config.example.yaml

# Server host/interface to bind to. Default is empty ("") to bind all interfaces (IPv4 + IPv6).
# Use "127.0.0.1" or "localhost" to restrict access to local machine only.
host: ""

# Server port
port: 8317

# TLS settings for HTTPS. When enabled, the server listens with the provided certificate and key.
tls:
  enable: false
  cert: ""
  key: ""

# Management API settings
remote-management:
  # Whether to allow remote (non-localhost) management access.
  # When false, only localhost can access management endpoints (a key is still required).
  allow-remote: false

  # Management key. If a plaintext value is provided here, it will be hashed on startup.
  # All management requests (even from localhost) require this key.
  # Leave empty to disable the Management API entirely (404 for all /v0/management routes).
  secret-key: ""

  # Disable the bundled management control panel asset download and HTTP route when true.
  disable-control-panel: false

  # Disable automatic periodic background updates of the management panel from GitHub (default: false).
  # When enabled, the panel is only downloaded on first access if missing, and never auto-updated afterward.
  # disable-auto-update-panel: false

  # GitHub repository for the management control panel. Accepts a repository URL or releases API URL.
  panel-github-repository: "https://github.com/router-for-me/Cli-Proxy-API-Management-Center"

# Authentication directory (supports ~ for home directory)
auth-dir: "~/.cli-proxy-api"

# API keys for authentication
api-keys:
  - "your-api-key-1"
  - "your-api-key-2"
  - "your-api-key-3"

# Enable debug logging
debug: false

# Enable pprof HTTP debug server (host:port). Keep it bound to localhost for safety.
pprof:
  enable: false
  addr: "127.0.0.1:8316"

# Enable local network mDNS / DNS-SD service discovery (RFC 6762 / RFC 6763).
# When enabled, broadcasts _ai-gateway._tcp with API protocol subtypes on the local network.
# NOTE for Docker users: Link-local multicast (224.0.0.251) cannot cross default NAT bridge networks.
# To broadcast across the physical LAN, run the container with host networking (--net=host / network_mode: host).
discovery:
  enabled: false
  # service-name: ""                  # Optional name prefix. Advertised as CPA-<ShortID> or <name>-<ShortID> (e.g. office-8F3B)
  # service-type: "_ai-gateway._tcp"  # Service type (RFC 6335 compliant)
  # subtypes:                         # DNS-SD API protocol subtypes to advertise
  #   - "_chat-completions"
  #   - "_responses"
  #   - "_messages"
  #   - "_generate-content"
  #   - "_interactions"
  # interfaces:
  #   include: []                     # White-listed interfaces (empty: auto-detect physical interfaces)
  #   exclude: []                     # Additional interface patterns to exclude
  # auth-required: true               # Advertise that credentials are required

# Credential concurrency is configured by Home in Home mode. The synthesized Home config is
# authoritative and local values, including the values below, are ignored. Do not use local
# configuration to override a Home concurrency policy.
# credential-concurrency:
#   lifecycle-config-revision: 1
#   observation-barrier-revision: 0
#   cpa-heartbeat-timeout: "3s"
#   cpa-cancel-bound: "5s"
#   reclaim-grace: "5s"
#   cleanup-interval: "5s"
#   release-flush-interval: 250ms
#   release-max-backoff: 2s
#   busy-retry-min: 250ms
#   busy-retry-max: 1s
#   max-limit: 1000000

# Credential in-flight observation snapshot contract.
# credential-in-flight:
#   snapshot-interval: 2s
#   stale-after: 10s
#   max-part-bytes: 262144
#   max-part-count: 64
#   max-revision-bytes: 16777216
#   max-aggregate-groups: 100000
#   max-details: 10000
#   max-string-bytes: 256
#   staging-retention: 1m

# Standard dynamic library plugins are trusted in-process code. They are disabled by default.
# Build Go examples with go build -buildmode=c-shared for the target GOOS/GOARCH.
# Other languages can implement the same C ABI and JSON method protocol.
# Plugin executors require a matching auth record with the same provider key.
# If the same provider is configured as OpenAI-compatible, the native executor wins.
# Plugin command-line flags and Management API routes are optional capabilities.
# Existing native flags/routes and higher-priority plugin flags/routes cannot be replaced.
# Plugin list Management API reads Logo and ConfigFields from plugin metadata for management UI display.
# Per-plugin enabled only controls plugins.configs.<pluginID>.enabled and does not implicitly change global plugins.enabled.
plugins:
  enabled: false
  dir: "plugins"
  # Additional plugin store registries. The built-in official registry is always included.
  # store-sources:
  #   - "https://example.com/cliproxy-plugins/registry.json"
  # Optional plugin store auth rules. Values are read from environment variables;
  # tokens are not written into plugin manifests or node status.
  # store-auth:
  #   - match: "https://example.com/cliproxy-plugins/"
  #     apply-to: ["registry", "artifact"]
  #     type: bearer
  #     token-env: "CLIPROXY_PLUGIN_STORE_TOKEN"
  configs:
    example:
      enabled: true
      priority: 1
      config1: true
      config2: "string"
      config3: 3
      mode: "safe" # enum example: safe, fast

# When true, disable high-overhead request logging and HTTP middleware features to reduce per-request memory usage under high concurrency.
commercial-mode: false

# When true, write application logs to rotating files instead of stdout
logging-to-file: false

# Maximum total size (MB) of log files under the logs directory. When exceeded, the oldest log
# files are deleted until within the limit. Set to 0 to disable.
logs-max-total-size-mb: 0

# Maximum number of error log files retained when request logging is disabled.
# When exceeded, the oldest error log files are deleted. Default is 10. Set to 0 to disable cleanup.
error-logs-max-files: 10

# When false, disable in-memory usage statistics aggregation
usage-statistics-enabled: false

# How long (in seconds) usage queue items are retained in memory for the Management API.
# The local Redis RESP usage output is disabled.
# Default: 60. Max: 3600.
redis-usage-queue-retention-seconds: 60

# Proxy URL. Supports socks5/http/https protocols. Example: socks5://user:pass@192.168.1.1:1080/
# Per-entry proxy-url also supports "direct" or "none" to bypass both the global proxy-url and environment proxies explicitly.
proxy-url: ""

# When true, unprefixed model requests only use credentials without a prefix (except when prefix == model name).
force-model-prefix: false

# When true, forward filtered upstream response headers to downstream clients.
# Default is false (disabled).
passthrough-headers: false

