openclacky 1.5.7 → 1.5.9

data/lib/clacky/message_format/anthropic.rb

@@ -181,10 +181,27 @@ module Clacky
           { id: tc["id"], type: "function", name: tc["name"], arguments: args }
         end
 
+        # Map Anthropic stop_reason → OpenAI-style finish_reason so downstream
+        # (llm_caller empty-response detector, agent loop, CostTracker) stays
+        # provider-agnostic. Each cluster shares the same semantic class:
+        #   stop     — natural end of turn (end_turn / pause_turn / stop_sequence)
+        #   length   — token or context-window limit hit (max_tokens /
+        #              model_context_window_exceeded; mapping the latter to
+        #              "length" also prevents a futile retry in llm_caller's
+        #              empty-response detector, which exempts "length")
+        #   content_filter — model declined (refusal)
+        #   other    — context compaction (compaction)
+        # Note: `compaction` is a real stop_reason but only emitted under the
+        # `compact-2026-01-12` beta header (server-side context compaction),
+        # so it is not yet in the stable SDK's StopReason literal list.
+        # See: https://platform.claude.com/docs/en/build-with-claude/compaction
+        # Unmapped values fall through unchanged.
         finish_reason = case data["stop_reason"]
-                        when "end_turn"   then "stop"
+                        when "end_turn", "pause_turn", "stop_sequence" then "stop"
                         when "tool_use"   then "tool_calls"
-                        when "max_tokens" then "length"
+                        when "max_tokens", "model_context_window_exceeded" then "length"
+                        when "refusal"    then "content_filter"
+                        when "compaction" then "other"
                         else data["stop_reason"]
                         end
 

data/lib/clacky/message_format/bedrock.rb

@@ -118,11 +118,16 @@ module Clacky
           { id: tc["toolUseId"], type: "function", name: tc["name"], arguments: args }
         end
 
-        # Map Bedrock stopReason → canonical finish_reason
+        # Map Bedrock stopReason → canonical finish_reason (same clusters as
+        # the Anthropic direct adapter — see lib/clacky/message_format/anthropic.rb).
         finish_reason = case data["stopReason"]
-                        when "end_turn"   then "stop"
+                        when "end_turn", "pause_turn", "stop_sequence" then "stop"
                         when "tool_use"   then "tool_calls"
-                        when "max_tokens" then "length"
+                        when "max_tokens", "model_context_window_exceeded" then "length"
+                        when "refusal"    then "content_filter"
+                        # `compaction` is a beta stop_reason (compact-2026-01-12);
+                        # see anthropic.rb for the full note.
+                        when "compaction" then "other"
                         else data["stopReason"]
                         end
 

data/lib/clacky/message_history.rb

@@ -286,7 +286,7 @@ module Clacky
     private def estimate_content_tokens(content)
       case content
       when String
-        ascii_chars = content.scan(/[ -~]/).length
+        ascii_chars = content.count(" -~")
         multibyte_chars = content.length - ascii_chars
         ((ascii_chars / 4.0) + (multibyte_chars / 1.5)).ceil
       when Array

data/lib/clacky/providers.rb

@@ -253,32 +253,34 @@ module Clacky
         "website_url" => "https://platform.deepseek.com/api_keys"
       }.freeze,
 
