Skip to content

Changelog

All notable changes to this project are documented here.

Format: Keep a Changelog. Versioning: Semantic Versioning.

1.3.1 (2026-06-29)

CI/CD

  • dependabot: auto-merge transitive patches and patch/minor security fixes (#139) (02eb9ea)

1.3.0 (2026-06-23)

Behavior change (error reporting): failed disease-tool calls now return an MCP error result (isError=true) instead of a successful result whose text content was an error-shaped JSON string ({"status": "error", ...}). Clients that parsed that JSON to detect failures should switch to checking isError on the result. Successful results and not-found / no-results negative results are unchanged. See #133; mirrors uniprot-mcp #88.

Features

  • disease: raise ToolError on tool failures so isError is set (mirrors #88) (66b9379)

Bug Fixes

  • storage: set SQLite busy_timeout in _open_db (#128) (4a1c3d1)

Documentation

CI/CD

  • scope workflow token permissions to job level (OpenSSF Scorecard) (#132) (b8fc96e)
  • stabilise the macOS/Python 3.11 test flake (scoped retry) (#131) (40bd931)

1.2.3 (2026-06-22)

Dependencies

  • security: bump msgpack 1.1.2 -> 1.2.1, addressing a high-severity use-after-free / out-of-bounds read on Unpacker reuse after a caught error (#127)
  • security: bump pydantic-settings 2.14.1 -> 2.14.2, addressing a path-traversal / link-following advisory in NestedSecretsSettingsSource (#126)
  • bump the production-minor-patch group with 3 updates (#125)
  • bump actions/checkout 6 -> 7 (#124)

Bug Fixes

  • restore the verbatim Apache-2.0 Section 9 text in LICENSE so the license is correctly detected (#123)

Documentation

  • add README motivation lead and align supply-chain, offline, and pLDDT claims with the code (#116) (5bd6b7f)

1.2.2 (2026-06-16)

Bug Fixes

  • align governance, supply-chain, and science claims with the implementation (#111) (6c170db)

Documentation

  • align CONTRIBUTING Python prerequisite with the 3.10+ support matrix (#112) (86db87f)
  • bump human-readable version stamps to 1.2.2 (#114) (245fb2e)
  • log the v1.2.1 hardening review in AUDIT.md; correct CITATION release date (#109) (4b0ff1d)

1.2.1 (2026-06-16)

Bug Fixes

  • align user-facing capability claims with the implementation; ship py.typed (#107) (590583e)
  • deps: bump starlette, python-multipart, cryptography (clears 9 Dependabot alerts) (#108) (a9a658d)

Testing

  • add metadata contract tests and Zenodo deposition metadata (#101) (70db7c0)

Refactoring

  • type-check the full source tree under mypy --strict (#102) (772dd3d)

Unreleased

1.2.0 - 2026-06-08

The "live-API validation" milestone. Every tool was exercised end-to-end against the real upstream APIs (Open Targets, ChEMBL, UniProt, Ensembl, ClinVar, gnomAD, HPO, AlphaFold DB); four that were non-operative were fixed, the local knowledge graph now seeds itself, and the documentation/examples were brought in line with what the code actually returns.

Added

  • The local knowledge graph seeds itself on first use. A curated seed dataset (the BCR-ABL/CML and BRCA1/breast-cancer entities used in the worked examples) is loaded when the graph is empty, so query_protein_database, query_variant_database, export_research_dataset, and find_drug_gene_network return representative results out of the box. Set AFSMCP_DISABLE_KG_SEED=1 to keep the graph empty. New relationship-write methods (store_protein_drug, store_protein_disease, store_variant_disease) back the seed.
  • EFO disease resolution + supporting client methods. DiseaseRecord.efo_ids (MONDO→EFO cross-reference), OpenTargetsClient.resolve_disease_efo (name→EFO via search), ChEMBLClient.molecule_names (bulk ID→name, chunked), and EnsemblClient.ncbi_gene_id (HGNC symbol→Entrez).

Fixed

  • map_disease_drug_landscape returned no drugs or targets for any disease. It passed a MONDO ID into ChEMBL's EFO-keyed drug_indication filter and into Open Targets. Now resolves MONDO→EFO (via the term's xref, falling back to an Open Targets name search for terms without one — e.g. "breast cancer"), backfills drug names, and retries Open Targets via the EFO ID. CML now returns 9 approved drugs (imatinib, dasatinib, nilotinib, ponatinib, bosutinib, asciminib, …) and ABL1/BCR/KIT targets.
  • Three phenotype tools returned HTTP 404. The HPO REST API moved from hpo.jax.org to ontology.jax.org; the client was repointed and its parsing rewritten. lookup_phenotype, get_gene_phenotype_profile, and phenotype_to_structures are operative again.
  • find_evolutionary_structural_shifts found no orthologs. Ensembl's homology-by-ID endpoint now requires the species in the path (/homology/id/{species}/{id}).
  • triage_variant_3d rejected standard HGVS. The parser now accepts the canonical ClinVar form (NM_…(GENE):c.…) and no longer mis-reads genomic accessions or chromosome names as gene symbols.
  • synthesize_protein_dossier provenance stamped the wrong field and omitted several sources; scientific-accuracy and source-honesty defects (licenses, ACMG PS1→PP5, MONDO/version claims) were corrected across the docs. It now also stamps alphafold_db (queried for pLDDT) in both the provenance footer and the data_sources map.
  • Seed scientific accuracy: the curated KG seed linked olaparib as a drug-target of BRCA1; olaparib targets PARP1 (synthetic lethality in BRCA-mutant tumours), so the edge was corrected to PARP1 (P09874).
  • KG read-only contract: the curated seed is now loaded once at server startup, not lazily on first tool access, so the readOnlyHint: true query/export tools never write. The emptiness check counts every entity table (not just proteins), so an existing variants-only DB is never overwritten.
  • Provenance version stamps refreshed to the live releases: ChEMBL v36→v37, Open Targets 24.06→26.03.
  • Relationship upserts (store_protein_disease / store_protein_drug / store_variant_disease) now refresh all mutable evidence columns on conflict.

