OwnMem 0.8.2

[0.8.2] - 2026-10-11

Fixed

• init no longer writes the instruction block into a CLAUDE.md that imports
  AGENTS.md. Claude Code inlines @AGENTS.md when it sits on a line of its
  own, so a repository that keeps one shared rule file already delivers the
  AGENTS.md block to Claude; the second copy in CLAUDE.md loaded the same
  instructions twice in every session. With both hosts selected and the import
  present, init --check reports a leftover block as drifted and
  init --update removes it, leaving the rest of the file byte for byte.
  Without Codex among the hosts, or with @AGENTS.md only mentioned inside a
  sentence, CLAUDE.md keeps its own block as before.
• The Edit/Write hook no longer drops an answer that repeats part of what the
  session already saw. When a later edit quotes a memory the session was
  already given and brings a new pointer with it, the daemon withholds the
  repeat and answers with the remainder, flagged deduplicated: true. The
  response contract pinned that flag to a null result, so the client rejected
  every partial repeat as invalid and failed open: nothing reached the agent,
  while the daemon had already recorded the answer as delivered. In the first
  ten days of 0.8.0 one repository lost 38 hook answers this way, logged as
  PRETOOL_FAIL_OPEN in .local-test/memory-hook/diagnostics.jsonl. The
  contract now reads deduplicated as "something was withheld" and lets the
  result be null or the remainder.
• A query that is one repository path and nothing else is classified as a
  path, whatever its segments spell. A path through a dated work-package folder
  (WORK-20261009-...) matched the error-code pattern, and a file named after
  a ruling matched the decision pattern, so the hook's path-only query turned
  mixed: every lexical lane ran, use-time trust validation fanned out through
  the graph, the hook's P95 rose from 31ms to 127ms on those edits, and the
  n-gram lanes handed back pointers that shared a syllable with the path. Such
  a query now takes the strict path plan like any other file.
• report no longer files an opened pointer as a full-text open with no
  delivery. The consumed-without-delivery data gap was keyed on the prose
  funnel, which a pointer-tier recall never enters because it is recorded as
  an abstention, so every pointer somebody opened was reported twice: once in
  the pointer funnel as an open, and once as an open nothing had delivered.
  The gap now counts only an open on a trace no surface delivered. The
  delivery.fulltext_open_without_delivery_events figure changes meaning
  accordingly; its name is unchanged.
• health splits the hook's unreachable memories into the ones that name an
  existing repository file in their text (anchorable, so a path anchor is
  missing) and the ones that name none (nothing to anchor). The report keys
  evidence.path_anchor_reach.unanchored_naming_file and
  unanchored_no_file replace unreachable.
• The instrumentation health check separates two readings it used to file as
  defects. A payload field whose producer was removed, such as
  recall.consumed.authority_followed, is listed as a retired field still
  present on historical rows, with its sample count taken from the rows that
  carry it rather than from every row of the event type. A component that
  emits only when something goes wrong or somebody acts, such as
  memory-outcome or memory-recall-previous, is listed as idle by design
  for the window, with the measurement that already covers it, instead of as a
  path that may never run. Both registers are printed with their reasons; an
  undeclared constant or silent component is still a finding.