Files
Alex Joaquim Pereira 773ed61443 Fix Codex's approval guidance for the inline tools table (R2.2) (#1918)
The hint for a tool Codex refused told the person to add a [mcp_servers.craft-anything.tools.<tool>] table, but setup writes tools as one inline table, and TOML refuses to extend an inline table that way, so Codex's config stopped parsing. The hint and docs/codex.md now say to add the tool inside tools, and a test checks that the old form fails and both valid forms parse. For #1867.
2026-11-16 00:38:07 +07:00

9.5 KiB

Using Codex with Craft Anything

Craft Anything reaches Codex in two ways, and both have to get past Codex's own tool approval.

Find and check Codex on macOS

An app opened from the Dock does not inherit the terminal's PATH. When PATH has no Codex, detection and execution also look for its CLI in known ChatGPT.app and Codex.app layouts, first under the user's Applications folder and then the system Applications folder. Only executable files are accepted; the desktop GUI program is never used as a coding agent. An explicitly selected program and PATH remain ahead of the bundle fallback. A saved bare codex command gets the same fallback. There is no shell-profile edit, global symlink, recursive scan or app launch.

In Settings → AI, add Codex under Found on this computer → Harnesses, then use Check in its configured row. The check runs the selected program's --version and login status on a worker. craft-anything-cli ai test <name> uses the same headless API. It sends no prompt, and its version_only flag remains true: installation and sign-in do not prove model entitlement or remaining usage. Unknown or failed status output leaves the version reply unchanged, without a signed-in claim. Every configured local agent CLI has the same Check action; its adapter determines whether sign-in can be verified, using that backend's selected configuration folder where supported.

Codex owns authentication. Run the selected Codex program with login in a terminal and choose Sign in with ChatGPT, or use Codex's supported API-key login. Craft Anything does not read, copy or store Codex's credential files or tokens. This is not native ChatGPT OAuth inside Craft Anything and does not change Codex's sandbox, approval policy or model selection. A known signed-out state and recognized authentication failures include sign-in guidance. Unrelated errors such as a failed login shell do not get authentication advice. SSH connection tests retain their existing behavior; the Settings action is local-only.

The bundle layouts, executable permissions, root order, login-status parsing, selected-CLI connection check and Settings result have isolated regressions. Fixtures use fake programs and temporary directories, never real credentials. Windows/Linux live-account behavior and model availability are not established by these tests.

Why a tool can be refused

codex exec is non-interactive: it has no one to ask, so its approval policy is never. A tool that an MCP server doesn't mark read-only (such as rebuild_start, which runs a coding agent) is refused before our server sees the call, with requires approval, but approval policy is never. Codex's own settings decide this; Craft Anything's consent is a second, separate check.

What we set

Only tools that ask the person themselves are approved on Codex's side (AGENT_APP_PREAPPROVED_TOOLS in the engine, today just rebuild_start; a test checks each one asks for consent). Every other tool keeps whatever approval policy you set in Codex.

So when a build's feed says Codex blocked another of our tools, such as build_command_run, running setup again changes nothing for it (#1867). To let Codex call it during a build, approve it yourself in Codex's config.toml; setup keeps the setting. Setup writes tools as one inline table, so add the tool inside that table:

tools = { rebuild_start = { approval_mode = "approve" }, build_command_run = { approval_mode = "approve" } }

A [mcp_servers.craft-anything.tools.build_command_run] table cannot be added after an inline tools table: TOML refuses it with "cannot extend value of type inline table with a dotted key". If your file already uses a [mcp_servers.craft-anything.tools] table, add build_command_run = { approval_mode = "approve" } there instead.

  • Connecting Codex to Craft Anything (craft-anything-cli setup --client codex --yes, or the Connect screen) writes the entry in ~/.codex/config.toml ($CODEX_HOME/config.toml if set):

    [mcp_servers.craft-anything]
    command = "…/craft-anything-cli"
    args = ["mcp"]
    startup_timeout_sec = 30
    tools = { rebuild_start = { approval_mode = "approve" } }
    

    Your other entries and comments stay as they are, the old file is kept as config.toml.craft-anything.backup, and settings you added to our entry (env, enabled_tools, approvals for other tools) are kept when setup updates it. Setting approval_mode for rebuild_start yourself is overwritten by the next setup run. craft-anything-cli doctor --client codex reports an entry without the approval as "a different copy or shape of the command"; running setup again fixes it.

