Skip to content

PLATFORM_SPEC.md — Mistral Vibe CLI (mistral-vibe)

New to gald3r on this platform? Start at Use gald3r with Mistral Vibe CLI for a short overview before the full breakdown below.

The "Mistral" coding surface is three distinct products, and gald3r integration targets the config-driven one — the Mistral Vibe CLI (mistral-vibe), an open-source terminal coding agent:

Surface What it is Config-driven?
Mistral Vibe CLI Open-source terminal coding agent (mistral-vibe); .vibe/ tree + config.toml ✅ Yes — the only gald3r-relevant surface
Mistral Code Closed in-IDE plugin (JetBrains + VSCode, enterprise bundle) ❌ No project config files
Le Chat Web/app chat with MCP connectors + memories ❌ Not a file-config surface

gald3r targets Mistral Vibe CLI — it is the only Mistral surface that reads project-level config files (.vibe/ + AGENTS.md). Mistral Code (IDE plugin) and Le Chat are documented here for honesty but are not gald3r config targets.

Authoring path: UPDATE. Verified 2026-06-02 against https://docs.mistral.ai/mistral-vibe/terminal + github.com/mistralai/mistral-vibe (see Verification Evidence). Latest version observed 2.13.0 (2026-05-29). This upgrades the prior spec: Agents are now NATIVE (custom agents + subagents), and Hooks moved from ❓ to ⚠️ (an experimental post-agent-turn lifecycle shipped v2.9.0, Apr 2026).

Instruction-file convention (platform-specific truth): Mistral Vibe reads AGENTS.md, NOT CLAUDE.md. There is no .mistral/ directory and no YAML config — config is TOML under .vibe/. Any prior body describing a .mistral/config.yaml scheme was fabricated; corrected here.

Docs-URL migration note: doc paths are migrating from /mistral-vibe/* to /vibe/code/cli/*; both resolve as of the 2026-06-02 scan.

1. Folder Hierarchy

Mistral Vibe reads config from a .vibe/ directory (NOT .mistral/), at both user and project scope. Project config layers over user config and only loads when the working directory is a trusted folder. Verified layout (configuration + skills + agents docs):

~/.vibe/                       ← user-global config
├── config.toml                ← main config (TOML): models, providers, tools, [[mcp_servers]]
├── .env                       ← API keys / provider credentials
├── AGENTS.md                  ← user-level instructions (all projects)
├── agents/<name>.toml         ← custom agent / subagent profiles (TOML)
├── prompts/<id>.md            ← custom system prompts (referenced by system_prompt_id)
├── skills/<name>/SKILL.md     ← user-level skills (Agent Skills spec)
└── trusted_folders.toml       ← trusted-execution allowlist (gates project .vibe/ loading)

<project-root>/
├── AGENTS.md                  ← project instructions (overrides user-level; closer files win)
└── .vibe/
    ├── config.toml            ← project config overlay
    ├── skills/<name>/SKILL.md ← project skills (also ./.agents/skills/)
    ├── agents/<name>.toml     ← project agent / subagent profiles
    ├── prompts/<id>.md        ← project prompts
    └── (hooks)                ← experimental post-agent-turn lifecycle; file location/schema undocumented

gald3r writes (project scope): AGENTS.md, .vibe/skills/<name>/SKILL.md, optionally .vibe/agents/*.toml, and .vibe/config.toml [[mcp_servers]] entries. Vibe owns: the .vibe/ namespace, the config.toml (TOML) schema, and the trusted-folders mechanism (trusted_folders.toml decides whether project .vibe/ — and its skills/agents/prompts/ hooks — loads at all).

2. AI Instruction File

Vibe reads AGENTS.md — the same cross-platform standard gald3r already uses as its primary instruction file (✅ direct parity). It does NOT read CLAUDE.md.

  • Resolution order: ~/.vibe/AGENTS.md (user-global) → project AGENTS.md files; files closer to the working directory override more distant ones. Instructions augment/override the default system prompt.
  • gald3r fit: gald3r already generates/merges AGENTS.md; Vibe consumes it directly with no transformation needed — the strongest parity point for this platform.
  • No .cursorrules-style legacy file; no .mistral-prefixed instruction file.
  • Source: https://github.com/mistralai/mistral-vibe

