Consults — one-off second opinions¶

Any Leo process can ask another model for a one-off opinion with the leo_consult MCP tool. Leo runs the selected template's harness/model as a headless one-shot process in the caller's workspace and returns its final answer directly as the tool result.

Usage¶

leo_consult(template: "codex", prompt: "Review this design: …")
template names a template, not a running agent — "consult fable" means the fable template, not an agent called fable. Asking an agent to consult a model is leo_consult; handing work to a running agent is leo_send_message. An unknown template comes back with the valid names.
Optional model overrides the template's model and is validated by that template's harness.
The call waits for the consultant to finish, like a normal subagent tool.
For a council, issue concurrent calls to several templates and reconcile the returned answers in the same turn.

Semantics and limits¶

Templates supply the harness, model, environment, and harness_options.
Claude runs headlessly through its print/task mode, Codex through codex exec --json, and OpenCode through opencode run --format json.
The consultant is advisory: a preamble instructs it to analyze without modifying files. This is not enforced; configured template permissions still apply.
One-shot only: no session is retained and no follow-up conversation exists.
Calls time out after 30 minutes. At most four run concurrently; additional calls wait and remain cancellable by their callers. Leo owns that deadline: every coding agent it launches is handed a longer per-tool MCP ceiling (32 minutes) through its harness's own knob, so a long consult is never cut short by the caller's harness — see Harnesses.
Supervised agents contribute their workspace. Other Leo callers are also supported and run from the daemon's working directory when no workspace can be resolved.

Watching a consult¶

Consults are started by agents, not by you, and a long one is otherwise a silent half-hour gap. Every consult records what it does, so you can watch it work:

$ leo consult list
ID          CALLER   TEMPLATE  MODEL          ELAPSED  STATUS
c-7f3a2b1e  leo      codex     gpt-5.3-codex  1:42     running
c-3c9d10a4  olympus  local     qwen3.6-35b    4:11     done

$ leo consult watch
[consult c-7f3a2b1e · codex/gpt-5.3-codex · from leo · running]
Review this design: …

   0:01  read     internal/consult/consult.go
   0:04  grep     CombinedOutput
   0:09  text     The dispatcher discards everything but the final text.
   0:11  bash     go test ./internal/consult/
[done after 1:58]

leo consult watch with no argument picks the newest running consult, falling back to the most recent one. An id may be abbreviated to any unique prefix.
Ctrl-C detaches. The consult keeps running; there is no way to cancel one from the CLI.
Consults that are queued behind the concurrency limit appear in list before they start, so a stalled call is distinguishable from a busy one.
Both commands accept --host and read the records on that host.

The feed shows the consultant's own text, the tools it invokes, and its final answer. Successful tool results are omitted — the call already says what it did, and the bodies would bury everything else — but failures are shown. Add --json to list for the raw records.

Where recordings live¶

Under <state>/consults/, two files per consult, both 0600 in a 0700 directory:

<id>.json — caller, template, harness, model, workspace, prompt, status, and timing. Status runs queued → running → done | failed | timeout | canceled.
<id>.ndjson — one line per harness event, {"t": <seconds since start>, "d": <the harness event verbatim>}. Output that was not JSON — a crashing harness, stderr chatter — is captured as {"t": …, "raw": "…"} instead of being lost. jq .d recovers the untouched harness stream.

The 20 most recent consults are kept; consults still in flight are never pruned. Recordings contain whatever the consultant read, the same trust boundary as task logs under <state>/logs.