Changed

  • All three worked examples re-captured from live runs and relabelled "Captured live": 01 (BRCA1 c.181T>G → ClinVar Pathogenic / ACMG PP3+PP5), 02 (EGFR druggability → HOT), 03 (CML drug landscape), each with the real output and transparent notes on the heuristic's limitations.
  • CI: Dependabot grouping + low-risk auto-merge; fetch-metadata v3.1.0.

Security

  • Forces patched pyjwt 2.13.0 into the runtime tree. pyjwt enters the install tree transitively through fastmcp → mcp 1.27.1, which pins pyjwt[crypto]>=2.10.1. That loose floor resolved to 2.12.1, which carries four advisories — PYSEC-2026-175, -177, -178, -179 — all fixed in 2.13.0 (confirmed by pip-audit). A direct pyjwt[crypto]>=2.13.0 floor was added to [project.dependencies]. Unlike a uv-only [tool.uv] constraint, a [project.dependencies] floor lands in the published wheel's Requires-Dist, so every installer (not just uv) is forced onto the patched release. No code path imports jwt; this is a security version floor on a package already in the tree, not a re-introduction of the direct functional dependency removed in v1.1.3. uv.lock regenerated: pyjwt 2.12.1 → 2.13.0, the only package that changed.

1.1.10 - 2026-06-05

Two reviewer-flagged residuals from the v1.1.9 audit, plus the release-workflow fix that was merged to main after the v1.1.9 tag but never shipped in a tagged release. The tool surface and tool count (29) are unchanged; the only behaviour change is the canonical form of the disease_mondo_id field returned by the Open Targets disease/target tools.

Fixed

  • server.json tags still advertised the five non-integrated sources. v1.1.9 corrected the data-source count (14 → 9) in the description and the README/manifest prose, but the server.json tags array still listed uniprot, rcsb-pdb, interpro, gene-ontology, and human-protein-atlas — none of which has a client. The nine integrated upstreams are AlphaFold DB, Open Targets, ChEMBL, Ensembl, ClinVar, gnomAD, MONDO, HPO, and DisGeNET. The five stale tags were removed so the manifest matches the corrected count.
  • Open Targets disease ids are now returned as canonical colon-form CURIEs. v1.1.9 normalised disease ids on the input path (so get_disease_targets returns results); the output disease_mondo_id still echoed the underscore form the live API returns (e.g. MONDO_0007254) or the raw caller input. A new idempotent _to_curie helper canonicalises the returned id to colon form (MONDO:0007254, EFO:0000305, …) at both construction sites (associated_targets and _row_to_score). It is prefix-agnostic, so EFO / Orphanet ids keep their own prefix rather than being forced to MONDO. Seven regression tests were added — the live underscore form was previously exercised only by colon-form mocks, so the gap shipped unnoticed.

Changed

  • Release workflow. v1.1.10 is the first tagged release built from the fully fixed release.yml: it ships the cosign v3 fix (one self-contained .sigstore bundle; the ignored legacy --output-signature / --output-certificate flags are gone) that was merged to main after the v1.1.9 tag.
  • Counts. 689 → 696 tests; 2,949 → 2,957 statements (790 → 794 branches); 100% line + branch coverage unchanged.

1.1.9 - 2026-06-04

Audit-and-polish work accumulated since v1.1.8, plus one functional fix: get_disease_targets now returns results (it previously returned an empty list for every disease). The remaining changes — residual version drift, a stale statement count, a lint-scope gap that let one style violation ship, a coverage-gate inconsistency, a data-source overclaim (docs advertised 14 sources; only 9 are actually queried), an MCP handshake that reported the framework version instead of the product version, and an ARCHITECTURE.md / documentation-site accuracy pass — do not change runtime behaviour. The tool surface and tool count (29) are unchanged.

Fixed

  • get_disease_targets returned no targets for any disease. The tool passed colon-form CURIEs (e.g. MONDO:0009061) to Open Targets, which keys disease records on the underscore form, so every query matched no disease and returned an empty list under a success status. The normalisation now lives in OpenTargetsClient.associated_targets, so all callers are robust. Verified against the live API: cystic fibrosis now returns CFTR (UniProt P13569) as its top target. (#52)
  • Residual version drift. examples/README.md still printed --version → 1.1.7; updated to 1.1.8 to match every other manifest and document (the v1.1.8 drift sweep missed this one file).
  • Stale statement count. LIMITATIONS.md L6 carried the v1.1.4 figure (2,868 statements) under an "as of v1.1.8" label. The verified pytest --cov count is 2,949 statements / 790 branches (this includes the one-line __version__ import added below; identical across Python 3.10–3.13, since no source uses version-conditional branches).
  • Unlinted entry point. src/alphafold_sovereign/__main__.py contained an E501 (over-length --self-test help string) that shipped because neither the CI ruff step nor noxfile's SRC_DIRS_LINT covered the file. The string is wrapped; the file is now linted.
  • Data-source overclaim corrected (14 → 9). The README, registry manifests (server.json, .well-known/mcp.json, smithery.yaml), the PyPI / pyproject description, CITATION.cff, STATUS.md, LIMITATIONS.md, docs/index.md, and the package docstrings advertised "AlphaFold DB + 13 other data sources." Only 9 upstreams are actually queried (AlphaFold DB, Open Targets, ChEMBL, Ensembl, ClinVar, gnomAD, MONDO, HPO, DisGeNET). RCSB PDB, Gene Ontology, and InterPro are never contacted; UniProt is used only as an identifier namespace; Human Protein Atlas is at most a transitive Open Targets score. The five were removed from the source lists, the count corrected to 9, and an explicit "not integrated in this release" note added to the README data-sources table.
  • MCP server version. The initialize handshake reported FastMCP's own version (3.3.1) instead of the product version, because server/app.py constructed FastMCP("alphafold-sovereign") with no version=. It now passes version=__version__, so a connecting client sees 1.1.8.

