Skip to content

README.md

Latest commit

 

History

History
266 lines (215 loc) · 11.7 KB

README.md

File metadata and controls

266 lines (215 loc) · 11.7 KB
Raw

Tool Restrictions Example

This example demonstrates how CAO resolves tool access per agent role — the mechanism documented in full in docs/tool-restrictions.md. It is not another orchestration-pattern example (see examples/assign/ for assign vs handoff mechanics); it exists purely to show, with real launch commands and real resolved tool lists, how role and allowedTools combine to decide what an agent can and cannot do.

Naming note

The profiles here are prefixed tool_restrictions_ (e.g. tool_restrictions_developer.md, not developer.md). CAO ships built-in profiles named plain developer and reviewer (src/cli_agent_orchestrator/agent_store/), and cao install stores a profile under its filename stem — installing a file literally named developer.md would silently overwrite your local copy of the built-in developer profile. examples/codex-basic/ avoids the same collision by naming its profiles codex_developer/codex_reviewer rather than bare developer/reviewer; this example follows that precedent.

Pattern Overview

Four profiles, four distinct tool-access patterns:

Profile Pattern role allowedTools
tool_restrictions_supervisor.md Role default — orchestration only supervisor (unset — inherited from role)
tool_restrictions_developer.md Role default — full access developer (unset — inherited from role)
tool_restrictions_reviewer.md Role default — read-only reviewer (unset — inherited from role)
tool_restrictions_custom_restricted.md Explicit override — no role (unset) ["fs_read", "fs_list", "execute_bash"]

The first three map a role to its built-in default via the table in docs/tool-restrictions.md. The fourth sets allowedTools directly with no role at all, proving allowedTools is a complete specification on its own.

Setup

# Start the CAO server
cao-server

# Install the agent profiles
cao install examples/tool-restrictions/tool_restrictions_supervisor.md
cao install examples/tool-restrictions/tool_restrictions_developer.md
cao install examples/tool-restrictions/tool_restrictions_reviewer.md
cao install examples/tool-restrictions/tool_restrictions_custom_restricted.md

Resolution Walkthrough

Each profile below shows: the launch command, the confirmation prompt CAO would show (per Launch Confirmation Prompt), and the resulting allowed/denied tools — both in CAO's own vocabulary and as native Claude Code tool names (using the translation table in Tool Vocabulary). Claude Code is a hard-enforcement provider — see Provider Enforcement — so what's "denied" below is physically blocked, not just discouraged.

tool_restrictions_supervisor — orchestration only

cao launch --agents tool_restrictions_supervisor --provider claude_code
Agent 'tool_restrictions_supervisor' launching on claude_code:
  Role:      supervisor
  Allowed:   @cao-mcp-server, fs_read, fs_list
  Directory: /home/user/my-project

Proceed? [Y/n]
Allowed Denied
CAO vocabulary @cao-mcp-server, fs_read, fs_list execute_bash, fs_write, web_fetch, discovery
Claude Code native Read, Glob, Grep, plus MCP tools (handoff, assign, send_message) Bash family (Bash, BashOutput, KillShell, Task, Agent, Monitor), file writes (Edit, Write, NotebookEdit), network (WebFetch, WebSearch) — all via --disallowedTools

tool_restrictions_developer — full access

cao launch --agents tool_restrictions_developer --provider claude_code
Agent 'tool_restrictions_developer' launching on claude_code:
  Role:      developer
  Allowed:   @builtin, fs_*, execute_bash, web_fetch, @cao-mcp-server
  Directory: /home/user/my-project

Proceed? [Y/n]
Allowed Denied
CAO vocabulary @builtin, fs_read, fs_write, fs_list, execute_bash, web_fetch, @cao-mcp-server discovery (not part of any built-in role — add it explicitly if this agent needs sibling discovery)
Claude Code native Everything: Read, Edit, Write, Glob, Grep, Bash, WebFetch, WebSearch, plus MCP tools Nothing — no --disallowedTools flags at all

tool_restrictions_reviewer — read-only

cao launch --agents tool_restrictions_reviewer --provider claude_code
Agent 'tool_restrictions_reviewer' launching on claude_code:
  Role:      reviewer
  Allowed:   @builtin, fs_read, fs_list, @cao-mcp-server
  Directory: /home/user/my-project

Proceed? [Y/n]
Allowed Denied
CAO vocabulary @builtin, fs_read, fs_list, @cao-mcp-server execute_bash, fs_write, web_fetch, discovery
Claude Code native Read, Glob, Grep, plus MCP tools Bash family (Bash, BashOutput, KillShell, Task, Agent, Monitor), file writes (Edit, Write, NotebookEdit), network (WebFetch, WebSearch) — all via --disallowedTools

