multi-result-handling

name	multi-result-handling
description	Internal guidance for presenting multi:* helper output back to the user
user-invocable	false

Multi-CLI Result Handling

When a multi:* subagent returns helper output (Codex, Cursor, Antigravity, OpenCode via multi, or any CLI added via multi-cli-anything):

Preserving the helper's structure

• Preserve the helper's verdict, summary, findings, and next-steps sections exactly. Do not paraphrase, re-order, or summarize them in your own words.
• If the helper produced a structured final report (markdown headers like ## Outcome, ## Files touched, ## Verification, ## Notes), surface those sections as the canonical answer. Your commentary, if any, should reference the report — not restate it.
• Use the file paths and line numbers exactly as the helper reports them. Do not normalize Windows backslashes to forward slashes or vice versa.
• Preserve evidence boundaries. If the helper marked something as an inference, an uncertainty, or an open question, keep that distinction in your reply.

When the helper's output is messy

Some upstream CLIs stream chain-of-thought tokens interleaved with the final answer. When you see a mix of reasoning prose and a structured final report:

• Present the structured final report as the answer.
• Optionally quote 1–2 short lines of reasoning if they explain a non-obvious choice — never paraphrase the whole stream.
• If there is no structured final section, fall back to presenting the helper's last coherent paragraph as the result and note that explicitly.

Touched files and verification

• If the helper made edits, say so explicitly and list the touched files when the helper provides them.
• If the helper ran shell commands or other verification steps, surface those exit codes and outputs.
• If the helper says it ran something but no exit code or output is visible, flag that gap rather than assuming success.

Failed runs

• If the helper reports a failed run (exit non-zero, or the failure line <CLI> <role> failed: ...), report the failure with the most actionable stderr lines and stop. Do not turn a failed helper run into a Claude-side implementation attempt.
• If the helper was never successfully invoked (binary missing, auth failure, sandbox block), do not generate a substitute answer at all. Direct the user to /multi:setup and stop.
• If the helper reports malformed output, include the most actionable stderr lines and stop there instead of guessing.

Review-style output

For research, review, or diagnosis output from multi:antigravity-researcher, multi:codex-review, multi:antigravity-explorer, multi:cursor-research, multi:opencode-researcher, multi:opencode-explorer, etc.:

• Present findings first, ordered by severity if the helper provided severity labels.
• If there are no findings, say that explicitly and keep the residual-risk note brief.
• CRITICAL: After presenting review findings, STOP. Do not make any code changes. Do not fix any issues. Explicitly ask the user which issues, if any, they want fixed before touching a single file. Auto-applying fixes from a review is strictly forbidden, even if the fix is obvious.

Background jobs

• This applies ONLY to a detached --background job (explicit fire-and-forget). Its rendered output shows a job ID; you cannot be re-woken when that detached worker finishes, so do not promise to "check back later" — direct the user to /multi:status <job-id> for progress and /multi:result <job-id> for the final output, then stop.
• This is NOT the same as running a multi:* subagent as a harness background task (Agent run_in_background: true). That IS harness-tracked: the subagent runs the companion in the foreground and the harness re-wakes the main thread with a <task-notification> on completion or failure — that path is how long offloads should normally run, and it is fine to expect that notification.