QECTOR v1.0.6

QECTOR Claude Plugin v1.0.6

Release date: 2026-08-23
Release type: Claude.ai marketplace compliance hardening + setup robustness
Full changelog: v1.0.5...v1.0.6

---

Why this release exists

v1.0.5 introduced the cross-platform launcher in a bin/ directory. When the
claude.ai-hosted marketplace pipeline (the one Claude Desktop and Claude Code
sync from on hosted/enterprise setups) reviewed the plugin, it flagged a
hard compliance problem: claude.ai-hosted plugins may not ship bin/
executables. Because bin/ entries are added to the command PATH when the
plugin loads, but are never shown on the admin approval surface, a plugin
shipping one represents a silent code-execution surface that administrators
cannot see or approve. This is an Anthropic platform policy, not a defect in
the launcher itself.

One more correctness gap surfaced during the same review window: the
/qec-setup command hard-required a local scripts/qector_system_setup.py.
On a normal machine that works. But in the sandboxed and remote execution
environments Claude also offers (claude.ai code execution, Cowork, containers
without a project checkout), that script does not exist, so the command
dead-ended. That is now fixed, not just patched.

---

Scope and impact

Who this release affects:

Audience	Impact
Everyone	Launchers relocated from bin/ to scripts/ in every artifact type. If you installed v1.0.5, reinstall the v1.0.6 bundle — do not reuse v1.0.5 cache files.
claude.ai-hosted / marketplace users	This was the direct trigger. The manifests are now canonical and the plugin loads in the strictest approval surfaces.
Windows users	No functional change. The Windows launcher is selected via the win32 override exactly as in v1.0.5 — it just lives in scripts/ now.
Sandboxed / remote users	/qec-setup no longer dead-ends; it falls back to native diagnostics and reports honestly.

---

What changed in v1.0.6

1. Claude.ai marketplace compliance (fixes a platform-policy rejection)

• Launchers moved from bin/ to scripts/. Both qector-python (POSIX
  sh) and qector-python.cmd (Windows) now live in scripts/, with the exec
  bit preserved through packaging (git mode 100755, archive attr 0o755).
  Every reference was repointed: plugin.json, .mcp.json, hooks/hooks.json
  (SessionStart + PostToolUse), the Desktop MCPB manifest, the win32
  platform_overrides block, the builder whitelists, and the bundle validator.
  Commit 2bd1aa4.
• Canonical marketplace manifests. marketplace.json now declares the
  plugin with the relative same-repo source ("source": "./"), the form every
  official Anthropic marketplace uses and the only form guaranteed resolvable
  across every sync path that clones the repository. All fields are from the
  documented schema (homepage, repository, keywords, owner url).
  Commits 5814242 and 7bb49f7.
• bin/ ban guard (regression lock). validate_plugin_bundle.py now
  hard-fails if a bin/ entry appears in any of the three release artifacts
  (plugin zip, source zip, Desktop MCPB). Commit 2bd1aa4.

2. Environment-agnostic setup

• /qec-setup works everywhere. The command now detects whether it is
  running on a real checkout (script present) or in a sandbox/remote/no-clone
  environment, and falls back to native diagnostics (Python version, pip,
  package versions, artifacts directory) instead of failing. It explicitly
  instructs Claude to never claim the script ran when it used the
  fallback — honest reporting by design. Commit c49b34f.

---

Interpreter pinning note (no action unless you set it in v1.0.5)

The userConfig.python_path / user_config.python_path block was removed
from plugin.json in this release. This was the newest and least-standard
manifest field, and the strict claude.ai approval schema does not recognize
it. The feature is not lost: interpreter pinning still works end to end —
set the QECTOR_PYTHON environment variable to an absolute Python 3.9-3.13
path and every launcher honors it before auto-resolution. If you previously
configured userConfig.python_path, migrate by setting QECTOR_PYTHON
instead; there is a one-line note in the README.

