`serena-hooks remind --client=codex` misses grep/read workflows inside `exec_command`

[TyceHerrman]

serena-hooks remind --client=codex misses grep/read workflows inside exec_command

serena-hooks remind --client=codex classifies pre-tool calls by tool_name / toolName. In Codex, shell work arrives as the generic tool name exec_command, with the actual command under tool_input.cmd.

That means common Codex exploration like this is invisible to the remind counter:

{
  "session_id": "demo",
  "tool_name": "exec_command",
  "tool_input": { "cmd": "rg \"SomeSymbol\" packages/server" }
}

Same for file-read style shell commands:

{
  "session_id": "demo",
  "tool_name": "exec_command",
  "tool_input": { "cmd": "sed -n '1,220p' packages/server/foo.ts" }
}

Current relevant matcher behavior:

• is_grep_tool(): for non-Claude clients, checks whether "grep" in tool_name
• is_read_file_tool(): for non-Claude clients, checks whether "file" in tool_name and one of read/view/open/show is also in tool_name

Since the tool name is only exec_command, neither path fires.

I am not sure Serena should parse arbitrary shell commands. This may be better solved by Codex exposing richer structured tool-call metadata to hooks, for example a shell command category or parsed command segments. But as-is, the Codex remind hook can run successfully before every tool call while missing the main grep/read behavior it is meant to detect.

Environment:

• Serena: serena-agent 1.1.2
• Client: Codex
• Hook: serena-hooks remind --client=codex

Written with assistance from Codex.

[MischaPanch]

Yes, we didn't expand the hook mechanism to cover these cases. Feel free to issue a PR that handles this for codex