ADR-052: Statusline Observability System

Status: Implemented Date: 2026-02-10 Authors: RuvNet, Claude Flow Team Version: 1.0.0 Related: ADR-051 (Infinite Context), ADR-048 (Auto Memory Integration), ADR-006 (Unified Memory), ADR-026 (3-Tier Model Routing)

Context

The Problem: Invisible System State

Claude Code operates with multiple concurrent subsystems — swarm agents, memory backends, neural learning, security scanning, context management — but provides no unified visibility into their state. Developers working in long sessions need to know:

• How much context window remains before compaction triggers
• Whether the swarm has active agents or is idle
• If the intelligence system is learning from patterns
• What security posture the project maintains
• Whether archived memory is being utilized

Without real-time feedback, developers make blind decisions about when to start new sessions, whether agents are drifting, and if the system is operating optimally.

What Claude Code Provides

Claude Code supports a statusLine configuration in .claude/settings.json:

{
  "statusLine": {
    "type": "command",
    "command": "<shell command>",
    "refreshMs": 5000,
    "enabled": true
  }
}

The command receives JSON on stdin with workspace metadata, model info, and optional context window data. Its stdout is displayed as a persistent status bar in the Claude Code terminal UI.

What We Built

A multi-tier statusline system with 4 implementations, a TypeScript generator for npx claude-flow init, and real-time data feeds from 8+ subsystems.

Decision

Implement a layered statusline architecture:

1. Active statusline (.claude/statusline.sh) — Bash script for the current project, read from settings.json statusLine.command
2. Generated statusline (.claude/helpers/statusline.cjs) — CommonJS script created by npx claude-flow init, comprehensive with 12+ metric panels
3. Lightweight statusline (.claude/statusline.mjs) — ES module for agentic-flow integration, compact pipe-separated format
4. Command statusline (.claude/statusline-command.sh) — JSON-input focused, shows swarm topology and task metrics
5. Generator (v3/@claude-flow/cli/src/init/statusline-generator.ts) — TypeScript that produces the .cjs script during project initialization

Architecture

System Context

┌─────────────────────────────────────────────────────────────────────┐
│                        Claude Code Terminal                          │
│                                                                     │
│  ┌─────────────────────────────────────────────────────────────┐   │
│  │ statusLine.command executes every refreshMs (5000ms)         │   │
│  │                                                              │   │
│  │  stdin: { workspace, model, context_window }                │   │
│  │  stdout: ANSI-colored multi-line status display             │   │
│  └──────────────────────┬──────────────────────────────────────┘   │
│                          │                                          │
│                          ▼                                          │
│  ┌──────────────────────────────────────────────────────────────┐  │
│  │              .claude/statusline.sh (active)                   │  │
│  │                                                               │  │
│  │  Reads 8 data sources:                                       │  │
│  │                                                               │  │
│  │  ┌─────────────────┐  ┌──────────────────────┐              │  │
│  │  │ autopilot-state  │  │ v3-progress.json     │              │  │
│  │  │ .json            │  │ (DDD domains)        │              │  │
│  │  │ Context: 27%     │  │ Domains: 5/5         │              │  │
│  │  └─────────────────┘  └──────────────────────┘              │  │
│  │                                                               │  │
│  │  ┌─────────────────┐  ┌──────────────────────┐              │  │
│  │  │ learning.json    │  │ audit-status.json    │              │  │
│  │  │ Intel: 86%       │  │ CVE: 3/3 CLEAN      │              │  │
│  │  └─────────────────┘  └──────────────────────┘              │  │
│  │                                                               │  │
│  │  ┌─────────────────┐  ┌──────────────────────┐              │  │
│  │  │ patterns.db      │  │ transcript-archive   │              │  │
│  │  │ (SQLite)         │  │ .db (SQLite)         │              │  │
│  │  │ Quality: 0.999   │  │ Entries: 35          │              │  │
│  │  └─────────────────┘  └──────────────────────┘              │  │
│  │                                                               │  │
│  │  ┌─────────────────┐  ┌──────────────────────┐              │  │
│  │  │ swarm-activity   │  │ ps aux (process)     │              │  │
│  │  │ .json            │  │ Node/MCP detection   │              │  │
│  │  │ Agents: 3/15     │  │ Memory: 2782MB       │              │  │
│  │  └─────────────────┘  └──────────────────────┘              │  │
│  │                                                               │  │
│  │  Output (4 lines + separators):                              │  │
│  │  ▊ Claude Flow V3  ● user  │  ⎇ branch  │  Model            │  │
│  │  ─────────────────────────────────────────                   │  │
│  │  🏗️  DDD Domains  [●●●●●]  5/5    ⚡ 1.0x → 2.49x-7.47x   │  │
│  │  🤖 Swarm ◉ [3/15] 👥 0   🟢 CVE 3/3  💾 2782MB            │  │
│  │       🛡️  27% 54.7K ⊘    🧠  86%                           │  │
│  │  🔧 Architecture  DDD ●79%  │  Security ●CLEAN              │  │
│  │       Memory ●AgentDB  │  Integration ●                     │  │
│  │  ─────────────────────────────────────────                   │  │
│  └──────────────────────────────────────────────────────────────┘  │
└─────────────────────────────────────────────────────────────────────┘

