@hsupu/copilot-api 0.7.20 → 0.7.22

package/config.example.yaml

@@ -5,12 +5,9 @@
 # ============================================================================
 # Proxy
 # ============================================================================
-# Proxy URL for all outgoing requests to GitHub / Copilot APIs.
-# Supports http://, https://, socks5://, socks5h:// schemes.
-# socks5h:// routes DNS through the proxy (recommended for privacy).
-# Authentication via URL credentials: socks5h://user:pass@host:port
-# Takes precedence over HTTP_PROXY/HTTPS_PROXY environment variables.
-# CLI --proxy flag takes precedence over this setting.
+# Proxy URL for all outgoing requests to GitHub / Copilot APIs. Supports http://, https://, socks5://, socks5h://.
+# socks5h:// routes DNS through the proxy (recommended for privacy). Auth via URL: socks5h://user:pass@host:port
+# Takes precedence over HTTP_PROXY/HTTPS_PROXY env vars. CLI --proxy flag takes precedence over this setting.
 # NOT hot-reloadable (requires restart).
 
 # proxy: "http://127.0.0.1:7890"
@@ -27,41 +24,44 @@
 #   - Short aliases: "opus", "sonnet", "haiku" (resolved to best available)
 #
 # Matching order: raw request name checked first, then resolved (normalized) name.
-# User overrides are deep-merged with built-in defaults (same key = user wins).
 #
 # Built-in defaults (always active unless overridden):
 #   opus   → claude-opus-4.6
 #   sonnet → claude-sonnet-4.6
 #   haiku  → claude-haiku-4.5
 #
-# If the target model is not in the available models list, it is treated as an
-# alias and resolved again. If still unavailable, the best model in the same
-# family is used as fallback.
+# If the target model is not in the available models list, it is treated as an alias and resolved again.
+# If still unavailable, the best model in the same family is used as fallback.
 
 model_overrides:
   opus: claude-opus-4.6-1m
-  sonnet: claude-opus-4.6-1m            # Redirect all sonnet requests to best opus
+  # Redirect all sonnet requests to best opus
+  # sonnet: claude-opus-4.6-1m
   haiku: claude-sonnet-4.6
   claude-opus-4.5: claude-opus-4.6-1m
   claude-opus-4.6: claude-opus-4.6-1m
   claude-opus-4-6: claude-opus-4.6-1m
-  claude-sonnet-4.5: claude-sonnet-4.6  # Latest sonnet
-  claude-haiku-4.5: claude-sonnet-4.5   # Upgrade haiku to sonnet
+  # Use latest sonnet
+  claude-sonnet-4.5: claude-sonnet-4.6
+  # Upgrade haiku to sonnet
+  claude-haiku-4.5: claude-sonnet-4.6
 
 # ============================================================================
 # Timeouts
 # ============================================================================
 # Timeout settings for upstream API connections. Apply to all streaming paths.
 
-stream_idle_timeout: 300       # Max seconds between SSE events (0 = no timeout).
-                               # Applies to all streaming paths (Anthropic, Chat Completions, Responses).
+# Max seconds between SSE events (0 = no timeout).
+# Applies to all streaming paths (Anthropic, Chat Completions, Responses).
+stream_idle_timeout: 300
 
-fetch_timeout: 300             # Seconds: request start → HTTP response headers (0 = no timeout).
-                               # Applies to all upstream API clients.
+# Max seconds from request start to HTTP response headers (0 = no timeout).
+# Applies to all upstream API clients.
+fetch_timeout: 600
 
-stale_request_max_age: 600     # Max seconds an active request can live before the stale reaper
-                               # forces it to fail (0 = disabled).
-                               # Safety net for requests that never complete/fail normally.
+# Max seconds an active request can live before the stale reaper forces it to fail (0 = disabled).
+# Safety net for requests that never complete/fail normally.
+stale_request_max_age: 600
 
 # ============================================================================
 # Shutdown
@@ -69,8 +69,10 @@ stale_request_max_age: 600     # Max seconds an active request can live before t
 # Control graceful shutdown timing.
 
 shutdown:
-  graceful_wait: 60   # Phase 2: seconds to wait for in-flight requests to complete naturally (default: 60)
-  abort_wait: 120     # Phase 3: seconds to wait after abort signal for handlers to wrap up (default: 120)
+  # Phase 2: seconds to wait for in-flight requests to complete naturally (default: 60)
+  graceful_wait: 300
+  # Phase 3: seconds to wait after abort signal for handlers to wrap up (default: 120)
+  abort_wait: 600
 
 # ============================================================================
 # History
@@ -78,8 +80,10 @@ shutdown:
 # Control history recording behavior.
 
 history:
-  limit: 200          # Maximum number of entries to keep in memory (0 = unlimited, default: 200)
-  min_entries: 50     # Minimum entries to keep even under memory pressure (default: 50)
+  # Maximum number of entries to keep in memory (0 = unlimited, default: 200)
+  limit: 200
+  # Minimum entries to keep even under memory pressure (default: 50)
+  min_entries: 50
 
 # ============================================================================
 # Anthropic
@@ -87,39 +91,68 @@ history:
 # Settings for Anthropic API tool handling and timeouts.
 
