Skip to content

armgit / Orkas

Public Imported from GitHub
0

Orkas is an open-source (MIT), local-first AI team for your desktop: a Commander turns your goal into to-dos and coordinates 16 built-in specialist agents plus Claude Code, Codex, OpenCode, OpenClaw and Hermes, with projects, scheduled automations and connectors. Free app for macOS & Windows; built-in models or your own keys.

agent-creator + skill-creator: add 'output language follows user UI language' Hard rule

wytbd44d77458cd
diff --git a/src/builtin/skills/agent-creator/SKILL.md b/src/builtin/skills/agent-creator/SKILL.md
index 286aa205..59366a36 100644
--- a/src/builtin/skills/agent-creator/SKILL.md
+++ b/src/builtin/skills/agent-creator/SKILL.md
@@ -22,6 +22,7 @@ You **must** consult before emitting any `<agent>` container. The protocol below
- **Mutation only via `<agent>` container.** Do NOT use `edit_file` / `write_file` / `bash` to mutate `agent.json` directly. Read for inspection is allowed; every write goes through the container — the sandbox physically blocks direct writes, the rule here keeps you from wasting tool calls probing it.
- **Do NOT dump the container as a workspace file** (e.g. `<name>-agent-definition.xml`). The container is a contract between the LLM and the server; the server parses it inline. An extra `write_file` only leaks an unused XML file.
- **One container per agent being created/edited this turn.** Several containers in one turn are legal only when the user's request spans distinct agents ("tighten the workflow on A and B"). Each container is parsed and applied independently. End the turn after — do NOT call `dispatch_to`.
+- **Output language follows the user's UI language** — `<name>`, `<workflow>` step titles + body, `<inputs>` `label` values, and the prose around the container all go in the user's current UI language (per the "User language" directive in the system prompt; that directive's coverage applies even though this file reaches you as a `read_file` result). `<description_zh>` / `<description_en>` are pinned by suffix. XML tag names, backticked tool / skill names, JSON keys, file paths, and `select` `value` strings stay as-is. The English used in this file is illustrative shape, not literal text to copy.
## Create vs edit decision
diff --git a/src/builtin/skills/skill-creator/SKILL.md b/src/builtin/skills/skill-creator/SKILL.md
index a42f1c76..1b3a434c 100644
--- a/src/builtin/skills/skill-creator/SKILL.md
+++ b/src/builtin/skills/skill-creator/SKILL.md
@@ -33,6 +33,7 @@ A skill is **an independent tool capability**, not a tutorial. When the LLM sees
- **Do NOT dump the container or any inner block as a workspace file.** The server parses them inline and persists to `<skill_dir>/<path>`.
- **Cross-skill writes are no longer supported.** Inside an inline edit chat, only the current skill's directory is writable. Do not try `<<<skill-file skill=...>>>` (deprecated).
- **One `<skill>` container per skill being created/edited this turn.** Several only when the user's request spans distinct skills. End the turn after — do NOT call `dispatch_to`.
+- **Output language follows the user's UI language** — every human-readable part of the SKILL.md you author (section titles, body prose, example labels, `## When to use` / `## How to call` / etc. — write them in the user's UI language, e.g. `## 何时使用` / `## 如何调用` for Chinese UI) plus the conversation prose around the `<skill>` container all go in the user's current UI language (per the "User language" directive in the system prompt; that directive's coverage applies even though this file reaches you as a `read_file` result). `description_zh` / `description_en` frontmatter fields are pinned by suffix. Code blocks, file paths, frontmatter field names (`name` / `description_zh` / `description_en`), `<<<skill-file>>>` syntax, and skill_id strings stay as-is. The English used in this file (including the section names listed in "Quality bar — SKILL.md body" below) is illustrative shape, not literal text to copy.
## Create vs edit decision