This release keeps dependencies and maintenance posture current for teams operating this tool.
✓ No known CVEs patched in this version
Topics
ReleasePort's take
Moderate signalThe release updates the backport policy to allow security fixes into the v1.x maintenance branch.
Why it matters: Security patches may now be applied to v1.x, ensuring continued protection for legacy deployments using that branch.
Summary
AI summaryUpdates Stage E verdict, 1964/2560, and 20/100 across a mixed release.
Changes in this release
| Type | Severity | Summary | CVE |
|---|---|---|---|
| Security | Critical |
Backport policy allows security fixes into v1.x maintenance branch. Backport policy allows security fixes into v1.x maintenance branch. Source: llm_adapter@2026-05-27 Confidence: high |
— |
| Feature | Medium |
F2 enforcement passed with zero failures and per-class delta ceiling at 10pp. F2 enforcement passed with zero failures and per-class delta ceiling at 10pp. Source: granite4.1:30b@2026-05-27-audit Confidence: low |
— |
| Performance | Medium |
Stage E1 dispatch sweep achieved 76.7% accuracy across 2560 outcomes. Stage E1 dispatch sweep achieved 76.7% accuracy across 2560 outcomes. Source: granite4.1:30b@2026-05-27-audit Confidence: low |
— |
| Performance | Medium |
Three new Phase F operation classes show low dispatch accuracy on Qwen3-8B-Q4. Three new Phase F operation classes show low dispatch accuracy on Qwen3-8B-Q4. Source: granite4.1:30b@2026-05-27-audit Confidence: low |
— |
| Bugfix | Low |
Stage E1 dispatch sweep had 5 spurious routes identified as errors. Stage E1 dispatch sweep had 5 spurious routes identified as errors. Source: granite4.1:30b@2026-05-27-audit Confidence: low |
— |
| Bugfix | Low |
E2 disposition test schema-bypass half passes all 15 cases. E2 disposition test schema-bypass half passes all 15 cases. Source: granite4.1:30b@2026-05-27-audit Confidence: low |
— |
| Refactor | Low |
Release is a stabilization commit with no surface or behavior changes vs v2.0.0rc1. Release is a stabilization commit with no surface or behavior changes vs v2.0.0rc1. Source: granite4.1:30b@2026-05-27-audit Confidence: low |
— |
Full changelog
Final cut after v2.0.0rc1 (PR #194). No surface or behavior changes vs rc1 — this is
the stabilization commit. See the [2.0.0rc1] section below for the full surface change
and migration table.
Stage E verdict
- E1 dispatch sweep (Task E1, full 512-probe Gate 0b set, 5 reps): 2560 outcomes
against Qwen3-8B-Q4 viachat.owl-atlas.ts.net, 0 errors. Overall dispatch accuracy
76.7% (1964/2560), 5 spurious routes. Baseline committed at
tests/dispatch_eval/runs/rc1__advanced__qwen3-8b-q4__2026-05-27T03-31-27Z.jsonl. - F2 enforcement (Task E3 Step 3): PASS (
f2_pass=true,failures=[]). Per-class
delta ceiling at 10pp holds for every new Phase F operation class on the primary cell.
(Haiku / Llama / Phi cells remain unavailable pergate_0b_decision.json
scope_limitations— same posture as Gate 0b and rc1.) - E2 disposition (Task E2): the dedicated 24-legal-probes test was not authored. The
schema-bypass half (Task D15,tests/test_phase_f_schema_bypass.py) passes 15/15. The
legal half is effectively covered by the 122 uniquezim_get-*dispatch probes from E1
against live Wikipedia plus the branch-level unit tests intests/test_zim_get.py. - E4 (migration conformance in CI):
tests/test_phase_f_migration.pyruns in the
default pytest suite; verified during PR #194 CI.
Known limitation — natural-language dispatch on three new operation classes
Three of the new Phase F operation classes showed low absolute dispatch accuracy on
Qwen3-8B-Q4 in the Stage E1 sweep — but F2 formally passes (these are new classes
with no b13 baseline to regress against):
| Class | Accuracy | Where the model goes instead |
| --- | --- | --- |
| zim_get-summary | 20% (20/100) | 80/100 → zim_query |
| zim_get-structure | 53% (56/105) | 49/105 → zim_query |
| zim_get-main-page | 76% (76/100) | 14/100 → zim_query, 10/100 → zim_metadata |
The model interprets natural-language phrasings ("give me a brief summary of X") as
query intent rather than direct-fetch intent. This is not a surface defect — when
zim_query is dispatched the user still gets a working answer via the natural-language
entry path. Description tuning and/or probe-set relaxation tracked at #199 for v2.5.
v1.x maintenance scope
Per the v1.x maintenance commitment,
the most recent v1.x tag is retained as a parallel maintenance branch until the FIRST of
{v2.5.0 ships, 6 calendar months after v2.0.0}.
- Accepted backports to v1.x: security fixes (always), data-corruption fixes (always),
pre-v2.0.0 crash bugs. - Rejected backports to v1.x: new features, new tools, performance work, refactors.
Weekly OSS security release digest.
The CVE patches and breaking changes that affected production tools this week. One email, every Sunday.
No spam, unsubscribe anytime.
Share this release
About cameronrye/openzim-mcp
Modern, secure MCP server for accessing ZIM format knowledge bases offline. Enables AI models to search and navigate Wikipedia, educational content, and other compressed knowledge archives with smart retrieval, caching, and comprehensive API.
Related context
Related tools
Earlier breaking changes
- v2.0.0a15 _attribute_sections falls back to first section when no section brackets located passage
- v2.0.0a13 canonical‑splice gate tightened to require exact path equality, fixing H2/H3 surface end‑to‑end behavior across all shapes.
- v2.0.0a11 Exposed `content_offset` as top-level `zim_query` parameter, validated >=0, threaded through options.
- v2.0.0a10 `get article M/<key>` now returns ZIM metadata entry rather than aliased C-namespace article body.
- v2.0.0a10 `metadata for <file>` returns concise metadata strings instead of full article bodies for new-scheme archives.
Beta — feedback welcome: [email protected]