Tool-First Onboarding

July 19, 2026 · View on GitHub

Start here by choosing the tool you are using now (今使っている tool). Claude Code Harness is still Claude-first, but Phase 73 makes the entry point explicit for every host so users do not mistake a candidate route for a proven install path.

Detailed commands live in install.md. Existing users should run the report-first migration path in migration.md before cleanup.

Support Tier Rule

Public wording must keep these tiers unchanged:

ToolPhase 73 tierStart here
Claude CodesupportedUse the Claude plugin install path in install.md.
Codex CLIsupportedUse scripts/setup-codex.sh --user in install.md.
Codex appcandidateUse the candidate smoke checklist in install.md; do not reuse Codex CLI proof.
OpenCodeinternal-compatibleUse scripts/setup-opencode.sh in install.md.
CursorsupportedUse scripts/setup-cursor.sh in install.md; containment is harness-side — see docs/CURSOR_INTEGRATION.md.
GroksupportedUse scripts/setup-grok.sh in install.md; Claude-envelope PreToolUse floor — not full Claude hook parity.
Hermes AgentcandidateUse the candidate boundary in install.md; manual symlink research only.
GitHub Copilot CLIcandidateUse the candidate boundary in install.md.
Antigravity CLIfuture/unsupportedUse the unsupported boundary in install.md.

not_observed != absent: missing local runtime evidence means the capability is not observed in the current artifact set. It does not prove the capability is absent.

Public 正式対応 maps to EN supported for Claude Code, Codex CLI, Cursor, and Grok (H8 pin). Remaining rows use 互換利用可 or 試験対応 per tier. OpenCode, the separate app candidate path, Hermes Agent, GitHub Copilot CLI, and Antigravity CLI must not borrow that wording until their own H1–H8 pass (docs/spec/planning-and-host-adapter.md).

EN tierJP public wording
supported正式対応
internal-compatible互換利用可 / 制限付き対応
candidate試験対応 / プレビュー
future/unsupported非対応 / 将来検討

Choose The Route

If you are using...Do this firstDo not claim
Claude CodeInstall the marketplace plugin, then run /harness-setup.Generic multi-host support beyond proven gates.
Codex CLIRun scripts/setup-codex.sh --user, restart Codex, then invoke $harness-plan.Direct Codex plugin install or Codex app parity.
OpenCodeRun scripts/setup-opencode.sh, start OpenCode, then ask for harness-plan.Claude hook parity or runtime auto-routing parity.
Codex appRecord candidate smoke evidence only.That app behavior is proven by Codex CLI docs or help output.
CursorRun scripts/setup-cursor.sh, reload Cursor, then invoke /breezing or /harness-plan.Top-tier Cursor adapter or Claude hook parity.
GrokRun scripts/setup-grok.sh, restart Grok, then invoke /harness-plan or /breezing.Public top-tier claim for the Grok adapter; Claude SessionStart/PreToolUse parity.
Hermes AgentTreat manual symlink exposure as candidate research only.Public Hermes support, Claude hook parity, or runtime workflow parity.
GitHub Copilot CLITreat it as CLI capability research only.Harness bootstrap support.
Antigravity CLIKeep it out of end-user install flow.Install support, setup support, or adapter support.

First Successful Session

A route is usable only when it has all of these:

  • first prompt,
  • first command,
  • verification command,
  • success look,
  • support tier and known asymmetry.

For candidate and unsupported hosts, the success look is not "installed". It is "the boundary stayed honest": evidence is recorded as candidate, future/unsupported, manual, or not observed.