Skip to content

v3.1.0: restores fixes the 3.0 rollback dropped

Choose a tag to compare

@kitfunso kitfunso released this 01 Nov 19:58
· 105 commits to master since this release
cc59e7b

Changed

  • sleep on the package root resolves after the embeds of the rows its auto-share copies, unless the Context sets the new optional detached keeper. detached takes an object whose detach(work) keeps each promise it is handed; settle them before you close the store. Its type, DetachedWork, is exported from the package root, hippo-memory/server and the entry/sleep and entry/eval-scripts entries.

Fixed

  • hippo promote, hippo share, hippo sync, hippo learn, hippo remember, hippo watch and capture wait for the embeds they start before they return. The command's store runtime now keeps that work and settles it before it closes, so it no longer closes under a running embed. A process run of these commands already waited, because Node stays up for a pending embed; the change shows where runCli resolves.
  • serve() and hippo serve no longer lose the global vector of a memory promoted just before stop(). stop() now waits up to 5 seconds for the embeds a request started, logs how many still run at that bound, and then closes the stores.
  • hippo mcp waits for the embeds its tool calls started before it exits when stdin closes. It exited once the calls in flight answered, which cut off the embed of a row hippo_share had copied to the global store. One 5 second bound now covers the calls and then that work, and a warning names how many tasks still run at the bound.
  • hippo goal push fails with its usage line on a name of control characters alone, and the log guard drops DEL and the C1 controls. A name such as a lone ESC used to become a goal named ?. Goal names now drop control characters instead of storing each as ?, and DEL and U+0080 to U+009F are dropped from goal names and from the hook and pre-compact logs too, since NEL breaks a line and CSI starts an escape in some terminals.
  • hippo import waits for its embedding backfill, and reports a failed one, before it returns. The command, and a caller that runs it in process through runCli, used to get control back while the backfill still ran, so the command's store runtime closed under it and, on Windows, the backfill held hippo.db open past the command.