Supported environments

• Python: 3.9 through 3.13. Launchers explicitly reject 3.14+ (the
  qector-decoder-v3 wheel matrix does not cover it) with guidance, rather
  than crashing at import time.
• qector-decoder-v3==1.0.0, the pinned decoder runtime.
• Claude Code: Windows, macOS, Linux (via the shipped launchers).
• Claude Desktop: Windows, macOS (Desktop MCPB, safe profile).
• Web / iOS / Android / Cowork: local stdio cannot reach those surfaces
  (Anthropic architecture); a hosted remote connector is on the 1.1.x
  roadmap.

---

What does NOT change

• Surface counts. 28 skills, 11 commands, 5 agents, 4 MCP servers, 8
  stable library tools, 32 research tools, 3 admin tools — identical to v1.0.5.
• All 16 reference theorems remain executable-proof-verified in the test
  suite (74 unit tests, 48 subtests across the math, protocol, and production
  readiness gates).
• The eight-tool safe Desktop profile is unchanged.
• Offline-by-default privacy binary is unchanged (no network calls unless
  explicitly requested).

---

Known limitations

• Sandbox / remote diagnostics depth. The native fallback path in
  /qec-setup reports the core audit (interpreter, pip, package versions,
  artifacts dir) but does not run the full math validation the local script
  performs (that needs the qector-decoder-v3 wheel, which sandboxes may not
  have). It is intentionally reduced scope, clearly labeled, and sessions are
  told to install and validate on a real machine when possible.
• Marketplace sync still needs a private repo on managed orgs. Per the
  Claude documents, if you use Claude Desktop's managed-org marketplace
  pipeline (the one that says "Sync automatically"), the repository must be
  private or internal and owned by the same org as the Claude GitHub App. A
  public repo can still be rejected with a generic sync error on that
  pipeline, even though every direct install path works.

---

Integrity

Artifacts are built deterministically with fixed timestamps, so the build is
byte-for-byte reproducible. SHA-256 hashes are published in SHA256SUMS,
sidecar files, and the MCP Registry descriptor, and the registry entry's
fileSha256 matches the published Desktop MCPB.

Artifact	SHA-256
qector-claude-desktop-1.0.6.mcpb	5b6b4c247ef6159dc92441023fd08a0dc800894a220ba8a61b4143bde92190ff
qector-claude-plugin-1.0.6.zip	e1660a45e87e62d5b74f561273ff7cac2a5367a70b8418f1f9c80ea69591de7f
qector-claude-plugin-source-1.0.6.zip	e0e04d94546799acda0d8f3258bc8911b8125c4bf2b78bb80c04d3e49588dc59

Verification summary: 832/832 source checks · bundle ALL CLEAR · 15/15
release metadata · 74/74 unit tests (+48 subtests) · claude plugin validate . --strict passes · fresh marketplace add + install exits 0.

---

Upgrade

Claude Code: /plugin marketplace update qector-tools then
/plugin install qector@qector-tools.

Claude Desktop: remove the old v1.0.5 extension, download
qector-claude-desktop-1.0.6.mcpb from the release page, install via the
extension UI, restart when prompted.

---

Traceability

Commit	Change
5814242	Canonical marketplace manifests
09ca506	Drop unknown license fields (strict sync)
2bd1aa4	Launchers bin/ → scripts/, bin/ ban guard
c49b34f	/qec-setup environment-agnostic fallback
af8a2ac	v1.0.6 release commit

---

Acknowledgments

Thanks to the claude.ai marketplace review pipeline for surfacing the bin/
approval-surface policy, which drove the relocation, and to everyone who
exercised the marketplace and /qec-setup paths from real Desktop and
sandboxed environments and reported the gaps. Straightforward, reproducible
feedback like that is what makes the resilience layer worth keeping.

---

Questions, edge-feature requests, or QEC research collaborations: open an
issue or reach out via the repository. QECTOR is source-available; the
reference manual ships separately via Zenodo DOI.

