# Features
> RTK token compression, OpenTelemetry observability, LSP code intelligence, and plugin hooks in workspace images.

## RTK Token Compression

[RTK](https://github.com/rtk-ai/rtk) intercepts Bash tool calls and compresses output from commands like `ls`, `find`, `git log`, and `tree`. This reduces the tokens Claude consumes reading command output.

### How It Works

RTK is initialized via `rtk init` in the container entrypoint. It creates shell aliases that wrap common commands with token-optimized formatting. Claude Code's Bash tool calls go through these aliases transparently.

### Measured Impact

Live A/B evaluation on an "explore repository" task:

| Metric | Without RTK | With RTK | Savings |
|--------|-------------|----------|---------|
| Context tokens | 61,117 | 28,663 | **53%** |
| Cost (USD) | $0.13 | $0.09 | **29%** |
| Conversation turns | 12 | 6 | **50%** |
| Bash tool calls | 10 | 5 | **50%** |

RTK reduces context consumption without changing agent behavior: the same tasks complete with fewer tokens.

## OpenTelemetry Observability

Workspace images export standard OTel metrics via OTLP, providing real-time token and cost tracking.

### Exported Metrics

| Metric | Type | Description |
|--------|------|-------------|
| `claude_code.token.usage` | Counter | Token count per turn (input, output, cacheRead, cacheCreation) |
| `claude_code.cost.usage` | Counter | Cumulative cost in USD |
| `claude_code.session.count` | Counter | Session counter |
| `claude_code.commit.count` | Counter | Git commit counter |

### Exported Log Events

| Event | Description |
|-------|-------------|
| `claude_code.api_request` | Per-API-call cost, model, cache tokens, duration, speed mode |
| `claude_code.api_error` | API error with status code and retry count |
| `claude_code.tool_result` | Tool execution completion (name, duration, success) |
| `claude_code.user_prompt` | User prompt submitted (length only, no content by default) |

### Configuration

OTel export is enabled by default in the image. The orchestrator auto-injects the collector endpoint when `COLLECTOR_URL` is configured:

```bash
# Set by the image (defaults, no user action needed)
CLAUDE_CODE_ENABLE_TELEMETRY=1
OTEL_METRICS_EXPORTER=otlp
OTEL_LOGS_EXPORTER=otlp
OTEL_METRIC_EXPORT_INTERVAL=5000  # 5s granularity

# Auto-injected by orchestrator when COLLECTOR_URL is set
OTEL_EXPORTER_OTLP_ENDPOINT=http://collector:8080
```

If no endpoint is provided, OTel export silently no-ops (no crash, no side effects).

### Two-Channel Observability

Workspace observability uses two complementary channels. They capture different data, neither can fully replace the other:

| Captured by Plugin Hooks only | Captured by OTel only | Captured by both |
|-------------------------------|----------------------|-----------------|
| Subagent lifecycle | Per-API-call cost + model | Token usage totals |
| Session end event | API errors (status code, retries) | Cost totals |
| Context compaction | Cache token breakdown | Tool completion |
| Git branch / merge / push | Active time tracking | User prompts |
| Tool start events | `prompt.id` correlation | |
| Permission requests | | |

See the [Observability guide](/docs/guide/observability) for the full event pipeline and how to verify both channels are flowing.

## LSP Code Intelligence

The image bundles language servers for Claude Code's LSP plugins:

| Language | Server | Plugin |
|----------|--------|--------|
| Python | pyright | pyright-lsp |
| TypeScript | typescript-language-server | typescript-lsp |
| Rust | rust-analyzer | rust-analyzer-lsp |

LSP servers are **lazy**: they only start when Claude encounters files in the matching language. An agent working on a Python-only repo will not start rust-analyzer or typescript-language-server.

## Pre-Bundled Plugins

Plugins are baked into `/opt/agentic/plugins/` and loaded at runtime via `--plugin-dir`. Available plugins:

- **sdlc**: Git operation hooks (commit metrics, branch tracking, push events)
- **workspace**: Compaction tracking, subagent lifecycle events

Plugin hooks emit JSONL events to stderr, captured by the orchestrator's event collection pipeline.