Changed

  • CI lint scope now matches the package. The lint job runs ruff check src/ (was four hand-listed subpackages), so server/ and __main__.py can no longer drift out of coverage.
  • CI type-check scope aligned with noxfile. The typecheck job now also covers server/ and __main__.py; the noxfile type session already did, so CI and local nox results agree as noxfile.py documents.
  • noxfile lint scope now includes __main__.py, matching its type session.
  • Coverage gate. [tool.coverage.report] fail_under raised from 99 to 100, matching the threshold CI and the nox cov session already enforce on the command line (--cov-fail-under=100) and the 100% figure advertised throughout the docs.
  • server.json capability. tools.listChanged corrected from false to true to match what the running FastMCP server advertises in the initialize handshake.
  • Dependency hygiene (Dependabot). Consolidated four routine updates, none touching the runtime dependency tree: dev tools ruff 0.15.13 → 0.15.15 and hypothesis 6.152.7 → 6.155.0 (lockfile), and the pinned CI-action SHAs for github/codeql-action (4.35.5 → 4.36.0) and codecov/codecov-action (6.0.0 → 6.0.1). Closes #44, #45, #46, #47.
  • Release workflow fix. release.yml references the SLSA provenance generator reusable workflow by its v2.1.0 tag rather than a commit SHA. The generator self-verifies its own ref (refs/tags/vX.Y.Z) and rejects SHA pins, so the earlier Scorecard "pin actions by hash" pass had silently broken the release pipeline; this restores a working PyPI publish + SLSA provenance + Sigstore-signed GitHub Release.

Documentation

  • ARCHITECTURE.md rewritten to match the shipped code. The previous module map described an aspirational system of roughly fifty modules, most of which do not ship — an HTTP transport with OAuth, MCP resources and prompts, a multi-layer Redis cache, federation, biothreat screening, ed25519 signing, and client modules for UniProt, RCSB PDB, InterPro, openFDA, ClinicalTrials, and PubMed. It now documents only what ships: the stdio transport, the single FastMCP application with 29 tools, the nine upstream clients plus the shared base, the domain types, and the SQLite knowledge graph. A clearly delimited "Roadmap (not yet shipped)" section records the unimplemented items, and the six empty namespace packages are named as reserved. Broken cross-references (docs/THREAT_MODEL.md, docs/adr/) were corrected.
  • Documentation site is link-clean. Corrected the residual "13 other data sources" claim in the mkdocs.yml site description and the "14 upstream APIs" figure in the threat-model diagram (both now 9), fixed the broken docs/adding-a-client.md reference in CONTRIBUTING.md, and repaired the cross-document links that resolved on GitHub but not in the rendered site (absolute URLs for the snippet-included root documents; in-site relative links for the example pages). mkdocs build now emits no broken-link warnings.
  • mkdocstrings. Disabled griffe's warn_unknown_params so the deliberate single-params-model docstring convention no longer produces spurious build warnings.
  • Social-preview cards. Enabled mkdocs-material's social plugin so the documentation site generates an Open Graph card per page (title plus the accurate, non-overclaiming site description); shared links now render a preview instead of a bare URL. Adds the imaging extra to the docs dependency group and the Cairo/Pango system libraries to the docs CI workflow.

1.1.8 - 2026-05-24

A metadata-consistency and validation-posture patch. Closes version drift between pyproject.toml / __init__.py (which were bumped to 1.1.8 in PR #43) and the 12+ documentation and manifest files that still read 1.1.7. Downgrades the maturity claim from Production/Stable to Beta to match the project's actual deployment and validation status.

Fixed

  • Version drift. 12 files (README.md, STATUS.md, LIMITATIONS.md, CITATION.cff, docs/index.md, docs/installation.md, mkdocs.yml, server.json status_note) still read 1.1.7 after the pyproject.toml / __init__.py / server.json / .well-known/mcp.json / smithery.yaml version fields were bumped to 1.1.8 in PR #43.
  • Stale test count. All references to "677 tests" updated to 689, the current pytest --collect-only count (includes parametrised expansions).
  • CONTRIBUTING.md overclaims. Removed unsubstantiated claims that the software is "used by pharmaceutical, clinical-research, and defense organizations" and that "real clinicians, researchers, and analysts use [it] to make consequential decisions." No evidence supports either claim; STATUS.md explicitly states no production deployment has occurred.
  • CONTRIBUTING.md nox sessions. The "How to Run the Test Pyramid" section listed 9 nox sessions (unit, property, contract, client, integration, benchmark, security, mutation, perf) that do not exist in noxfile.py. Replaced with the actual sessions.
  • CONTRIBUTING.md coverage gate. Stated "≥ 95% line, ≥ 90% branch" but noxfile.py enforces --cov-fail-under=100 and pyproject.toml sets fail_under = 99. Corrected to "100% line and branch".