# Number of additional credential retry rounds after the first round exhausts
# its eligible credentials. Round 0 is the initial round; round r only admits
# credentials whose effective request-retry is at least r. Explicit non-negative
# credential/provider overrides take precedence; omitted or negative overrides
# inherit this global value, and explicit 0 only admits round 0. New CPA nodes
# send retry_round=0 for the initial round and increment it for additional rounds;
# legacy dispatch methods omit the field and keep old semantics.
# Additional rounds apply to HTTP 403, 408, 429, 500, 502, 503, and 504 failures.
# Individual credential/provider overrides take precedence; 0 disables additional
# rounds, while an omitted or negative override inherits this global setting.
request-retry: 3

# Maximum number of different credentials to try in each credential retry round
# after per-credential round filtering. Set to 0 to try all available
# credentials. Credentials skipped by this cap still age with the global round,
# so the cap does not guarantee a fixed number of actual retries per credential.
max-retry-credentials: 0

# Maximum cooldown wait in seconds between retry rounds.
# Set to 0 or below to never wait for credential cooldown.
# Retry rounds that need no wait remain controlled by request-retry.
max-retry-interval: 30

# When true, disable auth/model cooldown scheduling globally (prevents blackout windows after failure states).
# A credential/provider disable-cooling value, when present, overrides this global value.
disable-cooling: false

# When true, persist per-auth cooldown status as .cds files next to auth files.
# Default is false; when false, cooldown status is kept in memory only.
save-cooldown-status: false

# Cooldown duration in seconds for transient upstream errors (408/500/502/503/504/520-526).
# Set to 0 to keep the legacy 60-second cooldown; set to -1 to disable transient error cooldowns.
transient-error-cooldown-seconds: 0

# When true, globally disable Claude request cloaking (the Claude Code CLI disguise and
# system prompt replacement), so the original system prompt is passed through to Claude as-is.
# Individual credentials can still override this: a claude-api-key entry via its "cloak.mode",
# or a Claude OAuth/token file via a "cloak_mode" value. Default false keeps the per-client
# "auto" behavior (cloak only non-Claude-Code clients).
disable-claude-cloak-mode: false

# Provider-wide Claude settings.
claude:
  # When true, scope quota cooldowns to the requested model rather than cooling down
  # the entire credential across all sibling models. Default is false.
  model-level-cooling: false

# Claude Code compatibility settings.
claude-code:
  # When true, return original model IDs in Anthropic model list responses instead of cloaked IDs.
  disable-cloaking-model-list: false

# disable-image-generation supports: false (default), true, "chat", or "passthrough".
# - true: disable image_generation everywhere (also returns 404 for /v1/images/generations and /v1/images/edits).
# - "chat": disable image_generation injection on non-images endpoints, but keep /v1/images/generations and /v1/images/edits enabled.
# - "passthrough": never inject or strip image_generation on non-images endpoints (forward the client payload unchanged); behaves like "chat" on /v1/images/* endpoints.
disable-image-generation: false

# Base model used by the legacy hosted image_generation tool path when a Codex image request is not proxied directly through the Image API.
# Must start with "gpt-" (case-insensitive). If unset or invalid, defaults to "gpt-5.4-mini".
# gpt-image-2-base-model: "gpt-5.4-mini"

# How long video IDs returned by /openai/v1/videos and xAI video creation stay bound
# to the credential that created them. Default: 3h.
video-result-auth-cache-ttl: "3h"

# Core auth auto-refresh and manual refresh-all worker pool size (OAuth/file-based auth token refresh).
# When > 0, overrides the default worker count (16).
# auth-auto-refresh-workers: 16

# Quota exceeded behavior
quota-exceeded:
  switch-project: true # Whether to automatically switch to another project when a quota is exceeded
  switch-preview-model: true # Whether to automatically switch to a preview model when a quota is exceeded
  antigravity-credits: true # Whether to use credits as last-resort fallback when all free-tier auths are exhausted for Claude models

# Routing strategy for selecting credentials when multiple match.
routing:
  strategy: "round-robin" # round-robin (default), weighted-round-robin, fill-first
  # weighted-round-robin uses each credential's integer weight (default 1, maximum 1,000,000).
  # Non-positive weights exclude the credential while this strategy is active.
  # For OAuth/file credentials, add a top-level numeric "weight" field to the auth JSON.
  # Enable universal session-sticky routing for all clients.
  # Explicit Claude Code, Codex, OpenCode, and pi session headers are preferred,
  # followed by prompt_cache_key, Responses conversation IDs, legacy body IDs,
  # execution or derived session identity, and the existing first-message hash fallback.
  # Automatic failover is always enabled when bound auth becomes unavailable.
  # An established binding outranks credential priority: once a session is bound, that
  # credential is kept even if a higher-priority credential recovers. Credential priority
  # still decides cold bindings, requests without a session, and post-failover rebinding.
  session-affinity: false # default: false
  # How long session-to-auth bindings are retained. Default: 1h
  session-affinity-ttl: "1h"
  # When true (default), subagents (child sessions with parent references) inherit and bind
  # to the parent's upstream credential across all providers (Claude, Codex, Antigravity, Gemini),
  # maximizing Prompt/KV Cache reuse and minimizing TTFT.
  # When false, subagents are distributed across the credential pool via the fallback selector
  # for maximum parallel concurrency.
  # Ignored when routing.session-affinity is false.
  session-affinity-subagents: true

