Skip to main content

Release 1.4.0

Release Number1.4.0
Release TypeFeature
Release DateNovember 17th, 2026
Tagv1.4.0

Release summary​

After upgrading, you can plan an Infrahub upgrade before you run it, with every release on the way checked for breaking changes that affect your repository and instance. You can also convert device types straight from a running NetBox instance, where before you had to export them manually first. When Infrahub is slow, you can also add the traces of the slow requests to the support bundle, and the analysis reports which service took the time. Guidance for generators, schemas, menus, and generated protocol files that described Infrahub behavior inaccurately has been corrected.

What to expect after upgrading

Nothing about how you invoke the skills has changed. Two items are worth a review on your side.

  • Review committed protocol files that were generated offline with --schemas.
  • Expect permission prompts for shell commands and file writes while a skill is active.

See Upgrade notes for full details.

Plan an Infrahub upgrade before you run it​

Plan a move from the Infrahub version you run to a target version with the planning-upgrades skill. You get a written plan; the skill does not run the upgrade (#159).

What changed:

  • Get a route of one hop per minor version, as the Infrahub upgrade guide requires, written to UPGRADE_PLAN.md with one section per hop.
  • Check each release's breaking changes against your repository and instance. Every finding records whether your deployment is affected (yes, no, or unknown) and the file, schema kind, query result, or probe output behind that verdict.
  • Plan upgrades within one edition. The skill does not plan a move from Community to Enterprise, or back.

Convert device types from a live NetBox instance​

Export device types from a running NetBox and convert them with the converting-netbox-device-types skill, without exporting the instance manually before conversion (#129).

What changed:

  • Run netbox_export_device_types.py with --url, --output-dir, and a NETBOX_TOKEN to pull device types and their component templates over pynetbox. Add --module-types to export module types and their templates as well. It writes them in the devicetype-library layout that the bundled converter already reads.
  • Narrow the export with --in-use, which keeps only types that NetBox counts at least one device for, or at least one module for when you pass --module-types, and with --manufacturer or --slug.
  • Load device types without silent overwrites: device types that share a model are now refused, where before they overwrote each other on load. An unknown --manufacturer exits with code 2.
  • Carry module ports for the generator with the new schema-library-module-ports.yml mapping profile.

Offer traces for a slow Infrahub instance​

When you report that Infrahub is slow, the collecting-diagnostics skill offers to add the traces of the slow requests to the support bundle (#234). Traces need infrahub-collect 2.5.0 or later.

What changed:

  • Add traces only after you agree. The skill asks whether to include them, for which window (--traces-since for a slowness that is still going on, --traces-start and --traces-end for one that ended), and which minimum duration of trace to keep (--traces-min-duration, default 1s).
  • Get more traces of a slowness you can reproduce. The skill explains how a user with the super_admin permission raises the trace sampling ratio for a limited time and removes it after the bundle is created. You run these calls; the skill does not.
  • Keep the API token out of the conversation. The skill says how INFRAHUB_TRACES_API_TOKEN reaches the command, and never asks for the token value.

Find the slow service in the traces of a bundle​

When a bundle has traces, the analyzing-diagnostics skill reads the trace files and reports which service and which operation took the time (#243).

What changed:

  • Get the slow operation from the bundle files. The skill reads the traces entry of bundle_information.json, then each file of bundle/traces/, from the slowest. For each span, it computes the duration and the self time: the duration minus the durations of the direct children of the span. A large self time points to work that has no span, such as a Git operation.
  • Keep the analysis on the bundle. The skill does not start or query a Tempo, Grafana, Prometheus or the Infrahub API, not even a local one. Outside the bundle, it reads only the opsmill/infrahub issues and docs.infrahub.app.
  • Know what to collect next. When the traces cannot answer the question, the report says what the next bundle needs: for example --include-traces with a window that covers the slowness, or a larger --traces-max-size when traces were left out.

