zeph 0.18.3

Lightweight AI agent with hybrid inference, skills-first architecture, and multi-channel I/O
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57

# Observability & Cost Tracking

## OpenTelemetry Export

Zeph can export traces via OpenTelemetry (OTLP/gRPC). Feature-gated behind `otel`.

```bash
cargo build --release --features otel
```

### Configuration

```toml
[observability]
exporter = "otlp"                        # "none" (default) or "otlp"
endpoint = "http://localhost:4317"       # OTLP gRPC endpoint
```

### Spans

| Span | Attributes |
|------|------------|
| `llm_call` | `model` |
| `tool_exec` | `tool_name` |

Traces flush gracefully on shutdown. Point `endpoint` at any OTLP-compatible collector (Jaeger, Grafana Tempo, etc.).

## Cost Tracking

Per-model cost tracking with daily budget enforcement.

### Configuration

```toml
[cost]
enabled = true
max_daily_cents = 500   # Daily spending limit in cents (USD)
```

### Built-in Pricing

| Model | Input (per 1M tokens) | Output (per 1M tokens) |
|-------|----------------------|------------------------|
| Claude Sonnet | $3.00 | $15.00 |
| Claude Opus | $15.00 | $75.00 |
| GPT-4o | $2.50 | $10.00 |
| GPT-4o mini | $0.15 | $0.60 |
| GPT-5 mini | $0.25 | $2.00 |
| Ollama (local) | Free | Free |

Budget resets at UTC midnight. When `max_daily_cents` is reached, LLM calls are blocked until the next reset.

Current spend is exposed as `cost_spent_cents` in `MetricsSnapshot` and visible in the TUI dashboard.

### Token Counting

Completion token counts use the `output_tokens` field from the API response (OpenAI, Ollama, and Compatible providers). Streaming paths retain a byte-length heuristic (`response.len() / 4`) as a fallback when the provider returns no usage data. Structured-output calls (`chat_typed`) also record usage so `eval_budget_tokens` enforcement reflects real token counts.