QECTOR v1.0.5

QECTOR Claude Plugin v1.0.5

General availability · August 23, 2026

QECTOR v1.0.5 delivers a verified quantum error correction workspace that
installs cleanly on every machine Claude Code and Claude Desktop support.
This release removes the last portability barriers between our software
and your hardware: install once, and the runtime locates, validates, and
uses the correct interpreter wherever your work takes you.

Engineered for every environment

Modern operating systems have diverged on how Python is presented, and
version 1.0.4 assumed more uniformity than the real world provides.
Version 1.0.5 replaces that assumption with engineered resolution.

Every entry point now starts through a launcher we ship and stand behind.
It honors an administrator supplied interpreter first, then discovers a
suitable system Python, verifies the version against the supported range,
and executes only a fully qualified candidate. When nothing qualifies,
the operator receives precise remediation guidance instead of a stack
trace. The supported window is Python 3.9 through 3.13, matching the
published native wheel matrix exactly.

Key improvements

• Universal launcher, shipped everywhere. Present in the plugin
  archive, the source distribution, and the Desktop bundle, with platform
  appropriate selection handled by manifest overrides.
• Interpreter governance on every surface. Administrators pin a
  specific interpreter through configuration on Claude Code and Claude
  Desktop alike; the pinned choice travels to the runtime and always
  supersedes discovery.
• First class Windows packaging. The Desktop bundle resolves the
  command shim on win32 automatically, while bundled runtime builds take
  complete control of interpreter selection.
• Integrity preserved end to end. Launchers are packaged executable
  with deterministic timestamps across every archive, so rebuilds are bit
  identical and verification never degrades.
• Stronger automated gates. Bundle validation now rejects version
  drift in auxiliary archives, confirms launcher presence in the current
  bundle, and the metadata gate resolves inherited versions statically
  with zero runtime prerequisites.

Verification and quality assurance

The release passed the complete gate suite before publication: eight
hundred thirty two structural assertions, seventy four mathematical and
protocol unit tests covering all sixteen reference theorems, source and
bundle validation, metadata cross checks, and deterministic rebuild
comparison. Artifact hashes are published with sidecars, a combined
checksum manifest, an SBOM, and provenance records binding each artifact
to its release commit and runtime pins.

Availability

• Claude Code: install from the marketplace with two commands, or
  update an existing installation and restart your session.
• Claude Desktop: download the Desktop bundle from this release page
  and install it through Extensions with a single click.
• Air gapped deployments: fully supported offline; the default
  configuration performs no network operations.

Web, iOS, Android, and Cowork arrive with the hosted remote connector on
the 1.1.x roadmap.

Licensing

QECTOR is proprietary software from iD01t Productions. The backend engine
remains free for personal, academic, educational, and non commercial
research. Commercial licensing, evaluation terms, and support are
available at qector.store.

Every correction QECTOR returns has been verified against the parity
relation H c = s mod 2 before it leaves the server. That guarantee is
the product, and version 1.0.5 delivers it everywhere you work.

QECTOR v1.0.4

QECTOR Claude Plugin v1.0.4

Production-ready release. Completes the 1.0.3 hardening pass and adds
process-local call budgets so an agent cannot loop mutating or
expensive tools without a ceiling.

What ships

4 MCP servers · 8 stable / 29 research / 3 admin tools · 11 commands
· 5 agents · 28 skills · zero-egress default

• qector-library — 8 stable tools, default-on, every result verified against H c = s (mod 2).
• qector-research — 29 provisional tools, opt-in (--with-research).
• qector-admin — 3 privileged tools, opt-in (QECTOR_ADMIN_ENABLED=1 + per-call confirm=true).
• qector-desktop-mcp — 8 stable tools re-exported by the Desktop safe-profile adapter.

The default install exposes only qector-library (or the Desktop safe
profile for Claude Desktop). Research and admin surfaces are added
explicitly.

