Access Log Fields

All supported variables for logging.access_log_format.

ONR access logs support an nginx-like format string via logging.access_log_format (or logging.access_log_format_preset).

Formatting rules

• Variables are referenced as $name (example: $status).
• If a variable is missing or empty, ONR prints - at that position.
• $$ renders a literal $.

Configuration

In onr.yaml:

logging:
  access_log: true
  access_log_path: ""            # empty = stdout
  access_log_format: ""          # non-empty overrides preset
  access_log_format_preset: ""   # onr_combined | onr_minimal
  access_log_rotate:
    enabled: false
    max_size_mb: 100
    max_backups: 14
    max_age_days: 14
    compress: false

Rotation notes:

• Rotation only applies when access_log_path is non-empty (file output).
• When access_log_rotate.enabled=true, access_log_path is required.
• Rotation triggers on either day boundary (local time) or size threshold.
• Archive file pattern: <access_log_path>.YYYYMMDD-HHMMSS.NNNNNNNNN (or .gz when compress=true).

Rotation knobs:

• max_size_mb: rotate when the next write would exceed this file size.
• max_backups: keep only the newest N archives.
• max_age_days: remove archives older than N days; 0 disables age-based cleanup.
• compress: gzip archive files after rotation.

Daily-first setup example:

logging:
  access_log: true
  access_log_path: "./logs/access.log"
  access_log_rotate:
    enabled: true
    max_size_mb: 102400   # keep very large if you want day-boundary rotation to dominate
    max_backups: 14
    max_age_days: 14
    compress: false

Note:

• Current behavior is day-boundary OR size-threshold (not an exclusive mode switch).

Field reference

Request line fields

Variable	Meaning
$time_local	Timestamp formatted as YYYY/MM/DD - HH:MM:SS.
$status	Downstream HTTP status code returned to the client.
$latency	End-to-end request latency as a Go duration string (example: 123ms, 1.2s).
$latency_ms	End-to-end request latency in milliseconds. If proxy result latency is available, ONR may use it.
$client_ip	Client IP as seen by the server.
$method	Downstream HTTP method (example: GET, POST).
$path	Downstream request path (URL path only, without query string).

Identity and routing

Variable	Meaning
$request_id	Request ID used for tracing.
$appname	Client app name for observability. Resolved from appname header or (optional) inferred from User-Agent.
$provider	Selected provider name (from DSL selection and routing).
$provider_source	Provider selection source (for example: forced header vs models routing).
$api	Normalized API category (example: chat.completions).
$stream	Whether the request is treated as streaming (true/false) when available.
$model	Requested/selected model.

Usage and cost

Variable	Meaning
$usage_stage	Where usage data comes from (for example upstream-provided vs best-effort estimation).
$input_tokens	Input tokens.
$output_tokens	Output tokens.
$total_tokens	Total tokens.
$cache_read_tokens	Cache read tokens (when available).
$cache_write_tokens	Cache write tokens (when available).
$cost_total	Total cost.
$cost_input	Input cost.
$cost_output	Output cost.
$cost_cache_read	Cache read cost (when available).
$cost_cache_write	Cache write cost (when available).
$billable_input_tokens	Billable input tokens (after provider-specific adjustments).
$cost_multiplier	Effective multiplier applied to pricing (if configured).
$cost_model	Pricing model used for cost calculation.
$cost_channel	Channel identifier used for cost attribution (if available).
$cost_unit	Cost unit (implementation-defined; depends on pricing config).

Flat usage extras:

• When usage_extract custom; uses usage_fact ... with non-legacy detail, ONR also emits flattened usage fields into the access log.
• These keys are appended directly as normal log variables and follow a snake_case convention derived from dimension + attributes + unit.
• Legacy scalar usage stays unchanged: input_tokens, output_tokens, total_tokens, cache_read_tokens, cache_write_tokens.
• Typical examples:
  • $cache_write_ttl_5m_tokens
  • $cache_write_ttl_1h_tokens
  • $server_tool_web_search_calls
  • $image_generate_images
  • $image_edit_images
  • $audio_tts_seconds
  • $audio_stt_seconds
  • $audio_translate_seconds
• For Anthropic non-stream /v1/messages, cache_write token attr.ttl="5m" and attr.ttl="1h" can therefore appear both as:
  • $cache_write_tokens
  • $cache_write_ttl_5m_tokens
  • $cache_write_ttl_1h_tokens
• The same flattening applies to canonical image/audio/tool facts extracted by ONR or produced by API-specific OpenAI presets such as openai_responses, openai_images_generations, or openai_audio_speech, for example:
  • image.generate image
  • image.edit image
  • audio.tts second
  • audio.stt second
  • audio.translate second
• These flattened keys are only present when the corresponding usage_fact matched or was produced with quantity 0 from an upstream response that included the field.

Upstream and streaming performance

Variable	Meaning
$upstream_status	Upstream HTTP status code (when an upstream request was made).
$finish_reason	Normalized finish reason when extractable.
$ttft_ms	Time-to-first-token in milliseconds (streaming).
$tps	Tokens per second (streaming), best-effort.