Skip to content

v1.8.3

Choose a tag to compare

@sathish-mg sathish-mg released this 16 Sep 09:28
· 206 commits to main since this release
250326d

Two Cloud-reported defects, both found by a real pack rather than by a test — plus the internal-requirement bump this patch needed because it adds public API.

Linux binaries carry a GLIBC 2.39 floor (the runner image); on an older distro the failure is a raw dynamic-linker error, not a Rust one. npm publishes under the scoped @areev/areev names.


Added

  • The selfimprove bench can open a Postgres memory (#250). selfimprove_aba --db PATH|postgres://…?schema=… — or $AREEV_BENCH_DB, the same variable the
    Python harnesses have honoured since 1.8.0 (#200) — moves the MEMORY out of
    the workdir while the transcripts and report.json stay in it. The A/B/A/B
    track is the one with no external dataset, a keyless deterministic floor and
    a programmatic scorer, which makes it the natural case for proving a
    provisioned schema answers like an embedded file; it was also the only track
    that could not target one, because the Rust binary hard-coded
    <workdir>/bench.db. With a DSN set no bench.db is created at all, the
    report records the memory redacted (a password must not ride into a published
    run directory) beside a db_backend field, ?provision=never is honoured so
    the run's role needs no CREATE, and "fresh memory" is judged by what the
    schema HOLDS rather than by whether a file exists. selfimprove_learn stays
    file-only and says so: its method is one fresh copy of the captured memory per
    pass, and a schema cannot be copied. Gated in CI's postgres job
    (cargo test -p areev-bench --features postgres --test selfimprove_pg);
    without the feature a DSN is refused by name, never treated as a filename.

  • One answer to "which grain do I use?" — docs/grains.md. The
    decision table for all thirteen types, the rule of thumb (true → Fact,
    happened → Event, measured → Observation, intended → Goal, procedure →
    Workflow, capability → Tool), the pairs people mix up, how each type is
    written on every surface, the two field traps, and add vs supersede in
    time. The guidance existed, but only inside the agent-building manual,
    three sections in; the quickstart showed nothing but a Fact; and the
    surfaces where the choice is actually made gave no help at all. Now each
    grain type carries a one-line purpose on the registry
    (GrainTypeMeta::purpose), DESCRIBE <type> reports it, the MCP
    areev_add description states the rule and enumerates the addable types
    from the registry instead of trailing off in an ellipsis, and the page's
    "Use it for" column is test-pinned to quote every registry row verbatim —
    the engine's answer and the doc's answer cannot drift. Its CAL examples
    are CI-parsed like the reference's. Every other doc that touched the
    question (the agent guide, FAQ, ARCHITECTURE §2.3, the CAL and MCP
    references, the quickstart, llms.txt) now defers to it.

Fixed

  • An abstract node could not call a tool with a dot in its name (#251).
    Anthropic and OpenAI forbid . in a tool name, so a pinned Definition called
    receipt.prepare is offered to the model as receipt_prepare — and the
    model, correctly, called back with what it was given. The scheduler's
    unknown-tool check compared that against the un-normalized set, so every
    call to a dotted tool from inside an abstract node failed the node after one
    corrective re-prompt the model could not possibly satisfy. The reverse map the
    normalizer's own doc comment promised ("the invoker reverse-maps on the return
    path") was never written. Now it is, once, where the model's answer enters the
    scheduler: exact match wins, then a unique normalized match, so the
    unknown-tool guard, the strict-argument validator (which had silently skipped
    every dotted tool, its schema keyed by the canonical name) and the dispatch
    all see the Definition's own tool_name. A name matching neither form still
    gets exactly one correction, now listing the tools the way the model was shown
    them. The outbound half is fixed with it: the tool-calling adapters render a
    replayed tool_use/tool_calls name — and a ToolChoice::Named — through the
    same normalizer, because a transcript naming a tool the request never offered
    is a 400 on the loop's second turn. Found by Areev Cloud's first
    model-backed capture pack; it removes the rule that a pack with an abstract
    node must name its tools without dots. One consequence worth stating: areev run verify on an ALREADY-FAILED pre-fix run (one whose journal records the
    model called unknown tool(s) failure) can now diverge, because the
    scheduler would take the dispatch the journal never contains. Runs that
    succeeded replay unchanged — the fix touches no journal shape.

