superaicore-dispatch

name

superaicore-dispatch

description

Delegate a coding task, code review, second opinion, or multi-model comparison to another locally-installed AI engine (Claude, Codex, Gemini, Copilot, Kimi, Qwen, Cursor, Grok, Kiro, SuperAgent) through the `superaicore` CLI. Use when the user asks for a second opinion / peer review from a different model, wants a task run on a specific other engine, or wants the same question compared across models.

superaicore-dispatch

Route a task to another local AI CLI via superaicore send, keep the conversation going with superaicore resume, and audit past dispatches with superaicore runs. You stay the deciding agent — external models provide input only.

Before dispatching

1. Read the user's preferences (scenario → model choices). If the user explicitly names a target or model, the user's choice wins.

   superaicore preferences show
   

2. Check the routing pool when unsure what a short name maps to:

   superaicore aliases            # full alias table
   superaicore aliases opus --json  # resolve one target exactly like send does
   

3. First-time setup problems? superaicore doctor shows which engines are installed/authenticated and whether aliases, preferences, and the run store are healthy.

Dispatch a task

superaicore send <target> "<task>" \
  --cwd "$PWD" --json-result --task-name <name>

• <target> is an alias (opus, sonnet, codex, gemini-pro, kimi, qwen, grok, …), a backend name (claude_cli), or a raw model id.
• Always pass --cwd "$PWD" for project work so the engine sees the repo.
• Review prompts must reference concrete sources: diffs, file paths, logs, or explicit line ranges — never "review the code".
• Use --prompt-file <path> for long prompts instead of a huge argv.
• Add --task-name for anything you may want to find again in runs list.
• Add --stream-progress to watch live engine output on stderr.

Interpret the result (--json-result)

Never assume the requested target answered — read the metadata:

field	meaning
ok / status	overall outcome (ok, failed, exhausted, no_candidates)
text	the model's answer
backend_used / model_used	who actually answered
requested_target / route_source	what you asked for and how it resolved
route_trace	every candidate tried, with per-step status + failure class
degraded / degrade_reason	true when a fallback candidate answered instead of the first choice
failure_class	on failure: quota, rate_limit, auth, network, tool_policy, validation, or null (runtime)
session_id	pass to resume for follow-ups
run_id	look it up later with superaicore runs show <id>

Quota/rate-limit/auth/network failures fall through to the next candidate automatically (degraded: true); runtime failures fail closed so you see the real error.

Follow-up questions

Resume only when the prior result carried a real session_id:

superaicore resume --session-id <id> "<only the new question>" \
  --json-result --task-name <name>-r2

Write only the delta — never re-paste conversation history. Chain further follow-ups off the latest result's session_id (engines may fork a fresh id per resume).

Audit

superaicore runs list             # newest dispatches first
superaicore runs show <run-id>    # full result contract + prompt excerpt