Per-process call budgets

The MCP processes enforce their own ceilings because there is no
external rate limiter. Exhaustion returns RESOURCE_LIMIT and the
process stays up. Override a single ceiling with
QECTOR_MCP_MAX_CALLS_<TOOL_NAME> (uppercase). Set -1 to disable.

Tool	Default	Why it is limited
threshold_sweep	8	writes hashed artifacts; large LER grid
decode_single	64	seeded decode loop; can be retried
decode_syndrome	256	per-call decode compute
build_code_from_matrix	32	arbitrary parity-check matrix
hot_path_microbench	4	machine-scoped; no portable claim
system_setup	2	can install packages, write dirs
configure_claude_desktop	2	rewrites Claude Desktop config
workbench_probe	2	launches a local executable

Input-size ceilings (QECTOR_MCP_MAX_DISTANCE, QECTOR_MCP_MAX_TRIALS,
QECTOR_MCP_MAX_SWEEP_POINTS, QECTOR_MCP_MAX_MATRIX_CELLS) still
apply on every call.

Identification and security

SECURITY.md carries:

• An Identification block with the canonical name, version, license,
  author, repository, reference-manual DOI, and privacy URL.
• A per-tool risk classification table that maps every tool to
  Read / Compute / Build / Write (local) / Launch (privileged).
• A "Static Scanner Notes" section that addresses PolicyLayer,
  MCP Registry, and CodeQL-style audits explicitly.
• A "Runtime Dependabot Advisories (status)" table that documents
  why the three open mcp==1.26.0 advisories (CVE-2026-59950,
  CVE-2026-52869, CVE-2026-52870) are non-applicable to this
  release — every QECTOR server uses stdio transport only, and
  the QECTOR wheel itself does not require mcp.

Canonical artifacts (this release)

Artifact	SHA-256
qector-claude-plugin-1.0.4.zip	6e0a36e96a52991d3f2bd9028f286aa29e51a78cd150da1ba46c052002410d12
qector-claude-plugin-source-1.0.4.zip	7c405144f367a449a446e2597e3fd8e0256f4dddc2b78065db49c498c30263ad
qector-claude-desktop-1.0.4.mcpb	e1301f88761addaa705408cb055ae8362b07f6d2825483f153104777aeee6a64

Per-file .sha256 sidecars, combined SHA256SUMS, SPDX-2.3 SBOM, and
provenance.json (with git commit + per-artifact hash) are also
attached.

Install

# Claude Code
claude plugin marketplace add GuillaumeLessard/qector-claude-plugin
claude plugin install qector@qector-tools

# Claude Desktop (Windows / macOS / Linux)
# The Desktop installer is included in the plugin zip and the
# qector-claude-desktop-1.0.4.mcpb is the artifact Claude Desktop's
# MCP Registry downloads automatically.

License

Proprietary — Copyright © 2026 Guillaume Lessard / iD01t Productions.
See LICENSE.md in the source archive. The qector-decoder-v3 backend
is free for personal, academic, educational, and non-commercial
research; commercial use requires a paid license
(qector.store/pricing).

v1.0.2 - Claude Code and Desktop App Connector

QECTOR Quantum Error Correction Plugin v1.0.2

This release brings 100% Native Claude Code and Claude Desktop (Windows App) integration.

🚀 Key Features

• Settings -> Connectors GUI: Fully automated, 1-click Windows installer that registers QECTOR natively into Claude Desktop's Extensions interface.
• Claude Code Readiness: Native .mcp.json\ mapping and \CLAUDE.md\ architecture grounding for CLI workflows.
• Zero-Egress Security: 100% local, bare-metal decoding with zero telemetry leakage.
• Mathematically Verified: Strict fail-closed c \equiv s \pmod 2$ enforcement across 10 code families.
• Statistical Rigor: Exact Wilson 95% binomial score intervals for all LER sweeps.

📦 Artifacts

• qector-claude-plugin-v1.0.2.zip: The complete extension, MCP server suite (37 tools), 28 skills, and 5 agents.
• qector-qector-core-skill.zip: The standalone core skill package.