v0.2.1

What's Changed

• build(deps): bump docker/setup-qemu-action from 3 to 4 by @dependabot[bot] in #6
• build(deps): bump docker/setup-buildx-action from 3 to 4 by @dependabot[bot] in #7
• build(deps): bump actions/deploy-pages from 4 to 5 by @dependabot[bot] in #8
• build(deps): bump actions/upload-pages-artifact from 3 to 5 by @dependabot[bot] in #9
• build(deps): bump actions/upload-artifact from 5 to 7 by @dependabot[bot] in #10

Full Changelog: v0.2.0...v0.2.1

0.2.0: alignment and evidence

Every public claim is now measured, printed at boot, or tightened to exactly what the code does.

• Public benchmark: six suites mapped onto the README claims, 54 tasks with recorded search-resistance evidence, mechanical judges only, Wilson intervals + noise bands, no aggregate score, no competitor table. First CI numbers (fresh install): depth 18/18, scholar graph 21/22, memory 10/10, cross-lingual 0/14 (that zero is the honest floor: the cross-lingual sense rides the accumulated retrieval memory and compounds with use).
• Weekly source-health feed on the health-data branch: every source probed from CI, rate_limited distinct from down, skipped never conflated with broken.
• Boot senses report: the server prints which senses are online and which are dormant, naming the extra that wakes each.
• Strictly verifiable privacy wording, capability tiers named (Core / Research / Walled), one vocabulary everywhere (retrieval memory + typed evidence graph), the demo infographic rebuilt after a four-reviewer fresh-eyes audit, 27 audited logic breaks repaired, and a hermetic smoke gate that turns any real network reach into a loud failure.

Full detail in CHANGELOG.md.

Full Changelog: v0.1.3...v0.2.0

OmniSeek 0.1.3

Polish release. Runtime strings speak only about the catalog (no private-era operator voice anywhere; a sync residue gate now makes that class unshippable). New: docs/sources.md, the source catalog as a self-regenerating document, and docs/case-study.md rebuilt around verbatim outputs from one real session.

PyPI: https://pypi.org/project/omniseek/0.1.3/

Full Changelog: v0.1.2...v0.1.3

OmniSeek 0.1.2

Fixes a first-run hang: on a fresh install, the FIRST ranked search could block forever (the ranking/recall import chain loaded lazily on the event-loop thread mid-call and could deadlock on Windows over stdio). Both entrypoints now warm the chain at boot; the first search returns within its budget.

Surfaced by a real Cline install test on launch day. PyPI: https://pypi.org/project/omniseek/0.1.2/ · Details: CHANGELOG.md

Full Changelog: v0.1.1...v0.1.2

OmniSeek 0.1.1

Registry-readiness patch.

• Official MCP Registry ownership markers: mcp-name in the PyPI long description, and the io.modelcontextprotocol.server.name label on the GHCR image.
• server.json at the repo root: PyPI + GHCR packages, streamable HTTP transport at http://localhost:8765/mcp.
• The package root docstring now speaks the product's own description, and __version__ is re-authored from pyproject.toml on every sync so it cannot drift from the released version.

PyPI: https://pypi.org/project/omniseek/0.1.1/ · Details: CHANGELOG.md

Full Changelog: v0.1.0...v0.1.1

OmniSeek 0.1.0

The first public release.

Your agent seeks what search can't find. A self-hosted perception MCP server: it hears speech nobody transcribed, reads images and video frames, crosses languages, reads login-walled pages with your own credentials, and remembers everything it reaches in a source-traced relation graph.

git clone https://github.com/Battam1111/omniseek.git && cd omniseek
docker compose up -d

or pip install omniseek (PyPI).

Highlights of this cut are in the CHANGELOG: the scheduler now runs on any deployment shape (docker included), the mcp 2.0 install break is pinned away, and the README's investigation is a real session, quoted as returned.

Full Changelog: https://github.com/Battam1111/omniseek/commits/v0.1.0