Skip to content

v4.12.26

Choose a tag to compare

@acn-ericlaw acn-ericlaw released this 26 Oct 17:02
· 69 commits to main since this release
Immutable release. Only release title and notes can be modified.
9a205c0

Version 4.12.26, 10/26/2026

The lock-step twin of Java 4.12.26, catching up from 4.12.24 (Java 4.12.25 closed the findings of the field's Sonar gate and had
no counterpart here). A JSONPath outside RFC 9535 is refused instead of answered with null, and a substituted value can no longer
change a query's structure (Changed 5 to 7, RFC-0010). Flows and graphs gain the date plugins f:addDays and f:daysBetween
(Added 4), the companion endpoint reports every refusal by the engine's own flag (Fixed 2), and the Playground gains mock model
variables, a PNG export, a favicon and the graph-set worked example. Three RustSec lockfile bumps are included: rustls,
rustls-webpki and tokio (Security 1 to 3). The Java release's dependency changes and its live invocation observers for an APM
agent have no counterpart in this engine. Upgrade notes: read Changed 5 and 6 before upgrading - a flow or a graph whose $
path is not RFC 9535 now fails at startup, and a value that carried query syntax into a $ path now fails the mapping. Changed 1
and Fixed 2 concern the dev-mode Playground and its companion driver only.

Security

  1. rustls 0.23.46 (Increment 180). The audit gate failed on 2026-10-16 on four RustSec advisories against rustls 0.23.45,
    published on 2026-10-14 and all patched in 0.23.46: RUSTSEC-2026-0346 (Debug impls for client session values expose TLS
    secrets), RUSTSEC-2026-0347 (buffered application data sent to the outer server after an ECH rejection), RUSTSEC-2026-0348
    (Brotli certificate decompression can allocate up to 1 GiB before authentication) and RUSTSEC-2026-0349 (an unauthenticated
    ServerHello can downgrade later connections to classical key exchange). The lockfile moves that one crate; its dependents are
    tokio-rustls (the platform's HTTP client) and redis (the four Redis extensions). The workspace builds and its tests pass
    unchanged. Upgrade note: none - a lockfile bump, no API change.
  2. rustls-webpki 0.103.16 (Increment 182). The audit gate failed on 2026-10-21 on RUSTSEC-2026-0382 against rustls-webpki
    0.103.15, published on 2026-10-20 and patched in 0.103.16: certificate verification may accept distrusted intermediates. The
    lockfile moves that one crate; it sits behind rustls 0.23.46, whose dependents are tokio-rustls (the platform's HTTP client)
    and redis (the four Redis extensions). The workspace builds and its tests pass unchanged. Upgrade note: none - a lockfile
    bump, no API change.
  3. tokio 1.54.1 (Increment 190). The weekly audit run of 2026-10-26 reported RUSTSEC-2026-0386 against tokio 1.52.3,
    published on 2026-10-13 and filed as unsound, not as a vulnerability: AsyncFd::new and three related constructors accept
    any file descriptor without proof of ownership, so a descriptor closed and reused while an AsyncFd is alive can lead to a
    use-after-free. No code in the workspace or its locked dependencies calls those constructors. The 1.52 line has no patched
    release; 1.54.1 is patched. The lockfile moves that one crate, the runtime under every crate of the workspace. The workspace
    builds and its tests pass unchanged. Upgrade note: none - a lockfile bump, no API change.