Data Sources

Statusline Implementations

1. Active Statusline — .claude/statusline.sh (432 lines)

The currently wired script in settings.json. Bash-based for maximum compatibility.

Display Layout:

Line 0: ▊ Claude Flow V3 ● user  │  ⎇ branch  │  Model
Line -: ─────────────────────────────────────────
Line 1: 🏗️  DDD Domains  [●●●●●]  5/5    ⚡ speedup → target
Line 2: 🤖 Swarm ◉ [N/15] 👥 sub  🟢 CVE X/3  💾 MEM  🛡️ CTX%  🧠 INT%
Line 3: 🔧 Architecture  DDD ●N%  │  Security ●STATUS  │  Memory ●AgentDB  │  Integration ●
Line -: ─────────────────────────────────────────

Metric Icons:

Color Thresholds:

Context Autopilot Integration (ADR-051):

The statusline reads from autopilot-state.json which is written by the context-persistence-hook.mjs on every UserPromptSubmit hook:

{
  "sessionId": "f1bd5b59-...",
  "lastTokenEstimate": 54700,
  "lastPercentage": 0.274,
  "pruneCount": 0,
  "warningIssued": false,
  "history": [{ "ts": ..., "tokens": ..., "pct": ..., "turns": ... }]
}

When autopilot is active:

• Icon changes from 📂 (folder) to 🛡️ (shield)
• Token count shown (e.g., 54.7K)
• Prune indicator: ⊘ (no prunes needed) or ⟳N (N prune cycles)

Intelligence Score Computation:

Base:     learning.json → intelligence.score (0-100)
+ Boost:  patterns.db → AVG(quality) × 20 (up to +20)
+ Boost:  transcript-archive.db → COUNT(*) / 10 (up to +10)
= Final:  capped at 100

2. Generated Statusline — .claude/helpers/statusline.cjs (1,193 lines)

Created by npx claude-flow init. CommonJS for ES module project compatibility.

12 Metric Panels:

Output Modes:

• Default: Full ANSI-colored 4-line display
• --json: Pretty-printed JSON of all metrics
• --compact: Minified JSON

3. Lightweight Statusline — .claude/statusline.mjs (110 lines)

ES module for agentic-flow integration. Compact pipe-separated format with 5-second cache TTL for swarm status.

4. Command Statusline — .claude/statusline-command.sh (177 lines)

JSON-input focused. Shows swarm topology configuration, CPU/memory (with color-coded thresholds), session duration, task success rate with streak tracking, and hooks activity status.