3. Agents Support — ✅ NATIVE

  • Built-in agents: default, plan, accept-edits, auto-approve, lean.
  • Custom agents: .toml files in ~/.vibe/agents/ or ./.vibe/agents/, selected via vibe --agent <name>. Format keys: display_name, description, safety, active_model, system_prompt_id, enabled_tools / disabled_tools, per-tool permission = ask|always.
  • Subagents: set agent_type = "subagent"; spawned by the model through the task tool, return text-only results to the parent, and cannot ask the user questions.
  • Format mismatch vs gald3r: Vibe agents are TOML behavior profiles; gald3r agents (g-agnt-*.md) are markdown — a real port requires markdown→TOML conversion (not done by the parity pipeline). safety (safe/neutral/destructive/yolo) is a visual hint only, not enforcement.
  • Source: https://docs.mistral.ai/vibe/code/cli/agents

4. Skills Support — ✅ NATIVE

  • Agent Skills (agentskills.io specification) — the same spec gald3r skills target, making skills portable across conforming agents. CHANGELOG confirms agentskills.io support added v1.3.0 (2025-12-23).
  • Discovery: ~/.vibe/skills/<name>/SKILL.md (user), ./.vibe/skills/<name>/SKILL.md (project), ./.agents/skills/, plus skill_paths in config.toml. Folder-per-skill SKILL.md ✅ matches gald3r exactly.
  • Frontmatter keys: name, description, license, compatibility, user-invocable, allowed-tools. Differs slightly from gald3r's subsystem_memberships / token_budget (a superset; Vibe ignores unknown keys but expects allowed-tools / user-invocable).
  • Management: enabled_skills / disabled_skills glob patterns in config.toml.
  • Source: https://docs.mistral.ai/vibe/code/cli/skills

5. Commands / Workflows — ⚠️ PARTIAL

  • No standalone command-file directory equivalent to .cursor/commands/*.md or .claude/commands/. Slash commands exist but only via the skills mechanism: set user-invocable: true in a skill's frontmatter and the skill name becomes available with /skill-name autocompletion in the CLI prompt.
  • gald3r gap: gald3r's command files have no native 1:1 landing zone; each would need wrapping as a skill (heavy) or describing in AGENTS.md (lossy — no executable invocation). A built-in /loop command (v2.9.5, 2026-05-06) exists for recurring prompt execution, but that is a shipped built-in, not a user-command-file surface.
  • Source: https://docs.mistral.ai/vibe/code/cli/skills

6. Hooks System — ⚠️ PARTIAL (experimental, minimal)

  • Experimental hooks shipped v2.9.0 (2026-04-28): "Experimental hooks system with post-agent-turn lifecycle." The configuration docs list hooks among what each .vibe/ path contributes to a session, but publish NO event taxonomy, file location, or schema. The README does not mention hooks and the dedicated /cli/hooks doc page 404s.
  • Only a single lifecycle point shipped (post-agent-turn), with no later CHANGELOG refinement through v2.13.0. Broader event hooks remain unshipped: issue #250 ("feat: Hooks" — before/after shell-command hooks, modeled on Claude Code) is CLOSED with no implementation; issue #531 ("feat: add BeforeTool hook") is still OPEN.
  • gald3r gap: gald3r's g-hk-*.py pre-tool / session-start / pre-commit hooks have no verified wiring target — the one shipped event (post-agent-turn) does not cover them and its config schema is undocumented. Tool permissions ([tools.bash] permission = "ask"|"always") and agent safety levels cover some guardrail ground hooks would otherwise serve, but they are not an event-hook system. Do NOT fabricate a hooks.json equivalent.
  • Source: https://github.com/mistralai/mistral-vibe/blob/main/CHANGELOG.md

7. Rules / Memory — ⚠️ PARTIAL

  • No scoped/glob rule auto-loader (no .cursor/rules/*.mdc equivalent). Persistent project instructions are injected via AGENTS.md resolution (closer files override) — a single layered instruction file, not per-file globbed or on-demand rules.
  • Custom system prompt: a file under ~/.vibe/prompts/<id>.md can replace the default system prompt via system_prompt_id in config.toml — another instruction-injection lever.
  • Custom compaction prompts (v2.11.1, 2026-05-27): user-definable context-compaction prompts.
  • Memory: Le Chat (the chat product) has a "Memories" feature, but that is the chat surface, not Vibe CLI project config — not a gald3r file target.
  • gald3r gap: gald3r's scoped g-rl-* rules (alwaysApply / globs / on-demand) collapse into a single AGENTS.md blob on Vibe — no per-file globs scoping, no per-rule on-demand loading.
  • Source: https://github.com/mistralai/mistral-vibe

