mirror of
https://github.com/agent0ai/agent-zero.git
synced 2026-05-19 07:59:34 +00:00
Split A0 remote workflow skills by affordance
Replace the combined A0 CLI remote workflow skill with separate text-editor and code-execution remote skills, update tool stubs to load the matching per-tool guide, and keep computer-use remote scoped to desktop control. Add prompt-gating coverage for the per-affordance skill split.
This commit is contained in:
Alessandro
parent
229de5166b
commit
a3d41e2ca1
7 changed files with 134 additions and 72 deletions
|
|
@ -1,7 +1,7 @@
|
|||
# code_execution_remote tool
|
||||
|
||||
Runs shell-backed execution on the machine where the subscribed A0 CLI is running.
|
||||
Load `a0-cli-remote-workflows` before using this tool for nontrivial local project work.
|
||||
Load `code-execution-remote` before using this tool for nontrivial local project work.
|
||||
|
||||
Current local access mode: `{{access_mode}}`
|
||||
|
||||
|
|
|
|||
|
|
@ -1,7 +1,7 @@
|
|||
# text_editor_remote tool
|
||||
|
||||
Reads, writes, and patches files on the machine where the subscribed A0 CLI is running.
|
||||
This is different from server-side file tools. Load `a0-cli-remote-workflows` before using it for edits.
|
||||
This is different from server-side file tools. Load `text-editor-remote` before using it for edits.
|
||||
|
||||
Current access mode: `{{access_mode}}`
|
||||
|
||||
|
|
|
|||
|
|
@ -1,64 +0,0 @@
|
|||
---
|
||||
name: a0-cli-remote-workflows
|
||||
description: Guide safe use of A0 CLI remote shell execution and remote file editing on the connected host machine. Load before using code_execution_remote or text_editor_remote for local project work through the CLI connector.
|
||||
version: 1.0.0
|
||||
author: Agent Zero Team
|
||||
tags: ["agent-zero", "a0", "cli", "connector", "remote-execution", "remote-files"]
|
||||
trigger_patterns:
|
||||
- "code_execution_remote"
|
||||
- "text_editor_remote"
|
||||
- "remote file editing"
|
||||
- "remote shell execution"
|
||||
- "edit my local files through a0 cli"
|
||||
- "run commands on the cli host"
|
||||
allowed_tools:
|
||||
- code_execution_remote
|
||||
- text_editor_remote
|
||||
---
|
||||
|
||||
# A0 CLI Remote Workflows
|
||||
|
||||
## Boundary
|
||||
|
||||
Use `code_execution_remote` and `text_editor_remote` only for work on the machine where A0 CLI is running. These paths, shells, runtimes, and files belong to the CLI host, not the Agent Zero server or Docker container.
|
||||
|
||||
If the task belongs inside Agent Zero's own runtime, use the normal server-side tools instead.
|
||||
|
||||
## Access Modes
|
||||
|
||||
- Read only: inspect files and poll/reset existing execution sessions. Do not attempt writes or mutating shell work until the user switches the CLI to Read&Write with F3.
|
||||
- Read&Write: shell-backed execution, writes, and patches may modify the CLI host. Keep changes narrow and intentional.
|
||||
- Execution may also be disabled locally in the CLI. If a remote tool returns a structured disabled/no-client error, explain the required CLI toggle instead of falling back to the server filesystem.
|
||||
|
||||
## Remote Execution
|
||||
|
||||
- Use `runtime=terminal` for shell commands, `runtime=python` for Python snippets, and `runtime=nodejs` for Node.js snippets.
|
||||
- Reuse the same integer `session` while continuing a workflow; session state is local to the CLI frontend.
|
||||
- Use `runtime=output` when a previous command is still running or returned before the shell reached a prompt.
|
||||
- Use `runtime=reset` when a session is stuck or a clean shell is safer.
|
||||
- Treat `runtime=input` as deprecated compatibility for sending one line to a running shell.
