From minimax
Internal guidance for presenting MiniMax output back to the user
How this skill is triggered — by the user, by Claude, or both
Slash command
/minimax:minimax-result-handlingThe summary Claude sees in its skill listing — used to decide when to auto-load this skill
How Claude should render and reason about MiniMax output after receiving it from `minimax-companion.mjs`. Applies to all `/minimax:*` commands.
How Claude should render and reason about MiniMax output after receiving it from minimax-companion.mjs. Applies to all /minimax:* commands.
The companion has already:
mini-agent -t <prompt>This skill is about what to do with the final text/JSON payload that companion wrote to stdout (or the error on stderr).
/minimax:review (Phase 3) the structured JSON is the primary payload; prose is commentary. For /minimax:ask, the response string is the primary payload.Starting MiniMax (cold start ~3s)... line. Do not add your own "MiniMax is slow because Python..." commentary. The UX contract is: user sees progress, Claude stays quiet about the 3-5s delay.MiniMax M2 leans into its native Chinese expressiveness more than Kimi or Gemini do. If the user asked in Chinese, the response will almost certainly be Chinese. Quote verbatim. Do NOT offer translation as an unprompted follow-up. /minimax:ask specifically forbids appending any commentary. If the user later asks "翻译一下" or similar, translate then.
Mini-Agent exposes bash + file-write tools. /minimax:rescue --sandbox (Phase 4) isolates the workdir but is not a security boundary -- the agent can still cd /, use absolute paths, etc. Therefore:
toolCalls[] (via --json mode or /minimax:rescue), scan for suspicious patterns before rendering:
rm\s+-rf\s+/ (any variant)>\s*/dev/ (device writes)curl\s+.*\|\s*sh or wget\s+.*\|\s*sh (pipe-to-shell)sudo\s+ (privilege escalation)chmod\s+0?777 (permission widening):\(\)\{\s*:\|:&\s*\};: (fork bomb)rm -rf /tmp/foo" as if it were routine output.Per-command rendering rules live in references/<command>-render.md. Read the matching reference for the command you're rendering:
/minimax:ask -> references/ask-render.md/minimax:review -> references/review-render.md (Phase 3)/minimax:rescue -> references/rescue-render.md (Phase 4)/minimax:adversarial-review -> references/adversarial-review-render.md (Phase 5)Command files (plugins/minimax/commands/<name>.md) remain the authoritative source of truth -- the reference docs capture cross-command patterns that wouldn't fit in a command file's frontmatter-bounded budget. When a command file and a reference disagree, the command file wins.
The companion maps Mini-Agent outcomes to a fixed status set (spec §4.1). Here is how Claude should open the message for each:
| Status | User-facing opener |
|---|---|
success | (just present response verbatim) |
success-but-truncated | "Note: response was truncated by model length finish reason." + response |
incomplete | "The agent stopped with pending tool calls." + whatever partial response existed |
auth-not-configured | "MiniMax isn't configured. Run /minimax:setup." |
config-missing | "MiniMax config file is missing. Run /minimax:setup." |
needs-socksio | "MiniMax needs the socksio Python extra. Reinstall with --with socksio." |
not-installed | "Mini-Agent CLI is not installed. Run /minimax:setup." |
llm-call-failed | "The LLM call failed after MiniMax's own retries. Check the log path shown." |
unknown-crashed | "Mini-Agent crashed unexpectedly. Check the log path and rerun." |
success-claimed-but-no-log | "Session finished but produced no parseable response. Rerun with --json for a diagnostic dump." |
Add diagnostic bundle (stderr head+tail) only when the user has not already seen it in the companion stderr. In /minimax:ask flow the user already sees stderr; just summarize.
If companion returns toolCalls[] with an unfamiliar name, do not guess its meaning. Tell the user: "MiniMax returned a <name> tool_call that this plugin version does not render. Raw arguments: ..."
If the companion returns an error status (non-zero exit), show it directly with context. Do NOT re-run automatically. Use the status -> message map above to choose the right user-facing opener.
references/review-render.md -- severity-sorted findings, JSON schema validation story (Phase 3)references/rescue-render.md -- multi-step progress rendering + sandbox boundary reminder (Phase 4)redactSecrets at the companion layer before stdout writes. By the time Claude reads the stdout, secrets are already ***REDACTED***./minimax:ask is independent.Creates, edits, and optimizes skills for Claude Code, including drafting, evaluating with test prompts, iterating on performance, and improving skill descriptions for better triggering accuracy.
npx claudepluginhub bbingz/minimax-plugin-cc --plugin minimax