Skip to content

IDE support

momentum is agent-agnostic. The same commands, rules, and workflow ship to every adapter; what changes is where the instruction file lives, how hooks attach, and which slash-command surface the IDE exposes.

AgentStatusPrimary instruction fileHook surfaceSlash commands
Claude CodeAvailableCLAUDE.md.claude/settings.json — PreToolUse, PostToolUse, SessionStartNative .claude/commands/*.md
CodexTesting pendingAGENTS.md.codex/hooks.json (PreToolUse / PostToolUse / SessionStart, apply_patch|Bash)Native skills at .agents/skills/<name>/SKILL.md
AntigravityTesting pendingAGENTS.md.agents/hooks.json (`run_commandview_file
CursorPlanned (Phase 19 — Reach).cursor/rules/TBDRules-based prompt fragments
Gemini CLIPlanned (Phase 19 — Reach)GEMINI.mdTBDEmbedded sections

Status legend. Available = installed adapter + live-runtime verified. Testing pending = adapter ships and unit-tests pass, but full live-runtime smoke against the actual CLI is still being validated (tracked as VAL-001 for Codex and VAL-002 for Antigravity). Planned = adapter not yet built.

Default adapter — --agent claude-code is implicit. The richest hook surface today.

Terminal window
npx @avinash-singh-io/momentum@latest init
  • CLAUDE.md at the project root — Claude Code reads this automatically at session start.
  • .agent/rules/project.md — the 13 autonomous rules (referenced from CLAUDE.md).
  • .claude/commands/*.md — all 15+ slash commands available natively.
  • .claude/settings.json — hook wiring for the three hook surfaces.
  • scripts/check-history-reminder.sh — PostToolUse hook; reminds about Rule 8 after edits to significant files.
  • scripts/brainstorm-gate.sh — PreToolUse hook; blocks Write/Edit on specs/ paths during /brainstorm-phase.
  • scripts/sessionstart-handoff.sh — SessionStart hook; auto-greets pending handoffs in .momentum/inbox/.
HookWhat it doesWhen it fires
PreToolUse (brainstorm-gate.sh)Blocks Write/Edit/MultiEdit calls targeting specs/ paths when .momentum/brainstorm-active existsBefore every tool call during a brainstorm session
PostToolUse (check-history-reminder.sh)Reminds the agent to log a history entry when significant files (status, backlog, phase tasks, decisions, architecture) are edited; opportunistically appends to ecosystem session logAfter every Write/Edit/Bash call
SessionStart (sessionstart-handoff.sh)Scans .momentum/inbox/ for pending handoffs, banners + [y/skip] promptAt each new session start

Hooks are the strongest discipline-enforcement layer: brainstorm-gate physically prevents premature disk writes, the history reminder closes the “I’ll log it later” failure mode, SessionStart eliminates the “did anyone hand work off to me?” check.

When you want the strongest discipline. The hooks make Rules 7 (brainstorm gate), 8 (history), and the handoff pickup all enforced rather than contractual.

Terminal window
npx @avinash-singh-io/momentum@latest init --agent codex
  • AGENTS.md — Codex’s primary instruction file (analogous to CLAUDE.md).
  • .agent/rules/project.md — the 13 rules.
  • .agents/skills/<name>/SKILL.md — each momentum recipe ships as a native Codex skill (skills docs) with name + description frontmatter. Codex auto-discovers skills at session start across CLI, IDE, and desktop app, so /brainstorm-phase, $brainstorm-phase, and natural-language “run brainstorm-phase” all route to the same SKILL.md.
  • .codex/agents/*.toml — TOML reviewer subagents (security, QA, architecture) with sandbox_mode = "read-only".
  • .codex/hooks.json — Codex hook config wiring the same three hook scripts as Claude Code. Hooks are stable + default-on in current Codex CLI; first run prompts you to trust each hook via /hooks.

Same three hook scripts as Claude Code — wired via .codex/hooks.json instead of .claude/settings.json. The discipline layer is the same; the configuration shape is per-adapter.

When you’re already using Codex and want momentum’s discipline without switching IDEs. Hook coverage matches Claude Code.

Terminal window
npx @avinash-singh-io/momentum@latest init --agent antigravity
  • AGENTS.md — shared format with Codex.
  • .agent/rules/project.md — the 13 rules.

Antigravity has no hook surface today. The brainstorm-gate discipline relies on the markdown contract — the agent still respects the rule, but there’s no PreToolUse hook to physically enforce it. Similarly, no SessionStart greet for handoffs (use /continue explicitly).

The contract still works — the brainstorm gate is documented in /brainstorm-phase’s body, and a well-behaved agent honors it. But the safety net is gone. If you accidentally tell the agent “just write the draft to disk so I can see it” mid-brainstorm, nothing physically stops it on Antigravity.

When you’re already on Antigravity and want momentum’s structure even without the hook enforcement layer. The discipline is still there; just trust-based.

Planned for Phase 19 — Reach. Cursor uses .cursor/rules/*.mdc files and doesn’t support slash commands the way Claude Code does. Commands will ship as rules-based prompt fragments — typing the command name will trigger the relevant rule.

Hook surface: TBD as part of Phase 14 design. Cursor’s editor hooks are different from Claude Code’s session hooks; the brainstorm-gate equivalent may need a different enforcement strategy (e.g., a file-watcher rule that warns if the sentinel exists).

Track progress: FEAT-007 in backlog.

Planned for Phase 19 — Reach. Gemini CLI uses GEMINI.md as the primary instruction file (single-file convention). Workflow prompts will embed as sections within GEMINI.md rather than as separate command files.

Hook surface: TBD. Gemini CLI’s plugin model may support before/after-tool hooks similar to Claude Code; if it does, the same three hook scripts will attach with minimal porting.

Track progress: FEAT-008 in backlog.

If you have a choice today, Claude Code gives you the strongest discipline because hooks physically enforce the brainstorm gate and surface handoffs at session start. Codex is the close second — AGENTS.md auto-loads and .codex/hooks.json enforces the same discipline. Antigravity works fine but relies on agent compliance with the markdown contract rather than hook enforcement.

If you’re already on Cursor or Gemini CLI, momentum will support both in Phase 14 — for now, you can wait, or use --agent claude-code to scaffold the structure and adapt it manually (the specs/ skeleton is agent-agnostic; only .claude/ and CLAUDE.md are Claude-specific).

If your team uses multiple agents (some on Claude Code, some on Codex), run momentum init --agent claude-code first, then momentum upgrade --agent codex to layer the Codex adapter overlay on top. Both CLAUDE.md and AGENTS.md will exist; both will reference the same specs/ and .agent/.

The adapter contract guarantees the overlays don’t conflict — installing both is additive. Each agent reads its own primary instruction file at session start; they share the per-project state.

Terminal window
momentum upgrade

The upgrade is marker-aware: anything under ## Project Extensions in CLAUDE.md / AGENTS.md is preserved across the upgrade. Default commands and rules update from the published template; project-specific content stays.

Known regression: see BUG-006 for an upgrade-related title-line issue. Fix is on the backlog.

Terminal window
momentum doctor

Walks the local layout: does the project look like a momentum-installed project? Are the rules referenced from the instruction file? Are the adapter files where they should be? Is the project an ecosystem member (if ecosystem.json is detectable in a sibling dir)?

Reports detected issues plus suggested fixes. Run doctor first whenever something feels off; it catches the common “the instruction file isn’t loading” class of problems.