Added

  1. The Playground's dry run mocks model variables (POST /api/mock/{id}?namespace=model; Increment 181). The twin of the Java
    engine's change. A deployed graph's model is partly set before the first node - model.cid from the request's correlation-ID
    header - and a dry run has no request, so the console's instantiate graph lines were the only way to set a model variable. The
    mock upload gains a third namespace: a JSON object whose keys are model paths (tier, limits.daily, items[0]) sets each as
    model.{key} and leaves the others in place, with the console line "Mock data loaded into 'model' namespace", in every member's
    instance of a collaborative session. A key with an empty path segment, a bad index or a null value is refused (HTTP 400). The
    regenerated bundle's Upload form gains Model variables rows - a name, a type (text, number, boolean or JSON) and the value -
    posted after the headers, and its labels read JSON payload (input body) and Input header (optional); help upload is
    mirrored from the Java repository. Upgrade note: none - a new namespace.
  2. The Playground exports the graph as a PNG image (Increment 183). The Graph tab's toolbar gains an Export PNG button that
    saves the whole graph, not only the part the viewport shows, at zoom 1 with a solid background and 2 image pixels per graph
    pixel, as <graph-id>.png; the minimap, the zoom controls and the selection outline stay out of the image, and the node cards
    are drawn without their shadow, which Safari paints as a solid grey block. There is no SVG
    export: the renderer's SVG keeps the node cards as HTML inside foreignObject, which vector editors and office suites show as a
    blank image. The change is in the Java repo's webapp (mercury-composable #571, with the npm dependency html-to-image 1.11.11,
    MIT); this repo receives the regenerated bundle, which carries both #570 and #571: index-CWNizJaH.js became
    index-Dhz_Pqc_.js and index-DaTDFwQo.css became index-BH572KKk.css; the vendor chunks and the 44 help pages are
    unchanged. Upgrade note: none - a new toolbar action in the dev-mode Playground.
  3. The Playground engine has a favicon - the Mercury atom. resources/public/favicon.ico holds 16, 32 and 48 px images of the white
    atom that the documentation site and the API playground use, on the API playground's deep purple. It is the same file as the Java
    engine's; every Playground page load used to answer /favicon.ico with 404. Upgrade note: none.
  4. f:addDays and f:daysBetween - calendar-date arithmetic in a flow or a knowledge graph (Increment 189). The twin of the
    Java engine's plugins (mercury-composable #591): f:addDays(date, days) returns the date a whole number of days later (earlier
    for a negative number), f:addDays(text(2027-01-15), int(15)) being 2027-01-30, and f:daysBetween(from, to) the whole days
    between two dates as an integer, the earlier date excluded and the later one included, in either order: 2027-04-10 to
    2027-05-11 gives 31, the same date twice 0. A date is an ISO-8601 calendar date written exactly as YYYY-MM-DD, with no time
    and no time zone. A date of another shape, a day the calendar lacks (2027-02-30), a fractional number of days, a null
    argument and a result outside the years 0000 to 9999 are refused with the Java engine's messages. Pinned by the shared
    date-plugin-vectors.json (65 vectors, byte-identical with the Java repository) and the graph fixture
    unit-test-date-arithmetic-1; the graph contract types the results (string, integer). The mirrored help page arrives with
    the regenerated bundle: index-D5Sx-GL2.js became index-MYdzjeK5.js. Upgrade note: none - two new plugins.

Changed

  1. The Playground processes mock headers as a request's headers (Increment 181). A header uploaded through
    POST /api/mock/{id}?namespace=header was stored as written. It is now processed as the HTTP edge and the deployed walker process
    a request's: every name in lower case (two names that differ only by case are refused), CR and LF removed from a value, the
    business correlation-ID header (http.correlation.id.header, X-Correlation-Id by default) setting model.cid - and, when the
    upload has none, the header carrying the instance's model.cid - and x-iteration-index setting model.iteration_index. The
    console names each derived value. Upgrade note: a dry run that read a mock header back under its written case now sees the
    lower-case name; a graph reading one header by name is unaffected.

  2. The Playground's graph view widens a node whose text would wrap heavily (Increment 183). Every node was 240 px wide, so a
    node with long mapping lines or a long description grew into a tall, narrow column. A node now starts at 240 px and widens in
    40 px steps, up to 480 px, while its property values would wrap onto more than 1.5 lines per row on average; every node in a
    layout column takes the width of the column's widest node, and thumbnail mode keeps every card at 240 px. The change is in the
    Java repo's webapp (mercury-composable #570); this repo receives the regenerated bundle (see Added 2). Upgrade note: none -
    a layout change in the dev-mode Playground only.