8. MCP Support — ✅ NATIVE

  • MCP servers configured as [[mcp_servers]] entries in config.toml. Transports: http, streamable-http, stdio. Keys: name, transport, url (http), command/args (stdio), headers, api_key_env, env, startup_timeout_sec, tool_timeout_sec.
  • Tool namespacing/filtering: tools are namespaced {server_name}_{tool_name} and filtered via enabled_tools / disabled_tools (exact, glob, or regex with re: prefix).
  • Limitation: "The CLI does not yet support MCP servers that require OAuth authentication." v2.13.0 (2026-05-29) improved MCP HTTP transport.
  • Le Chat separately offers MCP connectors invoked via /Connector_Name — that is the chat product, not Vibe CLI file config.
  • Source: https://docs.mistral.ai/vibe/code/cli/mcp-servers
[[mcp_servers]]
name = "my_http_server"
transport = "http"            # http | streamable-http | stdio
url = "http://localhost:8000"
headers = { "Authorization" = "Bearer my_token" }
api_key_env = "MY_API_KEY_ENV_VAR"
startup_timeout_sec = 15
tool_timeout_sec = 120

9. Other Extensibility

  • Trusted folders (trusted_folders.toml) gate whether project .vibe/ config (and its skills/ agents/prompts/hooks) loads at all.
  • Remote agents (powered by Mistral Medium 3.5) + a Vibe Code Web surface — orchestration features, not file-config primitives.
  • Plugin support is requested (Discussion #497) but not yet a shipped extensibility surface.

Known Gaps vs. Cursor Reference

# Gap Severity
1 Three products, one config-driven — only Vibe CLI reads config files. Mistral Code (IDE plugin) and Le Chat are not gald3r config targets. High (scope)
2 Hooks experimental & minimal — only a single post-agent-turn lifecycle point (v2.9.0), no published schema, no before-tool/session-start/pre-commit events; gald3r g-hk-*.py have no verified wiring target. (§6) High
3 No command-file directory — slash commands are skill-provided only; gald3r's command files have no native 1:1 landing zone. (§5) High
4 No scoped/glob rules — persistent instructions collapse into a single layered AGENTS.md (+ optional custom system prompt); no per-file globs or on-demand loading. (§7) Medium
5 Agent format mismatch — Vibe agents are TOML behavior profiles; gald3r agents are markdown. No automatic conversion. safety is visual-only, not enforcement. (§3) Medium
6 Skill frontmatter mismatch — folder-per-skill ✅ matches, but Vibe expects allowed-tools / user-invocable; gald3r skills emit subsystem_memberships / token_budget. Light adaptation needed. (§4) Low
7 Config is TOML, not YAML — the real surface is .vibe/config.toml; there is no .mistral/ and no config.yaml. (corrected)
8 MCP OAuth unsupported — the CLI does not yet support MCP servers requiring OAuth. (§8) Low
9 Parity pipeline does NOT target .vibe/ — neither gald3r sync nor gald3r platform install has a .vibe/ writer (and the historical platform_parity_sync.ps1 never existed on disk, C-NO-PS1). Any real install support is future work. Medium

Strongest parity points (not gaps): AGENTS.md is consumed directly (✅), and the Agent Skills SKILL.md folder-per-skill convention is shared (✅). Agents/subagents are now native (✅) but require markdown→TOML conversion.

Hook System

  • Type: experimental (single post-agent-turn lifecycle; file location/schema undocumented) ⚠️
  • Config file: .vibe/ (each trusted path contributes its AGENTS.md and .vibe/ config — tools, skills, agents, prompts, hooks — to the session); exact hook file/location NOT published as of the 2026-06-02 scan
  • Events available: post-agent-turn only (shipped v2.9.0, 2026-04-28). No before-tool / session-start / pre-commit events. [STUB — payload/schema undocumented]
  • Event payload format: [STUB] — undocumented
  • Limitations: tool permissions ([tools.bash] permission = "ask"|"always") and agent safety levels cover some hook-like ground but are NOT an event-hook system; broader hooks unshipped (issue #250 CLOSED unimplemented; issue #531 OPEN)
  • gald3r hook files: none verified — gald3r g-hk-*.py (pre-tool / session-start / pre-commit) have no confirmed target on Vibe; the one shipped event does not cover them [STUB]

