[F0rce]

Problem

mgrep install-codex appends a full skill file ("MANDATORY ... NEVER use grep") to the global ~/.codex/AGENTS.md. With it, Codex stops using any other way of searching. Without it, Codex ignores mgrep. The install also registers codex mcp add mgrep mgrep mcp, a server that exposes no tools, returns an invalid CallTool result and reports version 0.1.3; it only exists to start the watcher. Searches go through the shell CLI inside Codex's network sandbox, so each mgrep search asks for approval.

Expected: Codex uses mgrep where semantic search helps, keeps rg/grep for exact lookups, and searches without approval prompts.

Change

The Codex integration becomes a native Codex plugin, published from the same repository marketplace as the Claude plugin (.agents/plugins/marketplace.json, plugins/mgrep/.codex-plugin/plugin.json):

• MCP server mgrep mcp --expose-tools with one read-only search tool. Codex starts it in the session directory and stops it at session end. It syncs that directory in the background, so no watcher outlives the session. The first search waits up to 20 s for the initial sync, then searches the existing index and says so.
• SessionStart hook that only adds guidance: mgrep for conceptual or cross-file questions, rg/grep for exact strings, symbols and regexes. Codex asks the user to review it once.
• Skill with a short neutral description of the tool and the CLI.

mgrep install-codex adds the marketplace and installs mgrep@Mixedbread-Grep; mgrep uninstall-codex reverses it. Both remove the legacy AGENTS.md block and MCP registration.

Changes outside src/install/codex.ts that the MCP tool needs: search logic moves to src/lib/search.ts so the CLI and the tool share it; createStore({ interactive: false }) so the server never prompts; the version comes from package.json; startWatch no longer calls process.exit and reports whether watching started, so the server can wait for it. The Claude plugin is unchanged.

Scope and approval

There is no prior issue or discussion. The direction matches the unmerged maintainer branch chore/use-mcp-for-code (MCP tool instead of shell search). Happy to open a discussion first or split differently if you prefer.

Verification

Selective use, measured with codex exec (Codex CLI 0.160.1, gpt-6.1-sol, reasoning high) in fresh CODEX_HOMEs on a private 853-file TypeScript monorepo. 10 questions: 6 conceptual (no identifiers named) and 4 exact (a symbol, a literal error string, a config key, an env var). 5 runs per question and variant, 100 runs, none failed. Correct = the answer cites at least one expected file.

	Plugin without hook	Plugin with hook (this PR)
mgrep used, conceptual	8/30	30/30
mgrep used, exact	0/20	0/20
Correct	50/50	50/50
Mean cost index, conceptual (median)	68.7k (60.9k)	71.2k (61.7k)
Mean cost index, exact	10.1k	9.7k

Cost index = uncached input + 0.1 × cached input + 8 × output tokens. With the hook Codex makes about one mgrep call per conceptual question, reads the returned ranges and confirms with rg; it never uses mgrep for exact lookups. Cost and correctness stay within noise. This PR does not claim mgrep makes Codex cheaper on this repository.

A single-run screen of stronger wordings: an explicit decision rule led to up to 17 mgrep calls on one question; "always start with mgrep" also used mgrep on exact lookups and doubled their cost, which is the current AGENTS.md behaviour.

Other checks:

• test/codex.bats covers the exact codex invocations of install and uninstall, legacy AGENTS.md cleanup with surrounding content kept, idempotent reinstall, and the hook output. test/mcp.bats covers the tool schema, root resolution, waiting for the watcher, the 20 s bound and the fallback sync. 92/92 pass.
• Installed from a local checkout into a fresh CODEX_HOME: codex mcp list shows the server and app-server hooks/list shows the hook.
• In real codex exec sessions the MCP server starts in the session directory, mgrep.search runs without an approval prompt, and nothing keeps running after the session ends.

Not checked: interactive Codex TUI on a machine without a previous mgrep install, and Windows.

Model and harness: Claude Opus 5.5 (1M context) in Claude Code for design, review and benchmarking; GPT-6.1-Sol in Codex for implementation.