# Codex provider behavior.
codex:
  # Experimental full-duplex Codex Responses WebSocket steering (default: false).
  # Keeps one upstream account/model per socket; accepted input is never replayed.
  # See STEERING.md for limitations.
  response-steering: false
  # When true, and routing.strategy is fill-first or routing.session-affinity is true,
  # remap Codex prompt_cache_key and installation identity per selected auth.
  # Some superstitious users believe request tracking identifiers can be used
  # as evidence for TOS enforcement bans; this option only satisfies those odd concerns.
  identity-confuse: false
  # Disable forcing the official Codex User-Agent and Originator headers on HTTP/SSE and WebSocket requests.
  disable-codex-cloaking: false
  # Hold back the frames that carry nothing the client has seen - the handshake (response.created,
  # response.in_progress, the websocket metadata frames), keepalive heartbeats, and the *.added
  # announcements of an item or part that is still empty - until the upstream emits its first
  # generated event. Anything else, including a server-side tool item such as web_search_call and
  # any frame not on that list, releases the stream immediately. On the SSE transport the lines that
  # are not `data:` frames - comments, `event:` lines, blank separators - are held alongside them.
  # The hold is bounded by a frame budget and a 1 MiB byte budget, not by wall-clock time: an
  # upstream that trickles frames holds the response headers for as long as the client will wait.
  # Why: the upstream smuggles `server_is_overloaded` rejections *inside* an HTTP 200 stream,
  # right after those handshake events, instead of returning 503 on the wire. Buffering them
  # keeps the downstream response headers uncommitted long enough to transparently retry on
  # another credential. Only overload/rate-limit rejections trigger failover; every other terminal
  # failure is still delivered in-stream exactly as before. One pre-existing behaviour is worth
  # knowing, because this option makes the window it lives in much longer: if the stream ends while
  # the bootstrap is still holding, the attempt ends instead of the error reaching the client, and
  # whether another credential is then tried depends on how it ended. A clean end with no terminal
  # event is request-scoped on SSE, so nothing else is tried; a websocket close, or a transport error
  # such as a dropped connection on either transport, is not, so the request may be retried on
  # another credential. None of that is new, but before this option the hold ended at the first
  # keepalive, so it was rare; now it can span the whole reasoning phase.
  # Trade-off: response headers are delayed until generation starts. Because heartbeats are held
  # too, the delay now ends at the first generated token rather than at the first heartbeat. The
  # budget bounds how much is held, not how long: 48 SSE lines, which measures out as 15 heartbeats
  # of the three-line event:/data:/blank shape, 23 of a two-line ": keepalive" comment followed by
  # its blank separator, or 47 of a bare ": keepalive" with no separator - one short of the obvious
  # division, because the rejection frame's own event: line comes out of the same budget before its
  # data: line is parsed. An upstream that sends the rejection as a bare data: line with no event:
  # line gets the round number instead. At a 10s upstream heartbeat that is 150s, 230s or 470s of
  # uncommitted response headers depending on the framing the upstream picks, all of it past nginx's
  # 60s default proxy_read_timeout, so raise the timeouts in front of the proxy before enabling this.
  # Those seconds bound what the buffer holds, not when the response headers are committed. Headers
  # are committed on the first chunk the downstream translator actually produces, so for a format
  # that renders the handshake as nothing - Chat Completions and Gemini do - they stay uncommitted
  # past the budget, until the first generated token. That is not specific to the budget: any frame
  # that releases the buffer has always behaved this way.
  # The websocket transport spends the same 48 differently - one per message read, whether the loop
  # holds it or skips it, with no framing to divide by - so there it is 48 heartbeats, about 480s.
  # Two cases no budget covers: a websocket peer that sends only ping frames, which the library
  # answers inside the read call, leaving only the 5 minute websocket read deadline; and an upstream
  # that stops mid-line on SSE, where the scanner never returns and only the request context ends the
  # wait. Neither is made worse by this option - with it off the same stalled line stalls the
  # conductor instead, which waits on the same context - but a client-side request timeout, not
  # proxy_read_timeout, is what bounds them.
  # Default: false
  stream-bootstrap-buffering: false
  # Optional maximum duration to hold back uncommitted response headers during bootstrap buffering.
  # When set (e.g. "20s"), if the time ceiling is reached before the first generated token,
  # the stream is released to the client and further in-stream overloads are delivered rather
  # than failed over. This bounds how long trickled heartbeats or early empty frames can delay
  # response headers before streaming begins.
  # Recommendation: If your CPA is deployed behind a reverse proxy with a strict read timeout
  # (e.g. Nginx default proxy_read_timeout 60s or Cloudflare 100s timeout), setting this to "20s"
  # is strongly recommended to guarantee headers commit well before proxy timeouts while retaining
  # a comfortable 20s window to fail over early scheduling overloads.
  # Note: The timeout check is evaluated as upstream frames arrive; it never aborts the upstream
  # connection. Downstream headers are committed as soon as the downstream translator emits chunks
  # (for formats that render handshake/heartbeats, headers commit immediately on release; for
  # formats like Chat Completions and Gemini that render nothing until the first text token, headers
  # commit when that first token arrives).
  # When unset or set to "0", "0s", "none", "unlimited", "disabled", "off", or "never", no time ceiling
  # is applied, and buffering relies purely on the 48-frame and 1MB byte budget.
  # Default: "0" (unlimited)
  stream-bootstrap-timeout: "0"
  # When true, optimize Codex Desktop, codex-tui, and codex_cli_rs requests for multi-agent v2.
  # This refreshes Codex spawn_agent model details, removes message parameter encryption,
  # normalizes encrypted agent_message content for Codex, and converts agent_message input
  # into standard user messages for non-Codex upstream protocols.
  optimize-multi-agent-v2: false
  # When true, enable opt-in compatibility for orphan Codex delegation outputs.
  # Converts orphan function_call_output items from codex_app/create_thread and
  # codex_app/send_message_to_thread (which lack a valid call_id or matching function_call)
  # into standard user messages across Responses and non-Codex protocols when the request
  # header contains X-Openai-Subagent: collab_spawn.
  orphan-delegation-compatibility: false
  # When true, scope Codex usage_limit_reached quota cooldowns to the requested model rather
  # than cooling down the entire credential across all sibling models.
  # Default: false
  model-level-cooling: false
  # Terminate and relay Codex Live WebRTC audio and DataChannel traffic in this process.
  # This requires inbound UDP reachability. Keep disabled to preserve direct media behavior.
  live-media-relay:
    enabled: false
    # Maximum concurrent media sessions. Zero uses the default of 32.
    max-sessions: 32
    # Reject downstream SDP candidates that target private, loopback, link-local, or unspecified IPs.
    # Keep false for local or trusted-network Codex Desktop connections.
    disable-private-remote-ips: false
    # Public IPv4 or IPv6 address advertised when CPA is behind 1:1 NAT.
    public-ip: ""
    # Optional UDP allocation range. Both values must be set together and provide at least two ports per session.
    udp-port-min: 0
    udp-port-max: 0
    # Optional STUN/TURN servers. TURN credentials are never returned by the JSON config API.
    # Without a concrete global/per-auth proxy-url, WebRTC uses normal direct ICE/STUN/TURN connectivity.
    # With http, https, socks5, or socks5h proxy-url, the OpenAI-facing leg is forced through
    # authenticated ICE-TCP over that proxy and never falls back to UDP or a direct connection.
    # The Codex Desktop-facing leg remains direct, and configured ICE servers still apply to it.
    # ice-servers:
    #   - urls:
    #       - "stun:stun.example.com:3478"
    #   - urls:
    #       - "turn:turn.example.com:3478?transport=udp"
    #     username: "user"
    #     credential: "secret"

# Antigravity provider behavior.
# antigravity:
#   sensitive-words:             # optional: words to obfuscate with zero-width characters in system instructions
#     - "API"
#     - "proxy"
#   connection-pool:             # optional: upstream HTTP connection pool settings
#     enabled: false             # whether to enable connection pooling (default: false; set to true to enable connection pooling)
#     idle-conn-timeout: "30s"   # idle keep-alive timeout (default: 30s, capped at 210s)
#     max-idle-conns-per-host: 2 # max idle connections per host per credential (default: 2)

# Devin provider behavior.
# devin:
#   sensitive-words:             # optional: words to obfuscate with zero-width characters in system instructions and prompts
#     - "API"
#     - "proxy"

# xAI provider behavior.
xai:
  # When true, inject the native x_search tool when the request does not declare it.
  # The injected tool is also added to tool_choice.allowed_tools when applicable.
  inject-x-search: false

# When true, enable authentication for the WebSocket API (/v1/ws).
ws-auth: true

# When > 0, emit blank lines every N seconds for non-streaming responses to prevent idle timeouts.
nonstream-keepalive-interval: 0
# Streaming behavior (SSE keep-alives + safe bootstrap retries).
# streaming:
#   keepalive-seconds: 15   # Default: 0 (disabled). <= 0 disables keep-alives.
#   bootstrap-retries: 1    # Default: 0 (disabled). Retries before first byte is sent.