-      "minimax" => {
-        "name" => "Minimax",
-        "base_url" => "https://api.minimaxi.com/v1",
+      "glm" => {
+        "name" => "GLM (Z.ai / Zhipu)",
+        "base_url" => "https://open.bigmodel.cn/api/paas/v4",
         "api" => "openai-completions",
-        "default_model" => "MiniMax-M3",
-        "models" => ["MiniMax-M3", "MiniMax-M2.7", "MiniMax-M2.5"],
-        # MiniMax operates two regional endpoints with identical APIs & model
-        # lineup — mainland China (.com) and international (.io). Listing both
-        # lets find_by_base_url identify either one as provider "minimax",
-        # so capability checks (vision=false) fire correctly regardless of
-        # which endpoint the user configured.
+        "default_model" => "glm-5.2",
+        "models" => ["glm-5.2", "glm-5.1", "glm-5", "glm-5-turbo", "glm-5v-turbo", "glm-4.7"],
+        # Zhipu / Z.ai expose four functionally-equivalent endpoints:
+        # two regional sites (mainland open.bigmodel.cn + international api.z.ai)
+        # each with a general-billing and a Coding-Plan subpath. They share the
+        # same model lineup & identical capability profile, so a single preset
+        # with endpoint_variants is the right shape — one source of truth for
+        # vision/model_capabilities, four URLs recognised by find_by_base_url.
+        # Without this, users pointing at api.z.ai or the /coding/ path fell
+        # through to the conservative "assume vision=true" default and got
+        # hallucinated image descriptions on text-only GLM models (C-5563).
         "endpoint_variants" => [
-          { "label" => "Mainland China", "label_key" => "settings.models.baseurl.variant.mainland_cn",   "base_url" => "https://api.minimaxi.com/v1", "region" => "cn"   }.freeze,
-          { "label" => "International",  "label_key" => "settings.models.baseurl.variant.international", "base_url" => "https://api.minimax.io/v1",   "region" => "intl" }.freeze
+          { "label" => "Mainland · Pay-as-you-go",      "label_key" => "settings.models.baseurl.variant.mainland_cn_payg",    "base_url" => "https://open.bigmodel.cn/api/paas/v4",        "region" => "cn"   }.freeze,
+          { "label" => "Mainland · Coding Plan",        "label_key" => "settings.models.baseurl.variant.mainland_cn_coding",  "base_url" => "https://open.bigmodel.cn/api/coding/paas/v4", "region" => "cn"   }.freeze,
+          { "label" => "International · Pay-as-you-go", "label_key" => "settings.models.baseurl.variant.international_payg",  "base_url" => "https://api.z.ai/api/paas/v4",                "region" => "intl" }.freeze,
+          { "label" => "International · Coding Plan",   "label_key" => "settings.models.baseurl.variant.international_coding","base_url" => "https://api.z.ai/api/coding/paas/v4",         "region" => "intl" }.freeze
         ].freeze,
-        # MiniMax M2.5/M2.7 are text-only on this endpoint. M3 (released 2026-06-01)
-        # is natively multimodal — it accepts image input via OpenAI-style
-        # image_url content parts — so it overrides the provider-level
-        # vision=false below. M3 exposes a 1,000,000-token context window
-        # (M2.7: 204,800; M2.5: 40,960).
+        # GLM models are text-only except glm-5v-turbo which is vision-capable ("v" = visual).
         "capabilities" => { "vision" => false }.freeze,
         "model_capabilities" => {
-          "MiniMax-M3" => { "vision" => true }.freeze
+          "glm-5v-turbo" => { "vision" => true }.freeze
         }.freeze,
-        "default_ocr_model" => "MiniMax-M3",
-        "website_url" => "https://platform.minimax.io/"
+        "default_ocr_model" => "glm-5v-turbo",
+        "website_url" => "https://open.bigmodel.cn/console/overview"
       }.freeze,
 
       "kimi" => {
@@ -346,6 +348,34 @@ module Clacky
         "website_url" => "https://www.kimi.com/code"
       }.freeze,
 