 anthropic:
-  strip_server_tools: false  # Strip server-side tools (web_search, etc.) from requests
-  dedup_tool_calls: false    # false | "input" | "result" (true = "input" for compat)
-                             # "input": dedup by (name, input); "result": also require identical result
-  strip_read_tool_result_tags: false  # Strip <system-reminder> tags from Read tool results
-  # rewrite_system_reminders: false   # false = keep all (default), true = remove all
-  rewrite_system_reminders:           # Or provide rewrite rules (first match wins, top-down).
-                                      # Note: `model` field is NOT supported here (only in system_prompt_overrides).
+  # Strip server-side tools (web_search, etc.) from requests
+  strip_server_tools: false
+
+  # Dedup tool calls: false | "input" | "result" (true = "input" for compat)
+  # "input": dedup by (name, input); "result": also require identical result
+  dedup_tool_calls: false
+
+  # Strip <system-reminder> tags from Read tool results
+  strip_read_tool_result_tags: false
+
+  # Server-side context editing mode. Controls how Anthropic's context_management trims older context when input grows large.
+  #   off: disabled (default). No context_management sent.
+  #   clear-thinking: clear old thinking blocks.
+  #   clear-tooluse: clear old tool_use/result pairs.
+  #   clear-both: apply both clear-thinking and clear-tooluse.
+  # Only effective for models that support context editing.
+  context_editing: off
+
+  # Rewrite system-reminder tags in messages. false = keep all (default), true = remove all.
+  # Or provide rewrite rules (first match wins, top-down).
+  # Note: `model` field is NOT supported here (only in system_prompt_overrides).
+  # rewrite_system_reminders: false
+  rewrite_system_reminders:
     - from: "^Whenever you read a file, you should consider whether it would be considered malware"
-      to: ""                              # Empty = remove the tag
-    # - from: ".*"                        # Catch-all: keep unchanged (gms flags are automatic)
-    #   to: "$0"                          # $0 = original content
+      # Empty = remove the tag
+      to: ""
+    # - from: ".*"           # Catch-all: keep unchanged (gms flags are automatic)
+    #   to: "$0"             # $0 = original content
+
+# ============================================================================
+# Responses API
+# ============================================================================
+# Settings for OpenAI Responses API (/v1/responses) endpoint.
+
+openai-responses:
+  # Convert `call_xxx` IDs to `fc_xxx` in input. Required when clients send conversation history with
+  # Chat Completions-format tool call IDs to the Responses API (which requires `fc_` prefix). Default: false.
+  normalize_call_ids: true
 
 # ============================================================================
 # Rate Limiter
 # ============================================================================
-# Fine-tune the adaptive rate limiter behavior.
-# The rate limiter is enabled/disabled via --no-rate-limit CLI flag;
+# Fine-tune the adaptive rate limiter behavior. The rate limiter is enabled/disabled via --no-rate-limit CLI flag;
 # these settings control its parameters when enabled.
 
 rate_limiter:
-  retry_interval: 10        # Seconds to wait before retrying after rate limit error
-  request_interval: 10      # Seconds between requests in rate-limited mode
-  recovery_timeout: 10      # Minutes before attempting recovery from rate-limited mode
-  consecutive_successes: 5   # Consecutive successes needed to exit rate-limited mode
+  # Seconds to wait before retrying after rate limit error
+  retry_interval: 10
+  # Seconds between requests in rate-limited mode
+  request_interval: 10
+  # Minutes before attempting recovery from rate-limited mode
+  recovery_timeout: 10
+  # Consecutive successes needed to exit rate-limited mode
+  consecutive_successes: 5
 
 # ============================================================================
 # Auto-Truncate
 # ============================================================================
 # Control auto-truncation behavior when context limits are hit.
 
-# Compress old tool_result content before truncating messages.
-# When enabled, large tool_result content blocks are compressed to reduce
-# context size before resorting to message removal. Default: true.
+# Compress old tool_result content before truncating messages. When enabled, large tool_result content blocks
+# are compressed to reduce context size before resorting to message removal. Default: true.
 # compress_tool_results_before_truncate: false
 
 # ============================================================================
@@ -130,8 +163,8 @@ rate_limiter:
 #   2. prepend    — text inserted before the system prompt
 #   3. append     — text inserted after the system prompt
 #
-# All three work across all endpoints (Anthropic Messages, Chat Completions,
-# Responses API). For OpenAI formats, they target system/developer messages.
+# All three work across all endpoints (Anthropic Messages, Chat Completions, Responses API).
+# For OpenAI formats, they target system/developer messages.
 
 # ── Prepend ──
 # For Anthropic: inserted as the first text block or prepended to the string.
@@ -195,16 +228,13 @@ rate_limiter:
 #   from:   the pattern to match
 #   to:     the replacement text
 #   method: "line" or "regex" (default: "regex")
-#   model:  regex pattern to filter by resolved model name (optional, case-insensitive)
-#           When set, the rule only applies to requests whose resolved model matches.
-#           When omitted, the rule applies to all models.
+#   model:  regex pattern to filter by resolved model name (optional, case-insensitive).
+#           When set, the rule only applies to requests whose resolved model matches. When omitted, applies to all.
 #
 # Methods:
-#   line  — splits text by newlines; replaces any line whose trimmed content
-#           exactly matches trimmed `from`
+#   line  — splits text by newlines; replaces any line whose trimmed content exactly matches trimmed `from`
 #   regex — applies regex on the full text block with gms flags (default if omitted)
-#           (g=global, m=multiline ^$ match line boundaries, s=dotAll . matches \n)
-#           Supports capture group placeholders: $1, $2, etc.
+#           (g=global, m=multiline ^$ match line boundaries, s=dotAll . matches \n). Supports $1, $2, etc.
 
 # Example: Security Research Mode overrides (removes refusal patterns)
 # system_prompt_overrides: