# 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 portport: 8317 # TLS settings for HTTPS. When enabled, the server listens with the provided certificate and key.tls:  enable: false  cert: ""  key: "" # Management API settingsremote-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 authenticationapi-keys:  - "your-api-key-1"  - "your-api-key-2"  - "your-api-key-3" # Enable debug loggingdebug: 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 stdoutlogging-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 aggregationusage-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 # 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 behaviorquota-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:  # 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

Lines 1–500 of 989. Load more