Skip to content

PLATFORM_SPEC.md — Google Antigravity (agent-first IDE + CLI + SDK)

New to gald3r on this platform? Start at Use gald3r with Google Antigravity for a short overview before the full breakdown below.

Google Antigravity (DeepMind) is an agent-first development platform — a desktop IDE, a CLI, and an SDK that share one agent harness. As of the Antigravity 2.0 relaunch (agent-first; Subagents / Hooks / Scheduled Tasks / Agent Management announced at Google I/O 2026, May 2026), the IDE/CLI harness natively supports all six gald3r-relevant extension primitives — Workflows (slash commands), Rules, dynamic Subagents, Agent Skills (SKILL.md), lifecycle Hooks, and MCP. Critically for gald3r, Antigravity reads AGENTS.md (cross-tool) and GEMINI.md (Antigravity-specific, higher precedence), and discovers Anthropic-format SKILL.md packages, so gald3r's existing AGENTS.md + g-skl-*/SKILL.md artifacts are largely drop-in reusable.

Authoring path: UPDATE. Verified 2026-06-02 against https://antigravity.google/docs (see Verification Evidence). This supersedes the prior spec (last_doc_scan: 2026-05-28) which conservatively marked hooks/skills/rules as ❓/⚠️ — they are now confirmed NATIVE in the post-2.0 IDE/CLI harness.

Surface split (READ THIS): the capability ratings below reflect the Antigravity IDE / CLI product (one shared harness; docs at antigravity.google/docs). The hosted Gemini API "Managed Agents" surface (ai.google.dev/gemini-api/docs/antigravity-agent) is a separate product that currently states custom tools / function-calling and subagent delegation are "not yet supported." Do not conflate the two — gald3r targets the IDE/CLI harness, which does support hooks + subagents.

Instruction-file convention: Antigravity reads AGENTS.md (cross-tool, shared with Cursor / Claude Code) and GEMINI.md (Antigravity-specific). It does not rely on CLAUDE.md. Precedence: System Rules (immutable, DeepMind) > GEMINI.md > AGENTS.md.

1. Folder Hierarchy

Antigravity 2.0 splits config between project-local dirs and user-global state under ~/.gemini/ (shared with the Gemini CLI state tree).

<project-root>/
├── AGENTS.md                        ← instruction file Antigravity reads (cross-tool)
├── GEMINI.md                        ← Antigravity-specific instructions (higher precedence)
└── .agents/                         ← workspace customization root
    ├── rules/        *.md           ← Manual / Always On / Model Decision / Glob (≤12,000 chars each)
    ├── skills/       <name>/SKILL.md ← Agent Skills (Anthropic SKILL.md standard)  [path varies — see §4]
    ├── workflows/    *.md           ← project-relative Workflows (slash commands)  [confirmed  — see §5]
    └── hooks.json                   ← lifecycle hooks (workspace scope)
.antigravity/
    └── mcp.json                     ← project-local MCP config { "mcpServers": { ... } }

~/.gemini/                           ← user-global state (Gemini-namespaced)
├── GEMINI.md                        ← global Antigravity-specific rules
├── AGENTS.md                        ← global cross-tool rules (v1.20.3+)
├── config/                          ← global customizations (hooks.json customization dir)
└── antigravity/
    ├── skills/                      ← global Agent Skills (all workspaces)
    ├── mcp_config.json              ← global MCP config (same mcpServers shape)
    └── global_workflows/            ← saved-prompt Workflows (slash commands, /)

Path caveat (skills): the workspace skill dir is cited across official + community sources as .agents/skills/ AND .antigravity/skills/ AND .agent/skills/ (singular). Native skill discovery is unambiguous; the exact canonical path should be pinned by a live install test before being hard-coded into gald3r installers. Global-scope is ~/.gemini/antigravity/skills/.

  • gald3r writes: AGENTS.md, GEMINI.md (Antigravity-only directives), .agents/skills/<name>/SKILL.md, .agents/rules/*.md, .agents/workflows/*.md, .agents/hooks.json, .antigravity/mcp.json
  • Platform owns: ~/.gemini/ global state, IDE settings, the Orchestrator/subagent runtime, System Rules (immutable)

2. AI Instruction File

Antigravity reads, in precedence order: System Rules (immutable, DeepMind) → GEMINI.md (Antigravity-specific) → AGENTS.md (cross-tool, also read by Cursor / Claude Code). Global variants live at ~/.gemini/GEMINI.md and ~/.gemini/AGENTS.md (cross-tool global, v1.20.3+). No dedicated ANTIGRAVITY.md is required — gald3r's AGENTS.md is a first-class input. For Antigravity-only must-win directives, put them in GEMINI.md (it wins over AGENTS.md). Antigravity also has a durable "memories" concept (agent state) distinct from rule files.