  3. The Playground's command input states its keys once. The input box's placeholder repeated the hint printed below it; it now reads
    "Enter command", and the hint below uses symbols: "⏎ to send · ⇧⏎ for new line · ↑↓ for history". The change is in the Java repo's
    webapp (mercury-composable #571); this repo receives the regenerated bundle (see Added 2). Upgrade note: none.

  4. The Playground bundle is rebuilt with Vite 8.3.5 (Rolldown 1.2.15; Increment 184). The Java repo's webapp moved from Vite 8.1.5
    to close five Vite dev-server advisories (mercury-composable #572). The new Rolldown emits a different bundle - every JavaScript
    chunk changes, the CommonJS interop helper among them - so this repo receives the rebuilt files, byte-identical with the Java
    repo's: index-Dhz_Pqc_.js becomes index-CrqpNrFx.js, and the vendor and runtime chunks get new hashes. The stylesheet, the
    favicon and the 44 help pages are unchanged. Upgrade note: none - in Chrome and Firefox the rebuilt bundle behaves the same.

  5. A JSONPath outside RFC 9535 is refused, not answered with null (RFC-0010 WP3; Increment 185). The twin of the Java engine's
    change, which replaces Jayway with an in-house RFC 9535 engine. This engine already evaluated $ paths with serde_json_path,
    an RFC 9535 implementation, so the two engines now accept the same queries, give the same answers and refuse the same way.

    • At evaluation. A $ path the parser refuses used to answer a silent null, the same as a missing value. It now fails the
      mapping by name, as reason at position N of 'path', and the Playground's JSON-Path console prints that reason.
    • At deployment. The flow compiler drops a task whose $ source is not RFC 9535, as it does for any invalid mapping entry,
      with an ERROR that names the standard; CompileGraph, the graph-set packager and the Playground's run reject such a graph.
      The check covers a data mapping's source, a simple plugin's $ argument and a data dictionary's default value. A reference in
      the path ({model.key}, [model.i], {namespace.key}) is checked by where it lands; its value is still substituted first.
    • A {reference} never holds a JSONPath query in a knowledge graph: {$.input.body.amount} in a graph.math expression or
      a mapping source is refused.
    • A simple plugin's $ argument reaches the plugin whole. The argument splitter cut it at a comma inside its brackets or a
      string literal; it now ends a $ argument at the first comma outside them, and every other argument splits as before.
    • Evidence. The JSONPath Compliance Test Suite (cts.json, 715 cases, byte-identical with the Java engine's) runs against
      serde_json_path: every answer is the standard's and the 249 invalid queries are refused. Its normalized paths omit RFC 9535's
      escapes in 16 name-selector cases, which the engine never reads (pinned by name). The shared jsonpath-source-vectors.json (49
      vectors) and the parser and graph fixtures are byte-identical with the Java engine's, except one graph that differs where the
      two engines' tutorials differ (graph.math for graph.js).

    Upgrade note (READ): a flow or a graph whose $ path this engine answered with null, because the path is not RFC 9535, now
    fails at startup: the flow loses that task and the graph answers 404. Jayway-only syntax (in, =~, .length() and the like)
    was always null here; the standard's forms replace it (@.code == 'a' || @.code == 'b', match(@.code, 'a|b'),
    length(@.tags)). A {$...} reference is refused: map the value to a model variable first.

  6. A value substituted into a JSONPath cannot change the query's structure (RFC-0010 WP4; Increment 186). The twin of the Java
    engine's change. A reference in a $ path is still replaced before the query is parsed, and its value is now checked where it
    lands: in a name a key path (names joined by dots, each optionally with [n] indexes, as in items[0].name), in an index an
    integer, inside a string literal the author wrote its text escaped for that literal (its quote, its backslash and its control
    characters), and bare in a filter a number, a boolean or null as written and text as a string literal - Event Script keeping text
    that is already one literal (10, true, 'abc') as written, as it always has. A list or a map is never substituted. A value
    that fails its check fails the mapping by name (Cannot substitute {model.key} into '$.input.body.{model.key}' - '*' is not a key path). A simple plugin's $ argument gets the same checks. The 39 substitution vectors in the shared
    jsonpath-source-vectors.json give the same queries and messages on both engines. Upgrade note (READ): a value that carried
    query syntax into a $ path - a wildcard, a filter, .. or a quote - now fails the mapping instead of changing the query; in a
    knowledge graph, a text value inside a literal the author quoted ('{model.team}') is no longer quoted a second time, so that
    form now works.