# Signature cache validation for thinking blocks (Antigravity/Claude).
# When true (default), cached signatures are preferred and validated.
# When false, client signatures are used directly after normalization (bypass mode for testing).
# antigravity-signature-cache-enabled: true

# Bypass mode signature validation strictness (only applies when signature cache is disabled).
# When true, validates full Claude protobuf tree (Field 2 -> Field 1 structure).
# When false (default), only checks R/E prefix + base64 + first byte 0x12.
# antigravity-signature-bypass-strict: false

# Gemini API keys
# gemini-api-key:
#   - api-key: "AIzaSy...01"
#     weight: 5 # optional: weighted-round-robin share; omitted defaults to 1; maximum 1,000,000
#     prefix: "test" # optional: require calls like "test/gemini-3-pro-preview" to target this credential
#     disable-cooling: false # optional override: true disables cooling, false enables it; omit to inherit global
#     request-retry: 3 # optional per-auth override; 0 disables additional rounds; omit or set < 0 to inherit global
#     request-scoped-errors: # optional: custom rules to classify upstream errors by status and body patterns
#       - status: 400        # HTTP status code to match
#         match:             # optional: string contains matching
#           - "maximum_context_length"
#           - "context_length_exceeded"
#         match-regexr:      # optional: regular expression matching
#           - "maximum_context_length$"
#           - "^context_length_exceeded"
#         action: "stop"     # "stop" (return error, no cooling), "stop-and-cooldown" (return error and cool down),
#                            # "continue" (try next credential, no cooling), "continue-and-cooldown" (try next credential and cool down)
#     base-url: "https://generativelanguage.googleapis.com"
#     headers:
#       X-Custom-Header: "custom-value"
#       # Values starting with "$" dynamically copy the header value from downstream client requests.
#       # If the client did not send the specified header, the header is omitted.
#       # X-Claude-Code-Session-Id: "$ABC" # copies client's "ABC" header
#     proxy-url: "socks5://proxy.example.com:1080"
#     # proxy-url: "direct" # optional: explicit direct connect for this credential
#     models:
#       - name: "gemini-2.5-flash" # upstream model name
#         alias: "gemini-flash"    # client alias mapped to the upstream model
#         display-name: "Gemini Flash" # optional catalog display name
#         max-context-length: 1048576 # optional: override Codex client context window metadata
#         is-compat: false             # optional: preserve thinking blocks with empty signatures for compatible upstreams
#         thinking:                    # optional: exact thinking capability for this configured model
#           levels: ["high", "medium", "low", "none", "auto"]
#     excluded-models:
#       - "gemini-2.5-pro"     # exclude specific models from this provider (exact match)
#       - "gemini-2.5-*"       # wildcard matching prefix (e.g. gemini-2.5-flash, gemini-2.5-pro)
#       - "*-preview"          # wildcard matching suffix (e.g. gemini-3-pro-preview)
#       - "*flash*"            # wildcard matching substring (e.g. gemini-2.5-flash-lite)
#   - api-key: "AIzaSy...02"

# Native Interactions API keys
# These keys are used only for direct /v1beta/interactions execution. Regular gemini-api-key entries still
# send Gemini generateContent/streamGenerateContent requests when the client enters through the interactions API.
# interactions-api-key:
#   - api-key: "AIzaSy...03"
#     weight: 5 # optional: weighted-round-robin share; omitted defaults to 1; maximum 1,000,000
#     prefix: "native" # optional: require calls like "native/gemini-3-pro-preview" to target this credential
#     disable-cooling: false # optional override: true disables cooling, false enables it; omit to inherit global
#     request-retry: 3 # optional per-auth override; 0 disables additional rounds; omit or set < 0 to inherit global
#     request-scoped-errors: # optional: custom rules to classify upstream errors by status and body patterns
#       - status: 400
#         match:
#           - "invalid_argument"
#         action: "continue"
#     base-url: "https://generativelanguage.googleapis.com"
#     headers:
#       X-Custom-Header: "custom-value"
#       # Values starting with "$" dynamically copy the header value from downstream client requests.
#       # If the client did not send the specified header, the header is omitted.
#       # X-Claude-Code-Session-Id: "$ABC" # copies client's "ABC" header
#     proxy-url: "socks5://proxy.example.com:1080"
#     # proxy-url: "direct" # optional: explicit direct connect for this credential
#     models:
#       - name: "gemini-2.5-flash" # upstream model name
#         alias: "native-gemini-flash" # client alias mapped to the upstream model
#         max-context-length: 1048576 # optional: override Codex client context window metadata
#         is-compat: false             # optional: preserve thinking blocks with empty signatures for compatible upstreams
#         thinking:                    # optional: exact thinking capability for this configured model
#           levels: ["high", "medium", "low", "none", "auto"]
#     excluded-models:
#       - "gemini-2.5-pro"