Bug fixes​

  • client.create() in a generator drops any field the kind does not have, with no error, and ignores every keyword field when data is not empty. The managing-generators skill now warns about both, and the generators, transforms, common, and auditing-repo skills no longer claim that generated protocol classes type-check the field names passed to create() (#193).
  • The on_delete rule in the managing-schemas skill now states the default Infrahub applies when on_delete is omitted: cascade on kind: Component relationships and no-action on every other kind. Before, the rule said every kind defaulted to no-action, so Component peers were expected to survive when their owner was deleted (#190).
  • Generator targets: accept any group kind, not only CoreGeneratorGroup. The skill recommends reusing the group that holds the design objects and creating a CoreStandardGroup only when none exists, and the audit no longer flags a CoreStandardGroup target (#191).
  • The common skill now says to regenerate the committed protocol file from a running Infrahub server with the schema loaded, with --branch when the schema is on a branch, and to check the diff for classes or fields that disappear (#192).
  • Remove items from a loaded menu with the menus skill: a git sync deletes items dropped from a synced file, while items loaded with infrahubctl menu load stay until you delete them with infrahubctl object delete (#25).
  • The duplicate-issue search in the reporting-issues and reporting-skill-gaps skills returns results again. It no longer passes --state all, which gh search issues rejects (#136).
  • Audit findings from the auditing-repo skill list only live reference sites in sites: a file that nothing registers, includes, or imports is reported as dead code instead. A kind whose Dropdown has a different choice list, or whose relationship has a different peer, now counts toward a duplicate shape, and the difference is recorded as a feasibility blocker.
  • Analyze a pasted log excerpt with the analyzing-diagnostics skill without first being asked where the bundle is; what the excerpt lacks is listed under Open questions. Every report now states the bundle manifest, including when every collector succeeded, and writes the exact gh search issues line for each query, including queries that could not be run.
  • Lessons from the teaching-concepts skill name each node by its full kind, such as CampusSwitch, as the schema and the API do.
  • fetch_schema_limits.py in the managing-schemas skill reads the string-length limits from the public Infrahub schema again, such as 128 characters for a description. The public schema and /api/openapi.json name their definitions NodeSchemaWrite and so on, so the script returned no limits, and --check reported no over-limit field. The script now exits with an error when a source has no limits.
  • The managing-checks skill now states which repository commit a proposed change's checks run, and how to get a check fix merged to main into an open proposed change on a branch synced with Git (#194).
  • The output rule of the converting-netbox-device-types skill lists all five files the converter writes, including 04_module_types.yml and 05_module_templates.yml. Before, it listed three, so module types converted with the skill were not loaded (#135).

Minor changes​

Plugin

  • The skills no longer pre-approve tools through allowed-tools, and the SessionStart hook gives Claude the real plugin path instead of the literal ${CLAUDE_PLUGIN_ROOT} text. A plugin branch now carries only the files the plugin ships (#215).
  • claude plugin validate --strict passes: the SessionStart hook path is quoted so it works when the plugin root contains spaces (#212).
  • The Claude plugin directory shows the Infrahub logo, with OpsMill as the publisher (#214).

Documentation

  • The documentation site is organized around the Infrahub lifecycle, with a first-schema tutorial, a skill router page grouped by lifecycle stage, and a contributor section for the skill-change pipeline (#149).

Developer experience

  • scripts/check-skill-registration.py fails a user-invocable skill that is missing from any registration surface, or whose reference page lacks the standard sections (#160).
  • The skill-change pipeline searches other branches for a pull request that already covers a change before analysis or design starts (#181).
  • An evaluation trial whose output shows the agent read the grader is void and is run again (#184).

Reliability

  • Three evaluation graders that crashed before scoring now run, and the weekly scheduled evaluation run starts again (#209).
  • Pull requests with a missing or conflicting release bump label are rejected (#165).
  • Evaluation graders that scored correct answers 0.0 now grade them correctly: they find skill files where an evaluation trial places them, and grade what their rules require. Evaluation trials run one at a time per skill.

Upgrade notes​

Protocol files generated offline​

If: your repository commits a protocol file that you generated offline with infrahubctl protocols --schemas.