Changed

  • PyPI classifier downgraded from Development Status :: 5 - Production/Stable to Development Status :: 4 - Beta. The project has no production deployment experience (LIMITATIONS L4), no scientific validation (STATUS.md), and no external contributors. Production/Stable was not supported by evidence.
  • Maturity field in server.json, .well-known/mcp.json, and smithery.yaml changed from "stable" to "beta".
  • Language throughout changed from "engineering-grade" to "engineering-validated" to avoid implying a maturity level the project has not reached.

Added

  • STATUS.md validation matrix. A table distinguishing engineering validation (unit tests, coverage, static analysis, security scanning, release provenance) from scientific validation, clinical validation, and regulatory certification — all three marked as "Not performed" / "None".
  • REVIEWER.md. Step-by-step guide for a cold reviewer to install, self-test, run the offline test suite, inspect examples, and verify release provenance.

Verified

  • grep -rn '1\.1\.7' across all version-bearing files returns zero hits (only historical CHANGELOG entries remain).
  • grep -rn 'Production/Stable\|maturity.*stable' across pyproject.toml, server.json, .well-known/mcp.json, smithery.yaml returns zero hits.
  • 689 / 689 tests collected by pytest --collect-only.

1.1.7 - 2026-05-18

Ships the supply-chain hardening originally intended for v1.1.6.

Context — why this is v1.1.7 and not a re-tagging of v1.1.6

The v1.1.6 release pipeline encountered two distinct interacting issues that prevented the wheel from reaching PyPI and the signed assets from reaching the GitHub release page:

  1. The SLSA L3 reusable workflow ref was SHA-pinned alongside every other action. SLSA's trust model intentionally rejects SHA-pinned references because L3 attestations claim "this provenance was produced by a specific named version of the generator"; a commit SHA is not a named version. The provenance / generator job failed at validation. Fixed in PR #39 by reverting that single line to @v2.0.0.

  2. The recovery attempt via workflow_dispatch from main (rather than re-tagging) was incompatible with SLSA L3's build-identity requirement: the generator's final job needs the calling workflow to be triggered by a tag-push so the build identity can be tied to an immutable named ref. A workflow_dispatch from a branch fails this check. The provenance / final job failed as a result.

v1.1.7 ships the same workflow content (every action SHA-pinned except SLSA's reusable workflow, top-level minimum permissions on ci.yml) via the supported push: tags trigger that v1.1.3, v1.1.4, and v1.1.5 each used successfully. The v1.1.6 git tag and its incomplete GitHub release (only the auto-generated source-code zip + tar.gz assets, no wheel) are kept on the repo as an audit artifact of the diagnostic process. The canonical released version going forward is v1.1.7.

Changed (same content as the v1.1.6 supply-chain hardening)

  • Every uses: <action>@<tag> reference across ci.yml, docs.yml, release.yml, scorecard.yml is SHA-pinned to a 40-character commit hash with a # <tag> trailing comment (closes the OpenSSF Scorecard Pinned-Dependencies findings).
  • The SLSA L3 reusable workflow (slsa-framework/slsa-github- generator/.github/workflows/generator_generic_slsa3.yml) is kept on @v2.0.0 per the trust-model requirement above, with the underlying commit SHA in a trailing comment for audit traceability.
  • ci.yml declares a top-level permissions: contents: read block; the codeql job re-declares its own elevated permissions which take precedence (closes Scorecard Token-Permissions findings).

Verified

  • uv lock clean (only project version line changed)
  • All four workflow YAML files parse with yaml.safe_load
  • Every non-SLSA action ref is SHA-pinned (verified by grep -rn 'uses: ' .github/workflows/ | grep -v '@[0-9a-f]{40}' | grep -v slsa-framework returning zero hits)
  • ruff check + ruff format --check clean on CI-scoped paths
  • mypy --strict clean on CI-scoped paths (15 source files)
  • Full pytest suite 677 / 677 passing
  • All six authoritative version sources read 1.1.7; concept DOI unchanged (10.5281/zenodo.20134773).

After release

This release MUST be triggered by pushing the v1.1.7 git tag — not by workflow_dispatch. The release.yml workflow will: build the wheel; generate SLSA L3 provenance via the @v2.0.0 reusable workflow (now reachable through the supported trigger path); publish to PyPI via OIDC trusted publishing; generate CycloneDX + SPDX SBOMs; cosign keyless-sign each artifact; and attach all 13 assets to the GitHub release. Zenodo's GitHub integration will mint a fresh version-specific DOI under the concept DOI as on every prior release.

1.1.6 - 2026-05-18

A supply-chain hardening patch. Closes the actionable OpenSSF Scorecard findings (Token-Permissions × 3 and Pinned-Dependencies × ~40) by SHA-pinning every GitHub Action reference and tightening per-workflow permissions: blocks. No functional code or dependency-tree changes — the only edits to src/ and uv.lock are the version-string bump from 1.1.5 to 1.1.6 (in src/alphafold_sovereign/__init__.py and the project's own row in uv.lock); the dependency graph itself is unchanged. 677/677 repo-wide tests still pass.

