MESA Core · v0.9.3

Release policy

v0.9.3 source: docs/release.md

This guide has equivalent English and Turkish versions. The source sections behind this page are at the bottom.

Publication and production qualification#

These docs follow Core v0.9.3, tag commit 531209bc45805b478264f6dfa8a9e40a94e345cb. GitHub tag and PyPI wheel/sdist availability (not yanked) were verified on 17 November 2026. Alpha / production NO-GO remains.

Releases wait for CI#

The tagged changelog says tag-triggered PyPI releases wait for the complete reusable CI workflow before building and publishing. It reserves verification of the wiring for the first live tag run; this website refresh does not claim to have audited that run, signed-tag preflight or build attestations. See release notes and the pinned release procedure below.

Verified sources
Source reference for this page

The sections of the upstream file that back this page, translated to English where the original mixes in Turkish. The GitHub link shows the exact original.

The pinned source describes V4 package 0.9.3. The owner previously authorized the v0.8.1 source release exception recorded below. See the website release notes for verified v0.9.3 tag and PyPI publication. Production qualification

is still assessed as a release candidate. The production decision stays NO-GO

until Final MVP Certification is complete. Assigning a version number or a tag

does not on its own mean production GO. Old V3 tags/release records are read in their historical context.

Mandatory v4 gates#

Production GO can be given only when all of the following evidence exists for the

same commit. A source publication exception, if any, is recorded per release; it

does not mark these gates as met:

  • locked build, Ruff, Black and Mypy;
  • all core and benchmark contract tests;
  • v4 API/SDK/MCP contract parity;
  • Tier-3 rejection, retry, lease loss, projection crash and rollback tests;
  • Graph V2 identity/provenance/purge and two-way reconciliation;
  • cross-tenant/workspace/dataset negative security tests;
  • dataset-filtered RRF lane ablation and a vector-only comparison report;
  • production-like queue saturation, worker crash and concurrency;
  • offline backup → restore → rebuild → parity → cutover rehearsal;
  • boot/restart, cost and token evidence with the selected real provider/model;
  • a 24-hour soak.

The production-readiness decision stays NO-GO until every gate, including the last,

passes. CI success alone is not GO.

Local preflight#

On a clean checkout at the release commit:

bash
uv lock --check
uv sync --locked --python 3.11 --extra dev
uv pip check
uv run ruff check .
uv run mypy mesa_memory mesa_storage mesa_workers mesa_api mesa_client \
  --ignore-missing-imports --explicit-package-bases --follow-imports=skip
make test-local
# Separate adapter-inclusive gate: make test-all after installing dev + adapters
uv build --wheel

Compose contracts:

bash
MESA_API_KEY=release-placeholder \
MESA_PRINCIPAL_ID=release-principal \
MESA_MODEL_ENABLED=false \
MESA_EXTERNAL_PROVIDER_ENABLED=false \
docker compose -f docker-compose.v4.yml config --quiet

This model-disabled render only validates the deployment topology.

Full-cognitive evidence comes from a separate staging rehearsal with a real provider.

CI evidence#

These workflows must pass on the release commit:

  • MESA CI: quality, v4 contract, security, migrations, package and coverage;
  • Benchmark contracts and the dataset-filtered RRF ablation artifact:

the old Benchmark data quality workflow has been removed from Core; separate

mesa_benchmark evidence must be tied to the same release commit. Its absence is not a PASS;

  • MESA external release gates: v3 compatibility Compose, v4 combined Compose,

flow, bounded capacity and executable docs.

A 24-hour soak does not fit in a GitHub-hosted runner's time limit; it runs outside, in a

production-like environment. Start/end time, image digest, config hash,

health/parity series, restart/incident log and a result summary are kept as release evidence.

Version, build and tag#

After the gates pass:

  1. Apply the SemVer decision to pyproject.toml and CHANGELOG.md.
  2. Run the release commit through CI.
  3. Produce and verify the wheel/SBOM on a clean checkout.
  4. An authorized operator creates a signed annotated tag.
bash
git tag -s vX.Y.Z -m "MESA vX.Y.Z"
python scripts/release_preflight.py vX.Y.Z
git push origin vX.Y.Z

The Core CI tag job produces a reproducible wheel, a CycloneDX SBOM and a GitHub OIDC

attestation. The repository workflows do not create a GitHub Release or a production

deployment automatically.

On the CI tag checkout only SemVer, the pyproject.toml version and the CHANGELOG

heading are verified. Annotated tag and signature verification is done with the full

preflight command above, on a clean checkout that has the trusted operator keyring;

CI neither imitates nor weakens this signature check.

The release operator must download the OIDC build attestations produced for the wheel

and SBOM and verify them together with the artifact digests.

Migration and rollback package#

Release artifacts carry together:

  • wheel/image digest and SBOM;
  • the Alembic head and the migration rehearsal result;
  • the backup manifest format and restore verification;
  • the V4 upgrade/projection rebuild/parity report; separate ingestion verification if V3 content is moved;
  • the retained backup location and the cutover rollback procedure;
  • v4 API/OpenAPI, SDK/MCP contract test results;
  • benchmark and soak evidence.

Restore never overwrites an existing target. If a parity or health gate fails, there is

no cutover; the old storage and the v3 compatibility service are kept.

Release-specific publication exceptions#

The v0.8.1 source publication was explicitly approved by the repository owner on 18 October 2026.

Because no local GPG key was available, this release uses an annotated, unsigned tag;

signed-tag preflight success is not claimed. Real-provider and production

certification are not part of this source publication. Details are recorded in the

v0.8.1 release notes.

The v0.8.0 source publication approval, unsigned tag and missing certification evidence

are kept in the v0.8.0 release notes. This exception

does not carry over to later releases and does not count as production GO. This page defines

the general procedure; the release notes define the status of that publication.

PyPI trusted publishing#

The tag-triggered .github/workflows/release.yml builds wheel/sdist, runs Twine validation and uses GitHub Actions OIDC trusted publishing. One-time owner setup: PyPI project mesa-memory, GitHub owner Yasou13, repository MESA, workflow release.yml, environment pypi; create the matching GitHub environment. GPG tag signatures and OIDC build attestations are distinct evidence. The official v0.9.2 release records an SSH-signed annotated tag and passing full local signed-tag preflight. A configured workflow is not proof that a PyPI version exists.