Then: regenerate it from a running Infrahub server that has the schema loaded, adding --branch when the schema is on a branch, and check the diff for classes or fields that disappear.

Notes: offline output has no Profile or Template classes and no relationships added through extensions. Offline generation is still fine for a scratch file.

Permission prompts while a skill is active​

If: you relied on a skill running shell commands without asking, because every skill pre-approved Bash through allowed-tools.

Then: approve each tool when Claude Code asks, or add the commands you trust to your Claude Code permission settings.

Notes: the skills no longer list allowed-tools. While a skill is active, Bash, Write, and WebFetch now ask for permission, as they do outside a skill. The read-only commands in a skill's inline lines (find, head, cat, echo) still run without a prompt (#215).

Full changelog​

Added​

  • netbox_export_device_types.py in converting-netbox-device-types exports device types, module types, and component templates from a live NetBox instance (#129).
  • infrahub-planning-upgrades skill plans an upgrade one minor version at a time and checks every release's breaking changes against the repository and instance (#159).
  • managing-checks states which repository commit a proposed change's checks run (#194).
  • collecting-diagnostics offers infrahub-collect create --include-traces for a slow instance, after you agree, with a time window. Needs infrahub-collect 2.5.0 or later (#234).
  • analyzing-diagnostics reads the trace files of a bundle and reports the service and the operation that took the time; for a slowness without usable traces, it asks for a next bundle with --include-traces (#243).

Changed​

  • managing-generators warns that client.create() silently drops unknown fields and ignores keyword fields when data is not empty (#193).

Fixed​

  • Menus skill covers removing items from a loaded menu (#25).
  • Duplicate-issue search no longer passes --state all to gh search issues (#136).
  • managing-schemas states the on_delete default for Component relationships (#190).
  • managing-generators accepts any group kind as a generator target (#191).
  • infrahub-common regenerates committed protocol files from a running server (#192).
  • auditing-repo lists only live reference sites and counts differing kinds toward a duplicate shape.
  • analyzing-diagnostics analyzes a pasted excerpt, always states the manifest, and always writes the gh search issues line.
  • teaching-concepts names nodes by their full kind.
  • fetch_schema_limits.py in managing-schemas reads the *Write definitions of the public Infrahub schema and of /api/openapi.json, so --check finds over-limit fields again. It exits with an error when a source has no limits, instead of reporting no over-limit field.

Housekeeping​

  • Skill registration surfaces listed in one place and checked by scripts/check-skill-registration.py (#160).
  • Void reading added to the evaluation discrimination proof (#184).
  • NetBox device-type converter defects from an end-to-end run fixed (#129).
  • Three crashing evaluation graders fixed, and the weekly evaluation run starts (#209).
  • react and react-dom Dependabot updates grouped into one pull request (#189).
  • Documentation site restructured around the Infrahub lifecycle (#149).
  • Release bump labels standardized and enforced (#165).
  • Skill-change pipeline checks other branches for existing pull requests (#181).
  • Watch grader checks match whole path segments and identity tokens (#164).
  • Evaluation graders and tasks that scored correct answers 0.0 on the release evaluation gate fixed, and evaluation trials run one at a time.
  • Claude plugin directory validation findings cleared, and a plugin branch built with only the shipped files (#215).
  • claude plugin validate --strict passes (#212).
  • Infrahub logo added and OpsMill named as the plugin author (#214).
  • converting-netbox-device-types output rule lists all five output files, and a test fails if the rule and OUTPUT_FILENAMES disagree (#135).

Skills included​

  • infrahub-managing-schemas
  • infrahub-managing-objects
  • infrahub-managing-checks
  • infrahub-managing-generators
  • infrahub-managing-transforms
  • infrahub-managing-menus
  • infrahub-auditing-repo
  • infrahub-analyzing-data
  • infrahub-reporting-issues
  • infrahub-reporting-skill-gaps
  • infrahub-collecting-diagnostics
  • infrahub-analyzing-diagnostics
  • infrahub-importing-data
  • infrahub-teaching-concepts
  • infrahub-converting-netbox-device-types
  • infrahub-planning-upgrades