Repository navigation
Releases
Gantry v3.4.2 — Studio navigation and headless CAD security
Gantry v3.4.2 improves component-first Studio navigation and removes the blocked VTK dependency from the built-in CAD runtime.
- Navigate components to Work, Files and Requirements, retaining exact saved-state and CAD/code conversation anchors. Share in-progress updates separately from verification and formal adoption.
- Built-in STEP editing, collision/clearance screening and STEP/STL inspection use official OCP no-VTK 7.9.3.1.1, the same kernel version as before. VTK/CadQuery are no longer installed by the CAD extra.
- Reject conflicting full/headless OCP distributions and detect obsolete VTK left in existing environments. CI installs and tests the CAD extra on Linux/macOS.
CAD migration required: create a fresh virtual environment and install Gantry [cad] there, including the interpreter selected by GANTRY_CAD_PYTHON. In-place upgrades do not uninstall old VTK/CadQuery. External CadQuery scripts should use separately managed environments; this release does not patch those environments. See migration instructions.
No ledger migration or rewriting of old evaluation results. Native parsers still require customer-local sandbox/resource limits. Real static geometry and anonymous HTTP continuation were tested; no physical certification, licensed CAD qualification or speedup is claimed.
Install the attached wheel or use the v3.4.2 source tag. SHA256SUMS covers the wheel. The CAD extra downloads the official native dependencies separately.
Gantry v3.4.1
Gantry 3.4.1 hardens event storage integrity and cloud workspace credential isolation, and activates complete cross-platform CI.
- Caller metadata is preserved through lossless storage encoding. Existing event history remains readable without rewriting it.
- Cloud connector credentials use unambiguous workspace namespaces and are checked again before use.
- CI covers the complete Python, product JavaScript and website JavaScript suites on Linux/macOS and Python 3.11/3.12.
- Studio comments retain deterministic order when timestamps coincide.
- Includes the previously unreleased Studio onboarding, contextual Bot work and Crew recovery changes documented in CHANGELOG.md.
Cloud upgrade: provision the new credential environment names and update integration references using the migration instructions in docs/EXTERNAL-INTEGRATIONS.md. Old ambiguous cloud references fail closed. Local single-workspace references are unchanged.
Existing damaged histories are not silently repaired. Preserve them for incident assessment and recover only from verified backups. Tests use anonymous fixtures; this release does not attest to the integrity of a particular live deployment.
Install the attached wheel or use the v3.4.1 source tag. SHA256SUMS covers the wheel.
Gantry v3.4.0 — Shared product workspaces
Gantry 3.4.0 adds shared product workspaces: import an existing saved project, share unfinished CAD/code, and let another authorized participant restore the exact state and continue independently.
- Artifact-first Studio: CAD, Code and Checks; version-bound discussions and sharing.
- Local workspace import, sync and watch; frozen chunked uploads up to 10,000 scoped files / 1 GiB; interrupted and lost-response recovery.
- Recipient-owned credentials, reviewed member grants and session enrollment, revocation, operator provisioning and verified backup/restore.
- Integrated external/native engineering adapters and physical-evidence workflows, with baseline-bound acceptance and durable publication recovery.
- Opt-in Bot experience comparison and continuity work. No automatic paid inference or publisher telemetry.
Validation: 1,279 Python test invocations (1,271 passed, eight optional skips), 31 JavaScript passes, isolated-wheel installation, and actual local HTTP/CadQuery acceptance with two identities. Shared work remained unverified and unadopted until the separate verification/adoption paths. These checks do not demonstrate engineering-time reduction.
Important qualifications: native vendor APIs remain experimental and unverified in licensed applications; Zuken/Capital SDK drivers are not implemented. Customer SaaS authentication, realistic assemblies and live Linux/TLS deployment still need qualification. The included deployment recipe is for an operator-managed invited-team host, not an already provisioned cloud service. The static public demo does not store customer project data.