3. Agents Support — ✅ NATIVE (platform) / ⚠️ gald3r g-agnt-*.md convention does NOT auto-map

  • Dynamic Subagents: an Orchestrator (dispatch-only manager) decomposes the goal and spawns subagents on the fly (start_subagent tool) with isolated context; they run asynchronously in the background and communicate via unique agent IDs. Agent Management groups conversations by "project" with its own settings/permissions.
  • IMPORTANT gald3r nuance: subagents are dynamic — there is no documented file-based static-role discovery (no honored .agents/agents/ dir for g-agnt-*.md). gald3r's markdown agent-role files do NOT auto-map; fold critical agent guidance into rules/skills instead.
  • Orchestrator self-handoff: tracks cumulative subagent spawn count; on hitting a limit it dumps state to handoff files, kills background tasks, and spawns a successor that resumes — relevant to gald3r long-running autonomous loops.
  • Managed Agents caveat: the separate hosted Gemini API "Managed Agents" SDK currently states subagent delegation is not yet supported — that is the hosted-SDK limitation, distinct from the IDE/CLI harness which does spawn subagents.
  • Source: https://antigravity.google/docs/subagents

4. Skills Support — ✅ NATIVE

  • Agent Skills (Anthropic SKILL.md open format, progressive disclosure): a directory-based package containing a definition file (SKILL.md) + optional supporting assets. Frontmatter: name (optional / unique / lowercase-hyphen), description (mandatory — the semantic trigger); markdown body. Loaded on-demand when the agent deems it relevant.
  • "The SKILL.md format is the same standard used by Claude Code, Cursor, and every other compatible agent." gald3r g-skl-*/SKILL.md files are directly compatible.
  • Directory path varies across official + community sources: workspace-scope cited as .agents/skills/ and .antigravity/skills/ and .agent/skills/; global-scope ~/.gemini/antigravity/skills/. Native discovery is unambiguous; pin the exact path via install test. Same harness shared by the Antigravity CLI and IDE.
  • Source: https://codelabs.developers.google.com/getting-started-with-antigravity-skills

5. Commands / Workflows — ✅ NATIVE

  • Workflows (saved-prompt slash commands): markdown files invoked in Agent via a slash command /<workflow-name>. A workflow file contains a title, description, and a series of steps; files are limited to 12,000 chars. Created via + Global (all workspaces) or + Workspace; global workflows are stored at ~/.gemini/antigravity/global_workflows/. A dedicated /docs/command page also exists.
  • Project-relative directory CONFIRMED (reversing the downgrade, re-scanned 2026-07-18): "+ Workspace" Workflows resolve to a real, project-relative directory — <project-root>/.agents/workflows/*.md — distinct from the global ~/.gemini/antigravity/global_workflows/ path. Confirmed via the official first-party Google Codelab (literal shell instruction mkdir -p .agents/workflows .agents/skills and file-creation step Create .agents/workflows/startcycle.md:), cross-checked against a Google AI Developer Forum thread describing workflows as resolving "in the current workspace" (project-scoped, not global-only). Path caveat (same ambiguity as §4 Skills): third-party sources (a Google Antigravity blog write-up and a community rules aggregator) spell the directory singular (.agent/workflows/) rather than plural (.agents/workflows/); the first-party Google Codelab uses plural, and gald3r's own .agents/rules, .agents/hooks, .agents/skills entries for this platform are all plural, so .agents/workflows is adopted here for internal consistency — pin the exact spelling via a live install test before treating it as immutable, same caveat already carried for the skills path. layout_map.yaml now carries a real commands: {dir: ".agents/workflows", ext: ".md", prefix: "g-"} entry for antigravity (previously absent found only the undocumented global path and downgraded the matrix Commands cell to ⚠). Flipping the live the platform capability matrix antigravity Commands cell from ⚠ back to ✅ is a coordinator-owned .gald3r/ write (g-rl-33 Swarm Reconciliation Gate — bucket agents don't write shared .gald3r/ ledgers directly; same deferral pattern as the goose Commands cell), not included in this authoring pass.
  • gald3r @g-* / /g-* commands map directly to Workflows (the closest analogue).
  • Source: https://antigravity.google/docs/rules-workflows · https://codelabs.developers.google.com/autonomous-ai-developer-pipelines-antigravity (first-party Google Codelab — mkdir -p .agents/workflows .agents/skills; Create .agents/workflows/startcycle.md:) · https://discuss.ai.google.dev/t/agents-not-reading-workflow-when-directed-with-slash-commands/115448 (community forum — workflows exist "in the current workspace")