# Codex API keys
# codex-api-key:
#   - api-key: "sk-atSM..."
#     weight: 5 # optional: weighted-round-robin share; omitted defaults to 1; maximum 1,000,000
#     prefix: "test" # optional: require calls like "test/gpt-5-codex" to target this credential
#     disable-cooling: false # optional override: true disables cooling, false enables it; omit to inherit global
#     request-retry: 3 # optional per-auth override; 0 disables additional rounds; omit or set < 0 to inherit global
#     request-scoped-errors: # optional: custom rules to classify upstream errors by status and body patterns
#       - status: 400
#         match:
#           - "context_window_exceeded"
#         action: "stop-and-cooldown"
#     base-url: "https://www.example.com" # use the custom codex API endpoint
#     alpha-search: false # optional: allow this key to serve /v1/alpha/search via base-url + /alpha/search
#     headers:
#       X-Custom-Header: "custom-value"
#       # Values starting with "$" dynamically copy the header value from downstream client requests.
#       # If the client did not send the specified header, the header is omitted.
#       # X-Claude-Code-Session-Id: "$ABC" # copies client's "ABC" header
#     proxy-url: "socks5://proxy.example.com:1080" # optional: per-key proxy override
#     # proxy-url: "direct" # optional: explicit direct connect for this credential
#     models:
#       - name: "gpt-5-codex"   # upstream model name
#         alias: "codex-latest" # client alias mapped to the upstream model
#         display-name: "Codex Latest" # optional catalog display name
#         max-context-length: 1048576 # optional: override Codex client context window metadata
#         force-mapping: true    # optional: rewrite response model fields back to the alias
#         # When true and codex.optimize-multi-agent-v2 is also true, convert Codex
#         # MultiAgentV2 agent_message items into portable Responses message/user input
#         # for third-party Responses-compatible endpoints that reject agent_message.
#         # Default false keeps agent_message unchanged for native OpenAI/Codex endpoints.
#         # It also preserves thinking blocks with empty signatures for compatible upstreams.
#         is-compat: false
#         thinking:              # optional: exact thinking capability for this configured model
#           levels: ["xhigh", "high", "medium", "low"]
#     excluded-models:
#       - "gpt-5.1"         # exclude specific models (exact match)
#       - "gpt-5-*"         # wildcard matching prefix (e.g. gpt-5-medium, gpt-5-codex)
#       - "*-mini"          # wildcard matching suffix (e.g. gpt-5-codex-mini)
#       - "*codex*"         # wildcard matching substring (e.g. gpt-5-codex-low)

# xAI API keys
# Uses the native xAI executor, including its Responses namespace-tool handling.
# xai-api-key:
#   - api-key: "xai-..."
#     weight: 5 # optional: weighted-round-robin share; omitted defaults to 1; maximum 1,000,000
#     prefix: "xai" # optional: require calls like "xai/grok-4.5" to target this credential
#     disable-cooling: false # optional override: true disables cooling, false enables it; omit to inherit global
#     request-retry: 3 # optional per-auth override; 0 disables additional rounds; omit or set < 0 to inherit global
#     request-scoped-errors: # optional: custom rules to classify upstream errors by status and body patterns
#       - status: 400
#         match:
#           - "rate_limit_exceeded"
#         action: "continue-and-cooldown"
#     base-url: "https://api.x.ai/v1" # xAI-compatible Responses API endpoint
#     websockets: true # optional: use the xAI upstream websocket transport for downstream websocket requests
#     headers:
#       X-Custom-Header: "custom-value"
#       # Values starting with "$" dynamically copy the header value from downstream client requests.
#       # If the client did not send the specified header, the header is omitted.
#       # X-Claude-Code-Session-Id: "$ABC" # copies client's "ABC" header
#     proxy-url: "socks5://proxy.example.com:1080" # optional: per-key proxy override
#     # proxy-url: "direct" # optional: explicit direct connect for this credential
#     models:
#       - name: "grok-4.5"       # upstream model name
#         alias: "grok-latest"   # client alias mapped to the upstream model
#         display-name: "Grok Latest" # optional catalog display name
#         max-context-length: 1048576 # optional: override Codex client context window metadata
#         force-mapping: true     # optional: rewrite response model fields back to the alias
#         is-compat: false        # optional: preserve thinking blocks with empty signatures for compatible upstreams
#         thinking:               # optional: exact thinking capability for this configured model
#           levels: ["xhigh", "high", "medium", "low"]
#     excluded-models:
#       - "grok-4.1"             # exclude specific models (exact match)
#       - "grok-3-*"             # wildcard matching prefix
#
# Meta API keys (Muse Code)
# Uses the native Meta executor targeting https://api.meta.ai/v1.
# meta-api-key:
#   - api-key: "..."
#     weight: 5 # optional: weighted-round-robin share; omitted defaults to 1; maximum 1,000,000
#     prefix: "meta" # optional: require calls like "meta/muse-spark-1.3" to target this credential
#     disable-cooling: false # optional override: true disables cooling, false enables it; omit to inherit global
#     request-retry: 3 # optional per-auth override; 0 disables additional rounds; omit or set < 0 to inherit global
#     base-url: "https://api.meta.ai/v1" # optional: defaults to https://api.meta.ai/v1
#     proxy-url: "socks5://proxy.example.com:1080" # optional: per-key proxy override
#     models:
#       - name: "muse-spark-1.3" # upstream model name
#         alias: "muse-latest"   # client alias mapped to the upstream model
#     excluded-models:
#       - "muse-spark-1.1"

