Skip to main content

Documentation Index

Fetch the complete documentation index at: https://mintlify.com/swt-labs/vibe-better-with-claude-code-vbw/llms.txt

Use this file to discover all available pages before exploring further.

VBW spawns specialized agents for planning, development, and verification. Model profiles let you control which Claude model each agent uses, trading cost for quality based on your needs.

​
Model profiles

The model_profile setting selects a preset that assigns a Claude model to every agent. 
/vbw:config model_profile quality
/vbw:config model_profile balanced
/vbw:config model_profile budget

​
Built-in profiles

ProfileLeadDevQAScoutDebuggerArchitectEst. cost/phase
qualityopusopussonnetsonnetopusopus~$3.00
balancedsonnetsonnetsonnetsonnetsonnetsonnet~$1.50
budgetsonnetsonnethaikuhaikusonnetsonnet~$0.70
quality is the default. It gives maximum reasoning depth for critical work. QA and Scout stay on Sonnet because verification and research don’t need Opus overhead. budget keeps Dev and core agents on Sonnet as a quality baseline, but drops QA and Scout to Haiku. Good for exploratory work where you’re iterating fast.
Switch to balanced for roughly 50% cost savings on standard development without giving up meaningful quality. Switch to budget for the lowest cost on prototypes and spikes.

​
Effort vs. model profile

These two settings are independent:
  • Model profile controls which Claude model agents use (cost).
  • Effort controls how deeply agents plan and verify (workflow depth).
You can run thorough effort on the budget profile (deep workflow, cheaper models) or fast effort on the quality profile (quick workflow, expensive models). Most users match them naturally:
EffortModel profile
thoroughquality
balancedbalanced
fastbudget
Switch both at once with /vbw:profile production (thorough + quality) or /vbw:profile prototype (fast + budget).

​
Per-agent overrides

Override a single agent’s model without changing the whole profile: 
/vbw:config model_override dev opus
/vbw:config model_override qa sonnet
Or set model_overrides directly in .vbw-planning/config.json: 
{
  "model_overrides": {
    "dev": "opus",
    "qa": "haiku"
  }
}
Overrides take precedence over the active profile. Common patterns:
  • budget profile + Dev override to opus for complex implementation tasks
  • balanced profile + Lead override to opus for strategic planning phases
  • quality profile + QA override to haiku when verification is straightforward

​
Agent turn limits

Each agent has a base turn budget that scales with your effort level (thorough = 1.5×, balanced = 1×, fast = 0.8×, turbo = 0.6×).
AgentBase turns
Scout15
QA25
Architect30
Lead50
Dev75
Debugger80
Override specific agents via agent_max_turns: 
/vbw:config agent_max_turns '{"dev": 100, "qa": 50}'
Set a value to false or 0 to remove the cap for that agent entirely: 
{
  "agent_max_turns": {
    "dev": false,
    "debugger": 0
  }
}
You can also provide per-effort overrides using an object instead of a number: 
{
  "agent_max_turns": {
    "dev": { "thorough": 120, "balanced": 75, "fast": 50, "turbo": false }
  }
}

​
Runtime feature flags

These flags control optional runtime subsystems. All default to true. Set any flag to false to skip that subsystem entirely.
When true, enforces per-role character budgets on context passed to agents. The control plane truncates compiled context to the role’s max_chars limit before injection, preventing context window overflows. When false, context passes through untruncated.
When true, the execute protocol skips unnecessary agents based on effort level: Scout is skipped for turbo and fast (no research needed), Architect is skipped for non-thorough effort. Reduces token spend on simpler phases. When false, all agents are always included.
When true, after each task commit the Dev agent runs a two-phase verification: the artifact registry tracks all files written, then confirms the task’s contract was fulfilled before marking it complete. Rejected tasks trigger auto-repair. When false, tasks complete immediately after commit.
When true, VBW appends JSON events to .vbw-planning/.metrics/run-metrics.jsonl for cache hits/misses, context compilation, task/plan/phase execution timing, and gate policy decisions. Viewable with /vbw:status --metrics. When false, no metrics are collected.
When true, the execute protocol runs per-plan risk assessment and resolves a dynamic gate policy that overrides static effort-based tables for QA tier, plan approval, and teammate communication level. When false, static effort-based tables are used.
When true, VBW saves execution state snapshots to .vbw-planning/.snapshots/ at key lifecycle points (phase start, compaction, agent completion). On crash recovery, /vbw:resume can restore from the latest snapshot. Max 10 snapshots per phase; oldest are pruned automatically. When false, no snapshots are saved.

Build docs developers (and LLMs) love

Get started for freeTalk to us