+      "minimax" => {
+        "name" => "Minimax",
+        "base_url" => "https://api.minimaxi.com/v1",
+        "api" => "openai-completions",
+        "default_model" => "MiniMax-M3",
+        "models" => ["MiniMax-M3", "MiniMax-M2.7", "MiniMax-M2.5"],
+        # MiniMax operates two regional endpoints with identical APIs & model
+        # lineup — mainland China (.com) and international (.io). Listing both
+        # lets find_by_base_url identify either one as provider "minimax",
+        # so capability checks (vision=false) fire correctly regardless of
+        # which endpoint the user configured.
+        "endpoint_variants" => [
+          { "label" => "Mainland China", "label_key" => "settings.models.baseurl.variant.mainland_cn",   "base_url" => "https://api.minimaxi.com/v1", "region" => "cn"   }.freeze,
+          { "label" => "International",  "label_key" => "settings.models.baseurl.variant.international", "base_url" => "https://api.minimax.io/v1",   "region" => "intl" }.freeze
+        ].freeze,
+        # MiniMax M2.5/M2.7 are text-only on this endpoint. M3 (released 2026-06-01)
+        # is natively multimodal — it accepts image input via OpenAI-style
+        # image_url content parts — so it overrides the provider-level
+        # vision=false below. M3 exposes a 1,000,000-token context window
+        # (M2.7: 204,800; M2.5: 40,960).
+        "capabilities" => { "vision" => false }.freeze,
+        "model_capabilities" => {
+          "MiniMax-M3" => { "vision" => true }.freeze
+        }.freeze,
+        "default_ocr_model" => "MiniMax-M3",
+        "website_url" => "https://platform.minimax.io/"
+      }.freeze,
+
       "anthropic" => {
         "name" => "Anthropic (Claude)",
         "base_url" => "https://api.anthropic.com",
@@ -356,6 +386,70 @@ module Clacky
         "website_url" => "https://console.anthropic.com/settings/keys"
       }.freeze,
 
+      "openai" => {
+        "name" => "OpenAI (GPT)",
+        "base_url" => "https://api.openai.com/v1",
+        "api" => "openai-completions",
+        "default_model" => "gpt-5.5",
+        "models" => [
+          "gpt-5.5",
+          "gpt-5.4",
+          "gpt-5.4-mini",
+          "gpt-5.4-nano",
+          "o4-mini",
+          "o3"
+        ],
+        # GPT-5.x and o-series models are multimodal (text + image input).
+        "capabilities" => { "vision" => true }.freeze,
+        # Per-primary lite pairing: subagents use mini/nano for cheap/fast work.
+        # o4-mini and o3 are reasoning models without a lite-tier sibling here.
+        "lite_models" => {
+          "gpt-5.5" => "gpt-5.4-mini",
+          "gpt-5.4" => "gpt-5.4-mini"
+        },
+        # OpenAI's image generation model — same /v1/images/generations
+        # endpoint, so the OpenAICompat image provider handles it.
+        "image_models" => [
+          "gpt-image-2"
+        ],
+        "default_image_model" => "gpt-image-2",
+        "default_ocr_model" => "gpt-5.4-mini",
+        "website_url" => "https://platform.openai.com/api-keys"
+      }.freeze,
+
+      "qwen" => {
+        "name" => "Qwen (Alibaba)",
+        "base_url" => "https://dashscope.aliyuncs.com/compatible-mode/v1",
+        "api" => "openai-completions",
+        "default_model" => "qwen3.7-max",
+        "models" => [
+          "qwen3.7-max",
+          "qwen3.6-plus",
+          "qwen3.6-max",
+          "qwen3.6-27b",
+          "qwen3.6-flash",
+          "qwen-plus-latest",
+        ],
+        "endpoint_variants" => [
+          { "label" => "Mainland China",  "label_key" => "settings.models.baseurl.variant.mainland_cn",   "base_url" => "https://dashscope.aliyuncs.com/compatible-mode/v1",     "region" => "cn"   }.freeze,
+          { "label" => "Singapore",       "label_key" => "settings.models.baseurl.variant.international", "base_url" => "https://dashscope-intl.aliyuncs.com/compatible-mode/v1", "region" => "intl" }.freeze,
+          { "label" => "US (Virginia)",   "label_key" => "settings.models.baseurl.variant.us",            "base_url" => "https://dashscope-us.aliyuncs.com/compatible-mode/v1",   "region" => "us"   }.freeze
+        ].freeze,
+        "capabilities" => { "vision" => true }.freeze,
+        "model_capabilities" => {
+          "qwen3.7-max" => { "vision" => false }.freeze
+        }.freeze,
+        "default_ocr_model" => "qwen3.6-flash",
+        "lite_models" => {
+          "qwen3.7-max"      => "qwen3.6-flash",
+          "qwen3.6-plus"     => "qwen3.6-flash",
+          "qwen3.6-max"      => "qwen3.6-flash",
+          "qwen3.6-27b"      => "qwen3.6-flash",
+          "qwen-plus-latest" => "qwen3.6-flash"
+        },
+        "website_url" => "https://bailian.console.aliyun.com/?apiKey=1"
+      }.freeze,
+
       "mimo" => {
         "name" => "MiMo (Xiaomi)",
         "base_url" => "https://api.xiaomimimo.com/v1",
@@ -388,36 +482,6 @@ module Clacky
         "website_url" => "https://platform.xiaomimimo.com/"
       }.freeze,
 