# Claude API keys
# claude-api-key:
#   - api-key: "sk-atSM..." # use the official claude API key, no need to set the base url
#   - api-key: "sk-atSM..."
#     weight: 5 # optional: weighted-round-robin share; omitted defaults to 1; maximum 1,000,000
#     prefix: "test" # optional: require calls like "test/claude-sonnet-latest" to target this credential
#     disable-cooling: false # optional override: true disables cooling, false enables it; omit to inherit global
#     request-retry: 3 # optional per-auth override; 0 disables additional rounds; omit or set < 0 to inherit global
#     request-scoped-errors: # optional: custom rules to classify upstream errors by status and body patterns
#       - status: 400
#         match:
#           - "prompt is too long"
#         action: "stop"
#     base-url: "https://www.example.com" # use the custom claude API endpoint
#     headers:
#       X-Custom-Header: "custom-value"
#       # Values starting with "$" dynamically copy the header value from downstream client requests.
#       # If the client did not send the specified header, the header is omitted.
#       # X-Claude-Code-Session-Id: "$ABC" # copies client's "ABC" header
#     proxy-url: "socks5://proxy.example.com:1080" # optional: per-key proxy override
#     # proxy-url: "direct" # optional: explicit direct connect for this credential
#     models:
#       - name: "claude-3-5-sonnet-20241022" # upstream model name
#         alias: "claude-sonnet-latest"      # client alias mapped to the upstream model
#         display-name: "Claude Sonnet"      # optional catalog display name
#         max-context-length: 1048576         # optional: override Codex client context window metadata
#         force-mapping: true                 # optional: rewrite response model fields back to the alias
#         is-compat: false                    # optional: preserve thinking blocks with empty signatures for compatible upstreams
#         thinking:                           # optional: exact thinking capability for this configured model
#           levels: ["max", "xhigh", "high", "medium", "low", "minimal", "none", "auto"]
#     excluded-models:
#       - "claude-opus-4-5-20251101" # exclude specific models (exact match)
#       - "claude-3-*"               # wildcard matching prefix (e.g. claude-3-7-sonnet-20250219)
#       - "*-thinking"               # wildcard matching suffix (e.g. claude-opus-4-5-thinking)
#       - "*haiku*"                  # wildcard matching substring (e.g. claude-3-5-haiku-20241022)
#     rebuild-mid-system-message: false # optional: default is false; when true, move messages with role "system" into the top-level Claude system field
#     cloak:                         # optional: explicitly enable request cloaking for non-Claude-Code clients
#       mode: "auto"                 # "auto" (default inside this block): cloak only when client is not Claude Code
#                                    # "always": cloak every unconfirmed client; confirmed native Claude Code still passes through
#                                    # "never": never apply cloaking
#       # This "cloak" block applies to this claude-api-key entry only. For Claude OAuth
#       # credentials, set the same options in the auth/token JSON file via "cloak_mode" /
#       # "cloak_strict_mode" / "cloak_sensitive_words" / "cloak_cache_user_id". The top-level
#       # "disable-claude-cloak-mode: true" disables cloaking for all Claude credentials at once.
#       strict-mode: false           # false (default): legacy-model whitelist uses a user system-reminder;
#                                    # all other and future models use messages[].role=system
#                                    # true: strip caller prompts and keep only Claude Code billing and identity blocks
#       sensitive-words:             # optional: words to obfuscate with zero-width characters
#         - "API"
#         - "proxy"
#       cache-user-id: true          # optional: default is false; set true to reuse cached user_id per API key instead of generating a random one each request
#       # Every custom tool on a cloaked OAuth request automatically uses a caller-stable opaque mcp__<server>__<tool> alias.
#
#     # fingerprint-profile (optional, top-level on this claude-api-key entry; not a cloak sub-field):
#     # OAuth and API-key fingerprints are different contracts.
#     #   - Real Claude OAuth stays on the strict Claude Code CLI wire fingerprint.
#     #   - API keys (official Anthropic, custom gateways, Kimi) stay loose and
#     #     caller-owned unless this field is set.
#     #
#     # Default (omit / empty): keep the caller request fingerprint and headers.
#     # Official api.anthropic.com API keys do not add extra CLI betas/identity unless
#     # this field is set. Custom gateways and delegated providers are the same.
#     #
#     # Controls request fingerprint only on /v1/messages (and related Claude executor paths).
#     # Auth scheme stays API key (x-api-key on api.anthropic.com; Bearer on custom base-url).
#     # Does NOT enable OAuth refresh, profile fetch, or OAuth-cancellation semantics.
#     #
#     # Values:
#     #   omit / empty        = caller-owned API-key fingerprint (respects caller)
#     #   "claude-code-cli"   = same Messages fingerprint as Claude Code OAuth CLI,
#     #                         including official Anthropic API keys: OAuth Anthropic-Beta
#     #                         set, CCH signing on api.anthropic.com, stable CLI
#     #                         metadata.user_id / session_id / device identity.
#     #                         API keys seed identity from the key;
#     #                         delegated OAuth providers use stable auth ID instead of
#     #                         rotating access tokens. "oauth-cli" is a legacy alias.
#     #
#     # count_tokens keeps the native model/messages/tools shape for every origin, including
#     # Kimi opt-in. It does not send billing/CCH, currentDate, metadata, or diagnostics.
#     #
#     # CCH: the billing block may carry a per-request cch hash. CPA emits it exactly where
#     # Claude Code does, which is api.anthropic.com (first-party) and Vertex only. An opt-in
#     # on any other gateway (including Kimi) still sends the billing block, but without cch,
#     # so a per-request hash cannot bust that gateway's prompt cache. api.anthropic.com
#     # strips the block itself (0 tokens, no cache impact). Kimi drops the whole block by
#     # default and keeps it, unsigned, after an explicit fingerprint opt-in.
#     # A real Claude OAuth credential always signs, on every upstream: a downstream Claude
#     # Code pointed at CPA cannot produce that value itself.
#     #
#     # Example (official Anthropic or a custom Messages gateway):
#     #   - api-key: "your-key"
#     #     # base-url: "https://gateway.example"   # omit for api.anthropic.com
#     #     fingerprint-profile: "claude-code-cli"
#     #     cloak:
#     #       mode: "always"   # recommended when upstream rejects non-CLI clients
#     #
#     # Delegated Anthropic Messages OAuth files (Kimi, etc.) use "fingerprint_profile"
#     # in the auth JSON. Refresh keeps it. Example:
#     #   {
#     #     "type": "kimi",
#     #     "access_token": "...",
#     #     "refresh_token": "...",
#     #     "fingerprint_profile": "claude-code-cli"
#     #   }
#     #   Legacy "fingerprint-profile" credentials remain supported and are normalized at load time.
#     # fingerprint-profile: "claude-code-cli" # optional claude-api-key provider field; default is empty (caller-owned); uncomment to opt in
#     experimental-cch-signing: false # deprecated compatibility field; CCH is generated automatically
#                                     # for real Claude OAuth on any upstream, and for claude-code-cli profiles
#                                     # only on api.anthropic.com; Vertex keeps provider-native signing

# Anthropic-Beta is assembled per request rather than sent as a fixed list, matching
# Claude Code: context-1m sits right after claude-code, mid-conversation-system
# is added only for models that accept a role=system turn, advanced-tool-use only when
# the request uses tool search (deferred tools) or another advanced tool-use feature,
# afk-mode sits between fast-mode and extended-cache-ttl, and server-side-fallback / fallback-credit /
# structured-outputs trail effort. On direct api.anthropic.com a caller may only ask for
# betas real Claude Code also sends, and they are placed at their observed positions;
# anything else is dropped so the outgoing set stays one a real client could produce.
# Other Anthropic-compatible upstreams still forward caller betas verbatim.
#
# Default headers for Claude API requests. Update only after measuring a new Claude Code release.
# Unconfirmed clients use this CLI baseline. Verified native Claude Code CLI, sdk-cli,
# and VSCode requests preserve their measured entrypoint when the Claude Code version is a
# patch release in the configured major/minor line at or above this baseline. Older versions
# and newer major/minor lines fall back to the baseline until measured. In legacy mode,
# timeout is a fallback and
# verified native OS/arch values remain client-supplied. When stabilize-device-profile is
# enabled, OS/arch are pinned to the values below and cached profiles remain constrained to
# the same exact software baseline rather than learning newer client versions.
# claude-header-defaults:
#   user-agent: "claude-cli/2.1.258 (external, cli)"
#   package-version: "0.112.1"
#   runtime-version: "v26.3.0"
#   os: "MacOS"
#   arch: "arm64"
#   timeout: "600"
#   timezone: "Asia/Singapore"  # fallback IANA timezone for cloaked currentDate; a credential JSON "timezone" takes priority
#   stabilize-device-profile: false  # optional, default false; set true to enable per-auth/API-key fingerprint pinning

# Default headers for Codex OAuth model requests.
# These are used only for file-backed/OAuth Codex requests when the client
# does not send the header. `user-agent` applies to HTTP and websocket requests;
# `beta-features` only applies to websocket requests. They do not apply to codex-api-key entries.
# codex-header-defaults:
#   user-agent: "codex_cli_rs/0.114.0 (Mac OS 14.2.0; x86_64) vscode/1.111.0"
#   beta-features: "multi_agent"

