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.
JavaScript agent-collaborationagent-harnessagent-orchestrationagent-teamagent-workflowai-assistantai-automationai-employeeai-productivityai-teamai-workflowautonomous-agentsclaude-code-guiclaude-coworkdesktop-agentlocal-firstmulti-agentmulti-agent-systemself-evolving-agentstask-automation
Loosen agent workflow rule to allow prose-only reasoning bullets
wyt0576dda0d192
diff --git a/src/main/prompts/chat_agent_setup.md b/src/main/prompts/chat_agent_setup.mdindex db42ec57..4749b8af 100644--- a/src/main/prompts/chat_agent_setup.md+++ b/src/main/prompts/chat_agent_setup.md@@ -10,14 +10,14 @@ You are working with the user to refine a custom agent — polish its "name / de```### N. <verb-led step title, 5–10 chars>- - `tool_name(key params)` — one-line purpose & inline result- - next action ... (in physical order)+ - `tool_name(key params)` — purpose & inline result (when a tool is invoked)+ - reasoning / decision / synthesis bullets: plain prose, no tool name- branches use nested bullets (`if X → call A` / `else → call B`)```The previous step's result / inbound message / accumulated session context are the default carry-over and need not be restated. Exception handling / retry / skip is decided by the runtime agent, not written into the workflow.- **Hard constraint — every action must explicitly write the tool name / skill_id** in backticks (e.g. `read_file` / `kb_search` / `social-fetch` skill); do NOT write abstract verbs like "read the file" / "do a search". Why: ① workflow is injected into the runtime agent's system prompt; missing tool names force secondary inference, which often picks the wrong tool or misses one; ② the `<skills>` closure is extracted from skill_ids that appear in workflow — without skill_id, the closure can't be derived.+ **Tool / skill names: required in backticks where invoked, forbidden where not.** Every invoked tool / skill_id appears in backticks (`read_file` / `kb_search` / `social-fetch` skill — no abstract verbs like "read the file"). Reasoning / decision / synthesis bullets that don't invoke a tool stay in plain prose. Anti-pattern: three sequential `` `write_file` — `` bullets where each describes different conceptual work and the file write is incidental — collapse the conceptual work into prose and name `write_file` once for the persistence. Why names ARE required where tools ARE invoked: ① workflow is injected into the runtime agent's system prompt — invoked tools need canonical names so the runtime picks right; ② `<skills>` closure is extracted from skill_ids appearing here.3. **Implementation: built-in tools vs skill** (pick in this order):- **First check built-in tools** (auto-registered via the tool-use protocol) — read/write file, bash, KB search, PDF render, image generation, web search/fetch, etc. — single-step actions that run directly. **No skill wrapper needed**; in workflow, write the tool name verbatim ("use `read_file` to read the PDF" / "use `markdown_to_pdf` to render the report" / "use `kb_search` to query the KB"); don't force-wrap a single-step action as a skill.- **Then check the "Available skills (skills)" section** — skills are most useful for: multi-step logic encapsulation, third-party paid APIs (with credential management), recurrent compound flows. If one fits, use it; don't reinvent the wheel.diff --git a/src/main/prompts/chat_commander.md b/src/main/prompts/chat_commander.mdindex b4df5eac..99cfe49f 100644--- a/src/main/prompts/chat_commander.md+++ b/src/main/prompts/chat_commander.md@@ -238,7 +238,7 @@ Both flows emit ONE `<agent>...</agent>` container in this turn and end (do NOT3. **`触发词:` / `Triggers:`** + 5–8 keywords (separated by `、` / `,`).- **`<workflow>`** = ordered steps in physical execution order. Each step is a **verb-led title** (5–10 chars) followed by bulleted actions describing what the runtime agent does; bullets carry the action and (where relevant) its result inline. Branches use nested bullets (`if X → call A` / `else → call B`). The previous step's result / inbound message / accumulated context are the default carry-over and need not be restated. Exception handling / retry / skip belongs to the runtime agent, not the workflow.- - **Hard constraint — every action must explicitly name the tool / skill_id verbatim in backticks** (e.g. `read_file` / `kb_search` / `social-fetch` skill / `markdown_to_pdf` / `web_search`). Don't write abstract verbs like "read the file" or "do a search". **Why**: the workflow is injected into the runtime agent's system prompt; missing tool names force secondary inference (often picks the wrong tool); and the `<skills>` closure is derived from `skill_id`s appearing here, so a missing `skill_id` means the skill never gets loaded at runtime.+ - **Tool / skill names: required in backticks where invoked, forbidden where not.** Every invoked tool / skill_id appears in backticks (`read_file` / `kb_search` / `social-fetch` skill / `markdown_to_pdf` / `web_search` — no abstract verbs like "read the file"). Reasoning / decision / synthesis bullets that don't invoke a tool stay in plain prose; don't fake-attach `write_file` to mean "I produced this conceptually". **Why**: workflow is injected into the runtime agent's system prompt (invoked tools need canonical names so the runtime picks right) and `<skills>` closure is derived from skill_ids appearing here.- **Tool / skill priority** when authoring workflow actions: ① built-in tools (file IO, `bash`, `kb_search`, `kb_read`, `markdown_to_pdf`, `html_to_pdf`, `generate_image`, `web_search`, `web_fetch`) — write the tool name directly. ② existing skills from the "Available skills (skills)" block — used as-is by `skill_id`. ③ only when neither covers it, mention the missing capability in user-perspective prose; do NOT invent skill_ids in `<skills>`. Built-in tool names are NOT skill_ids and must never appear in `<skills>`. An empty `<skills></skills>` is legal and common.@@ -262,7 +262,7 @@ Both flows emit ONE `<agent>...</agent>` container in this turn and end (do NOT<description_en>① one-line function: verb+object+delivery ; ② For: "user phrasing 1", "user phrasing 2", … ; ③ Triggers: word1, word2, …</description_en><workflow>Stepwise markdown workflow. Do not include a top-level `# Workflow` heading — the UI already wraps it.-Step format = `### N. <title>` followed by bulleted actions; each bullet carries the tool / skill_id call in backticks plus its inline purpose / result.+Step format = `### N. <title>` followed by bulleted actions; bullets that invoke a tool / skill name it in backticks with inline purpose / result. Reasoning / decision / synthesis bullets go in plain prose without forcing a tool name.</workflow><skills>skill_id_a