Skip to content

deessee0/workflows

v1.3.0MIT

State-of-the-art imitation of Claude Code's Dynamic Workflows for Mistral Vibe: dynamic orchestration with parallel subagents, independent verification, an adversarial pass, convergence loops, and resumable checkpoints.

GitHubOpen Website

vibe-dynamic-workflows

State-of-the-art imitation of Claude Code's Dynamic Workflows for Mistral Vibe — as a native Vibe plugin (Agent Plugins 1.0).

One orchestrator skill plans your task, fans it out across parallel subagents, re-verifies every finding independently, has adversaries try to break the result, iterates until the answers converge, and hands you a single coordinated report. Nothing unverified reaches you.

you -> /workflows:run audit the auth service for input-validation bugs
       |
       |- plan      6 workers - independent verification - adversarial pass
       |- confirm  [you approve the plan before it runs]
       |- run      -> wf-bug-hunt-01 -> 34 endpoints checked,
       |               5 confirmed issues, 2 refuted, 1 uncertain, report written

What it gives you

ComponentRole
/workflows:run <task>Full dynamic workflow: plan, fan-out, verify, attack, converge, report
/workflows:stress-test <plan>"A plan you want stress-tested from every angle before you commit to it": feasibility, correctness, security, cost, alternatives
workflows:workerSubagent executing one narrow mandate, replies in summaries with evidence
workflows:verifierRe-checks one finding independently from scratch: CONFIRMED / REFUTED / UNCERTAIN
workflows:adversaryTries to break the result: BREAKER / RISK / NIT objections
Hooks (ai.mistral.vibe/hooks.toml)Native enforcement: spawn budget cap, spawn audit log, shell guard against irreversible commands
Knowledge folderThe research this imitates, in the plugin, for any agent to consult

Behaviors faithful to the original feature:

  • Confirmation before the first fan-out — you see the plan (workers, mandates, verification approach) before it runs.
  • Verification before delivery — a finding enters the report only after an independent verifier confirms it.
  • Adversarial pass — dedicated agents try to refute what survived.
  • Convergence loop — holes get targeted re-waves, until convergence or a hard budget of 3 rounds.
  • Resumable checkpoints — all run state in the scratchpad (workflows/<run-id>/), so an interrupted run resumes instead of restarting.
  • Failed workers are dropped, not fatal.
  • Native enforcement — plugin hooks make the budgets real: a pre_tool hook denies workflow spawns past WORKFLOWS_SPAWN_CAP (default 40), a post_tool hook counts executed spawns and writes an audit trail to ${VIBE_HOME:-~/.vibe}/workflows-hooks/, and a shell guard blocks irreversible commands (rm -rf, git push, git reset --hard, git clean -f) in command position.

Install

git clone https://github.com/deessee0/vibe-dynamic-workflows.git \
  ~/.vibe/plugins/workflows

Or for one project only: clone into <project>/.vibe/plugins/workflows/ (trusted folder required). Then run /reload inside Vibe, and check /plugins — the plugin should appear with no diagnostics.

Vibe 3.0.x parses plugin hooks but only executes hooks declared in ~/.vibe/hooks.toml / .vibe/hooks.toml. If you cloned the repository (rather than an archive), run ./install.sh from it once: it copies the plugin without .git and merges the three enforcement hooks into ~/.vibe/hooks.toml under workflows:-prefixed names (idempotent; skip with WORKFLOWS_NO_HOOKS=1). User-scope hooks apply to every session — delete the workflows: blocks from that file to remove them.

Requires Vibe 2.x+ (plugins, agents.spawn, scratchpad).

Use

/workflows:run find every endpoint in src/auth that trusts client input
/workflows:run stress-audit: dead code across the whole service
/workflows:stress-test @docs/migration-plan.md

Or just ask: "run a workflow to " — the skills describe when the agent should start one on its own judgment (big, parallelizable, verifiable tasks only).

When NOT to use it

Small, sequential, or unverifiable tasks. A workflow multiplies token usage — every worker, verifier, and adversary is its own conversation. Start with a scoped task to get a feel for the cost, like the original docs recommend.

Not Mistral's "Workflows"

Mistral's own product pages called Workflows (docs.mistral.ai/studio/workflows, plus "Run Workflows" under Vibe Work docs) are a durable-orchestration platform in Studio: workflows in Python, activities for side effects, replay from history, pauses on human signals. This plugin shares only the name. It is Claude Code-style dynamic agent orchestration inside a Vibe session: conversational subagents, no platform, no durability beyond the scratchpad.

How it maps to Claude Code's Dynamic Workflows

Claude CodeThis plugin
"Create a workflow" / ultracode/workflows:run, /workflows:stress-test
Dynamic plan + parallel fan-outOrchestrator skill + tools.agents.spawn in waves of 4-6
Verification of every findingworkflows:verifier
Adversarial refutationworkflows:adversary
Iterate until convergenceConvergence loop, max 3 rounds
First-run confirmationPlan approval gate in Phase 1
Progress saved during the runScratchpad checkpoints per run id
Enforcement of budgets / safety railsPlugin hooks: spawn cap, spawn audit, shell guard

Full research notes and sources live in ai.mistral.vibe/knowledge/dynamic-workflows/KNOWLEDGE.md.

Honest limitations vs the original

  • Scale: waves of 4-6 workers, not tens to hundreds of agents.
  • Orchestration runs inside the main conversation, not in an external daemon; the scratchpad gives resumable state, not crash recovery.
  • No ultracode auto-trigger.
  • Plugin hooks are parsed but not executed by Vibe 3.0.x — install.sh works around this by merging them into ~/.vibe/hooks.toml (user scope, every session). Revisit when plugin-hook execution ships.

For the full external-daemon experience see open-dynamic-workflows (no Vibe adapter as of October 2026).

License

MIT