# OpenAI compatibility providers
# openai-compatibility:
#   - name: "openrouter" # The name of the provider; it will be used in the user agent and other places.
#     disabled: false # optional: set to true to disable this provider without removing it
#     prefix: "test" # optional: require calls like "test/kimi-k2" to target this provider's credentials
#     base-url: "https://openrouter.ai/api/v1" # The base URL of the provider.
#     support-prompt-cache-key: false # optional: derive prompt_cache_key for requests from all input protocols
#     disable-cooling: false # optional provider override: true disables cooling, false enables it; omit to inherit global
#     request-retry: 3 # optional per-provider override; 0 disables additional rounds; omit or set < 0 to inherit global
#     request-scoped-errors: # optional: custom rules to classify upstream errors by status and body patterns
#       - status: 400
#         match:
#           - "maximum_context_length"
#           - "context_length_exceeded"
#         match-regexr:
#           - "maximum_context_length$"
#           - "^context_length_exceeded"
#         action: "stop" # "stop", "stop-and-cooldown", "continue", "continue-and-cooldown"
#     headers:
#       X-Custom-Header: "custom-value"
#       # Values starting with "$" dynamically copy the header value from downstream client requests.
#       # If the client did not send the specified header, the header is omitted.
#       # X-Claude-Code-Session-Id: "$ABC" # copies client's "ABC" header
#     api-key-entries:
#       - api-key: "sk-or-v1-...b780"
#         weight: 5 # optional: weighted-round-robin share; omitted defaults to 1; maximum 1,000,000
#         proxy-url: "socks5://proxy.example.com:1080" # optional: per-key proxy override
#         # proxy-url: "direct" # optional: explicit direct connect for this credential
#       - api-key: "sk-or-v1-...b781" # without proxy-url
#     models: # The models supported by the provider.
#       - name: "moonshotai/kimi-k2:free" # The actual model name.
#         alias: "kimi-k2"               # The alias used in the API.
#         display-name: "Kimi K2"         # optional catalog display name
#         max-context-length: 1048576    # optional: override Codex client context window metadata
#         image: false                   # optional: set true to allow this model on /v1/images/generations and /v1/images/edits (not chat/responses image input)
#         input-modalities: [text, image] # optional: declare /v1/chat/completions and /v1/responses multimodal input for Codex clients. Use [text] for upstreams that reject multimodal tool result content.
#         output-modalities: [text]       # optional: declare output modalities when known
#         is-compat: false                # optional: preserve Claude thinking blocks for compatible upstreams
#         use-max-completion-tokens: false # optional: default false (uses max_tokens); set true to use max_completion_tokens
#         thinking:                      # optional: omit to default to levels ["low","medium","high"]. Exceeding levels (e.g. "max", "xhigh") are clamped to "high"; explicitly declare levels here if the upstream supports them
#           levels: ["low", "medium", "high"]
#       # You may repeat the same alias to build an internal model pool.
#       # The client still sees only one alias in the model list.
#       # Requests to that alias will round-robin across the upstream names below,
#       # and if the chosen upstream fails before producing output, the request will
#       # continue with the next upstream model in the same alias pool.
#       - name: "deepseek-v3.1"
#         alias: "claude-opus-4.66"
#       - name: "glm-5"
#         alias: "claude-opus-4.66"
#       - name: "kimi-k2.5"
#         alias: "claude-opus-4.66"

# Vertex API keys (Vertex-compatible endpoints, base-url is optional)
# vertex-api-key:
#   - api-key: "vk-123..."                        # x-goog-api-key header
#     weight: 5                                    # optional: weighted-round-robin share; omitted defaults to 1; maximum 1,000,000
#     prefix: "test"                              # optional: require calls like "test/vertex-pro" to target this credential
#     disable-cooling: false                       # optional override: true disables cooling, false enables it; omit to inherit global
#     request-retry: 3                             # optional per-auth override; 0 disables additional rounds; omit or set < 0 to inherit global
#     base-url: "https://example.com/api"         # optional, e.g. https://zenmux.ai/api; falls back to Google Vertex when omitted
#     proxy-url: "socks5://proxy.example.com:1080" # optional per-key proxy override
#     # proxy-url: "direct" # optional: explicit direct connect for this credential
#     headers:
#       X-Custom-Header: "custom-value"
#       # Values starting with "$" dynamically copy the header value from downstream client requests.
#       # If the client did not send the specified header, the header is omitted.
#       # X-Claude-Code-Session-Id: "$ABC" # copies client's "ABC" header
#     models:                                     # optional: map aliases to upstream model names
#       - name: "gemini-2.5-flash"                # upstream model name
#         alias: "vertex-flash"                   # client-visible alias
#         display-name: "Vertex Flash"            # optional catalog display name
#         thinking:                                # optional: exact thinking capability for this configured model
#           levels: ["high", "medium", "low", "none", "auto"]
#       - name: "gemini-2.5-pro"
#         alias: "vertex-pro"
#     excluded-models:                            # optional: models to exclude from listing
#       - "imagen-3.0-generate-002"
#       - "imagen-*"

# Global OAuth model name aliases (per channel)
# These aliases rename model IDs for both model listing and request routing.
# Supported channels: vertex, aistudio, antigravity, claude, codex, kimi, xai, meta.
# NOTE: Aliases do not apply to gemini-api-key, interactions-api-key, codex-api-key, xai-api-key, meta-api-key, claude-api-key, openai-compatibility, or vertex-api-key.
# NOTE: Because aliases affect the merged /v1 model list and merged request routing, overlapping
# client-visible names can become ambiguous across providers. For strict backend pinning, use
# unique aliases/prefixes or avoid overlapping names.
# You can repeat the same name with different aliases to expose multiple client model names.
# Optional per-entry fields:
#   fork: true                  # keep the upstream model and also expose the alias as a separate client-visible model
#   display-name: "Model Name" # override the human-readable name shown in model catalogs
#   force-mapping: true         # rewrite upstream response model fields back to the client-visible alias (example below uses antigravity only)
# Per-auth OAuth aliases can also be stored in an OAuth auth JSON file as "model_aliases".
# Legacy "model-aliases" credentials remain supported and are normalized at load time.
# They apply only to that selected auth and take precedence over global aliases for the same client-visible alias.
# Example auth JSON:
# {
#   "type": "codex",
#   "email": "user@example.com",
#   "model_aliases": [
#     {"name": "gpt-5.3-codex-spark", "alias": "gpt-5.5"},
#     {"name": "gpt-5.3-codex-spark", "alias": "gpt-5.4"}
#   ]
# }
# oauth-model-alias:
#   vertex:
#     - name: "gemini-2.5-pro"
#       alias: "g2.5p"
#   aistudio:
#     - name: "gemini-2.5-pro"
#       alias: "g2.5p"
#   antigravity:
#     - name: "gemini-pro-agent"                   # upstream Antigravity model id
#       alias: "gemini-3.1-pro-preview"            # client-visible id (Gemini 3.1 Pro Preview)
#       display-name: "Antigravity Gemini 3.1 Pro" # optional catalog display name
#       fork: true
#       force-mapping: true
#   claude:
#     - name: "claude-sonnet-4-5-20250929"
#       alias: "cs4.5"
#   codex:
#     - name: "gpt-5"
#       alias: "g5"
#   kimi:
#     - name: "kimi-k2.5"
#       alias: "k2.5"
#   xai:
#     - name: "grok-4.3"
#       alias: "grok-latest"
#   meta:
#     - name: "muse-spark-1.3"
#       alias: "muse-latest"
#   sample-provider:                  # plugin provider keys are supported for OAuth plugins
#     - name: "sample-model-latest"
#       alias: "sample-latest"

