Changelog

Notable changes to the WhisperGraph skills. Format follows Keep a Changelog; versions follow semantic versioning as defined in CONTRIBUTING.md.

A changelog matters more here than in most repositories: once the plugin is listed, updates reach installed users without them asking, and this is the only place they can see what changed in the instructions their assistant is following.

[2.0.0] — unreleased

Everything in the previous release was written against a tool surface that no longer exists. This is a rewrite, not an edit.

Breaking

• Skills moved from the repository root into skills/. The documented copy command changes from cp -r whisper-skills/whisper-* ~/.claude/skills/ to copying out of skills/, and npx skills add whisper-sec/whisper-skills now does it for you on any client.
• Removed four tool names that no longer exist on the connector. The skills previously instructed the assistant to call a history tool, a domain-variants tool, and two schema-introspection tools. All four are gone; the current surface answers the same questions through the schema tool, the workflow gallery, and procedures called inside a query. A first run against the old skills failed.
• Removed scripts/format_sweep.py. It hand-rolled what a single gallery workflow now does server-side, and it imposed a Python runtime requirement that no client enforces and several do not satisfy. Report formatting moved into the skill body.
• Removed the sixty-pattern cookbook and the bundled schema dump. Both were snapshots of a moving target and neither was tested. Replaced by a maintained pattern set that CI plans against the live graph on every change, and by an instruction to read the schema live.

Added

• whisper-bulk-triage — a fourth skill for a list from a SIEM, EDR or spreadsheet. Splits results into assessed, never-seen and check-failed so an unobserved indicator is never counted as clean.
• A Claude Code subagent of the same name, so a large run happens in an isolated context and returns only the table.
• /whisper-setup — connects the connector, or works out why it is not answering.
• Plugin packaging: a Claude Code plugin manifest and marketplace entry, plus an Agent Plugins 1.0.0 manifest for cross-vendor plugin surfaces.
• Coverage semantics throughout. Every skill now reads the coverage fields before reporting a verdict, and states plainly that "never seen" is not "clean".
• An identical untrusted-data block in every skill, with a check that fails the build if the copies drift apart.
• CI that runs with no credentials, so a fork's pull request gets the same verdict a maintainer's does: frontmatter validation against the Agent Skills specification, a live check that every tool name, resource and workflow slug still exists, a live check that every Cypher block still plans, a lint that forbids graph statistics in prose, link checking, and manifest agreement.
• A nightly drift job that opens an issue when the live surface stops matching the repository.

Changed

• All three surviving skills rewritten against the current connector: the workflow gallery, the schema tool, host identity, documentation on demand, the typed error model, the self-correction contract, pagination and parameter binding.
• Corrected three claims that were repeated from the documentation and turned out to be wrong when measured against the live graph, all three of which would have made an assistant misread a result: a variable-length pattern does follow a query-time edge and return rows; the threat-listing edge carries no properties, so recency comes from a verdict's sources[] rather than from Cypher; and a verdict row can arrive with no coverage qualification at all, or with a zero score alongside a populated recoveredScore — both of which now have explicit handling rather than being read as clean.
• whisper-brand-protection now runs the gallery's typosquat workflow instead of a five-step hand-rolled sweep, and separates registered from weaponised so a defensive registration is never reported as an attack.
• Frontmatter reduced to the six fields in the Agent Skills specification. A runtime-specific field loads in Claude Code but is rejected outright by claude.ai upload and the Skills API, which would have quietly made these Claude-Code-only.
• Every reference to a graph statistic removed from prose, and a lint added to keep it that way.

[1.0.0] — 2026-05-14

Initial release: three skills for threat investigation, Cypher authoring, and brand protection.