Skip to content

counterparts 0.3.42

Choose a tag to compare

@mlapeter mlapeter released this 31 Oct 02:56
· 818 commits to master since this release

0.3.42 — 2026-10-30

The wake is cleaner: it no longer lists bare ids, "Might be relevant" is gone once a store has its page, and where this folder stands is one block, led by the newest handoff. Yesterday's day summary shows again when a late write-up of another day used to hide it. Turns are faster on a large store and a late one is shown rather than lost. A folder set off, paused or to observe now stays out of the night and the catch-up, and a folder setting file that won't read stops the night and says so instead of thinning the page. No store-format change (store v15, cache v7), and nothing to run after upgrading beyond the usual restart of the dashboard and every Claude Code session. The first background run after updating builds recall's new index by itself, about half a second, no migration, and the cache file grows by about 15 MB. Going back to 0.3.41 needs no store restore: counterparts disconnect, 0.3.41's install, then counterparts connect.

  • The work after each turn takes about a third of the time. On a store of about 5,000 memories
    the background process each turn starts took 3.4 s and now takes about 1 s, mostly from no longer
    opening 60 full memory cards to write the sidebar's file. The file says the same. counterparts log now shows where each of those runs spent its time (runner.steps).
  • Turning a folder off now reaches the sessions already captured there. The night's catch-up,
    counterparts catch-up and the write-up pointer used to take every owed session whatever its
    folder's setting, so a later night could write memories from a folder you had since set off. A
    session in a folder that is off, observer or paused (a subfolder of one included) is now left out,
    isn't counted as owed in doctor or by the night, and shows as such in counterparts coverage. A paused
    folder you resume is owed again until its text ages out. Memories already written stay as they are.
    While scopes.json won't read, write-ups stop, and doctor, counterparts coverage, counterparts catch-up
    and the session-start line all say so and name the file.
  • counterparts verify --rebuild --keep-vectors no longer reports a false mismatch. It counted
    archived and superseded rows as unaccounted for and exited with an error; no row was ever missing.
  • The wake says where this folder stands in one place. The last chapter here, the last
    thing said here and the handoffs are now one block, grouped by session. The newest handoff
    leads, up to five lines and always with what it says comes next, with a newer chapter named
    under it, and the next two handoffs get a line each.
  • The wake no longer lists bare ids, and "Might be relevant" has left it. A count points
    at a recall that lists the items by title: still open now also lists what's coming up,
    and a new handoffs here lists this folder's handoffs to expand or retire. Coming up shows
    the oldest overdue items by title and how long ago they were due. Recall already offers
    what might be relevant on every turn, with ratings. A new store keeps the lane until it has
    its first self page, since that is all its wake has of its memories.
  • Recall stops throwing away finished answers. About one turn in twenty showed no memories because
    recall had passed its time limit, even though the answer was already done and showing it cost about
    5 ms. Those turns now show it, and the record says how late they ran.
  • Recall is faster on a large store, with the same results. On about 5,000 memories, recall itself
    went from about 415 ms to about 110 ms when the store hasn't been read for a minute or so, and the
    whole step before your message is sent from about 725 ms to about 400 ms. The first background run
    after updating builds an index for this, about half a second, and the cache file grows by about 15 MB.
  • Yesterday's summary isn't hidden by a late write-up of another day. A session's write-up filed
    days later, from a folder since set off, used to make every wake that morning fall back to the
    chapter list. The summary now holds back only for what day's end is handed: that day's chapters
    and memories, and what it cites, from a folder set off, paused or to observe. When it
    does hold back, it leaves a row saying what kind of thing held it, with no words and no folder.
  • The night no longer reads a folder set off, paused or to observe. Day's end, the dream, the
    reflection and the tagging aren't shown its memories or chapters, can't find them through recall, and
    can't merge them, so what's closed stays out of the page, the day's summary and new merges. What's
    already stored stays stored, and your own sessions' recall hasn't changed.
  • A folder setting file that won't read now stops the night instead of thinning it. With scopes.json
    broken, the night would have read almost nothing and could have rewritten your page from that. It now holds:
    no page, dream or reflection, doctor goes red and names the file, and it runs again once the file is fixed.
  • A reflection sent back to rewrite part of its work still can't read a folder set off, paused or to
    observe.
    When a session ran its own reflection, those folders opened to its recall at the first
    finish, even with a page or share still to send again. They now stay closed until the part lands or the
    session has been quiet for half an hour. A reflection with nothing sent back gives the session its
    normal recall at once.
  • The night can leave one line of its own. After the reflection revises the page, it may add one
    or two sentences for whoever wakes tomorrow: what stays from the night that the page doesn't say,
    told as what happened. recall "last night" shows it; the wake doesn't yet. It's never cut: past
    300 bytes it's asked for once more, then left empty, and an empty line is fine.
  • Chapter summaries about work fade sooner. A new summary of an unmarked or work chapter starts
    at 0.25 instead of 0.4, so one nobody ever uses stops turning up in ambient recall after about 3
    lived days instead of about 24. A chapter marked about you, the two of you or the assistant keeps
    0.4. Replaying the last 15 days loses no day a memory was used. Summaries already written keep what
    they have, and the dream, the reflection and the wake's chapter lines read summaries by chapter, so
    they see the same ones. On a store with no self page yet, "Might be relevant" no longer lists an
    unmarked or work summary.
  • The log says more about what recall and the wake decided. Each recall turn names up to three
    memories that almost showed, and what each shown memory's words, meaning and bar came to. A session
    start records the bytes of each wake section and how much of the page it printed. Wake renders list
    the ids each section kept, and refresh renders no longer carry the list of trimmed ids. Ratings
    and reflections count why a return was refused, and a dream's opening row says how far down the
    queue the night reached. Ids and numbers only, on rows that were already being written.
  • A memory the dream folds or merges keeps the memory ids its parts pointed to. A note that named
    another memory by id would lose that pointer when the night folded it, so the memory it pointed to stopped
    coming along. New folds and merges now keep those ids on a last line, up to four, and it comes along as
    See also again. What's already folded stays as it is.
  • A memory that names another by id no longer brings along one from a folder set off, paused or to
    observe
    , unless you're working in that folder. Your own recall of it hasn't changed.