Documentation:
- Connect and share a workspace
- Membership and cloud operations
- Verification and remaining limits
- Exact native CAD coverage
Install the attached wheel in a security-patched Python 3.11+ virtual environment on macOS/Linux. Core has no mandatory runtime dependencies. Optional CAD/native/feedback dependencies are separate. Back up existing ledgers and preserve credentials separately before upgrading. Existing state and event history are preserved through additive contracts; no destructive migration is required.
Gantry v3.3.0 — Requirements across development and execution
Gantry v3.3.0 — requirements across development and execution
Additive update to the public preview. Requirements can now be traced from an
adopted revision through development states, changes, Mentor proposals, Crew work
and physical evidence. Existing records are preserved; no ledger migration is required.
Added
- Optional pinned requirement references and explicit verification check IDs show
which requirements a change serves and which checks remain unmet or unverified. - API and State UI distinguish historical evidence from current applicability.
Work completion, verification, sharing and formal human adoption remain separate. - Mentor proposals, consultations and delegated child tasks retain their requirement
subset. Explicitly unlinked exploratory work remains supported. - Requirement changes hold reviews and new execution. The local Runner requests
cancellation during execution; stale results remain recorded but cannot satisfy
the changed baseline. Reassess the state and create a new assignment to resume. - CAD and physical plans, reports and checkpoints carry the same requirement
references. Missing or mismatched links cannot produce a passing assessment. - Crew can publish physical evidence through the fenced
publish_physicalaction
using Bot credentials, without administrator capabilities.
Installation
Download the attached wheel and install it in a patched Python environment:
python -m pip install --upgrade ./gantry_ledger-3.3.0-py3-none-any.whl
# Optional local CAD support:
python -m pip install './gantry_ledger-3.3.0-py3-none-any.whl[cad]'See requirement traceability for the state, work and
physical-report contracts. Legacy free-text requirements remain unlinked until
explicitly mapped; no requirement fulfillment is inferred from their text.
Verification and limits
Release checks: 683 Python regression cases (one skipped, no failures), nine
JavaScript rendering tests, installed-wheel imports/CLI checks and package-layout
inspection passed. The 50 pinned optional CAD/runtime dependencies had no known
vulnerabilities reported by pip-audit on 2026-11-02. This is a point-in-time package
advisory check, not a native-library or operating-system security certification.
Local regression includes real CadQuery STEP geometry screening through requirement
coverage, Bot-scoped physical publication, stale-review holds, reassigned-work
resumption and cancellation of an actual Python child process after a requirement
change. Review/model responses and non-geometric measurements are synthetic fixtures.
Detailed checks are recorded in the implementation record.
This release does not certify physical safety, run a strength solver, or establish
live commercial-CAD compatibility. Runner cancellation is cooperative software
process control, not a hardware emergency stop. Customer production LLMs and a
deployed browser service were not exercised. No paid inference or customer project
data is included.
Upgrade and rollback
Stop workers and back up the ledger and artifact store before upgrading. History
is not rewritten. To roll back, stop workers and reinstall the v3.2.0 wheel;
retain all backups and use a separate pre-upgrade copy for older workers. Older
versions do not enforce the new requirement guards and must not execute work that
depends on them. An older runtime's compatibility with newly written records is
not guaranteed. Updating Gantry does not patch Python or the operating system.
Gantry v3.2.0 — Local CAD and physical evidence
Gantry v3.2.0 — local CAD and physical evidence
Additive update to the public preview. No ledger migration is required.
Added
- Optional CadQuery adapter: STEP import, bounded declarative editing and export.
- Static interference/clearance reports bound to the exact CAD inputs.
- Capture external strength, tolerance and prototype measurements against a fixed acceptance plan.
- Publish raw artifacts and physical check outcomes as immutable development checkpoints available to Mentor.
- Missing, stale or incompatible evidence stays unknown; sharing and review never imply formal adoption.
