Skip to content

Orion 0.5.2

Choose a tag to compare

@zug11 zug11 released this 22 Oct 10:23

Orion 0.5.2

22 October 2026. Orion 0.5.2 improves adaptive web research and the delivery of
findings in Space voice. Whisper Small and Medium editions have identical application features.
The application code is committed as
d148c82.
The subsequent publication commit contains README and release records only.

Changes

  • Web research chooses quick, standard or deep investigation from the question
    and the evidence still needed, including a more capable model and larger
    reasoning allowance for deep research. Voice remains automatic when useful;
    typed Chat retains its explicit Web button.
  • Removed the fixed three-paid-lookup quota and the overall web-research turn
    and time ceilings. A productive investigation can continue while pursuing
    the current question. Individual requests retain deadlines and cancellation.
  • Repeated failures or unchanged source evidence prompt a supported answer
    with remaining uncertainty. Cached lookups do not consume another paid
    request or extend a stalled assignment; explicit retries remain available.
  • Checkpoint recovery retains cumulative lookup counts and recent-query
    receipts while bounding the working evidence and result cache.
  • Completed worker findings supply substantive context to the main voice,
    reducing holding responses such as “I'm still thinking about what you said.”
    Findings still respect active speech, user turns and cancellation, with a
    bounded fallback when a transcript tail never receives closing punctuation.

Verification

The implementation passed npm run check with 2,906 renderer tests passed
and one skipped
, plus TypeScript and the production renderer build.
Native Rust tests passed 222 tests with one ignored; MCP Rust tests passed
114 tests. Both Cargo formatting checks and git diff --check passed.

Regression coverage includes eight sequential web lookups in typed Chat and
one voice work order, a four-lookup batch, 36-lookup checkpoint recovery,
productive simulated research beyond fifteen minutes, duplicate-evidence
stopping, explicit retry, cancellation during a seventh lookup, and a stalled
individual request. Voice tests cover finding delivery and spoken-turn guards.

These are automated fixtures and build checks. Live microphone delivery and
paid-provider research quality, latency and adaptive depth still require a
separate acceptance run.

Both final installers are Developer ID signed with secure timestamps and
Hardened Runtime, Apple-notarized, stapled and Gatekeeper-accepted. Each final
DMG was mounted read-only and its app independently copied out after
notarization. The copied apps passed deep/strict signature and Gatekeeper
checks, reported version 0.5.2 and the production app.orion.knowledge
identity, and contained the expected Whisper model with exact byte count and
SHA-256. Both bundled connectors report 0.5.2.

The release pipelines rebuilt the renderer and native app with the coordinated
0.5.2 metadata, reran the MCP tests, and verified all 57 connector tool schemas
and 23 library tool contracts from the copied-out installers. Both Cargo
formatting checks and the Git whitespace check also passed after the metadata
update. No installed app was replaced or launched for packaging checks.

Downloads

Edition File Size Apple submission
Whisper Small Orion-0.5.2-Apple-Silicon.dmg 537,149,925 bytes 744aba17-6b8f-49df-8b77-2c5328980335
Whisper Medium Orion-0.5.2.m-Apple-Silicon.dmg 1,505,163,439 bytes bb735d56-5154-43b0-b7ae-89dc05489ffe

See SHA-256 checksums and
artifact and notarization records for the exact hashes,
source and renderer fingerprints, and Apple acceptance records. Detailed local
logs are retained in outputs/release-qa-0.5.2/. Published assets are the exact
verified files, without repackaging or re-signing during upload.