Agent Tools — Cross-CLI Compatibility Matrix¶
This page maps agent tool capability across the seven LazyAI targets: how each CLI lets an agent/subagent restrict its tools, the native tool names/casing, and what LazyAI emits today vs. the correct shape. It is the reference for the agent-tools alignment work tracked in the epic linked at the bottom.
Root finding¶
Canonical agents (packages/cli/library/canonical/agents/*.md) now express capability via a machine-readable tools: grant list (#569). Each agent carries a 7-token vocabulary (read, edit, shell, search, web, mcp, spawn); read-only roles (researcher, reviewer, evidence-verifier) declare tools: [read, search], while full-capability roles carry the full grant set. The legacy mode:/role: self-contradiction (e.g. researcher.md declaring mode: all) has been removed. Every adapter derives per-agent restrictions from this single source via frontmatter.ParseAgentToolGrants, so read-only roles are emitted restricted on every target that has a native mechanism.
Per-agent tool model by target¶
| Target | Native per-agent restriction | Built-in tool names (casing) | Allow/deny semantics |
|---|---|---|---|
| Claude Code | tools: (allow) + disallowedTools: (deny) in agent .md frontmatter |
Capitalized: Read, Write, Edit, Bash, Grep, Glob, Agent, WebFetch, WebSearch, Skill |
allowlist; disallowedTools wins |
| OpenCode | tools:{bash:true,…} gate map + permission: + mode: in agent frontmatter / opencode.json agent.<name> |
lowercase built-in gates: bash, edit, write, plus permission keys |
per-tool boolean + ask/allow/deny |
| Copilot | tools: list in .agent.md |
lowercase set: read, search, edit, shell (+ *) |
allowlist |
| Pi | none (CLI flags --tools/--exclude-tools; experimental skill allowed-tools) |
lowercase: read, bash, edit, write, grep, find, ls |
n/a per-agent |
| OMP | tools: (CSV/YAML subset) in .omp/agents/<name>.md + spawns, thinkingLevel, autoloadSkills, read-summarize |
lowercase: read, bash, edit, write, search, task, … |
allowlist (subset of built-ins) |
| Kiro | tools (whitelist) + allowedTools (auto-approved) in agent JSON |
aliases: fs_read/read, fs_write/write, execute_bash/shell, use_aws/aws |
allowlist + auto-approve |
| Antigravity | define_subagent enable flags (enable_mcp_tools, enable_write_tools, enable_subagent_tools) — not a tool-name list |
distinct: view_file, write_to_file, replace_file_content, run_command, grep_search, search_web, read_url_content, invoke_subagent |
capability flags + permissions engine |
Canonical capability → target mechanism¶
| Canonical capability | Claude | OpenCode | Copilot | Pi | OMP | Kiro | Antigravity |
|---|---|---|---|---|---|---|---|
| file read | Read |
(default) | read |
read |
read |
read/fs_read |
view_file |
| file write/edit | Write/Edit |
edit:true / permission.edit |
edit |
write/edit |
write/edit |
write/fs_write |
write_to_file/replace_file_content (enable_write_tools) |
| shell/exec | Bash |
bash:true / permission.bash |
shell |
bash |
bash |
execute_bash/shell |
run_command |
| search/grep | Grep/Glob |
(default) | search |
grep/find |
search |
(built-in) | grep_search/find_by_name |
| web fetch/search | WebFetch/WebSearch |
(built-in) | (n/a) | extension only | web_search |
(built-in) | search_web/read_url_content |
| MCP tools | mcp__<srv>__<tool> |
.mcp.json servers |
VS Code/CLI MCP | none | mcp__<srv>_<tool> |
mcpServers/autoApprove |
mcp(server/tool) (enable_mcp_tools) |
| subagent spawn | Agent |
(built-in) | (n/a) | subagent ext. | task/spawns |
(agent profiles) | invoke_subagent (enable_subagent_tools) |
