The goal is behavioral parity, not a byte-for-byte copy of Gemini configuration or chat history. Start with the few workflows that are both hardened and worth repeating.
Current host boundary: consumer Gemini CLI stopped serving free and Google AI Pro/Ultra requests on June 18, 2026 and transitioned to Antigravity CLI. Standard/Enterprise and paid API-key Gemini CLI use remains supported. This guide can recover useful Gemini artifacts and supports portable-skill discovery for continuing Gemini CLI, but it does not claim that Antigravity has the same skill paths, permissions, hooks, or lifecycle. Google says initial parity is not 1:1; see the official transition announcement.
- Run
/migrate-geminito inventory instructions, commands, skills, hooks, and MCP configuration. - As an alternative, explicitly invoke Claude Code's interactive
/import gemini --dry-runwhen the installed version/provider supports it. A custom skill cannot invoke that slash command for you. - Define a parity case for each selected workflow before applying changes.
- Run
/mine-gemini-workflowsonly when session evidence is needed to reconstruct the workflow. - Migrate one workflow, validate it, and review the diff before moving to the next.
Reusable provider-neutral instructions belong in .agents/skills/<name>/SKILL.md, which Codex and continuing enterprise/API-key Gemini CLI can discover. Claude-specific slash-command adapters live in .claude/commands/ and can load the same portable skill. Antigravity discovery has not been validated here. Hooks, tool names, MCP servers, and permission rules stay in provider-specific adapters.
This keeps the workflow logic in one place without pretending each agent runtime has the same execution model.
Compare observable invariants:
- required sections and artifacts
- tool/capability use
- approval boundaries
- citations or evidence
- forbidden behavior
- validation results
Exact wording is not a useful parity target. Use docs/templates/workflow-parity.json as the review contract.
Gemini session recordings can contain source code, local paths, prompts, tool arguments, and credentials. The included miner defaults to metadata-only output, requires an explicitly chosen session directory, folds append-log rewinds and metadata updates, and never emits thought/reasoning fields. Free-form summaries, sanitized paths, and observable content each require explicit session IDs and opt-in flags. Redaction is best-effort, so every opt-in report remains sensitive and needs review before sharing or committing.
Generated inventories belong under .context-os/migrations/, which is ignored by git.