|
||||
- Match the remote host shell syntax. A Windows CLI may need PowerShell syntax even when Agent Zero runs on Linux.
|
||||
|
||||
## Remote File Editing
|
||||
|
||||
- Start with `read` when inspecting a file or preparing line-based edits.
|
||||
- Use `write` only when replacing or creating the whole file is truly the right operation.
|
||||
- Prefer `patch` with `patch_text` for context-anchored edits, especially after inserts/deletes or when line numbers may have shifted.
|
||||
- Use `patch` with `edits` only for small line-range edits based on the latest remote read.
|
||||
- If freshness-aware line patching rejects an edit as stale, reread the file and retry with updated ranges.
|
||||
|
||||
## Patch Text Rules
|
||||
|
||||
- `patch_text` supports update hunks for one file.
|
||||
- Use one `@@ existing line` anchor, then `+new line` entries for insertion.
|
||||
- For replacement, use `@@ before target` followed by `-old` and `+new`, or use `@@ old target` followed by the same replacement pair.
|
||||
- Do not repeat the same old line as both context and deletion in one hunk.
|
||||
- Every non-header content line must begin with exactly one prefix: space for context, `+` for additions, or `-` for removals.
|
||||
- Do not stack multiple `@@` anchors for one insert.
|
||||
|
||||
## Failure Handling
|
||||
|
||||
- If no CLI is connected or subscribed, ask the user to connect A0 CLI to this chat.
|
||||
- If writes are blocked, tell the user to switch local access to Read&Write with F3.
|
||||
- If execution is disabled, tell the user to enable remote execution in the CLI.
|
||||
- If a request times out or the CLI disconnects, poll once if a session may still be running; otherwise summarize the failure and wait for reconnection.
|
||||
45
plugins/_a0_connector/skills/code-execution-remote/SKILL.md
Normal file
45
plugins/_a0_connector/skills/code-execution-remote/SKILL.md
Normal file
|
|
@ -0,0 +1,45 @@
|
|||
---
|
||||
name: code-execution-remote
|
||||
description: Guide safe use of code_execution_remote for shell-backed execution on the connected A0 CLI host machine.
|
||||
version: 1.0.0
|
||||
author: Agent Zero Team
|
||||
tags: ["agent-zero", "a0", "cli", "connector", "remote-execution", "shell"]
|
||||
trigger_patterns:
|
||||
- "code_execution_remote"
|
||||
- "remote shell execution"
|
||||
- "run commands on the cli host"
|
||||
- "run python on the cli host"
|
||||
- "run node on the cli host"
|
||||
allowed_tools:
|
||||
- code_execution_remote
|
||||
---
|
||||
|
||||
# Code Execution Remote
|
||||
|
||||
## Boundary
|
||||
|
||||
Use `code_execution_remote` only for shell-backed execution on the machine where A0 CLI is running. Shells, runtimes, and paths belong to the CLI host, not the Agent Zero server or Docker container.
|
||||
|
||||
If the task belongs inside Agent Zero's own runtime, use the normal server-side execution tool instead.
|
||||
|
||||
## Access Modes
|
||||
|
||||
- Remote execution can be disabled locally in the CLI. If the tool returns a disabled/no-client error, explain the required CLI toggle instead of falling back to the server runtime.
|
||||
- Read&Write local file access allows mutating runtimes such as `terminal`, `python`, and `nodejs`.
|
||||
- Read only local file access blocks mutating runtimes. `runtime=output` and `runtime=reset` remain available for existing sessions.
|
||||
|
||||
## Runtimes
|
||||
|
||||
- Use `runtime=terminal` for shell commands, `runtime=python` for Python snippets, and `runtime=nodejs` for Node.js snippets.
|
||||
- Reuse the same integer `session` while continuing a workflow; session state is local to the CLI frontend.
|
||||
- Use `runtime=output` when a previous command is still running or returned before the shell reached a prompt.
