Supported agents
Claude Code and OpenAI Codex: where Argus reads each one from, what it never opens, and how each is counted.
Argus reads two coding agents: Claude Code and OpenAI Codex. There is nothing to configure. On start it checks which agents' data exist on this machine and watches those. The startup log names them, for example Detected adapters: claude_code, codex.
Every session is stored with its agent. Session ids look like claude_code:<id> or codex:<id>, and the dashboard labels each session with its agent and version (for example Claude Code 2.1.94 or Codex 0.159.2). Once you have sessions from both, the Sessions page gains an Agent filter and column.
Started using an agent after Argus?
Argus decides which agents to watch when it starts. If you start using Codex (or Claude Code) while Argus is already running, restart it: argus start again, or argus daemon restart if you use the daemon.
Claude Code
| Details | |
|---|---|
| Reads from | ~/.claude/: session transcripts under projects/<project>/<session>.jsonl, their sub-agent transcripts, and the prompt history file history.jsonl |
| Location | Fixed at ~/.claude/; there is no flag or environment variable to move it |
| One turn | One API message (Claude Code writes a streamed reply as several lines; Argus folds them into one turn) |
| Tool calls | Each tool_use block; an error is a tool_result the transcript flags as one |
| Sub-agents | Each Task / Agent delegation's transcript, kept as its own sub-session under the parent |
| Resume a session | claude --resume <id> (the id chip on a session page copies it) |
Claude Code deletes its own session files after 30 days by default (cleanupPeriodDays in ~/.claude/settings.json). Argus keeps every session it has ingested regardless.
OpenAI Codex
| Details | |
|---|---|
| Reads from | $CODEX_HOME if set, otherwise ~/.codex/, and only the session rollouts under sessions/ and archived_sessions/ |
| Never opens | auth.json (cached credentials), config.toml, the SQLite state databases, memories, logs, plugins, skills and history.jsonl |
| One turn | One model response (Codex's per-response usage record) |
| Tool calls | Completed actions: shell commands, apply_patch file changes, MCP calls as mcp__<server>__<tool>, web.search, view_image, plus model calls such as spawn_agent and send_message |
| Sub-agents | Each spawned agent's rollout, kept as a sub-session under the thread that spawned it |
| Resume a session | codex resume <id> |
A few things that behave differently from Claude Code:
CODEX_HOMEset to a folder that doesn't exist means Codex is treated as absent. Argus does not fall back to~/.codex/.- A tool call is an error only on an explicit failure signal: a non-zero exit code or failed status for
shell, a failed file change, or an MCP result marked as an error. Codex's internalexecandwaitwrapper calls are not counted; the actions they ran are. - One Codex thread can span several files. When you edit and resend a turn, Codex continues the thread in a new rollout file; Argus merges them into one session. A rollout moved to
archived_sessions/is recognised as the same session, so nothing is counted twice. - Sub-agents at any depth are listed under the top-level thread. Codex copies the parent's history into each child's file; Argus skips that copied part, so the parent's turns are not counted twice.
- Older Codex versions that don't write per-response usage records are still counted, from their
token_countevents. - Codex reasoning is never stored. It is encrypted in the rollout anyway.
- WSL is a separate home. Codex inside WSL uses the Linux
~/.codex, not the Windows one. Argus reads the home of the system it runs on.
Costs
Codex sessions are priced with OpenAI's published standard-tier API rates. That is an API-equivalent estimate: Codex usage on a ChatGPT plan is billed by subscription, not per token.
Known gaps
gpt-5-codexis not on OpenAI's current price list, so its turns show tokens but $0 cost.- The higher long-context rate (over 272K input tokens in one response) isn't applied.
- Codex's own prompt history file (
history.jsonl) isn't imported. Your prompts are still captured from the sessions themselves and are searchable through Transcripts once indexing is on. - A sub-agent's progress appears when its parent thread next writes to its rollout.
Privacy
Both agents are read-only sources: Argus never writes to ~/.claude/ or the Codex home, and every read is checked to stay inside the agent's own folders, so a symlink can't redirect it elsewhere. Transcript text is only stored if you turn on indexing. Nothing leaves your machine.
Verified against argus-code 0.7.0.