wrong level.
• Fixedsession_report counts what compact_select served. It was filed at nothing, and that is the call the usage contract tells an agent to make first — so the one call an exploring session always makes was the one call the accounting could not see. Any reduction figure that report showed was better than the truth by whatever the exploration had cost. Figures you took from earlier versions were flattering; this one counts it.