Nuance: supervisor and reviewer deny the exact same native Claude Code tools — both lack execute_bash/fs_write/web_fetch. The only CAO- vocabulary difference between them is @builtin, which the translation table marks (internal) for every provider: it isn't a real --disallowedTools entry, so it never shows up as a difference at the Claude Code enforcement layer. The difference between the two roles is about intent and MCP-tool availability (both actually keep @cao-mcp-server per the built-in role table), not about a distinct Claude Code denylist. The Task/Agent/Monitor/BashOutput/KillShell grouping under "Bash family" is deliberate, not an approximation — see Known Limitations #1 for why the subagent tool is folded into execute_bash rather than gated separately.

tool_restrictions_custom_restricted — explicit override, no role

cao launch --agents tool_restrictions_custom_restricted --provider claude_code
Agent 'tool_restrictions_custom_restricted' launching on claude_code:
  Role:      (not set)
  Allowed:   fs_read, fs_list, execute_bash
  Directory: /home/user/my-project

Proceed? [Y/n]
Allowed Denied
CAO vocabulary fs_read, fs_list, execute_bash fs_write, web_fetch, @cao-mcp-server, discovery
Claude Code native Read, Glob, Grep, Bash family (Bash, BashOutput, KillShell, Task, Agent, Monitor) File writes (Edit, Write, NotebookEdit), network (WebFetch, WebSearch) — and no MCP tools at all, since this profile never configures the cao-mcp-server server

Note this does not show the "no role or allowedTools set" reminder that Default Behavior describes — that reminder only fires when neither is set. Here allowedTools is set, so it's a complete, deliberate specification, not a fallback.

Override Hierarchy

How Overrides Work ranks five levels, highest priority first: --yolo, the --allowed-tools CLI flag, allowedTools in the profile, role in the profile, then the developer fallback. This example exercises the middle two directly and the CLI flag on top of them:

  • role alone — tool_restrictions_supervisor, _developer, and _reviewer set only role; CAO fills in allowedTools from the built-in role table.
  • allowedTools alone, no role — tool_restrictions_custom_restricted sets allowedTools directly with no role at all. allowedTools is a complete specification; it doesn't need a role to fall back on.
  • allowedTools beats role when both are set — this example doesn't ship a fifth file for this (the scope here is four profiles), but you can see it by editing tool_restrictions_custom_restricted.md to add role: developer above its allowedTools line: the resolved tools stay exactly fs_read, fs_list, execute_bash — role: developer would normally add fs_write/web_fetch/@cao-mcp-server/@builtin, but the explicit allowedTools list wins and those stay denied. This is the same override the docs demonstrate with their own restricted_developer example — see allowedTools — The Precise Way.
  • The --allowed-tools CLI flag beats both, at launch time, without editing any file: 
    cao launch --agents tool_restrictions_custom_restricted --allowed-tools fs_read
    This launches with only fs_read allowed — fs_list and execute_bash from the profile's own allowedTools are dropped, because the CLI flag is a higher-priority override than the profile.

Cross-Agent Delegation: Children Resolve Their Own Profile

Install and launch the supervisor, then give it a small task:

cao launch --agents tool_restrictions_supervisor
Implement is_valid_email(s) and have it reviewed.

tool_restrictions_supervisor (role: supervisor → @cao-mcp-server, fs_read, fs_list)
  │
  ├─ handoff("tool_restrictions_developer")
  │    → Developer profile resolves role: developer → full access
  │    → Claude Code launched with no --disallowedTools
  │
  └─ handoff("tool_restrictions_reviewer")
       → Reviewer profile resolves role: reviewer → read-only
       → Claude Code launched with --disallowedTools Bash Edit Write WebFetch WebSearch

The supervisor itself can't write files or run bash — but it can still handoff to tool_restrictions_developer, which gets full write/execute access, because that access comes from the developer's own profile, not from anything the supervisor holds or grants. The same is true in reverse: tool_restrictions_developer has full access, but the tool_restrictions_reviewer it hands off to next is still read-only, because the reviewer resolves its own restrictions too. Nothing is inherited up or down the delegation chain — every handoff/assign target gets its tool access from its own profile, resolved fresh at launch. See Cross-Provider Inheritance for the same pattern in the docs.

Usage

In the supervisor terminal:

Implement a Python function is_valid_email(s) that validates email format,
then have it reviewed.

Expected flow:

  1. Supervisor calls handoff(tool_restrictions_developer, ...) and blocks.
  2. Developer writes is_valid_email, returns it.
  3. Supervisor calls handoff(tool_restrictions_reviewer, ...) with that code and blocks.
  4. Reviewer reads and critiques the code (cannot edit it), returns comments.
  5. Supervisor presents the implementation plus the review to you.

You can also launch any profile standalone to see its restrictions in isolation — tool_restrictions_custom_restricted in particular isn't part of the delegation chain above; it's a self-contained example of the override pattern:

cao launch --agents tool_restrictions_custom_restricted

See Also

  • docs/tool-restrictions.md — the full spec for role, allowedTools, --yolo, override priority, and per-provider enforcement.
  • examples/assign/ — assign vs handoff orchestration mechanics (this example deliberately keeps that part simple).
  • examples/cross-provider/ — the same delegation pattern across multiple CLI providers.