# Proxitor — lightweight proxy for routing CLI requests to OpenRouter
#
# Copy this file to proxitor.config.yaml and edit as needed.
# CLI flags override config file values.

# Server settings
host: '0.0.0.0'
port: 8828

# OpenRouter settings
openrouterKey: '' # Or set OPENROUTER_API_KEY env var
openrouterBaseUrl: 'https://openrouter.ai/api'
# Authentication type: "bearer" (default) or "oauth" for custom proxy providers
# authType: bearer
# Custom URL for fetching provider/model data (defaults to openrouterBaseUrl).
# Set this if your custom API doesn't support OpenRouter data endpoints.
# Fallback to https://openrouter.ai/api is automatic.
# openrouterDataUrl: 'https://openrouter.ai/api'

# Provider routing — controls which upstream provider handles requests
# provider:
#   # Allow only specific providers (string or array)
#   only: "deepinfra"
#   # only:
#   #   - "openai"
#   #   - "azure"
#
#   # OR: try providers in this order, optionally allow fallbacks
#   order: "anthropic"
#   allowFallbacks: true
#
#   # Ignore specific providers
#   ignore: "deepinfra"
#
#   # Sort providers by price, throughput, or latency
#   sort: "throughput"
#   # sort:
#   #   by: "price"
#   #   partition: "none"
#
#   # Filter by quantization level
#   quantizations:
#     - "fp8"
#
#   # Maximum pricing to accept
#   maxPrice:
#     prompt: 1       # $/M tokens
#     completion: 2   # $/M tokens
#
#   # Only use providers supporting all request parameters
#   requireParameters: true
#
#   # Data collection policy: "allow" or "deny"
#   dataCollection: "deny"
#
#   # Zero Data Retention enforcement
#   zdr: true
#
#   # Performance preferences (soft thresholds, not hard limits)
#   preferredMinThroughput:
#     p90: 50  # tokens/sec
#   preferredMaxLatency:
#     p90: 3   # seconds

# Custom headers to add to all proxied requests
# headers:
#   X-Custom-Header: "my-value"

# Per-model config overrides. Keys are exact model names or prefix patterns.
# Overrides layer on top of global settings (provider replaces, headers merge).
# modelOverrides:
#   "claude-sonnet-4-6":
#     provider:
#       only: "anthropic"
#   "claude-*":
#     provider:
#       order:
#         - "anthropic"
#         - "deepinfra"
#   "gpt-*":
#     provider:
#       only:
#         - "openai"
#         - "azure"
#     headers:
#       X-Model-Family: "gpt"
#   # Any field can also override its global counterpart per-model — e.g.
#   # normalizeVolatileSystem, cacheControl, cacheControlTtl, sessionId.
#   "qwen-*":
#     normalizeVolatileSystem: true

# Request body size limit
# bodyLimit: "50mb"

# OpenRouter attribution
# attributionReferer: "https://github.com/neiromaster/proxitor"
# attributionTitle: "proxitor"

# Prompt caching: inject cache_control for automatic caching.
# "auto" = Anthropic models only (safe default), "always" = all models,
# "skip" = passthrough (leave the client's cache_control untouched, inject nothing)
# cacheControl: auto

# Cache TTL: control how long cached prompts persist (Anthropic endpoints only).
# "5m" = 5 minutes (Anthropic default), "1h" = 1 hour (higher write cost),
# "omit" = strip the ttl field (guarantee no TTL, even client-sent),
# "skip" = passthrough (preserve client ttl, ignore inherited). Default: absent = passthrough.
# Only has effect when caching is active (cacheControl is "auto"/"always").
# cacheControlTtl: '1h'
# rewriteBlockTtl: skip   # auto | always | skip — normalize block cache_control ttl to cacheControlTtl

# Session sticky routing: inject session_id for provider stickiness from the first request.
# "auto" = passthrough client session ID, generate only if missing
# "always" = always generate proxy session ID (ignores client-provided)
# "skip" = passthrough (leave client session headers untouched)
# sessionId: auto

# Normalize Claude Code's volatile hashes in the system prompt to constants:
# the per-turn `cch=…` hash and the per-build `cc_version=<semver>.<hash>` hash
# (the readable semver is kept). Both drift every turn and invalidate the prefix
# cache for non-Anthropic providers (qwen/glm/etc.); normalizing them keeps the
# cached prefix stable.
# true = rewrite the hashes (default false), also settable per-model in modelOverrides.
# normalizeVolatileSystem: false

# Normalize Responses-api (/v1/responses) bodies so they satisfy OpenRouter's
# strict input schema: tag message items with type:"message", lift role:"system"
# into the top-level instructions field, and add id/status on assistant history.
# Some clients omit these and OpenRouter rejects the request with
# "400 invalid_prompt | Invalid Responses API request".
# "auto" = on for /v1/responses (default), "always" = all endpoints, "skip" = off.
# normalizeResponses: auto

# Enable verbose logging
verbose: false

# Cache observability: per-request cache-outcome logging and optional body dumps.
# Every proxied request logs an outcome line (HIT/PARTIAL/MISS/COLD/NOUSAGE with
# hit %) at info level. Set PROXITOR_DUMP_BODY=1 to additionally write paired
# request+response dumps (with provider routing metadata) to the dump directory.
# observability:
#   # Send the x-openrouter-metadata header to capture provider routing. Default true.
#   routerMetadata: true
#   # hitPct >= this counts as a full HIT (otherwise PARTIAL). Default 80.
#   hitThreshold: 80
#   # A request with no tools and max_tokens <= this is classified as a [side] call.
#   sideMaxTokens: 4096
#   # Bounded LRU of remembered sessions for warm/cold detection. Defaults: 4096 / 10m.
#   sessionMaxEntries: 4096
#   sessionTtlMs: 600000