Changelog

All notable changes to the Junis kit. The format follows Keep a Changelog; versions are the release tags users pin with JUNIS_KIT_REF.

[Unreleased]

Added

• Wiki: junis_wiki and wiki gain verify (a person confirms a page is correct today — text and version stay, its review timer restarts; hosted POST /api/external/management/wiki/verify) and lint (read-only check of the whole wiki — pages past their review date, broken [[links]], orphans, empty pages, missing annotations, oversized pages — with a findings count; hosted GET /api/external/management/wiki/lint). lint takes no file_key. references/memory-and-identity.md describes soul version history and restore (get_soul_history, restore_soul_version), memory kept per (user, organization), and the wiki review period.
• Knowledge: references/knowledge.md gains "Live dashboards and custom front-ends" — building a team's own screen on an organization's Knowledge with the hosted server's new knowledge_changes (what changed since a cursor; SSE form GET /api/external/brain/changes/stream), knowledge_entities_update_attrs (state updates by type + external_id with no re-import and no credits), knowledge_map_limits / knowledge_map_set_limits (the map caps, with confirm_cost), knowledge_graph_snapshot(as_of=…), and the External-only OpenAPI document at GET /api/external/openapi.json. Browser tokens (POST /api/external/brain/browser-token, read-only, ≤15 min, bound to the page origin) let the page call the read endpoints directly. knowledge_geo_near / knowledge_geo_within (entities near a point or in a box, from a geo_point attribute) and the AG-UI form of the run events (GET /api/external/sessions/<id>/events?format=agui) for screens. knowledge_changes_webhook / _set / _delete: a signed webhook that pushes the change feed every minute (Standard Webhooks headers, secret named by an Org Key).
• examples/live-dashboard/: a single-file reference front-end (browser token, paged first draw, SSE follow, map dots from a geo_point location) and the platform's new ops knowledge-map preset it is written against.
• wiki get|put|append: safe wiki page writes from the command line — empty bodies are refused, the current page is backed up under the kit home (one file per write), the write carries the version that was read (a 409 comes back as conflict with the backup path), and the page is read back and compared before success is reported. The Claude Code hook asks before wiki put like it does before api PUT.
• Actions: the skill's new reference references/actions.md covers the platform's typed, approvable operations on an organization's own systems (connections, drafts, OpenAPI import, dry runs, assignment to agents, runs and the run log) through the hosted server's actions_* tools, including the simulate connection for trying the flow with nothing connected. The kit drafts and runs; approving a draft, disabling an action and approving a waiting run stay with a person in the web app.
• Five behaviour eval cases for Actions (not yet run).
• Knowledge: references/knowledge.md covers set questions over the graph (knowledge_graph_around) and verified graph questions (knowledge_graph_templates, knowledge_graph_template_create, knowledge_graph_template_test; approval stays with a person in the web app), qualifiers on neighbor edges, what a kind's property definitions check, and the web-only settings for restricted kinds and member browsing.

Security

• Confirmation now also covers Actions: actions_run (it acts in a system outside Junis; dry_run=true is not asked), actions_agent_set, actions_connection_update, actions_connection_delete, actions_draft_delete, actions_draft_update (it replaces the whole draft in place, as the External PUT already was asked), actions_import_openapi and actions_import (only when it really writes: the tool previews by default and is asked with dry_run=false; the route imports by default and is asked unless dry_run=true), and through junis_api POST /actions/run, POST /actions/import/openapi and POST /actions/import.

• junis_wiki (actions get, put, append) gives agents the same safe wiki operations through the MCP bridge; put on an existing page needs confirmed=true.

Changed

• A write whose body sets instruction or content to an empty string is refused before it leaves the kit, on every path: api, tools call and the MCP bridge (junis_api, junis_call_tool). Typical cause: instruction=@file with a file that was never written.
• api --file with a missing or empty payload file is now a clear file_not_found / empty_payload error (before: a raw ENOENT or a generic bad_json).
• actions_run is given 420 seconds before the kit reports a timeout, because the platform holds the call open while the action executes.
• Browser login asks for 29 permissions: the 26 as before plus actions:read, actions:run and actions:manage. A platform that does not know one of them yet refuses the list (400 invalid_scopes); the kit then asks once more for the permissions that platform offers, so the key is the same as before this change instead of the platform's narrower default set. Permissions named with --scopes are still never replaced.
• The hint for a missing permission says that browser login requests the permissions the platform offers, rather than every permission.

Fixed

• restore_soul_version now asks before it runs, like update_soul_content and rollback_agent: it replaces the soul or identity text for every agent of the organization.

• wiki verify on a platform that does not have the route yet answers unsupported instead of not_found (which read as "the page is missing").

• The catalogue fixture carries the hosted server's 214 tools, including update_soul_content's expected_version and summary.

• tools describe <wrong name> suggests tools by name words (make_agent → create_agent, clone_agent, …) instead of by description search, which stopped naming create_agent once the catalogue grew.

• The catalogue fixture the tests check documentation against now carries all 202 platform tools (actions_* and knowledge_* were never checked before). doctor decodes the bridge's output as one UTF-8 stream.

• doctor reported the local bridge as broken on macOS once its tool list grew past 8 KiB: it parsed the second answer before the pipe had delivered all of it. It now waits for whole lines.

[0.1.0] - 2026-11-05

First tagged release.

Security

• URL overrides (JUNIS_API_BASE, JUNIS_MCP_URL, JUNIS_WEB_URL) and saved addresses are accepted only for junis.ai, its subdomains and this machine. Any other host needs JUNIS_ALLOW_CUSTOM_HOST=1, and is then named by status, doctor and the bridge's start-up log (host_not_allowed otherwise).
• The Claude Code Bash hook also reads the kit's shape under another script name (node <script> api …, node <script> tools call …), so a copy or link of junis.mjs is asked about. A hook event that cannot be parsed but names the kit is asked about instead of passing.
• On the command line, a destructive api or tools call run without a terminal stops with exit code 2 until it is repeated with --yes.
• install.sh refuses a JUNIS_KIT_REF or JUNIS_KIT_REPO that is an option or several words, passes both to git after --, and switches an existing clone to a changed JUNIS_KIT_REPO instead of ignoring it.

Added

• login --browser --scopes read-only|operator|all|<list> asks for a narrower key. A list the platform refuses is an error, never widened.
• CI on GitHub Actions (Node 18 and 20), code owners, a bug report template, this changelog and a "Supported versions" section in SECURITY.md.

Changed

• Documentation: the hook's limits (name and shape, wrapper scripts not caught, 10-second fail-open), that login under an override saves the address, the masked key returned by status, Windows file permissions, and wording about the platform's own junis-magic-auto key.

Initial (2026-09-30)

The kit as first published: the junis skill, the stdio MCP bridge (mcp, mcp --lazy), the command line (status, login, tools, api, run, session, lint, snapshot, doctor) and the installer for Claude Code, Codex, Cursor, Gemini CLI, OpenCode and Copilot.