-      "glm" => {
-        "name" => "GLM (Z.ai / Zhipu)",
-        "base_url" => "https://open.bigmodel.cn/api/paas/v4",
-        "api" => "openai-completions",
-        "default_model" => "glm-5.2",
-        "models" => ["glm-5.2", "glm-5.1", "glm-5", "glm-5-turbo", "glm-5v-turbo", "glm-4.7"],
-        # Zhipu / Z.ai expose four functionally-equivalent endpoints:
-        # two regional sites (mainland open.bigmodel.cn + international api.z.ai)
-        # each with a general-billing and a Coding-Plan subpath. They share the
-        # same model lineup & identical capability profile, so a single preset
-        # with endpoint_variants is the right shape — one source of truth for
-        # vision/model_capabilities, four URLs recognised by find_by_base_url.
-        # Without this, users pointing at api.z.ai or the /coding/ path fell
-        # through to the conservative "assume vision=true" default and got
-        # hallucinated image descriptions on text-only GLM models (C-5563).
-        "endpoint_variants" => [
-          { "label" => "Mainland · Pay-as-you-go",      "label_key" => "settings.models.baseurl.variant.mainland_cn_payg",    "base_url" => "https://open.bigmodel.cn/api/paas/v4",        "region" => "cn"   }.freeze,
-          { "label" => "Mainland · Coding Plan",        "label_key" => "settings.models.baseurl.variant.mainland_cn_coding",  "base_url" => "https://open.bigmodel.cn/api/coding/paas/v4", "region" => "cn"   }.freeze,
-          { "label" => "International · Pay-as-you-go", "label_key" => "settings.models.baseurl.variant.international_payg",  "base_url" => "https://api.z.ai/api/paas/v4",                "region" => "intl" }.freeze,
-          { "label" => "International · Coding Plan",   "label_key" => "settings.models.baseurl.variant.international_coding","base_url" => "https://api.z.ai/api/coding/paas/v4",         "region" => "intl" }.freeze
-        ].freeze,
-        # GLM models are text-only except glm-5v-turbo which is vision-capable ("v" = visual).
-        "capabilities" => { "vision" => false }.freeze,
-        "model_capabilities" => {
-          "glm-5v-turbo" => { "vision" => true }.freeze
-        }.freeze,
-        "default_ocr_model" => "glm-5v-turbo",
-        "website_url" => "https://open.bigmodel.cn/usercenter/apikeys"
-      }.freeze,
-
       # Volcengine Ark (Doubao) — ByteDance's model platform, OpenAI-compatible.
       # Exposes three functionally-equivalent endpoints (Pay-as-you-go / Coding
       # Plan / Agent Plan) that share the same model lineup and capability
@@ -472,7 +536,8 @@ module Clacky
         "model_capabilities" => {
           "glm-5.2"           => { "vision" => false }.freeze,
           "deepseek-v4-pro"   => { "vision" => false }.freeze,
-          "deepseek-v4-flash" => { "vision" => false }.freeze
+          "deepseek-v4-flash" => { "vision" => false }.freeze,
+          "minimax-m2.7"      => { "vision" => false }.freeze
         }.freeze,
         "default_ocr_model" => "doubao-seed-2.0-lite",
         "website_url" => "https://console.volcengine.com/ark/region:cn-beijing/overview"
@@ -539,68 +604,57 @@ module Clacky
         "website_url" => "https://ollama.com/settings/keys"
       }.freeze,
 