Security — Pinned-Dependencies

Every uses: <action>@<tag> reference across the four workflow files is now uses: <action>@<40-char-SHA> # <tag>. This closes the OpenSSF Scorecard Pinned-Dependencies findings (roughly 40 of them, one per uses: line) and prevents a hijacked-tag supply-chain attack: an attacker who compromises an action's tag to point at a malicious revision can no longer cause our workflow to run that revision. The # <tag> trailing comment preserves human readability; Dependabot's uv + github-actions ecosystem config (added in v1.1.3) keeps pinned SHAs up-to-date when new upstream versions release.

Actions SHA-pinned (commit SHA → tag for traceability):

Action SHA Tag
actions/checkout 34e114876b0b11c390a56381ad16ebd13914f8d5 v4
actions/download-artifact d3f86a106a0bac45b974a628896c90dbdf5c8093 v4
actions/upload-artifact 043fb46d1a93c77aae656e7c1c64a875d1fc6a0a v7
astral-sh/setup-uv 37802adc94f370d6bfd71619e3f0bf239e1f3b78 v7
anchore/sbom-action e22c389904149dbc22b58101806040fa8d37a610 v0
codecov/codecov-action 57e3a136b779b570ffcdbf80b3bdc90e7fab3de2 v6
github/codeql-action 9e0d7b8d25671d64c341c19c0152d693099fb5ba v4
ossf/scorecard-action 4eaacf0543bb3f2c246792bd56e8cdeffafb205a v2.4.3
pypa/gh-action-pypi-publish cef221092ed1bacb1cc03d23a2d87d1d172e277b release/v1
sigstore/cosign-installer 398d4b0eeef1380460a10c8013a76f728fb906ac v3
slsa-framework/slsa-github-generator 5a775b367a56d5bd118a224a811bba288150a563 v2.0.0

Security — Token-Permissions

ci.yml now declares a top-level permissions: contents: read block, which Scorecard's Token-Permissions check requires. The codeql job inside ci.yml re-declares its own elevated permissions (security-events: write, actions: read, contents: read) — that job-level block takes precedence for codeql alone. release.yml already had tight per-job permissions blocks (introduced earlier in the release-readiness work); docs.yml already had a minimum contents: write (the only privilege mkdocs gh-deploy needs); scorecard.yml already had appropriately scoped per-job permissions.

After this change, all four workflows declare their permissions explicitly. Scorecard's Token-Permissions findings should close on the next weekly Scorecard scan.

Verified

  • uv lock clean (no dependency changes; project version line only)
  • All four workflow YAML files parse cleanly
  • grep -rn "uses: " .github/workflows/ | grep -v "@[0-9a-f]\{40\}" returns zero hits (every action is SHA-pinned)
  • ruff check + ruff format --check clean on CI-scoped paths
  • Full pytest suite 677 / 677 passing
  • All six authoritative version sources read 1.1.6; concept DOI unchanged (10.5281/zenodo.20134773).

Scorecard findings that remain after this release

The following Scorecard findings are deliberately not addressed here, with documented reasons:

  • Code-Review (high) — requires a co-maintainer with a separate-approver workflow. Structurally unfixable while the project is solo-maintained; see STATUS.md "Project maturity".
  • Fuzzing (medium) — OSS-Fuzz integration is out of scope for a project of this size composed mostly of thin API wrappers; most code paths are exercised exhaustively by the existing 677-test hermetic suite with respx-mocked upstreams.
  • Branch-Protection (high) — already addressed by the "Protect main" ruleset (https://github.com/smaniches/alphafold-sovereign-mcp/settings/rules) active since 2026-05-17; this finding will close on the next Scorecard scan when Scorecard re-reads the repo's protection rules.

The real vulnerability scanner — CodeQL security-extended — continues to report zero findings on the shipped surface.

1.1.5 - 2026-05-17

A dependency-hygiene patch. Closes all ten Dependabot PRs that landed within hours of .github/dependabot.yml going live in v1.1.3, by consolidating their lockfile / workflow updates into a single coordinated release. No code, runtime-behaviour, or scientific output changes — same src/ tree as v1.1.4, 677/677 tests still pass.

Changed (Python lockfile bumps — within existing >= constraints)

uv lock --upgrade ran cleanly and produced the following moves, all verified by the full test suite passing:

  • black 26.3.1 → 26.5.0 (dev formatter; Dependabot PR #29)
  • fastmcp 3.2.4 → 3.3.1 (runtime; Dependabot PR #26)
  • hypothesis 6.152.5 → 6.152.7 (dev test framework; Dependabot PR #25)
  • mypy 2.0.0 → 2.1.0 (dev type-checker; Dependabot PR #27)
  • pymdown-extensions 10.21.2 → 10.21.3 (docs build; Dependabot PR #28)

Plus the following transitives also advanced inside the lockfile but were not flagged by Dependabot: cachetools 7.1.1→7.1.2, click 8.3.3→8.4.0, cyclopts 4.11.2→4.13.0, fonttools 4.62.1→4.63.0, idna 3.14→3.15, jaraco-functools 4.4.0→4.5.0, numpy 2.4.4→2.4.5, python-discovery 1.3.0→1.3.1, python-multipart 0.0.28→0.0.29, requests 2.33.1→2.34.2, ruff 0.15.12→0.15.13, sse-starlette 3.4.2→3.4.4, uvicorn 0.46.0→0.47.0, virtualenv 21.3.1→21.3.3.