Atypical Handling

  • Config tree is .vibe/ with config.toml (TOML, not YAML); AGENTS.md is the instruction file (NOT CLAUDE.md); Agent-Skills via SKILL.md.
  • Trusted-folders gate: project .vibe/ config (skills/agents/prompts/hooks) loads only when the working directory is trusted (trusted_folders.toml).
  • Hooks are experimental — a single post-agent-turn event exists, but the file location and schema are unpublished; existence without a usable contract for gald3r's hook needs.
  • Tool permissions + agent safety levels are the only verified gating mechanism for guardrails.

gald3r Integration Notes

  • Ship gald3r's AGENTS.md — Vibe consumes it directly (no transformation). This is the cheapest high-value install footprint today.
  • gald3r skills (g-skl-*/SKILL.md) load natively (agentskills.io); frontmatter needs light adaptation (allowed-tools / user-invocable).
  • gald3r hooks have no verified wiring target — the only shipped event (post-agent-turn) does not cover pre-tool / session-start / pre-commit, and the schema is undocumented. Do NOT assume .vibe/hooks/ works.
  • Re-crawl Mistral Vibe docs on the next @g-platform-scan-docs mistral specifically for a hook event list + file location; until then Hooks stays ⚠️ (crawl_max_age_days: 14).

Capability Summary

Hooks Rules Skills Commands MCP Docs Fresh
⚠️ ⚠️ ✅ ⚠️ ✅ ✅

Legend: ✅ verified working · ⚠️ partial / format-mismatch · ❌ not supported · ❓ untested.

  • Hooks ⚠️: experimental post-agent-turn lifecycle only (v2.9.0); no schema, no pre-tool/ session-start/pre-commit — cannot wire gald3r hooks.
  • Rules ⚠️: AGENTS.md injection (+ optional custom system prompt); no scoped .mdc rule system.
  • Skills ✅: native Agent Skills spec (agentskills.io), folder-per-skill SKILL.md (frontmatter needs light adaptation).
  • Commands ⚠️: slash commands exist but only via the skill mechanism (user-invocable: true); no command-file directory.
  • MCP ✅: .vibe/config.toml [[mcp_servers]] (http / streamable-http / stdio); no OAuth yet.
  • Agents ✅ (not in 5-col summary): native custom agents + subagents (.vibe/agents/*.toml).
  • Docs Fresh ✅: doc scan performed 2026-06-02 (docs.mistral.ai + GitHub README + CHANGELOG).

Verification Evidence (docs crawl 2026-06-02, docs.mistral.ai + github.com/mistralai/mistral-vibe)

Capability How verified
.vibe/ + config.toml (not .mistral//YAML) /vibe/code/cli/configuration — .vibe/ (user ~/.vibe/, project ./.vibe/); project layers over user; loads only in trusted folders
AGENTS.md instruction file + resolution (NOT CLAUDE.md) GitHub README — AGENTS.md files; closer-to-cwd overrides; ~/.vibe/AGENTS.md user-global; augments/overrides system prompt
Agents/subagents (TOML profiles) /vibe/code/cli/agents — built-ins default/plan/accept-edits/auto-approve/lean; custom .toml; vibe --agent; agent_type="subagent" via task tool, text-only
Skills (Agent Skills, SKILL.md) /vibe/code/cli/skills — agentskills.io spec; ~/.vibe/skills, ./.vibe/skills, ./.agents/skills, skill_paths; CHANGELOG v1.3.0 (2025-12-23)
Slash commands = skill-provided /vibe/code/cli/skills — user-invocable: true → /skill-name autocompletion; no standalone command-file directory
Hooks experimental (post-agent-turn) CHANGELOG v2.9.0 (2026-04-28) "experimental hooks, post-agent-turn lifecycle"; no schema; /cli/hooks 404s; issue #250 CLOSED, #531 OPEN
MCP [[mcp_servers]] (TOML) /vibe/code/cli/mcp-servers — http/streamable-http/stdio; {server}_{tool} namespacing; enabled_tools/disabled_tools; no OAuth; v2.13.0 HTTP improvements
Le Chat MCP connectors (separate surface) Le Chat connectors invoked via /Connector_Name — chat product, not Vibe CLI file config
Mistral Code = closed IDE plugin JetBrains/VSCode enterprise plugin; no project config files
No .vibe/ writer in parity pipeline Verified absence — gald3r sync / gald3r platform install target .cursor//.claude/ family, not .vibe/

Reference page — canonical source: g-skl-platform-mistral/PLATFORM_SPEC.md in the engine repo, generated by scripts/generate_platform_support_docs.py. Pages marked curated are hand-assessed and not yet live-verified.