-      "openai" => {
-        "name" => "OpenAI (GPT)",
-        "base_url" => "https://api.openai.com/v1",
+      "orcarouter" => {
+        "name" => "OrcaRouter",
+        "base_url" => "https://api.orcarouter.ai/v1",
         "api" => "openai-completions",
-        "default_model" => "gpt-5.5",
+        "default_model" => "openai/gpt-5.5",
+        # Curated default lineup. OrcaRouter exposes 190+ models from OpenAI,
+        # Anthropic, Google, DeepSeek, Qwen, MiniMax, Zhipu and others behind a
+        # single OpenAI-compatible endpoint. Shipping a small list of the
+        # mainstream Claude + GPT entries gives users a working dropdown out
+        # of the box; users can still type any other OrcaRouter model id
+        # manually (e.g. "orcarouter/auto" for request-level auto-routing).
         "models" => [
-          "gpt-5.5",
-          "gpt-5.4",
-          "gpt-5.4-mini",
-          "gpt-5.4-nano",
-          "o4-mini",
-          "o3"
+          "anthropic/claude-sonnet-5",
+          "anthropic/claude-opus-4.8",
+          "anthropic/claude-haiku-4.5",
+          "openai/gpt-5.5",
+          "openai/gpt-5.4",
+          "openai/gpt-5.4-mini",
+          "google/gemini-3.5-flash",
+          "deepseek/deepseek-v4-flash",
+          "z-ai/glm-5.2",
+          "orcarouter/auto"
         ],
-        # GPT-5.x and o-series models are multimodal (text + image input).
-        "capabilities" => { "vision" => true }.freeze,
-        # Per-primary lite pairing: subagents use mini/nano for cheap/fast work.
-        # o4-mini and o3 are reasoning models without a lite-tier sibling here.
+        # Per-primary lite pairing — Claude family pairs with Haiku, GPT
+        # family pairs with the mini variant. Mirrors the openrouter preset
+        # so subagents on OrcaRouter get a sensible cheap/fast sidekick.
         "lite_models" => {
-          "gpt-5.5" => "gpt-5.4-mini",
-          "gpt-5.4" => "gpt-5.4-mini"
+          "anthropic/claude-sonnet-5" => "anthropic/claude-haiku-4.5",
+          "anthropic/claude-opus-4.8" => "anthropic/claude-haiku-4.5",
+          "openai/gpt-5.5"            => "openai/gpt-5.4-mini",
+          "openai/gpt-5.4"            => "openai/gpt-5.4-mini"
         },
-        # OpenAI's image generation model — same /v1/images/generations
-        # endpoint, so the OpenAICompat image provider handles it.
-        "image_models" => [
-          "gpt-image-2"
-        ],
-        "default_image_model" => "gpt-image-2",
-        "default_ocr_model" => "gpt-5.4-mini",
-        "website_url" => "https://platform.openai.com/api-keys"
-      }.freeze,
-
-      "qwen" => {
-        "name" => "Qwen (Alibaba)",
-        "base_url" => "https://dashscope.aliyuncs.com/compatible-mode/v1",
-        "api" => "openai-completions",
-        "default_model" => "qwen3.7-max",
-        "models" => [
-          "qwen3.7-max",
-          "qwen3.6-plus",
-          "qwen3.6-max",
-          "qwen3.6-27b",
-          "qwen3.6-flash",
-          "qwen-plus-latest",
-        ],
-        "endpoint_variants" => [
-          { "label" => "Mainland China",  "label_key" => "settings.models.baseurl.variant.mainland_cn",   "base_url" => "https://dashscope.aliyuncs.com/compatible-mode/v1",     "region" => "cn"   }.freeze,
-          { "label" => "Singapore",       "label_key" => "settings.models.baseurl.variant.international", "base_url" => "https://dashscope-intl.aliyuncs.com/compatible-mode/v1", "region" => "intl" }.freeze,
-          { "label" => "US (Virginia)",   "label_key" => "settings.models.baseurl.variant.us",            "base_url" => "https://dashscope-us.aliyuncs.com/compatible-mode/v1",   "region" => "us"   }.freeze
-        ].freeze,
+        # Per-model API type overrides. OrcaRouter proxies Claude through a
+        # native Anthropic /v1/messages endpoint (https://api.orcarouter.ai/v1/messages)
+        # in addition to its OpenAI-compatible /chat/completions endpoint.
+        # Routing "anthropic/*" via the native endpoint preserves cache_control
+        # fidelity, matching what Claude Code CLI does internally (same rationale
+        # as the openrouter preset). Non-Claude models keep the OpenAI shim.
+        "model_api_overrides" => {
+          /\Aanthropic\// => "anthropic-messages"
+        }.freeze,
+        # Most models on OrcaRouter are vision-capable; DeepSeek / GLM-5.2 and
+        # the "orcarouter/auto" router are text-only.
         "capabilities" => { "vision" => true }.freeze,
         "model_capabilities" => {
-          "qwen3.7-max" => { "vision" => false }.freeze
+          "deepseek/deepseek-v4-flash" => { "vision" => false }.freeze,
+          "z-ai/glm-5.2"               => { "vision" => false }.freeze,
+          "orcarouter/auto"            => { "vision" => false }.freeze
         }.freeze,
-        "default_ocr_model" => "qwen3.6-flash",
-        "lite_models" => {
-          "qwen3.7-max"      => "qwen3.6-flash",
-          "qwen3.6-plus"     => "qwen3.6-flash",
-          "qwen3.6-max"      => "qwen3.6-flash",
-          "qwen3.6-27b"      => "qwen3.6-flash",
-          "qwen-plus-latest" => "qwen3.6-flash"
-        },
-        "website_url" => "https://bailian.console.aliyun.com/?apiKey=1"
+        "default_ocr_model" => "google/gemini-3.5-flash",
+        "website_url" => "https://www.orcarouter.ai"
       }.freeze
 
