Advanced Features

Multi-Agent Coordinator

Coordinator Mode enables orchestrating multiple agents working in parallel. The coordinator model doesn't execute work directly — it delegates to specialized workers and synthesizes the results.

Recommended workflow

1. Research (parallel)
   Agent(worker A) → explore authentication
   Agent(worker B) → explore database
   Agent(worker C) → explore tests
          │
          ▼
2. Synthesis (coordinator)
   Reads findings from A, B, C
   Writes implementation spec with paths, line numbers, details
          │
          ▼
3. Implementation (sequential per code area)
   Agent(worker D) → implement auth (src/auth/)
   Agent(worker E) → implement db (src/db/)
          │
          ▼
4. Verification (fresh agent, no bias)
   Agent(verifier) → confirms nothing was broken

How to use

The coordinator agent uses the Agent tool to spawn workers:

Spawn two workers in parallel to research the authentication bug:
- Worker A: analyze src/auth/ and identify where tokens are validated
- Worker B: check authentication tests in src/__tests__/auth/

Each worker receives a self-contained prompt (doesn't see the coordinator's conversation).

Resume workers

Use SendMessage to continue a worker with context:

SendMessage(
  to: "agent-a1b2c3",
  message: "Validation is in src/auth/validate.ts:42 — investigate the expired token case"
)

Worktree isolation

For parallel work on the same repository without conflicts:

Agent(
  isolation: "worktree",
  prompt: "Implement feature X on an isolated branch"
)

The worktree is created automatically and cleaned up if the agent makes no changes.

Context Compaction

Verboo Code automatically manages the context window to prevent prompt_too_long errors.

Compaction types

Full compaction — Summarizes the entire conversation in ~20K tokens:

/compact

Partial compaction — Summarizes only a part, preserving prompt cache:

• up_to: summarizes up to a point, preserves the end
• from: summarizes from a point, preserves the beginning

Microcompaction — Removes old tool result content (default: >36h), replacing with [Old tool result content cleared]. Saves tokens without losing significant context.

Auto-compact

Verboo Code monitors token usage and triggers automatic compaction before reaching the limit:

Context window limit
  │
  ├── Warning: "20K tokens remaining"
  │
  ├── Auto-compact threshold: compacts automatically
  │
  └── Hard limit: blocks new input

Circuit breaker: If compaction fails 3 consecutive times, auto-compact is disabled for the session to prevent infinite loops.

Post-compaction

After compacting, Verboo Code automatically reinjects:

• Recent files (up to 5, 50K token budget)
• Current plan file (if in plan mode)
• Background agent statuses
• Skills invoked in the session

Memory System

The memory system persists information between sessions in ~/.verboo/projects/<project>/memory/.

Memory types

File structure

~/.verboo/projects/<project-hash>/memory/
├── MEMORY.md          ← index (max 200 lines, always loaded)
├── user_role.md
├── feedback_testing.md
├── project_auth_rewrite.md
└── reference_linear.md

MEMORY.md is automatically loaded in every session. Topic files are loaded on demand.

Memory file format

---
name: feedback_terse_responses
description: User prefers short responses without end summary
type: feedback
---

Do not add a summary paragraph at the end of responses.

**Why:** User said "I can read the diff".
**How to apply:** Every response — end directly after the content.

Save a memory

The agent saves memories automatically when it learns something relevant. You can also request explicitly:

remember that I prefer semantic commits with scope (feat(auth): ...)

Voice Mode

Voice Mode enables voice interaction with the agent via STT (Speech-to-Text).

Requirements

• OAuth authentication with claude.ai (doesn't work with simple API key)
• Feature flag enabled on your account

Enable

/voice

Voice Mode is controlled by:

1. VOICE_MODE feature gate in the build
2. GrowthBook feature flag (tengu_amber_quartz_disabled as kill-switch)
3. Valid OAuth token presence

Buddy System

The Buddy is a deterministic visual companion generated from the userId.

Characteristics

• Deterministic: The same userId always generates the same companion
• Attributes: Species, eyes, hat, rarity (common → epic → legendary), shiny flag
• Stats: Distributed based on rarity
• Reactions: Responds to name mentions and the /buddy command

The Buddy doesn't affect agent functionality — it's purely a UI personalization element.

Feature Flags

Verboo Code uses two feature flag systems:

Build-time macros

Enabled/disabled in scripts/build.ts and eliminated from the bundle at compile time:

GrowthBook (runtime)

Dynamic feature flags updated without rebuild. Examples:

• tengu_compact_cache_prefix — shared prompt cache during compaction
• tengu_amber_quartz_disabled — voice mode kill-switch
• tengu_auto_background_agents — automatic agent backgrounding

Plan Mode

Plan Mode restricts the agent to read-only operations, enforcing a planning cycle before implementation.

Enter plan mode

EnterPlanMode

Or the agent enters automatically when receiving complex tasks.

In plan mode

• Only read-only tools (Read, Glob, Grep, WebSearch, read-only Agent)
• The agent can write only to the plan file (.verboo/plans/<name>.md)
• Ideal for exploring the codebase and aligning approach before any changes

Exit plan mode

ExitPlanMode

The user approves or rejects the plan. After approval, the agent executes with full context from the plan.

Advanced Diagnostics

Full verification

bun run hardening:strict

Runs: typecheck → build → smoke test → doctor runtime.

Privacy verification

bun run verify:privacy

Confirms the build makes no unauthorized calls to external servers.

PR security scan

bun run security:pr-scan --base origin/main

Analyzes branch changes for problematic security patterns.