Release v0.26.0 · Bevel-Software/Hexis

Upgrade notes

The agent guide is no longer a file in your repository. On the first start of this version the platform removes the AGENTS.md copies earlier versions wrote to every protected branch (only copies that still carry the platform's own header; a file your organisation edited is left as it is) and takes the rule that hid them out of .bevelignore. A root AGENTS.md is your own page from now on: never overwritten, never hidden, and served to agents before the platform's guide. The Agent guide file setting is retired; the guide is read as AGENTS.md on every deployment.
Distributions: the guide is composed from sections the platform ships. Shape it through the CorePorts.agentGuide hook (append, replace by id, drop — the shared file-rules section must come back with text in it). A kb-template/AGENTS.md a distribution still ships is not seeded any more. Removed from the public packages: agentsFilePointerSentence, retargetAgentsFilePointer, mentionsAgentsFile (shared); POINTER_GUIDE_NAME_BUDGET, withPointerSentence, TemplateFilesStep's fourth constructor argument, DeploymentSettingsService.resolveAgentsFileLink (core-backend). IWorkflowService.listChangeRequestsByState takes an optional workspaceId.
No database migrations in this release.

get_agent_guide returns the platform's guide — whole, or one section at a time via section — and the whole-guide answer lists every section. Every tool of the platform's own now opens its description with one sentence: call get_agent_guide first. read_file on AGENTS.md at the knowledge-base root answers with the same guide, after your organisation's own AGENTS.md when it has one, and grep searches it there too.
Five read tools over change requests, shaped after the GitHub pull-request API and answering in Hexis's own vocabulary: list_change_requests, get_change_request, list_change_request_files, list_change_request_reviews, list_change_request_comments. Every field is filtered by what the caller may read; an applied request is read from its own merge commit; a declined one answers its author alone. Summaries carry the last apply failure, with its reason only for a caller who may read every file.
Personal plugins hold skills and tools only. my_plugin and the guide now say that notes and knowledge go under the knowledge folder, private ones included (a folder there can be restricted to one person), and that an agent asked to file a document into a personal plugin should decline and offer the knowledge root. Guidance only; no write is refused.

A connected server's input schema reaches AI clients as the server sent it — deeply nested schemas are no longer truncated on the way. A tool whose schema is genuinely invalid JSON Schema is kept off every agent surface instead of being dropped silently by the client, and the reason (place and message) is shown to the people who manage that server: on the tool page and in list_tool_setup as hiddenTools.

The local server signs in as the agent that runs it — named by the client's handshake, or guessed from the process tree (Claude, Cursor, VS Code, Windsurf, Zed, Codex, Gemini CLI, Cline) — and otherwise as "Unknown agent · local server on ", never as "hexis-mcp". The sign-in line on stderr says where the name came from. stdout is guarded so a dependency's logging cannot corrupt the JSON-RPC channel.

The Docker image carries the agent-guide sections, and a test pins every asset folder against the npm file list and the Dockerfile.
Review fixes from cubic's pass over this release (Hexis #377): schema sanitising budgets, merge-commit recognition for requests merged by earlier versions, tool-description caps, and documentation corrections.