6. Hooks System — ✅ NATIVE

  • Lifecycle Hooks (added in Antigravity 2.0 / Google I/O 2026, May 2026): execute custom local shell scripts at critical stages of the agent's execution cycle. Configured in hooks.json in the customization dir (.agents/ workspace, or ~/.gemini/config/ global; workspace precedence). Hooks receive input via stdin as JSON and return output via stdout as JSON.
  • Hook points: before_tool_call, after_tool_call (logging), before_model_call (inject system instructions), after_model_call (override exit rules), on_loop_stop, on_error.
  • gald3r g-hk-*.py hooks (python <path>, post- Python port; no PowerShell involved) can be wired as the executed shell scripts: before_tool_call ~ preToolUse, on_loop_stop ~ stop, before_model_call enables session-start-style instruction injection.
  • Source: https://antigravity.google/docs/hooks

7. Rules / Memory — ✅ NATIVE

  • Rules are plain Markdown files of constraints. Activation types: Manual (at-mention), Always On, Model Decision (NL description), Glob (file pattern). Each rule file is limited to 12,000 chars. Global rules: ~/.gemini/GEMINI.md; workspace rules: .agents/rules/ of the workspace/git root.
  • A full always-on guarantee exists (Always On activation) — stronger than the partial picture in the prior gald3r spec. Precedence: System Rules (immutable, DeepMind) > GEMINI.md > AGENTS.md. A durable "memories" concept also exists.
  • gald3r g-rl-* map to Always On (for alwaysApply: true) or Model Decision (for description:-scoped) rules.
  • Source: https://antigravity.google/docs/rules-workflows

8. MCP Support — ✅ NATIVE

  • Model Context Protocol: connect local tools, databases, and external services. Config shape { "mcpServers": { ... } } at .antigravity/mcp.json (project) or ~/.gemini/antigravity/mcp_config.json (global; Settings → Customizations → Open MCP Config). Both a JSON-file config and a GUI MCP Store ("..." dropdown → Browse & Install & Authenticate). Trusted/owned-workspace security gating for write-capable servers.
  • gald3r MCP servers (e.g. example_app) integrate via the mcpServers JSON.
  • Source: https://antigravity.google/docs/mcp

9. Scheduled Tasks — ✅ NATIVE (Antigravity-native bonus)

  • Scheduled Tasks (Antigravity 2.0): define crons to trigger agent invocations on a schedule. Adoptable for gald3r heartbeat / scheduled jobs.
  • Source: https://antigravity.google/docs/hooks + I/O 2026 deep-dive https://antigravity.google/blog/google-io-2026-feature-deep-dive

Parity vs. Cursor Reference

Antigravity now reaches near-full parity with the Cursor reference (g-skl-platform-cursor): native commands (Workflows), rules, agents (dynamic subagents), skills, hooks, and MCP, plus an Antigravity-native Scheduled Tasks bonus with no Cursor analog. The one structural caveat is that subagents are dynamic (Orchestrator-spawned) — there is no file-based g-agnt-*.md role discovery, so gald3r agent-role files must be folded into rules/skills rather than dropped into an agents dir.

Reuse note (important): because Antigravity reads AGENTS.md + GEMINI.md and discovers Anthropic-format SKILL.md packages, gald3r's AGENTS.md + g-skl-*/SKILL.md artifacts are largely reusable on Antigravity without a separate port. The exact skill directory path (.agents/skills vs .antigravity/skills vs .agent/skills) and the hook payload schemas should be pinned by a live install test before being hard-coded into installers.

Hook System

  • Type: native (hooks.json, shell scripts, stdin/stdout JSON)
  • Config file: hooks.json in the customization dir — .agents/ (workspace) or ~/.gemini/config/ (global); workspace precedence
  • Events available: before_tool_call, after_tool_call, before_model_call, after_model_call, on_loop_stop, on_error
  • Event payload format: JSON via stdin; result returned as JSON via stdout
  • Command extensions: local shell scripts (official examples are .sh); gald3r g-hk-*.py wired via { "type":"command", "command":"python <path> …" } (post- Python port; no PowerShell involved)
  • gald3r hook files: g-hk-*.py map to before_tool_call (pre-tool guards), on_loop_stop (stop), before_model_call (session-start-style instruction injection)
  • OS note: hooks invoke the local shell directly; python <path> runs cross-platform with no PowerShell dependency — pin payload schema via install test

Atypical Handling

  • Two products, one name: the IDE/CLI harness (full extensibility — this spec) vs the hosted Gemini API "Managed Agents" (separate; no subagents / custom tools yet). Do not conflate.
  • Instruction file: AGENTS.md + GEMINI.md (NOT CLAUDE.md); GEMINI.md outranks AGENTS.md; immutable System Rules sit above both.
  • Dynamic subagents: Orchestrator spawns subagents on the fly — no static g-agnt-*.md dir.
  • Global state is Gemini-namespaced under ~/.gemini/ (shared with the Gemini CLI tree).
  • Skill dir path is not yet canonical across sources — pin via install test.
  • Workflows dir path: confirmed project-relative (.agents/workflows/ per the first-party Google Codelab); same singular/plural spelling ambiguity as the skill dir across third-party sources — pin via install test.

