ask-agent 0.27.0 → 0.28.0

checksums.yaml

@@ -1,7 +1,7 @@
 ---
 SHA256:
-  metadata.gz: f67174ee80499aad6a21ff621f02c341daf0e992392996d577218aec2d8987ab
-  data.tar.gz: 25b21cc738b1e106730330b3ec28b118545f96a296898e4e9b3e337b4a2bf9a0
+  metadata.gz: 36e26d74ffa1c3e3e97b8e172f4fef2ada07c1c194488a8f550e8b77b7473ee1
+  data.tar.gz: 82568a4cc3ed19617293416d10f2f3194c1a18079a1426431a2d878afee75fa0
 SHA512:
-  metadata.gz: 74fdca8fecf20a3707bf162a319a0b4fa910f3718745fa8276235200e66e98ddd9a005cd7d4ce4bb5fc7d74321a1a86983934cc17041e6a20e58bce48156ce75
-  data.tar.gz: 78250a88b07fe2265e276ab5f156dd1c712315585eb906d81a198825c63450cf470ec48702bfbab2ac4b4d757523ff53a86d28646293e21809891626af4616c7
+  metadata.gz: cfd202f83bb95e3349282c3771e591f93062976b58c964d65e928e09a6408818ef969200c10b68c4a307e0c0d9fabde112aaf20b9776110a5040590370ecd317
+  data.tar.gz: 7d3ae82b872659e9e27217bc98cf7dfffee28f5209ee5f351dcc944564a63a777888fdad780de218956019d38f223996ebd6372c7be84cf33fdbf9a1eeb4e5a7

data/CHANGELOG.md

@@ -1,3 +1,44 @@
+## [0.28.0] — 2026-08-06
+
+### Changed (breaking)
+
+- **`Ask::Agent::Extensions` is now `Ask::Agent::Policies`.** The
+  tool-lifecycle policy classes — `ApprovalPolicy`, `Permissions`,
+  `RateLimiter`, `AuditLog` — moved from `lib/ask/agent/extensions/` to
+  `lib/ask/agent/policies/` and are namespaced under `Ask::Agent::Policies`.
+  The folder is now named after its seam (like `middleware`,
+  `stream_transforms`, `persistence`) instead of "extra stuff".
+  `Ask::Agent.load_extensions` → `Ask::Agent.load_policies`.
+  Update references: `Ask::Agent::Extensions::X` → `Ask::Agent::Policies::X`.
+
+  **What this means for the taxonomy:** policies are opt-in, replaceable
+  implementations of the tool-lifecycle hook seam — the agent loop runs
+  without them, and users can swap in their own implementations. Core
+  mechanisms are unchanged and stay on `Session`: the approval queue, the
+  `:pending` result status, and the `approval: true` option are core;
+  `Policies::ApprovalPolicy` is the reference classification policy wired on
+  top of them.
+
+### Added
+
+- **`Ask::Agent.load_policies`** — replaces `load_extensions` (same
+  behavior: eagerly requires every policy in the policies directory).
+
+## [0.27.1] — 2026-08-06
+
+### Fixed
+
+- **`chat.ask` / `chat.stream.ask` events now measure real LLM latency.** The
+  event was emitted after the call without a block, so `event.duration` was
+  ~0ms and duration metrics (e.g. `ask_llm_duration_seconds`,
+  `llm.duration_ms` spans) were meaningless. The provider call now runs
+  inside the instrument block; tokens/cost/tool_calls are enriched through a
+  shared nested `usage` payload hash (known only after the call returns) and
+  subscribers read it from there. Instrumentation failures can no longer
+  fail an `ask` — a wrapper error before the call falls through and runs the
+  call without telemetry, and a subscriber error after success returns the
+  response.
+
 ## [0.27.0] — 2026-08-06
 
 ### Added

data/lib/ask/agent/chat.rb

@@ -73,8 +73,6 @@ module Ask
           }.compact
         )
 
-        emit_instrumentation(stream, response_msg)
-
         response_msg
       end
 
@@ -208,20 +206,35 @@ module Ask
           begin
             req = build_request(stream)
 
-            result = if @middleware_pipeline
-              @middleware_pipeline.invoke(provider, req) do
+            # The chat.ask event wraps the actual provider call so
+            # event.duration measures the true LLM latency. Tokens/cost are
+            # only known once the call returns, so the (mutable) payload is
+            # passed into the block and enriched there — before the event's
+            # finish fires.
+            response = instrument_llm_call(stream) do |payload|
+              result = if @middleware_pipeline
+                @middleware_pipeline.invoke(provider, req) do
+                  call_provider(req, calls_acc, &block)
+                end
+              else
                 call_provider(req, calls_acc, &block)
               end
-            else
-              call_provider(req, calls_acc, &block)
-            end
 