     }.freeze

data/lib/clacky/server/channel/adapters/feishu/ws_client.rb

@@ -71,6 +71,12 @@ module Clacky
               ssl_context = OpenSSL::SSL::SSLContext.new
               ssl_context.set_params(verify_mode: OpenSSL::SSL::VERIFY_PEER)
               ssl = OpenSSL::SSL::SSLSocket.new(tcp, ssl_context)
+              # Feishu/Lark WebSocket endpoints are served behind CDN edges that
+              # require SNI to select the correct certificate/backend. Some
+              # Ruby/OpenSSL combinations do not infer the hostname for manually
+              # created SSLSocket instances, which can make the peer abort the TLS
+              # handshake with `tlsv1 alert internal error`.
+              ssl.hostname = uri.host if ssl.respond_to?(:hostname=)
               ssl.sync_close = true
               ssl.connect
               ssl

data/lib/clacky/server/http_server.rb

@@ -184,7 +184,7 @@ module Clacky
         - 将大目标拆解为可执行的小步骤
       MD
 
-      def initialize(host: "127.0.0.1", port: 7070, agent_config:, client_factory:, brand_test: false, sessions_dir: nil, socket: nil, master_pid: nil)
+      def initialize(host: "127.0.0.1", port: 7070, agent_config:, client_factory:, brand_test: false, sessions_dir: nil, projects_file: nil, socket: nil, master_pid: nil)
         @host           = host
         @port           = port
         @agent_config   = agent_config
@@ -197,7 +197,7 @@ module Clacky
         @restart_script = File.expand_path($0)
         @restart_argv   = ARGV.dup
         @session_manager = Clacky::SessionManager.new(sessions_dir: sessions_dir)
-        @project_manager = Clacky::Server::ProjectManager.new
+        @project_manager = Clacky::Server::ProjectManager.new(projects_file: projects_file)
         @registry        = SessionRegistry.new(
           session_manager:  @session_manager,
           session_restorer: method(:build_session_from_data),
@@ -1840,6 +1840,9 @@ module Clacky
             requested_model: state["requested_model"],
             configured: state["configured"]
           }
+          if entry && entry["base_url"] && state["source"] != "custom"
+            out[t][:saved_custom] = { model: entry["model"], base_url: entry["base_url"] }
+          end
         end
 