gald3r Integration Notes

  • Ship gald3r's AGENTS.md + g-skl-*/SKILL.md — Antigravity reads/discovers them natively.
  • Ship gald3r's @g-* / /g-* commands as Workflows at .agents/workflows/*.md ( — project-relative, layout_map.yaml commands entry now authored).
  • Put Antigravity-only must-win directives in GEMINI.md (outranks AGENTS.md).
  • Hooks fire natively (shell commands; g-hk-*.py via python <path>, post- Python port; no PowerShell involved) — before_model_call covers session-start injection, on_loop_stop covers stop, before_tool_call covers .gald3r/ guards.
  • Fold g-agnt-* guidance into rules/skills (no file-based subagent discovery).
  • Adopt Scheduled Tasks (crons) for gald3r heartbeat jobs.
  • Evidence caveat: antigravity.google/docs/* is a JS SPA (WebFetch returned title only); capability evidence comes from search-engine extractions quoting the official pages verbatim, cross-checked against multiple 2026 sources + Google codelabs. Pin skill path + hook payload schema via a live install test before hard-coding.
  • New, high-churn platform (relaunched ~2026-05-19) — re-verify on the next @g-platform-scan-docs antigravity (crawl_max_age_days: 7).

Capability Summary

Hooks Rules Skills Commands MCP Docs Fresh
✅ ✅ ✅ ✅ ✅ ✅

Legend: ✅ verified working · ⚠️ partial / Cursor-generic · ❌ not supported · ❓ untested.

Agents (6th primitive): ✅ native dynamic subagents at the platform level; ⚠️ the gald3r g-agnt-*.md file convention does not auto-map (no static-role discovery) — fold into rules/skills.

Verification Evidence (docs crawl 2026-06-02, https://antigravity.google/docs)

Capability How verified
Commands / Workflows /docs/rules-workflows — Workflows = markdown saved-prompts, /workflow-name, title/description/steps, ≤12,000 chars, Global vs Workspace; also /docs/command. (2026-07-18): Workspace scope is project-relative .agents/workflows/*.md — confirmed via codelabs.developers.google.com/autonomous-ai-developer-pipelines-antigravity (first-party Google Codelab, literal mkdir -p .agents/workflows) + discuss.ai.google.dev forum thread 115448 (workflows resolve "in the current workspace"); singular/plural spelling ambiguity noted, same as the skills path below
Rules /docs/rules-workflows — Markdown rules; activation Manual / Always On / Model Decision / Glob; ≤12,000 chars; ~/.gemini/GEMINI.md (global) + .agents/rules/ (workspace); precedence System Rules > GEMINI.md > AGENTS.md
Agents /docs/subagents — dynamic Orchestrator-spawned subagents (start_subagent), async/background, agent IDs, Agent Management; no file-based g-agnt-*.md discovery
Skills codelabs (getting-started-with-antigravity-skills) — Anthropic SKILL.md standard ("same as Claude Code, Cursor"); YAML frontmatter (name optional, description mandatory); on-demand load; path .agents/.antigravity/.agent skills (pin via install test)
Hooks /docs/hooks — hooks.json shell scripts, stdin/stdout JSON; events before_tool_call / after_tool_call / before_model_call / after_model_call / on_loop_stop / on_error; global + workspace (workspace precedence); added Antigravity 2.0 / I/O 2026
MCP /docs/mcp — { "mcpServers": {...} } at .antigravity/mcp.json (project) / ~/.gemini/antigravity/mcp_config.json (global); MCP Store GUI; owned-workspace gating
Instruction file /docs/rules-workflows — AGENTS.md (cross-tool, also Cursor/Claude Code) applied after GEMINI.md; global ~/.gemini/GEMINI.md + ~/.gemini/AGENTS.md (v1.20.3+)
Scheduled Tasks /docs/hooks + blog/google-io-2026-feature-deep-dive — cron-style agent invocations
Surface separation ai.google.dev/gemini-api/docs/antigravity-agent — hosted "Managed Agents": custom tools + subagent delegation "not yet supported" → distinct from IDE/CLI harness rated here
Evidence method antigravity.google/docs/* is a JS SPA (WebFetch → title only); ratings from search-engine extractions quoting official pages verbatim, cross-checked vs multiple 2026 sources + Google codelabs; confirmed pages: /docs/rules-workflows, /docs/mcp, /docs/subagents, /docs/hooks, /docs/agent, /docs/command, /docs/task-list

Reference page — canonical source: g-skl-platform-antigravity/PLATFORM_SPEC.md in the engine repo, generated by scripts/generate_platform_support_docs.py. Pages marked curated are hand-assessed and not yet live-verified.