Changed (GitHub Actions bumps)

  • astral-sh/setup-uv v4 → v7 across ci.yml, docs.yml, release.yml (Dependabot PR #24).
  • actions/upload-artifact v4 → v7 across ci.yml, scorecard.yml, release.yml (Dependabot PR #20).
  • github/codeql-action/* v3 → v4 in ci.yml (init + analyze) and scorecard.yml (upload-sarif) (Dependabot PR #22).
  • codecov/codecov-action v4 → v6 in ci.yml (Dependabot PR #23).
  • ossf/scorecard-action v2.4.0 → v2.4.3 in scorecard.yml (Dependabot PR #21).

Added — documentation

  • CHANGELOG [1.1.5] section (this entry) enumerating which Dependabot PRs were absorbed.

Verified

  • uv lock clean (no unexpected transitive drift beyond what was triggered by the named bumps)
  • pytest 677/677 passing
  • ruff check + ruff format --check clean on CI-scoped paths
  • mypy --strict clean on mypy 2.1.x (CI-scoped paths)
  • All six authoritative version sources read 1.1.5; concept DOI unchanged (10.5281/zenodo.20134773).

Closes Dependabot PRs

All ten open Dependabot PRs (#20–#29) superseded by this consolidated release. Their lockfile / workflow updates are folded into this commit and verified together by a single CI run, rather than landing as ten separate merges.

1.1.4 - 2026-05-17

An accuracy patch. Closes one bibliographic error, two stale-claim errors, and seven smaller drifts found during a self-audit of every "claim" file in the repo. No code, runtime, or dependency changes.

Fixed

  • Zenodo concept DOI. Six metadata fields cited 10.5281/zenodo.20134774, which is the v1.1.0-rc1 version-specific DOI rather than the project's concept DOI. The actual concept DOI (the one that always redirects to the latest archived version) is 10.5281/zenodo.20134773. Citations of the previous value were silently pinned to v1.1.0-rc1; the correct concept DOI now resolves to v1.1.3 today and will resolve to v1.1.4 once this release archives. The CITATION.cff header comment that mislabelled the rc1 DOI as "the concept DOI minted on the v1.1.0 release" is also rewritten to describe Zenodo's concept-vs-version model accurately. Files corrected: CITATION.cff, README.md (DOI badge + bibtex), STATUS.md, server.json, .well-known/mcp.json.
  • README.md uvx install example still showed uvx --from alphafold-sovereign-mcp alphafold-sovereign-mcp, the pre-#17 form. PR #17 simplified this to uvx alphafold-sovereign-mcp in docs/installation.md per a Gemini suggestion but missed the README copy. Now consistent.
  • STATUS.md Zenodo claim. The earlier wording "minted automatically on every GitHub Release" was historically inaccurate: the GitHub-Zenodo webhook was installed only for v1.1.0-rc1, broke between then and v1.1.3, and was reinstalled today before v1.1.3 published. Rewritten to describe what is now true and verifiable: the concept DOI redirects to the latest archived version, and the integration mints a new version-specific DOI on each GitHub Release.
  • STATUS.md module count. The "20 modules" line was approximate; the actual surface is 5 subpackages with 18 substantive .py modules (10 client modules, 1 domain module, 4 tool modules, 1 storage module, 2 server modules). Reworded to match.
  • AUDIT.md audit log. The previous table listed only the PR #2 Gemini review, undersurfacing the cumulative engineering review history. Expanded to enumerate the seven PRs reviewed by gemini-code-assist[bot] between 2026-05-11 and 2026-05-17 (PRs #2, #6, #16, #17, #18, #19 + the consolidation PR #15) and the commits that resolved each batch of findings. Also clarified the threat-model row (the maturity expectation for an external STRIDE audit was unclear; now stated as "defer until external security audit"). This brings AUDIT.md into honest agreement with the actual review trail visible in the PR archive.
  • LIMITATIONS.md self-references. The "as of v1.1.0" anchor in the preamble was three releases stale (now "as of v1.1.4"). The "~3,000 statements" approximation in L6 is replaced by the verified exact count from pytest --cov output (2,868 statements).
  • Stale "Last updated" dates in AUDIT.md (2026-05-11), STATUS.md (2026-05-16), and LIMITATIONS.md (2026-05-11) are all refreshed to today, 2026-05-17.
  • AUDIT.md as-of line and INCIDENT_RESPONSE.md as-of line re-anchored from v1.1.3 to v1.1.4.

Verified correct (no change)

For traceability, the following claims were verified against the codebase during this audit and do not need correction:

  • 677 tests (pytest --collect-only returns 677)
  • 100% line and branch coverage (2868/2868 statements, 776/776 branches per pytest --cov on this commit)
  • 29 MCP tools (12 disease + 6 precision medicine + 6 structure intelligence + 5 knowledge graph; confirmed via grep -c "^@mcp.tool" per module)
  • 14 data sources (AlphaFold DB + 13 others, per the README table)
  • Version 1.1.4 consistent across pyproject.toml, src/alphafold_sovereign/__init__.py, server.json, .well-known/mcp.json, smithery.yaml, CITATION.cff

1.1.3 - 2026-05-17

A dependency-trimming patch that reduces the runtime surface and closes the open Dependabot alert.

