entity-loom-v0.3.8

Fixed

• Memory review UI: the Save button now persists the edited content of a daily
  or significant memory. Previously the textarea was looked up with
  querySelector('#mem-edit-${type}-${filename}') against an ID that was never
  set and would have been parsed as a class selector anyway (filenames contain
  .md). Save silently fell through to the load-from-disk branch and discarded
  the user's edits. Save and Cancel are now separate functions; successful Save
  re-renders read mode with the just-saved content; Cancel reloads read mode
  from disk.
• PUT /api/memories/daily/:filename and
  PUT /api/memories/significant/:filename now accept empty-string content.
  Clearing the textarea to start a memory over no longer returns 400.
• Resume and Purge buttons on the setup screen no longer corrupt Windows paths.
  The package directory was rendered into an inline
  onclick="resumePackage('${ dir }')" string literal, where \ in paths
  like H:\Psycheros\... was parsed as a JS escape character, corrupting the
  path before the handler ever saw it. Package rows now carry the dir/name in
  data-* attributes (which don't parse backslashes) with click handling
  delegated to the list container. The same render helper is now used for both
  initial load and post-purge refresh, eliminating the duplicated row template.
• Stale running checkpoint no longer bricks a resumed package. If the process
  died mid-stage (laptop closed, terminal killed, power loss), the persisted
  checkpoint kept status: "running" for the background stage even though
  nothing was in memory — on the next resume, the UI trusted the stale value and
  hid the Start button with no way to continue. On resume, if no stage is
  actually running in this process, the significant/daily/graph stages are now
  normalized to status: "aborted" (preserving processedItems and
  failedItems) so the existing resumable UI path offers the Continue button.
  Setup and convert are excluded — setup is synchronous, convert is cheap to
  re-run.
• LLM client now handles providers that hard-require a specific temperature
  (e.g. Moonshot/Kimi on certain models). Previously, providers that rejected
  any temperature other than a fixed value (commonly 1) would fail every
  request with no workaround, because entity-loom uses task-specific values (0.2
  for graph JSON, 0.3 for significant memories, 0.7 for daily prose). The client
  now sends the task-specific temperature first; if the provider rejects with an
  explicit fixed-temperature error, it learns the required value, retries once
  with it, and uses that value for the life of the LLMClient instance so later
  calls don't pay the retry cost. Range errors ("temperature must be between 0
  and 2") and multi-value errors ("temperature: expected one of [0, 0.7, 1]")
  are not guessed — guessing wrong would just produce another rejection.
  Task-specific temperatures are preserved for every provider that allows them.

---

How to install

This package ships as a scoped source release. Download entity-loom-v*.tar.gz
(or .zip) — those archives contain the entity-loom package tree only.

The archives prefixed Psycheros- are GitHub's auto-attached full-monorepo
source — they contain the whole workspace, useful if you want the companion
packages too.