Skip to content

enzyme v0.13.0

Choose a tag to compare

@jshph jshph released this 23 Oct 02:10
· 34 commits to main since this release

Enzyme release notes - v0.13.0

Short version

Search is now hybrid and fast. enzyme catalyze ranks notes by keyword and
meaning together from a persistent full-text index (median search time on a
~5,600-note vault fell from about 1 s to 63 ms), and enzyme petri --query
ranks the map by what the query is about instead of by recency. Markdown
sources can declare how they date notes with new date rules. Hosted catalyst
generation splits large jobs so they stop timing out. enzyme serve and the
installers got security fixes, and Enzyme now builds and releases from this
public repository.

What shipped

Search: keyword and meaning together

  • catalyze now fuses BM25 keyword ranking with semantic similarity by
    default ("search_strategy": "hybrid"). Names, titles, plain words, and
    concepts all work as queries, with no model call. --routing catalysts
    keeps the old route-through-catalysts behavior. (fed6177)
  • Each result carries the catalyst that best connects that note to the query
    (via_catalyst_text), chosen per result rather than one catalyst for all.
    (823e7e3, fed6177)
  • The keyword side reads a persistent SQLite FTS5 index built during init
    and kept current by refresh, instead of rebuilding an index on every call.
    On a ~5,600-note vault, median catalyze time fell from 1,054 ms to 63 ms
    (p90 1,074 ms to 89 ms). The index adds about 12% to the database. (dc700e3)
  • Measured on 36 real agent queries, judged blind: +0.18 mean relevance over
    semantic-only search (95% CI +0.11 to +0.25), and +0.39 over catalyst
    routing; about twice the strongly relevant notes on cross-domain questions.

The map: petri

  • petri --query ranks entities and catalysts by keyword and meaning
    (0.6 / 0.4) with no recency weighting, and entities the query names rank
    first. Previously the most recently active folder topped almost every query.
    On 96 queries naming an entity, it now ranks that entity first 90 times
    (previously 9). (3a9af35)
  • The relevance field in petri --query JSON is now that blended 0–1 score,
    not raw cosine similarity.
  • Catalyst questions appear without the stored list of source notes, in both
    the tree and the JSON text field. (823e7e3)
  • Prepared catalyze --target folders are listed relative to the vault when
    they are inside it, and dropped once their folder is gone; --target accepts
    those vault-relative paths from any directory. (8535d14)

Date rules (enzyme-spec 0.5.0)

  • A source markdown block can say how its notes are dated, ahead of the
    built-in parsing (4877ebc, 72909f6):

    source markdown "notes" {
      path "~/notes"
      date notes in "journal" by filename as "YYMMDD"
      date entries in "journal" by headings as "YYMMDD"
      date notes by field "day" or "created" as "YYMMDD"
      date notes in "imports" by field "ts" as unix seconds
    }
    

    notes sets created dates, edits modified dates, entries the dated
    entries of an appended log. Rules can be scoped to a folder, and only
    turns the built-in parsing off for that target and scope. The setup skill
    now suggests a rule instead of rewriting files when a vault uses its own date
    convention.

  • A workspace that declares more than one Markdown source cannot declare date
    rules yet.

Catalyst generation

  • Built-in profiles ask plainly about what was decided, changed, or left open,
    in the user's own voice. (f3d57d0)
  • Large hosted jobs are split into calls of at most 30 questions, each seeing
    every excerpt but assigned its own periods. On a test vault at a budget of
    80, delivery rose from 1,036–1,241 to 1,343–1,395 of 1,396 planned questions,
    with no failed entities. Timeouts now apply per call (120 s hosted, 300 s
    local); ENZYME_CATALYST_CALL_TIMEOUT_SECS and
    ENZYME_CATALYST_MAX_QUESTIONS_PER_CALL override them. A split job keeps the
    catalysts from the calls that succeed. (798c44f)
  • Hosted generation forwards the reasoning effort chosen by the hosted config
    to the model. (8efa0e0)

Agent guidance

  • The runtime skill, Hermes/OpenClaw skill, and READMEs describe petri as the
    map and catalyze as local hybrid search. Agents are told to map the arc of
    the user's thinking before anchoring on one note, plan on at least four
    searches for questions about the user's own thinking, carry a catalyst phrase
    into the query when a turn allows only one or two searches, and never quote a
    catalyst as evidence. (b62cc99, 26e9cb2)

Security and distribution

  • enzyme serve binds to loopback only, drops permissive CORS, rejects
    requests not addressed to localhost (DNS rebinding), and refuses absolute or
    parent-relative note paths on save. Previously any page or host on the
    network could write files into the vault. (21f682b)
  • install.sh and the Python package verify the release checksum and refuse
    to install on a missing or mismatched checksum. (439b511)
  • The updater, model registry, installers, and package metadata use
    github.com/byenzyme directly instead of relying on the old useenzyme
    redirect. (f851c64)
  • Enzyme is now developed and released from this repository. Install, skill,
    and agent files are served from the repository root, and the retired
    Claude/Codex marketplace plugin is removed. (e1dc68b, 1c359c9)

Behavior changes to check

  • A bare frontmatter number below 100,000,000 is no longer read as Unix
    seconds; those notes take their date from the filename, path, or file time.
  • Vaults initialized before this release search through an in-memory keyword
    index (correct but slower) until the next enzyme refresh or init builds
    the persistent one.
  • Vaults that declare date rules re-read their Markdown on the first refresh
    after upgrading.
  • Marketplace plugin installs no longer update. Install the CLI with
    install.sh and remove the plugin (claude plugin uninstall enzyme@enzyme
    or codex plugin remove enzyme@enzyme).

Validation and limits

  • Full workspace test suite: 928 passed, 0 failed. New tests cover hybrid
    ranking and FTS sync/escaping/fallback, petri ranking, date rules (grammar
    and indexing), job splitting, and target paths.
  • Release checks: install matrix 13/13, sync contract 4/4, vault lifecycle
    19/19 including the hosted-bootstrap tier (init, catalyst generation, petri,
    refresh).
  • Search-quality numbers come from one personal vault, judged blind; the
    petri ranking comparison used a single grader.

Upgrade

curl -fsSL https://raw.githubusercontent.com/byenzyme/enzyme/main/install.sh | bash
enzyme refresh   # builds the persistent search index