Removed

  • Seven pre-positioned runtime dependencies that no code path imported. Each was annotated in pyproject.toml with a comment describing future work (# JWT — OAuth 2.1 bearer tokens, # Cryptography (signing, FIPS), # Optional: OpenTelemetry (graceful no-op if not installed), # Optional: Prometheus metrics) that was never written. Removed: mcp (still pulled in transitively via fastmcp), pydantic-settings, cryptography, python-jose[cryptography], opentelemetry-sdk, opentelemetry-exporter-otlp-proto-grpc, prometheus-client. The HTTP transport, observability layer, and bearer-token handling will declare their own dependencies at the time they actually ship, when the choice can be made against real requirements.
  • The dead jose.* entry from the [[tool.mypy.overrides]] block that matched the removed python-jose package.

Security

  • Resolves the GHSA-w8m6-9963-pmpv "Minerva timing attack on P-256 in python-ecdsa" Dependabot alert. ecdsa was a transitive of python-jose; removing the latter drops the former from the install tree.

Changed

  • uv.lock regenerated. Thirteen packages drop out of the install tree: ecdsa, python-jose, rsa, pyasn1, prometheus-client, grpcio, protobuf, googleapis-common-protos, opentelemetry-exporter-otlp-proto-common, opentelemetry-exporter-otlp-proto-grpc, opentelemetry-proto, opentelemetry-sdk, opentelemetry-semantic-conventions. pydantic-settings and opentelemetry-api remain in the lockfile as legitimate transitives of mcp and openapi-pydantic respectively.
  • New .github/dependabot.yml adds weekly automated dependency-update PRs for both the Python ecosystem and GitHub Actions, closing the OpenSSF Scorecard Dependency-Update-Tool finding.
  • AUDIT.md and INCIDENT_RESPONSE.md re-anchor the still-true "as of " claims to v1.1.3.

1.1.2 - 2026-05-17

A metadata-coherence patch on top of v1.1.1. The v1.1.1 published artifacts (PyPI wheel, GitHub Release, Smithery / MCP-registry scrape) still carried release-candidate-era manifest fields because the audit landed in PR #17 that merged after the v1.1.1 tag was cut. v1.1.2 rebuilds the wheel and re-publishes so the public-facing metadata matches the actual stable status. No runtime behaviour changes.

Changed

  • pyproject.toml trove classifier raised from Development Status :: 4 - Beta to Development Status :: 5 - Production/Stable so the PyPI surface matches the "stable" maturity declared in server.json, .well-known/mcp.json, and smithery.yaml.
  • server.json "maturity" is "stable" and the registry install command drops the --pre flag (the wheel published for v1.1.1 still carried pip install --pre alphafold-sovereign-mcp).
  • .well-known/mcp.json "maturity" is "stable".
  • smithery.yaml gains an explicit maturity: stable field (the registry manifest only carried it implicitly via the description before) and the description coverage figures are aligned with the rest of the manifests (677 tests, 100% line and branch).
  • docs/installation.md no longer pitches the project as a release candidate, drops the --pre / --prerelease=allow install instructions, simplifies the uvx example to uvx alphafold-sovereign-mcp, and removes the obsolete ## Stable-only pip install section.
  • AUDIT.md and INCIDENT_RESPONSE.md re-anchor the still-true "as of " claims (no external audit, no postmortems) to v1.1.2.
  • Version strings raised to 1.1.2 across pyproject.toml, src/alphafold_sovereign/__init__.py, server.json, .well-known/mcp.json, smithery.yaml, CITATION.cff, STATUS.md, README.md, docs/index.md, mkdocs.yml, docs/installation.md, and examples/README.md.

1.1.1 - 2026-05-17

A maintenance release that hardens the 1.1.0 surface. It resolves every outstanding automated code-review comment, verifies the codebase against the live AlphaFold DB v6 response schema, and makes the structure tools behave consistently when AlphaFold DB has no model for an accession.

Fixed

  • EnsemblClient._first_uniprot returned the string "None" for a list whose first element was not a string; it now type-checks before coercion.
  • AlphaFoldClient.get_pae could return a non-dict (a JSON null body or a null first element of a single-element array); it now always returns a dict.
  • _validate_af_file_url rejected an uppercase HTTPS scheme; the scheme comparison is now case-insensitive per RFC 3986.
  • ClinVarClient exact-match ranking and RefSeq-accession detection were case-sensitive; both are now case-insensitive.
  • _fetch_af_plddt raised TypeError when AlphaFold DB returned a null uniprotSequence; the value is now coerced to an empty string.
  • Structure-intelligence AlphaFold fetchers are routed through AlphaFoldClient and adapt to the AlphaFold DB v6 schema: the removed meanPlddt field is replaced by globalMetricValue.
  • The AlphaMissense UniProt accession is resolved from the Ensembl VEP swissprot consequence field, removing a redundant gene lookup.

Changed

  • The structure tools now return one consistent, human-readable response when AlphaFold DB has no model for an accession, including a structure_available flag. A stale coverage claim was removed.
  • score_binding_pocket_geometry docstring and methodology string now describe the additive druggability index and the pLDDT threshold the code actually applies.
  • Self-description accuracy pass: stale version strings and overstated docstrings were corrected across the repository.
  • Test surface: 677 tests at 100% line and branch coverage, up from 623 tests at 99%. The noxfile.py coverage gate is raised to 100% to match CI.

Security

  • The offline-mode air-gap pre-check resolved its target URL with a manual branch that could inspect a different host than the request reached for a protocol-relative or uppercase-scheme path. It now uses httpx.URL.join, the same RFC 3986 resolution httpx applies internally, so the pre-check inspects the host the request will actually contact.

1.1.0-rc1 — 2026-05-11

First release candidate that ships the refactored modular tree alongside the full audit-readiness kit (STATUS.md, LIMITATIONS.md, AUDIT.md, threat model, examples, mkdocs site).