Installation
Download the attached wheel and install it using an updated Python environment:
python -m pip install --upgrade ./gantry_ledger-3.2.0-py3-none-any.whl
# Optional local CAD support:
python -m pip install './gantry_ledger-3.2.0-py3-none-any.whl[cad]'
python -m gantry.hardware --helpCadQuery requires a compatible platform/Python and native dependencies. Tested on macOS arm64, Python 3.12.15, CadQuery 2.8.0. See docs/HARDWARE-ADAPTER.md.
Verification and limits
582 Python regression cases (one skip, no failures), seven JavaScript tests, real STEP edit/collision/recheck and checkpoint publication. A separate authenticated loopback HTTP run used real Codex subscription inference for specialist and coordinator reviews; missing strength evidence produced a conditional verdict, not adoption. See docs/HARDWARE-VERIFICATION.md.
This is not an FEA solver, tolerance simulation engine, physical test-bench driver, continuous-motion verifier or live Onshape/Fusion connector. External measurements must come from those systems. The live verification used an operator-specified CAD edit, not autonomous CAD redesign. Native CAD parsing runs locally and requires an appropriately resource-limited environment for untrusted files.
Upgrade and rollback
Back up the ledger and artifact store before upgrading. No stored history is rewritten. To roll back, stop workers and reinstall the v3.1.1 wheel; preserve the ledger and artifacts. New adapter commands are unavailable on the older version. Updating Gantry does not update Python or OS security patches.
Gantry 3.1.1 — security maintenance
Security maintenance for the public preview. No ledger migration.
- Reject unsupported HTTP content encodings at the API and encrypted feedback receiver.
- Do not expose Python input-parsing exception details in HTTP responses.
- Raise the setuptools build floor to 78.1.1.
- Add
python -m gantry.security_check, an offline, dated interpreter/tooling baseline check. - Enable dependency alerts and security updates; add weekly Dependabot configuration.
- Document Python runtime and certificate/installer updates. PHP and Brotli are not dependencies of this distribution.
Validation: CPython 3.12.15, 552 unittest cases (4 skipped, no failures); 7 JavaScript tests; installed-wheel CLI smoke check; resolved dependency audit with no known vulnerabilities reported. See verification and limitations.
Updating Gantry does not update Python, OS certificates, standalone tools or proxies. Follow the security update guide. This is not a security certification; the CI workflow remains a template.
The release contains only public Gantry code and documentation. Customer project data is not included.
Gantry 2.1.0 — recovery and tool views
Gantry 2.1.0 — recovery and tool views
Early public prerelease. This extends the existing project connection and review
workflow; it does not introduce a new agent organization or a paid model service.
Added
- Recover a stopped Mentor worker using its original private journal and delegated
identity. Core checks current inputs, permissions and execution conditions before
rotating the lease. Saved model output, applied edits and completed verification
can be reused without repeating the work. - Resume an expired project review through
gantry project mentor-recoveror
project_mentor_recoverin MCP. The original review context and saved answer
remain tied to the same evidence. - Render selected evidence fields through explicit, versioned JSON output
profiles, retaining source references and conversion provenance. Supported unit
conversions do not overwrite native data; missing information stays explicit. - Optional CAD screening for hash-pinned STEP files. Each part's bounding box is
computed once per pose, with exact checks for unresolved pairs and measurements
of calculation counts and time.
Fixed
- Reconnecting to a review no longer changes its saved model input merely because
the server clock or lease metadata changed. - Recovery reconciles ambiguous acknowledgements using durable request IDs. Known
live processes, changed premises and revoked permissions prevent unsafe reuse. - Interrupted verification is collected as an unknown outcome, never inferred to
pass or automatically rerun. A new inference attempt requires explicit selection
and consumes the existing allowance; no paid API fallback is introduced.
Validation and limits
The Python suite ran 327 cases: 320 passed and 7 optional-environment cases skipped.
