llm-cli-gateway: Skill for Claude Code

Provider: Mistral Vibe CLI

Process guidance for keeping the gateway's Mistral (vibe) integration aligned with the upstream CLI. This skill does not define argv/env behaviour; the single mechanical source of truth is UPSTREAM_CLI_CONTRACTS.mistral in src/upstream-contracts.ts, enforced by validateUpstreamCliArgs / validateUpstreamCliEnv. Never re-encode flags, agent modes, output formats, the env-model rule, or resume rules here or in TOML.

Identity

These values mirror UPSTREAM_CLI_CONTRACTS.mistral.upstreamMetadata and docs/upstream/provider-sources.dag.toml ([providers.mistral]). The TypeScript metadata is authoritative; the TOML is scanner input only.

When to use

• A Mistral Vibe release lands and you need to check for contract drift.
• A mistral request fails the upstream contract check before spawn.
• npm run upstream:scan -- --live reports a change on the Vibe releases page.

How LLM agents should use Mistral Vibe through the gateway

1. Discover the live gateway/provider surface before relying on Vibe-specific controls:

   provider_tool_capabilities({cli:"mistral"})
   

   For a cached read-only resource, use provider-tools://mistral.
2. Use mistral_request for normal turns and mistral_request_async for long-running review, analysis, or fifth-perspective checks. Sync calls may auto-defer; poll llm_job_status and fetch with llm_job_result when that happens.
3. Omit model unless the caller explicitly asked for a specific variant; the gateway selects the resolved model through VIBE_ACTIVE_MODEL.
4. The gateway emits --agent <mode> explicitly and defaults programmatic callers to accept-edits. Use approvalStrategy:"legacy" for Mistral; callers may use permissionMode:"plan" or another supported Vibe mode when they need a different behavior. Mistral rejects mcp_managed before launch, and approvalPolicy has no effect.
5. allowedTools maps to repeated --enabled-tools flags and disallowedTools maps to repeated --disabled-tools flags. Check provider_tool_capabilities before relying on either provider-native control.
6. mcpServers is descriptive metadata only; Vibe owns its MCP configuration.
7. Vibe supports output format, trust, working directory/additional directories, workspace/worktree, session, and max-turn/price/token controls as reported by provider_tool_capabilities. It does not support gateway effort / reasoningEffort.
8. Vibe continuity is real via sessionId, resumeLatest, and createNewSession. Continuation, trust:true, and a non-empty addDir retain the provider's native behavior under legacy. Current Vibe defaults session logging on; check doctor --json for an explicit [session_logging] enabled = false warning before assuming continuity transcripts are retained.
9. Bind concurrent repository work with workingDir or a verified workspace. Use gateway worktree only with an explicit provider-native sessionId that is not overridden by createNewSession; fresh, createNewSession, and resumeLatest-only worktree requests fail closed. An unscoped child uses a fresh neutral temporary cwd, and cwd-scoped resumeLatest fails closed without a stable target. Vibe's current prompt contract is argv-bound, so oversized UTF-8 input returns non-retryable input_too_large without truncation. All other caller-controlled argv values are admitted in their final encoded form before spawn. Embedded NUL bytes return non-retryable invalid_input without exposing the rejected value.