Reference
Harness support
The agents Wisdom records, the versions tested, and what each one reports.
A harness is the coding agent Wisdom records. Each has its own adapter; none is central, and all of them read the same way.
| Claude Code | Codex | Pi | |
|---|---|---|---|
| Last tested | 2.1.319 | 0.166.0 (0.164.0 minimum) | 1.2.0 |
| Reports through | Hooks | Hooks, once trusted | Extension |
| Connected settings | .claude/settings.local.json |
.codex/hooks.json |
.pi/settings.json |
| Connected: start in | The project’s top folder | Any folder in the project, not a git worktree | The project’s top folder |
| Tracked home variable | CLAUDE_CONFIG_DIR |
CODEX_HOME |
PI_CODING_AGENT_DIR |
| Tracked first run | wisdom track claude, log in |
wisdom track codex --setup-hooks, trust, log in |
wisdom track pi, /login |
| Thinking | When recorded, else redacted |
Usually encrypted, redacted |
When recorded |
| Cost | Total per session | Not reported; the dashboard estimates it | Per message |
Optional MCP ledger tool |
Through wisdom track |
Through wisdom track |
Not available |
wisdom doctor says when an agent is newer than the last tested version. A newer version
usually works, but an agent can change its transcript format.
Claude Code
Section titled “Claude Code”- Its hooks are the main evidence; its transcript format is internal, so the copy backs them up.
- A background
/forkis recorded as its own session and can outlive its terminal. - A finished background command wakes the agent; that turn shows as
in (harness notice).
- Codex runs a hook only after you review it in
/hooks. Until then, sessions are read from their own files, markednot-recorded-live. - Its thinking is usually encrypted, and it reports tokens but not cost.
- Calls marked
script stepare Codex’s own scripting (exec,wait); theBashorapply_patchcalls beside them are the real actions.
- Recorded through an extension. Inside Pi,
/wisdom-statusreports the session. - In a tracked home, Wisdom chooses where Pi stores sessions, so
--session-diris refused.
Commit guides
Section titled “Commit guides”Reading a commit works the same for all three. The model that
writes guides runs through Claude Code, whichever agent did the work. Without Claude Code
installed and logged in, use the plain guide (wisdom guide off), or name another writer in
guides.command.
Finding the agent
Section titled “Finding the agent”Wisdom finds claude, codex and pi on your PATH, as your shell does. On Windows it
follows npm’s .cmd shims to the real program. --claude-bin, --codex-bin and --pi-entry
name another installation.