  • A dead registry row that lied about required fields. Each grain type
    carried a required_add_fields list in types/registry.rs that nothing
    read
    — the CAL JSON builder kept its own copy, and that copy is what
    DESCRIBE and VAL-E001 have always used. Unread, the registry's version
    had drifted: observation claimed to need observer_id/observer_type
    when the write path demands content, consent was missing user_id,
    and workflow claimed to need nodes though an empty container is legal
    by design. No output was ever wrong — the wrong copy was simply the one a
    reader would find first, and the new purpose sentence now sits in that
    same row, so it had to become true rather than merely unused. The
    builder's required_fields now reads the registry, the existing
    required_fields_match_the_validator test pins the row to what the
    builder arms enforce, and DESCRIBE output is byte-identical to before.

  • A macOS-only CI flake in the credential broker's binding test. The
    Python stub upstream in areev-py's #201 test answered without ever
    reading the POST body it was sent, so socketserver closed the connection
    over two queued bytes — and a close over unread data sends an RST rather
    than a FIN, which on macOS discards what the peer has already buffered.
    The broker read the status line but lost the body it had been promised,
    and reported the admitted call as a 200 with an empty body: its documented
    behaviour for a mid-body read failure, and indistinguishable from an
    upstream that answered with nothing. Harmless on an idle machine and
    roughly even money on a loaded one — 10 of 20 runs failed under CPU
    saturation, 0 of 20 after, with the leftover bytes confirmed in the kernel
    queue on every failing run and absent on every passing one. The stub now
    drains to Content-Length, which the Rust stub for the same test
    (areev-cli/tests/common/egress201.rs) always did and Node's http
    server does on its own. Test-only: no shipped behaviour changed.

Changed

  • The crypto stack moves to the current RustCrypto generation, with the
    compatibility actually proven
    (#240). sha2 0.10→0.11, argon2 0.5→0.6,
    getrandom 0.2→0.4, hkdf 0.12→0.13 and aes-gcm 0.10→0.11 — five majors
    at once, across passphrase key derivation, the CAS sidecar's AEAD and every
    nonce and token in the tree. The code changes are small: getrandom renamed
    its one function, hybrid-array deprecated from_slice in favour of the
    infallible From<[u8; N]>, and a digest no longer renders itself as hex.

    The risk was never the compile. A round trip seals and opens with the same
    build, so the whole 2,800-test suite would have stayed green if a derivation
    had shifted — and every passphrase-encrypted memory and every encrypted blob
    in the world would have become unopenable.
    There were no committed
    fixtures to catch it either. So the reference values were computed with the
    exact crates 1.8.2 shipped and are now pinned as known-answer tests: the
    Argon2id key for a fixed passphrase, salt and parameters; the HKDF-SHA256
    blob subkey; and a ciphertext produced by aes-gcm 0.10.3, which still
    decrypts with its address binding intact. All three hold. Those tests stay,
    so the next bump that does move one fails loudly instead of silently — and
    if it ever fires, the bump is a migration with a re-key path, not an upgrade.

    One cost recorded rather than hidden: areev-store named aes-gcm directly
    because the storage engine already pulled the same version, keeping the blob
    sidecar on the identical primitive as the database pages. The engine stays on
    0.10.3, so a default build now links both. Taken deliberately — holding a
    security primitive back to match an upstream pin is the worse trade — and it
    reverses on its own when the engine catches up. The manifest says so now
    instead of asserting something false.