Init System Integration

Generator — v3/@claude-flow/cli/src/init/statusline-generator.ts (1,317 lines)

Produces the .cjs script during npx claude-flow init:

function generateStatuslineScript(options: InitOptions): string {
  // Generates 1,193-line CommonJS script with all metric panels
  // Configurable via StatuslineConfig interface
}

function generateStatuslineHook(options: InitOptions): string {
  // Generates shell integration hook for bash/zsh/starship
}

Settings Generator — v3/@claude-flow/cli/src/init/settings-generator.ts

Wires the statusline into .claude/settings.json:

function generateStatusLineConfig(options: InitOptions): object {
  return {
    type: 'command',
    command: 'node .claude/helpers/statusline.cjs',
    refreshMs: config.refreshInterval,  // Default: 5000
    enabled: config.enabled,
  };
}

Executor — v3/@claude-flow/cli/src/init/executor.ts

During init:

1. Copies advanced statusline files (.sh, .mjs) from package source if available
2. Falls back to generating .cjs + hook via statusline-generator.ts
3. On upgrade: force-updates statusline while preserving metrics

Types — v3/@claude-flow/cli/src/init/types.ts

interface StatuslineConfig {
  enabled: boolean;           // Enable statusline
  showProgress: boolean;      // V3 DDD progress
  showSecurity: boolean;      // CVE status
  showSwarm: boolean;         // Swarm agent count
  showHooks: boolean;         // Hooks metrics
  showPerformance: boolean;   // Performance targets
  refreshInterval: number;    // Refresh ms (default: 5000)
}

Init Presets

File Inventory

Performance

Security Considerations

1. No secrets exposed: Statusline only reads metric files, never credentials
2. Read-only: All data sources are read-only file access or process inspection
3. No network: statusline.sh makes no network calls (gh API is cached from init)
4. Graceful failure: Every metric read is wrapped in error suppression (2>/dev/null)
5. No code execution: Statusline never evaluates dynamic code from data files

Consequences

Positive

1. Real-time visibility: Developers see context %, intelligence, swarm, and security at a glance without switching windows or running commands
2. Context Autopilot feedback: The 🛡️ 27% 54.7K ⊘ display confirms compaction is blocked and shows exact token usage — no surprises
3. Intelligence tracking: The 🧠 86% score motivates pattern storage and shows the system is learning from session interactions
4. Multi-implementation flexibility: Bash for current use, CJS for init portability, MJS for agentic-flow, each optimized for its context
5. Zero-dependency active script: statusline.sh uses only bash builtins, jq, and sqlite3 — no npm install required

Negative

1. 4 implementations: Maintenance burden of 4 separate statusline files. Mitigation: the generator produces the CJS version programmatically; the bash scripts are stable
2. Process detection heuristics: ps aux | grep is fragile for counting agents. Mitigation: swarm-activity.json from the monitor daemon is preferred when available
3. SQLite dependency: Intelligence score reads from patterns.db and transcript-archive.db via sqlite3 CLI. Mitigation: falls back to learning.json if sqlite3 is not installed

Neutral

1. 5-second refresh: Default refreshMs: 5000 balances freshness with CPU cost
2. ANSI colors: Terminal-dependent rendering. Claude Code's terminal handles ANSI well; raw SSH sessions may vary
3. JSON data coupling: Statusline depends on specific JSON key paths in metric files. Changes to metric formats require statusline updates

Future Enhancements

1. Unified statusline: Consolidate 4 implementations into a single TypeScript module that compiles to CJS, with bash as a thin wrapper
2. WebSocket feed: Replace file polling with event-driven updates from the MCP server for sub-second refresh
3. Customizable layout: Allow users to configure which panels appear and in what order via StatuslineConfig
4. Sparkline graphs: Show context growth trend as a sparkline character sequence (e.g., ▁▂▃▅▇ for rising usage)
5. Team statusline: When Agent Teams are active, show teammate status inline

References

• Claude Code statusLine config: @anthropic-ai/claude-agent-sdk settings schema
• ADR-051: Infinite Context via Compaction-to-Memory Bridge (autopilot-state.json)
• ADR-048: Auto Memory Integration (learning patterns)
• ADR-006: Unified Memory Service (AgentDB)
• ADR-026: 3-Tier Model Routing (intelligence scoring)