  7. The guides describe RFC 9535 JSONPath (RFC-0010 WP5; Increment 187). The Event Script syntax guide's JSON Path section states
    the grammar, the result shape, the refusal before a flow or a graph runs, a replacement for each Jayway-only form (which this
    engine answered with null) and the checks on a substituted value; the flow grammar, the knowledge graph's command reference, the
    commands catalog and the AI agent guide's pre-send checklist follow it. Seven claims pin the behaviour against this engine's
    tests: json-path-result-shape and mapping-source-dynamic-variables are updated, and five are new. The result-shape test is
    now json_path_source_takes_the_mercury_shape. The Playground bundle is redeployed from the Java repo with its help
    (help graph-data-mapper, help tutorial 8): index-CrqpNrFx.js became index-D5Sx-GL2.js, the vendor chunks, the stylesheet
    and the other 42 help pages are unchanged, and every deployed file is byte-identical with the Java repo's. Upgrade note: none.

  8. The Playground example deploys tutorials 2 to 14 as a graph set, and an import from a deployed model prints one line less
    (#387, the twin of the Java engine's #567).
    The example is now the worked example of graph sets: tutorial-1 and
    support-triage stay loose graphs, and tutorials 2 to 14 deploy as the set tutorials (resources/graph/tutorials.pack,
    version 1.0, byte-identical with the Java example's), unpacked into file:/tmp/unpack. When import graph falls back to a
    deployed graph model, the Playground no longer adds "Please export an updated version and re-import to instantiate an instance
    model" after "Found deployed graph model in {location}": the next line, "Graph model imported as draft", already says what
    happened, and the five tutorial help pages that quote the output follow. The unpack examples in help package, the
    starter-graph template's commented manifest and the AI agent guide used a folder under /tmp/graph, which the graph-set loader
    refuses because it is the Playground's temporary folder (location.graph.temp); they now use file:/tmp/unpack and state the
    rule, and the configuration reference's refusal sentence names a folder inside the temporary folder too. The help pages and the
    bundle are deployed from the Java repo. Upgrade note: none - an example's deploy layout, one line of dev-mode output and the
    docs.

Fixed

  1. A text value with a backslash keeps it in a knowledge graph's quoted expression (Increment 186). In a graph.math expression
    (IF, CONDITION, COMPUTE), a {reference} to a text value is quoted as a string literal, and the quoting escaped the quote but
    not the backslash. The expression lexer reads a backslash as an escape, so C:\temp became a tab and a value that ended with \
    escaped the closing quote and failed the statement. The backslash is now escaped first; the Java engine has the same fix.
    Upgrade note: none - an expression that compared such a value read the wrong text or failed.
  2. The companion endpoint reports every refusal as a failure, by the engine's own flag (Increment 188). The twin of the Java
    engine's fix (mercury-composable #586). The synchronous companion endpoint (POST /api/companion/{id}/sync) decided ok by
    matching each console line against a list of error words, so a refusal worded otherwise came back ok: true and a driver that
    trusted ok believed the command had worked: create node on a node the model already holds (node {name} already exists), an
    import graph that the importer refused after resetting the draft, import node of a node the file does not hold, an export
    refused for its root's name, a delete connection between nodes that are not connected, and list with an unknown kind. The
    word list also failed commands that had worked: help run, help schema and seven tutorials answered ok: false for prose
    that mentions an abort or a "not found". The command service now sends each refusal with an error status, as the traveler
    already sent a traversal's abort, and the endpoint reports the first such line as its error, or the dispatch's own failure.
    import node from a file that does not exist, and session with an unknown argument, now say so instead of printing nothing.
    The WebSocket console shows the same lines. Upgrade note: a driver that sent create node to replace a node now sees the
    failure it always was; update node replaces a node.