| read-only role (researcher/reviewer) | tools: minus Write/Edit/Bash, or disallowedTools:[Write,Edit,Bash] |
permission:{edit:deny,bash:deny} + mode |
tools:["read","search"] |
--no-builtin-tools / runtime |
tools:[read,search,...] |
tools minus write/exec |
omit enable_write_tools |
What LazyAI emits today (gap status)¶
| Target | Current emission | Respects per-agent tools? | Gap |
|---|---|---|---|
| Claude Code | agent .md via RewriteAgentForClaudeCode; read-only agents emit disallowedTools: Edit Write Bash (PascalCase), full-capability agents none |
✅ disallowedTools from canonical grants |
none (done in #570) |
| OpenCode | canonical read-only agents emit permission: {edit: deny, bash: deny} via RewriteAgentForOpenCode; full-capability agents get no permission block (OpenCode default); no tools: gate map emitted |
✅ permission from canonical grants |
none (done in #572) |
| Copilot | .agent.md with per-agent tools derived from canonical grants (copilotAgentMarkdownContent/ParseAgentToolGrants); read-only → ["read","search"] |
✅ per-agent Copilot tool list | none (done in #571) |
| Pi | agents copied; no tools field (Pi has no mechanism) | ✅ correct by design | none (document intentional non-mapping) |
| OMP | RewriteAgentForOMP transform: tools (OMP allowlist from canonical grants), thinkingLevel, autoloadSkills (from skills:); LazyAI-only fields dropped |
✅ read-only agents restricted (tools: ["read","search"]); full-capability agents get OMP equivalents |
✅ closed by #573 |
| Kiro | canonical agents → .kiro/agents/<name>.json via RewriteAgentForKiro; tools/allowedTools from ParseAgentToolGrants (#574) |
✅ tools + allowedTools from canonical tools: |
none (done in #574) |
| Antigravity | no agent/subagent files (skills-only); emits subagent blueprint rules, workflow skills, hook expansion | ⚠️ blueprint only (rules doc, not enforced by file format) | #575 implemented: subagent capability blueprint (.agents/rules/lazyai-subagents.md), workflow skills (workflow-*/SKILL.md), write-guard + PreInvocation hooks; commands surface: n/a |
Evidence (file:line)¶
- Canonical:
packages/cli/library/canonical/agents/researcher.md:5-7(tools: [read, search]); all canonical agents carrytools:grants as of #569. - Claude:
agent_transform.go:103RewriteAgentForClaudeCode+claudeDisallowedTools(disallowedToolsfrom grants);claudecode_frontmatter_test.go:364TestRewriteAgentForClaudeCode_ToolGrants(#570). - OpenCode:
agent_transform.go:194RewriteAgentForOpenCode(permission:{edit:deny,bash:deny}fromParseAgentToolGrants);opencode_adapter_test.go:109TestOpenCodeAdapter_CanonicalReadOnlyAgentsGetPermission(#572). - Copilot:
copilot.go:300copilotAgentMarkdownContent+ParseAgentToolGrants(copilot.go:358);TestCopilotAgentMarkdownContent_ReadOnly/_FullCapNonCopilotGrantsOmitted(#571). - OMP:
agent_transform.goRewriteAgentForOMP;omp.go(transform-based copy viaCopyLibraryDirectoryOption.Transform);omp_frontmatter_test.go(#573). - Kiro:
kiro.go;agent_transform.goRewriteAgentForKiro;docs/ai-cli-tools/tool-systems/kiro.md(JSON required, confirmed). - Pi:
pi.go,docs/ai-cli-tools/tool-systems/pi.md(no per-agent mechanism). - Antigravity:
antigravity.go(no agent emission; blueprint atantigravity/subagents-blueprint.md; workflow emission viaCopyLibraryDirectory; hooks inantigravity/hooks.json).
Tracking¶
This matrix backs the cross-CLI agent-tools alignment epic #568 and its per-target child issues (#569 canonical model, #570 Claude, #571 Copilot, #572 OpenCode, #573 OMP, #574 Kiro, #575 Antigravity). Update the "gap status" column as each child closes.