|
||||
- Use `runtime=reset` when a session is stuck or a clean shell is safer.
|
||||
- Treat `runtime=input` as deprecated compatibility for sending one line to a running shell.
|
||||
- Match the remote host shell syntax. A Windows CLI may need PowerShell syntax even when Agent Zero runs on Linux.
|
||||
|
||||
## Failure Handling
|
||||
|
||||
- If no CLI is connected or subscribed, ask the user to connect A0 CLI to this chat.
|
||||
- If execution is disabled, tell the user to enable remote execution in the CLI.
|
||||
- If mutating runtimes are blocked, tell the user to switch local file access to Read&Write with F3.
|
||||
- If a request times out or the CLI disconnects, poll once if a session may still be running; otherwise summarize the failure and wait for reconnection.
|
||||
51
plugins/_a0_connector/skills/text-editor-remote/SKILL.md
Normal file
51
plugins/_a0_connector/skills/text-editor-remote/SKILL.md
Normal file
|
|
@ -0,0 +1,51 @@
|
|||
---
|
||||
name: text-editor-remote
|
||||
description: Guide safe use of text_editor_remote for reading, writing, and patching files on the connected A0 CLI host machine.
|
||||
version: 1.0.0
|
||||
author: Agent Zero Team
|
||||
tags: ["agent-zero", "a0", "cli", "connector", "remote-files", "file-editing"]
|
||||
trigger_patterns:
|
||||
- "text_editor_remote"
|
||||
- "remote file editing"
|
||||
- "edit my local files through a0 cli"
|
||||
- "read files on the cli host"
|
||||
- "patch files on the cli host"
|
||||
allowed_tools:
|
||||
- text_editor_remote
|
||||
---
|
||||
|
||||
# Text Editor Remote
|
||||
|
||||
## Boundary
|
||||
|
||||
Use `text_editor_remote` only for file work on the machine where A0 CLI is running. These paths and files belong to the CLI host, not the Agent Zero server or Docker container.
|
||||
|
||||
If the task belongs inside Agent Zero's own runtime, use the normal server-side file tools instead.
|
||||
|
||||
## Access Modes
|
||||
|
||||
- Read&Write: reads, writes, and patches may modify the CLI host. Keep changes narrow and intentional.
|
||||
- Read only: inspect files only. If writes are blocked, tell the user to switch local file access to Read&Write with F3.
|
||||
|
||||
## Editing Flow
|
||||
|
||||
- Start with `read` when inspecting a file or preparing line-based edits.
|
||||
- Use `write` only when replacing or creating the whole file is truly the right operation.
|
||||
- Prefer `patch` with `patch_text` for context-anchored edits, especially after inserts/deletes or when line numbers may have shifted.
|
||||
- Use `patch` with `edits` only for small line-range edits based on the latest remote read.
|
||||
- If freshness-aware line patching rejects an edit as stale, reread the file and retry with updated ranges.
|
||||
|
||||
## Patch Text Rules
|
||||
|
||||
- `patch_text` supports update hunks for one file.
|
||||
- Use one `@@ existing line` anchor, then `+new line` entries for insertion.
|
||||
- For replacement, use `@@ before target` followed by `-old` and `+new`, or use `@@ old target` followed by the same replacement pair.
|
||||
- Do not repeat the same old line as both context and deletion in one hunk.
|
||||
- Every non-header content line must begin with exactly one prefix: space for context, `+` for additions, or `-` for removals.
|
||||
- Do not stack multiple `@@` anchors for one insert.
|
||||
|
||||
## Failure Handling
|
||||
|
||||
- If no CLI is connected or subscribed, ask the user to connect A0 CLI to this chat.
|
||||
- If writes are blocked, tell the user to switch local file access to Read&Write with F3.
|
||||
- If a request times out or the CLI disconnects, summarize the failure and wait for reconnection.