The UI suite passed all 7 cases. Coverage includes HTTP/service restart, lease
expiry, stale inputs, lost acknowledgements, process checks, project MCP recovery,
output mappings and native STEP checks with the optional CAD runtime installed.
Model responses in these regression tests are fixtures, not live-model evaluation.
The built wheel was installed into a clean environment and exercised through the
synthetic project example: saved failure, review, correction, unchanged passing
test, new-client resume and matching audit/replay. Static scanning reported no
high-severity findings; the two existing medium findings remain the allowlisted
SQL query and the sandbox's temporary mount described in the connection review.
Secret-scanner candidates were the same five existing negative-test fixtures.
The checked optional feedback dependencies (PyNaCl 1.6.2, cffi 2.1.1 and pycparser
3.0) had no known advisories at publication review. These checks are not an
independent security certification.
Recovery requires the same identity and private journal; it does not launch an
arbitrary user agent or transfer credentials. Tool profiles are selected JSON
mappings, not universal CAD/electrical converters. CAD screening covers explicit
pairs at a declared pose, not continuous motion or physical certification.
Existing evaluators must explicitly adopt the helper and verify equivalent
results. No end-to-end development speedup is claimed by this release.
See usage, contracts and remaining limits.
Upgrade
Back up the ledger, artifacts and private worker journals; stop active workers,
install the new version, then restart servers/workers. No destructive schema
migration is required, and previous events and artifacts are retained. New journal
fields are additive. Recovery neither extends expired credentials nor changes
feedback consent or formal adoption.
If rollback is needed, stop workers and restore the matching pre-upgrade backup
with its previous executable. Do not mix old workers with newer recovery journals
or assume an older reader supports events written by a newer version.
The preceding release's connection behavior and qualification limits remain
documented in the 2.0.3 connection review.
Gantry 2.0.3 — scoped project connection
Gantry 2.0.3 — scoped project connection
Connect an existing local project with one reviewed delegation. gantry connect . combines saved-file discovery, the existing approval audit chain, a project zone, development state and client MCP setup. Work-capable and record-only connections are explicit. Existing low-level APIs remain supported.
The connected agent can read/edit within scope, run exact approved commands in a sandbox, checkpoint, submit to Mentor, apply a current review's proposed correction, save results and resume. Independent candidates and provisional assumptions are recorded separately from formal adoption. Mentor can use the current client model or an explicitly selected official Codex subscription adapter; no maintainer-funded inference is used.
Validation and limits are in the connection review. macOS command execution and four installed MCP configurations were exercised; GUI permission dialogs, arbitrary CAD runtimes and Linux execution are not claimed as qualified. The bridge does not automatically start or continue arbitrary user agents, nor sandbox their direct host operations.
The change adds event-backed connection and operation records. Existing events, artifacts, principal policies and feedback consent remain unchanged. connect does not silently extend old agent permissions or import old private ledgers. Back up data before upgrading; once new connection events exist, older readers that do not recognize them cannot replay that ledger. See the agent guide for setup, limits, interruption and revocation.
Gantry 2.0.2 — public preview
Gantry 2.0.2 — capture and authorization hardening
- Check committed Git snapshots and historical diffs for common credential patterns and sensitive filenames before upload. Bound snapshot sizes and require a commit-valued diff base.
- Enforce credential expiry and command restrictions before direct Python API GitHub sync, including cached responses.
- Validate feedback receipts against the actual encrypted envelope sent.
Validation: 271 Python cases (268 passed, 3 optional-dependency skips), 7 JavaScript tests passed. Regression tests include denied capture before upload, expired/restricted sync identities and mismatched upload receipts. The tested feedback dependencies have no known vulnerabilities in the release-time lookup.
No data migration is required and existing feedback consent is unchanged. This remains an early local-use preview, without an independent security audit, OS execution sandbox or native encryption at rest. GitHub Actions is not active; release checks ran locally. See the publication review for evidence and limitations.