HN 표시: Codex 자동/검토 루프

Run Codex app-server reviews of uncommitted changes via an MCP tool. Compared to the built-in /review , this keeps review context clean while fixes happen in the main session that retains implementation knowledge, reducing regressions and enabling longer autonomous runs with better code quality. - Codex CLI installed and authenticated. - uv installed. Codex loads MCP servers from ~/.codex/config.toml and supports configuring them via the codex mcp CLI. Profile example (gpt-5.5, xhigh reasoning effort): # ~/.codex/config.toml [profiles.review] model = "gpt-5.5" model_reasoning_effort = "xhigh" AGENTS.md example instruction: for verification call `review_uncommitted_changes: runs=1` until no issues Set tool_timeout_sec for the MCP server in ~/.codex/config.toml to a value larger than the review timeout (--timeout-seconds passed to the server command). Example: if the review timeout is 2700 seconds, set tool_timeout_sec = 3000 . This registers a stdio MCP server that Codex launches when a session starts. codex mcp add codex-code-review -- \ uv run -m mcp_code_review.server --parallelism 1 --concurrency-mode auto --timeout-seconds 2700 \ --profile review The codex mcp add workflow is the supported way to add MCP servers from the CLI. After adding the server, set tool_timeout_sec in ~/.codex/config.toml (see Option B) so it is higher than --timeout-seconds . Additional review instructions are enabled by default. Existing configurations that include --enable additional_review_instructions still work for backward compatibility. To disable additional review instructions, add: --disable additional_review_instructions Add an MCP server entry in ~/.codex/config.toml : [mcp_servers.codex-code-review] command = "uv" tool_timeout_sec = 3000 args = [ "run", "-m", "mcp_code_review.server", "--parallelism", "1", "--concurrency-mode", "auto", "--timeout-seconds", "2700", "--profile", "review" ] Codex reads MCP server entries from the mcp_servers table in ~/.codex/config.toml . If you prefer not to clone locally, you can run the server directly from the Git repository using uvx . Repository: https://github.com/Szpadel/codex-mcp-code-review codex mcp add codex-code-review-uvx -- \ uvx --from git+https://github.com/Szpadel/codex-mcp-code-review \ python -m mcp_code_review.server --parallelism 1 --concurrency-mode auto --timeout-seconds 2700 \ --profile review [mcp_servers.codex-code-review-uvx] command = "uvx" tool_timeout_sec = 3000 args = [ "--from", "git+https://github.com/Szpadel/codex-mcp-code-review", "python", "-m", "mcp_code_review.server", "--parallelism", "1", "--concurrency-mode", "auto", "--timeout-seconds", "2700", "--profile", "review" ] { "mcpServers": { "codex-code-review-uvx": { "command": "uvx", "args": [ "--from", "git+https://github.com/Szpadel/codex-mcp-code-review", "python", "-m", "mcp_code_review.server", "--parallelism", "1", "--concurrency-mode", "auto", "--timeout-seconds", "2700", "--profile", "review" ], "env": {} } } } If you want Codex to run the server directly from this source checkout, point uv at the project path. codex mcp add codex-code-review-dev -- \ uv run --project /absolute/path/to/codex-mcp-code-review -m mcp_code_review.server \ --parallelism 1 --concurrency-mode auto --timeout-seconds 2700 --profile review [mcp_servers.codex-code-review-dev] command = "uv" tool_timeout_sec = 3000 args = [ "run", "--project", "/absolute/path/to/codex-mcp-code-review", "-m", "mcp_code_review.server", "--parallelism", "1", "--concurrency-mode", "auto", "--timeout-seconds", "2700", "--profile", "review" ] - Tool name: review_uncommitted_changes . - Uses the native app-server review target uncommittedChanges (includes untracked files). - Additional review instructions are enabled by default, exposing the additional_developer_instructions tool argument. - When that argument is set, the server switches that run to a custom review target prompt built from a synced copy of Codex's native uncommitted-changes instructions and appends an Additional review instructions: section. - Existing --enable additional_review_instructions usage is still accepted for backward compatibility. - When disabled with --disable additional_review_instructions , the tool schema does not advertiseadditional_developer_instructions , and requests that send it anyway are rejected. - Default runs: 4 (override by setting --parallelism on the MCP server config). - Sandbox: read-only; approval policy: never.