Added — release-ready surface

  • examples/ with three documented end-to-end transcripts: variant triage (BRCA1 c.5266dupC), target characterisation (EGFR), and drug-discovery (Imatinib → BCR-ABL).
  • mkdocs-material documentation site deployed to GitHub Pages at https://smaniches.github.io/alphafold-sovereign-mcp/.
  • AUDIT.md, docs/threat-model.md (STRIDE × MCP server surface), INCIDENT_RESPONSE.md.
  • smithery.yaml and server.json so the server is discoverable in the Smithery registry and the official MCP server registry.
  • .well-known/security.txt (RFC 9116) and .well-known/mcp.json.
  • noxfile.py with lint, type, test, cov, mutate, docs, build sessions.
  • --self-test and --version CLI flags. alphafold-sovereign --self-test boots the server in offline mode and verifies the deterministic logic of generate_variant_clinical_report on a known variant.
  • Pre-registered 10-prompt benchmark under benchmarks/ with SHA-256-committed prompt manifest and a deterministic re-run verifier.
  • Mutation testing via mutmut; per-module mutation scores published to docs/quality/mutation-scores.md.
  • Release workflow with SLSA L3 provenance attestations and cosign-signed artefacts.
  • OpenSSF Scorecard workflow; badge in README.
  • Zenodo DOI integration via GitHub releases; DOI added to CITATION.cff.

Changed — language scrub for open release

  • Removed marketing and overclaim language from the README, docstrings, and module headers. Specifically:
  • Dropped claims of "drift tensor R²=0.9992" — there is no benchmark in the repository that substantiates this number.
  • Dropped "patent-pending TOPOLOGICA methodology" framing from the structure-intelligence module. The patent status is described soberly in PATENTS.md.
  • Dropped "Defense-grade sovereignty stack (FedRAMP-aligned, FIPS, SBOM, SLSA L3)" comparison line. None of those certifications has been audited; only an SBOM is actually emitted today.
  • Renamed _wasserstein_distance → _fingerprint_distance because the function computes L2 distance between length-normalised fingerprint vectors, not a Wasserstein distance on persistence diagrams. Updated callers, tests, and the tool's annotation title accordingly.
  • The ACMG/AMP output of the variant tools is now consistently described as a "draft surface of the upstream evidence", not a clinical interpretation.
  • The druggability tier is consistently described as a "heuristic".
  • Removed the LICENSE-COMMERCIAL.md file and the "Enterprise Edition" wording everywhere. The project is licensed under Apache 2.0, full stop.
  • Dropped enterprise / commercial contact email addresses from the docs. Contact channels are: GitHub Discussions for questions, GitHub Issues for bugs, GitHub Security Advisories for vulnerabilities.

Changed — test surface

  • 623 unit tests, hermetic (respx-mocked upstreams), runs in under 15 seconds on a laptop.
  • Coverage on the shipped surface: 99% line + branch (99.52% measured), with 19 of 20 modules at 100%. The remaining gap is 1 statement and 2 partial branches in tools/knowledge_graph_tools.py in a defensive sync-fallback path.
  • CI matrix: Python 3.10, 3.11, 3.12, 3.13 × Ubuntu, macOS.
  • CodeQL security-extended runs on every push to a public repo (free tier of GitHub-hosted runners).

Fixed

  • SQL injection (CWE-89) in storage/knowledge_graph.py: parameterised LIMIT clauses and added the _ALLOWED_TABLES allow-list guarding export_to_dict(tables=...).
  • __init__.py no longer eagerly imports legacy parsers/core/ features/topology modules. import alphafold_sovereign now succeeds without numpy, so the --no-deps wheel-import check in CI passes.
  • 24 mypy strict errors across the client modules (proper return-type casts for Any-typed upstream responses; removal of stale # type: ignore comments).
  • 5 bandit / CodeQL B608 hardcoded-SQL flags eliminated.
  • _traverse_network in tools/knowledge_graph_tools.py was awaiting a sync method (kg._fetchall); calling it synchronously now.

Removed

  • _archive/legacy/ — the original monolithic alphafold_mcp.py (5,840 LOC) and its supporting modules moved here. Not packaged in the wheel; excluded from lint, type, coverage, and security tooling. Tracked for deletion in v2.0.
  • tests/verify_config.py — a one-off helper with a hardcoded Windows path; not a test.

Known limitations (see LIMITATIONS.md)

  • ACMG criterion mapping has not been reviewed by an independent clinical geneticist (L1).
  • Druggability tier thresholds are unvalidated heuristics (L2).
  • Upstream API schemas are not pinned (L3).
  • No production deployment experience yet (L4).
  • macOS Python 3.11 test flake (L5).
  • Single-maintainer bus factor (L6).
  • No real-world correctness telemetry (L7).

Unreleased

1.0.0 — 2026-03-01 (Phase 1 baseline)

Added

  • 25 MCP tools across structure, enrichment, and analysis modules.
  • FastMCP server over stdio transport.
  • Local-first AlphaFold structure cache with AlphaFold DB fallback.
  • Persistent homology via Vietoris-Rips (Betti numbers β₀, β₁, β₂).
  • PAE matrix extraction and domain detection.
  • Intrinsic-disorder prediction.
  • GO semantic similarity (Resnik, Lin, Jiang).
  • UniProt metadata enrichment.
  • Multi-mode local cache (sovereign / readonly / disabled).
  • HTML5 API documentation.