|
||||
|
|
@ -14,7 +14,6 @@ trigger_patterns:
|
|||
- "native ui"
|
||||
allowed_tools:
|
||||
- computer_use_remote
|
||||
- code_execution_remote
|
||||
---
|
||||
|
||||
# Computer Use Remote
|
||||
|
|
@ -25,6 +24,8 @@ Load this skill before using `computer_use_remote` for local desktop and native
|
|||
|
||||
If the task is browser-only and the user is flexible, prefer direct browser tooling because it is usually more reliable and token-efficient than screenshot-driven desktop control.
|
||||
|
||||
If the task needs shell execution on the CLI host, load `code-execution-remote` separately rather than treating desktop control and shell execution as one affordance.
|
||||
|
||||
## Core Loop
|
||||
|
||||
1. Call `start_session` first.
|
||||
|
|
|
|||
|
|
@ -198,8 +198,8 @@ def test_computer_use_enabled_cli_adds_computer_stub():
|
|||
assert "code_execution_remote tool" not in prompt
|
||||
|
||||
|
||||
def test_remote_workflow_skills_parse():
|
||||
connector_skill = _parse_skill_frontmatter(
|
||||
def test_remote_affordance_skills_parse():
|
||||
legacy_connector_skill = (
|
||||
PROJECT_ROOT
|
||||
/ "plugins"
|
||||
/ "_a0_connector"
|
||||
|
|
@ -207,11 +207,40 @@ def test_remote_workflow_skills_parse():
|
|||
/ "a0-cli-remote-workflows"
|
||||
/ "SKILL.md"
|
||||
)
|
||||
text_editor_skill = _parse_skill_frontmatter(
|
||||
PROJECT_ROOT
|
||||
/ "plugins"
|
||||
/ "_a0_connector"
|
||||
/ "skills"
|
||||
/ "text-editor-remote"
|
||||
/ "SKILL.md"
|
||||
)
|
||||
code_execution_skill = _parse_skill_frontmatter(
|
||||
PROJECT_ROOT
|
||||
/ "plugins"
|
||||
/ "_a0_connector"
|
||||
/ "skills"
|
||||
/ "code-execution-remote"
|
||||
/ "SKILL.md"
|
||||
)
|
||||
computer_skill = _parse_skill_frontmatter(
|
||||
PROJECT_ROOT / "skills" / "computer-use-remote" / "SKILL.md"
|
||||
)
|
||||
|
||||
assert connector_skill["name"] == "a0-cli-remote-workflows"
|
||||
assert connector_skill["description"]
|
||||
assert not legacy_connector_skill.exists()
|
||||
assert text_editor_skill["name"] == "text-editor-remote"
|
||||
assert text_editor_skill["allowed_tools"] == ["text_editor_remote"]
|
||||
assert code_execution_skill["name"] == "code-execution-remote"
|
||||
assert code_execution_skill["allowed_tools"] == ["code_execution_remote"]
|
||||
assert computer_skill["name"] == "computer-use-remote"
|
||||
assert computer_skill["description"]
|
||||
assert computer_skill["allowed_tools"] == ["computer_use_remote"]
|
||||
|
||||
|
||||
def test_remote_tool_stubs_point_to_per_tool_skills():
|
||||
text_stub = (PROMPT_ROOT / "agent.connector_tool.text_editor_remote.md").read_text(encoding="utf-8")
|
||||
exec_stub = (PROMPT_ROOT / "agent.connector_tool.code_execution_remote.md").read_text(encoding="utf-8")
|
||||
|
||||
assert "Load `text-editor-remote`" in text_stub
|
||||
assert "Load `code-execution-remote`" in exec_stub
|
||||
assert "a0-cli-remote-workflows" not in text_stub
|
||||
assert "a0-cli-remote-workflows" not in exec_stub
|
||||
|
|
|
|||
Loading…
Add table
Add a link
Reference in a new issue