         # Surface what the current default model can offer, even when the
@@ -1949,50 +1952,78 @@ module Clacky
       end
 
       def api_update_media_config(kind, req, res)
-        body = parse_json_body(req) || {}
+        update_sidecar_config(res, kind, parse_json_body(req) || {})
+      end
+
+      private def update_sidecar_config(res, kind, body)
         source = body["source"].to_s
         unless %w[off auto custom].include?(source)
           return json_response(res, 422, { error: "invalid source" })
         end
 
-        @agent_config.models.reject! { |m| m["type"] == kind }
+        existing = @agent_config.models.find { |m| m["type"] == kind }
 
         case source
         when "off"
-          @agent_config.models << {
-            "id"       => SecureRandom.uuid,
-            "type"     => kind,
-            "disabled" => true
-          }
-        when "auto"
-          override = body["model"].to_s.strip
-          unless override.empty?
+          if existing
+            existing["mode"] = "off"
+          else
             @agent_config.models << {
-              "id"    => SecureRandom.uuid,
-              "type"  => kind,
-              "model" => override
+              "id"   => SecureRandom.uuid,
+              "type" => kind,
+              "mode" => "off"
             }
           end
+        when "auto"
+          override = body["model"].to_s.strip
+          if existing
+            existing.delete("disabled")
+            existing["mode"] = "auto"
+            existing["model"] = override unless override.empty?
+          else
+            unless override.empty?
+              @agent_config.models << {
+                "id"    => SecureRandom.uuid,
+                "type"  => kind,
+                "mode"  => "auto",
+                "model" => override
+              }
+            end
+          end
         when "custom"
           model    = body["model"].to_s.strip
           base_url = body["base_url"].to_s.strip
           api_key  = body["api_key"].to_s
-          if model.empty? || base_url.empty? || api_key.empty? || api_key.include?("****")
+          if api_key.empty? || api_key.include?("****")
+            api_key = existing ? existing["api_key"].to_s : ""
+          end
+          if model.empty? || base_url.empty? || api_key.empty?
             return json_response(res, 422, { error: "model, base_url, api_key are required" })
           end
 
-          @agent_config.models << {
-            "id"               => SecureRandom.uuid,
-            "model"            => model,
-            "base_url"         => base_url,
-            "api_key"          => api_key,
-            "anthropic_format" => body["anthropic_format"] || false,
-            "type"             => kind
-          }
+          if existing
+            existing["model"]            = model
+            existing["base_url"]         = base_url
+            existing["api_key"]          = api_key
+            existing["anthropic_format"] = body["anthropic_format"] || false
+            existing["mode"]             = "custom"
+            existing.delete("disabled")
+          else
+            @agent_config.models << {
+              "id"               => SecureRandom.uuid,
+              "model"            => model,
+              "base_url"         => base_url,
+              "api_key"          => api_key,
+              "anthropic_format" => body["anthropic_format"] || false,
+              "type"             => kind,
+              "mode"             => "custom"
+            }
+          end
         end
 
         @agent_config.save
-        json_response(res, 200, { ok: true, state: @agent_config.media_state(kind) })
+        state = kind == "ocr" ? @agent_config.ocr_state : @agent_config.media_state(kind)
+        json_response(res, 200, { ok: true, state: state })
       rescue => e
         json_response(res, 422, { error: e.message })
       end
@@ -2016,6 +2047,9 @@ module Clacky
           configured:     state["configured"],
           primary:        state["primary"] || false
         }
+        if entry && entry["base_url"] && state["source"] != "custom"
+          out[:saved_custom] = { model: entry["model"], base_url: entry["base_url"] }
+        end
 