  • Building with Codex from the app starts codex exec with the same setting on its -c flags, so nothing in your config is needed. Chats and builds also start the Craft Anything MCP server with --all-tools: Codex doesn't refresh its callable tool list after enable_toolset, so every tool permitted by the current mode is listed from startup. Clean room restrictions still apply.

Who still asks

Approving rebuild_start in Codex doesn't remove Craft Anything's own consent: it runs a coding agent, so our server asks before it runs. Who can answer depends on how the server was started:

  • Builds started from the app run the server with the project's inbox, so the person answers in the app (Allow / Don't allow).
  • The entry from setup starts craft-anything-cli mcp on its own, with no app attached, so there is no approval card: the call returns consent_required and nothing runs. To let Codex start builds this way, add "--allow-run" to the entry's args (every tool that runs code is then treated as allowed, so only do it for a Codex you trust), or start the build from the app.

In Clean room mode the tools that touch the original stay refused either way.

If it still fails

A refusal for one of our tools now says what was blocked and how to fix it, instead of showing null; another MCP server's refusal keeps Codex's own words, because our setup can't change that server's policy. To approve more of our tools by hand, add <tool> = { approval_mode = "approve" } inside the tools table in Codex's config.toml (the inline form setup writes), and start the chat again. Approving all of them (default_tools_approval_mode = "approve") also approves tools with no consent check of their own, such as project_inbox_answer, which can answer a pending approval, so we don't do it.

Startup tool discovery was verified with Codex CLI 0.166.0, a loopback model and the real MCP server: Codex received 107 Clean-room-permitted tools, including non-Core tools, while Deep-only tools stayed hidden. No real-account build was run.

Continuing a Codex session

The adapter saves Codex's reported thread id and uses codex exec resume <id> for later turns, including a project reopened with its saved agent session. The run's model, sandbox and MCP configuration are passed again. Exec-only flags stay before the resume subcommand, the id is protected as positional data, and the prompt continues to arrive on stdin. It never selects the most recent unrelated session.

Argument regressions and a local stand-in program cover a fresh turn, a second turn and reopening a saved conversation. To explicitly check an installed Codex against a real account, build the CLI and run the manual example:

cargo build -p craft-anything-cli
PATH="$PWD/target/debug:$PATH" cargo run -p craft-anything-ai --example codex_resume_smoke -- <model>

The model argument is optional. The example uses the installed Codex's sign-in and model allowance; it is not part of automated tests. It ignores Codex's user configuration for this check, calls the real core_targets_list MCP tool, creates one file in a temporary workspace using the workspace-write sandbox, then verifies both an in-memory follow-up and a reopened conversation keep the thread and marker. The temporary workspace is removed afterwards; Codex keeps its usual session record. It does not test a complete Study → Build → Check → Ready project or GUI behavior.

This check passed on macOS with Codex CLI 0.169.0-alpha.11.1 and gpt-6.1-sol on 2026-11-14, using an existing ChatGPT sign-in.

Resume argument order (including --sandbox before resume) has been live-checked only on Codex 0.169.0-alpha.11.1. Older releases, including 0.167, have not been verified with this adapter; passing --sandbox after resume is not supported by the checked CLI. The overlapping PR #1805 reports a config-override approach on 0.167.

A missing saved thread is reported on stderr as no rollout found for thread id. The adapter marks the session missing, clears the saved id and explains that asking again starts a fresh conversation. Other stderr failures do not clear the session. To include this recovery in the manual live check, append --missing-thread after the model argument. It first tries a deliberately nonexistent id and then runs the normal fresh-turn, follow-up and reopened-session checks. This extended check passed with 0.169.0-alpha.11.1 on macOS on 2026-11-14.

The native adapter refuses a backend's --worktree option before starting any turn, because Codex 0.169 rejects it with resume. Select an existing worktree's folder as the project folder instead.

The native adapter also refuses --ephemeral before starting a turn: Codex does not persist those conversations, so their reported ids cannot be resumed later. Remove that option to use the project's saved conversation. A two-turn subprocess regression checks that both attempts explain the refusal, start no Codex process and save no session id.