# OAuth provider excluded models
# oauth-excluded-models:
#   vertex:
#     - "gemini-3-pro-preview"
#   aistudio:
#     - "gemini-3-pro-preview"
#   antigravity:
#     - "gemini-3-pro-preview"
#   claude:
#     - "claude-3-5-haiku-20241022"
#   codex:
#     - "gpt-5-codex-mini"
#   kimi:
#     - "kimi-k2-thinking"
#   xai:
#     - "grok-3-mini"
#   meta:
#     - "muse-spark-1.1"

# OAuth provider request-scoped error rules (custom error classification for OAuth credentials)
# oauth-request-scoped-errors:
#   vertex:
#     - status: 400
#       match:
#         - "maximum_context_length"
#         - "context_length_exceeded"
#       match-regexr:
#         - "maximum_context_length$"
#         - "^context_length_exceeded"
#       action: "stop" # options: "stop", "stop-and-cooldown", "continue", "continue-and-cooldown"
#   aistudio:
#     - status: 400
#       match:
#         - "invalid_argument"
#       action: "stop"
#   antigravity:
#     - status: 500
#       match:
#         - "internal_server_error"
#       action: "stop-and-cooldown"
#   claude:
#     - status: 400
#       match:
#         - "prompt is too long"
#       action: "stop"
#   codex:
#     - status: 400
#       match:
#         - "context_window_exceeded"
#       action: "stop"
#   kimi:
#     - status: 400
#       match:
#         - "length_limit"
#       action: "stop"
#   xai:
#     - status: 400
#       match:
#         - "max_tokens_exceeded"
#       action: "stop"
#   meta:
#     - status: 400
#       match:
#         - "context_length_exceeded"
#       action: "stop"

# Optional payload configuration
# payload:
#   default: # Default rules only set parameters when they are missing in the payload.
#     - models:
#         - name: "gemini-2.5-pro" # Supports wildcards (e.g., "gemini-*")
#           protocol: "gemini" # restricts the rule to a specific protocol, options: openai, gemini, claude, codex, antigravity
#           from-protocol: "responses" # restricts the rule to the source protocol, options: openai, responses, gemini, claude
#           headers: # all configured request headers must match; values support "*" wildcards
#             X-Client-Tier: "tenant-*-region-*"
#           match: # all payload JSON paths must equal the configured values
#             - "metadata.client": "codex"
#           not-match: # payload JSON paths must not equal the configured values
#             - "metadata.mode": "dev"
#           exist: # all payload JSON paths must exist and not be null
#             - "tools.#(type==\"web_search\").type"
#           not-exist: # all payload JSON paths must be missing or null
#             - "metadata.disable_payload"
#       params: # JSON path (gjson/sjson syntax) -> value
#         "generationConfig.thinkingConfig.thinkingBudget": 32768
#   default-raw: # Default raw rules set parameters using raw JSON when missing (must be valid JSON).
#     - models:
#         - name: "gemini-2.5-pro" # Supports wildcards (e.g., "gemini-*")
#           protocol: "gemini" # restricts the rule to a specific protocol, options: openai, gemini, claude, codex, antigravity
#       params: # JSON path (gjson/sjson syntax) -> raw JSON value (strings are used as-is, must be valid JSON)
#         "generationConfig.responseJsonSchema": "{\"type\":\"object\",\"properties\":{\"answer\":{\"type\":\"string\"}}}"
#   override: # Override rules always set parameters, overwriting any existing values.
#     - models:
#         - name: "gpt-5.4-fast"
#           protocol: "codex"
#         - name: "gpt-5.5-fast"
#           protocol: "codex"
#       params:
#         service_tier: priority
#     - models:
#         - name: "gpt-*" # Supports wildcards (e.g., "gpt-*")
#           protocol: "codex" # restricts the rule to a specific protocol, options: openai, gemini, claude, codex, antigravity
#       params: # JSON path (gjson/sjson syntax) -> value
#         "reasoning.effort": "high"
#     - models:
#         - name: "claude-fable-5-1"
#           protocol: "claude"
#       params:
#         max_tokens: 128000 # override the final Claude Messages output budget after protocol translation
#   override-raw: # Override raw rules always set parameters using raw JSON (must be valid JSON).
#     - models:
#         - name: "gpt-*" # Supports wildcards (e.g., "gpt-*")
#           protocol: "codex" # restricts the rule to a specific protocol, options: openai, gemini, claude, codex, antigravity
#       params: # JSON path (gjson/sjson syntax) -> raw JSON value (strings are used as-is, must be valid JSON)
#         "response_format": "{\"type\":\"json_schema\",\"json_schema\":{\"name\":\"answer\",\"schema\":{\"type\":\"object\"}}}"
#   filter: # Filter rules remove specified parameters from the payload.
#     - models:
#         - name: "gemini-2.5-pro" # Supports wildcards (e.g., "gemini-*")
#           protocol: "gemini" # restricts the rule to a specific protocol, options: openai, gemini, claude, codex, antigravity
#       params: # JSON paths (gjson/sjson syntax) to remove from the payload
#         - "generationConfig.thinkingConfig.thinkingBudget"
#         - "generationConfig.responseJsonSchema"
#     - models:
#         - name: "gpt-*" # Supports wildcards (e.g., "gpt-*")
#           protocol: "codex"
#       params:
#         # Flat tools array (standard OpenAI format or Codex CLI < 0.154):
#         - 'tools.#(name=="apply_patch")'
#         # Nested additional_tools structure in input[] (Codex CLI >= 0.154 "Responses Lite"):
#         - 'input.#(type=="additional_tools")#.tools.#(name=="functions")#.tools.#(name=="apply_patch")#'