         # Auto-mode preview: surface what the OCR sidecar *would* be if the
         # user flipped to "auto" — derived from the same provider as the
@@ -2037,58 +2071,8 @@ module Clacky
       # PATCH /api/config/ocr
       # Body: { source: "off"|"auto"|"custom", model?, base_url?, api_key?,
       #         anthropic_format? }
-      # Mirrors api_update_media_config but for the single "ocr" type.
       def api_update_ocr_config(req, res)
-        body = parse_json_body(req) || {}
-        source = body["source"].to_s
-        unless %w[off auto custom].include?(source)
-          return json_response(res, 422, { error: "invalid source" })
-        end
-
-        @agent_config.models.reject! { |m| m["type"] == "ocr" }
-
-        case source
-        when "off"
-          @agent_config.models << {
-            "id"       => SecureRandom.uuid,
-            "type"     => "ocr",
-            "disabled" => true
-          }
-        when "auto"
-          override = body["model"].to_s.strip
-          unless override.empty?
-            @agent_config.models << {
-              "id"    => SecureRandom.uuid,
-              "type"  => "ocr",
-              "model" => override
-            }
-          end
-        when "custom"
-          model    = body["model"].to_s.strip
-          base_url = body["base_url"].to_s.strip
-          api_key  = body["api_key"].to_s
-          if api_key.include?("****")
-            existing = @agent_config.models.find { |m| m["type"] == "ocr" && m["api_key"] }
-            api_key = existing ? existing["api_key"].to_s : ""
-          end
-          if model.empty? || base_url.empty? || api_key.empty?
-            return json_response(res, 422, { error: "model, base_url, api_key are required" })
-          end
-
-          @agent_config.models << {
-            "id"               => SecureRandom.uuid,
-            "model"            => model,
-            "base_url"         => base_url,
-            "api_key"          => api_key,
-            "anthropic_format" => body["anthropic_format"] || false,
-            "type"             => "ocr"
-          }
-        end
-
-        @agent_config.save
-        json_response(res, 200, { ok: true, state: @agent_config.ocr_state })
-      rescue => e
-        json_response(res, 422, { error: e.message })
+        update_sidecar_config(res, "ocr", parse_json_body(req) || {})
       end
 
       # POST /api/config/ocr/test
@@ -2562,7 +2546,7 @@ module Clacky
       # the same public marketplace here.
       def api_store_extensions(req, res)
         brand  = Clacky::BrandConfig.load
-        result = brand.search_extensions!(query: req.query["q"], sort: req.query["sort"])
+        result = brand.search_extensions!(query: req.query["q"], sort: req.query["sort"], page: req.query["page"], per_page: req.query["per_page"])
 
         if result[:success]
           installed = installed_extension_containers
@@ -2574,7 +2558,7 @@ module Clacky
               "installed_version" => container&.dig(:version)
             )
           end
-          json_response(res, 200, { ok: true, extensions: extensions })
+          json_response(res, 200, { ok: true, extensions: extensions, meta: result[:meta] })
         else
           json_response(res, 200, {
             ok:         true,
@@ -4521,7 +4505,11 @@ module Clacky
         upload_meta = Clacky::BrandConfig.load_upload_meta
         shadowed    = @skill_loader.shadowed_by_local
 
-        skills = @skill_loader.all_skills.reject(&:brand_skill).map do |skill|
+        skills = @skill_loader.all_skills.reject(&:brand_skill).reject do |skill|
+          # Third-party extension skills (installed/local) are managed from the
+          # extension panel, not the skills panel.
+          @skill_loader.loaded_from[skill.identifier] == :extension
+        end.map do |skill|
           source = @skill_loader.loaded_from[skill.identifier]
           meta   = upload_meta[skill.identifier] || {}