-            # Flush any buffered stream transforms (e.g. TextBuffer)
-            if block && @transform_pipeline
-              flush_transforms(&block)
+              # Flush any buffered stream transforms (e.g. TextBuffer)
+              if block && @transform_pipeline
+                flush_transforms(&block)
+              end
+
+              response = build_response_from_result(result, calls_acc, stream)
+              usage = payload[:usage] ||= {}
+              usage[:input_tokens] = response.input_tokens
+              usage[:output_tokens] = response.output_tokens
+              usage[:cost] = response.cost
+              usage[:tool_calls] = response.tool_call?
+              response
             end
 
-            return build_response_from_result(result, calls_acc, stream)
+            return response
           rescue Ask::RateLimitError => e
             raise if attempt >= MAX_CHAT_RETRIES - 1
 
@@ -357,28 +370,47 @@ module Ask
         nil
       end
 
-      def emit_instrumentation(stream, response_msg)
-        return unless defined?(Ask::Instrumentation)
-
+      # Run the provider call inside the chat.ask event so event.duration
+      # measures the real LLM latency. The event payload is yielded to the
+      # block so the caller can enrich it (tokens, cost) before the event
+      # finishes.
+      #
+      # Instrumentation must never break the chat loop: a subscriber error
+      # after the call succeeded is swallowed (the response is returned);
+      # an error before the block ran falls through and runs the call
+      # without telemetry. Only real LLM errors propagate.
+      def instrument_llm_call(stream)
         payload = {
           model: @model_id,
           provider: @model_info.provider,
-          input_tokens: response_msg.input_tokens,
-          output_tokens: response_msg.output_tokens,
-          cost: response_msg.cost,
-          tool_calls: response_msg.tool_call?,
           stream: stream,
           middleware: @middleware_pipeline&.configured?,
-          stream_transforms: @transform_pipeline&.configured?
+          stream_transforms: @transform_pipeline&.configured?,
+          # Tokens/cost are only known once the call returns, and
+          # instrument() copies the payload shallowly — this nested hash is
+          # shared with the event, so in-block enrichment is visible to
+          # subscribers at finish time.
+          usage: {}
         }.compact
 
-        if stream
-          Ask::Instrumentation.instrument("chat.stream.ask", payload)
+        called = false
+        result = nil
+        if defined?(Ask::Instrumentation)
+          Ask::Instrumentation.instrument(stream ? "chat.stream.ask" : "chat.ask", payload) do
+            called = true
+            result = yield(payload)
+          end
         else
-          Ask::Instrumentation.instrument("chat.ask", payload)
+          called = true
+          result = yield(payload)
         end
+        result
       rescue StandardError
-        nil
+        return result if called && !result.nil?
+
+        raise if called
+
+        yield({})
       end
     end
   end

data/lib/ask/agent/{extensions → policies}/approval_policy.rb

@@ -2,7 +2,7 @@
 
 module Ask
   module Agent
-    module Extensions
+    module Policies
       # Approval policy hook: classifies tool calls as approval-required and
       # routes them into an {Ask::Agent::ApprovalQueue}.
       #
@@ -15,7 +15,7 @@ module Ask
       #
       # @example
       #   queue = Ask::Agent::ApprovalQueue.new
-      #   policy = Ask::Agent::Extensions::ApprovalPolicy.new(queue: queue)
+      #   policy = Ask::Agent::Policies::ApprovalPolicy.new(queue: queue)
       #   session = Ask::Agent::Session.new(
       #     model: "gpt-4o",
       #     tools: [SendEmail],

data/lib/ask/agent/{extensions → policies}/audit_log/active_record_writer.rb

@@ -3,7 +3,7 @@ require "time"
 
 module Ask
   module Agent
-    module Extensions
+    module Policies
       class AuditLog
         # ActiveRecord adapter for the audit log.
         # Auto-creates the +ask_audit_logs+ table on first write using

data/lib/ask/agent/{extensions → policies}/audit_log.rb

@@ -5,7 +5,7 @@ require "time"
 
 module Ask
   module Agent
-    module Extensions
+    module Policies
       # Event-driven audit log for agent sessions.
       #
       # Subscribes to all session events and writes them to a configurable
@@ -93,16 +93,16 @@ module Ask
 
           case adapter
           when :active_record
-            require "ask/agent/extensions/audit_log/active_record_writer"
+            require "ask/agent/policies/audit_log/active_record_writer"
             AuditLog::ActiveRecordWriter.new
           when Hash
             resolve(adapter[:adapter] || adapter[:writer])
           when Symbol, String
             # Try to load adapter by convention:
-            # :active_record → ask/agent/extensions/audit_log/active_record_writer
+            # :active_record → ask/agent/policies/audit_log/active_record_writer
             name = adapter.to_s
             begin
-              require "ask/agent/extensions/audit_log/#{name}_writer"
+              require "ask/agent/policies/audit_log/#{name}_writer"
               klass_name = name.split("_").map(&:capitalize).join
               klass = AuditLog.const_get(klass_name)
               klass.new

data/lib/ask/agent/{extensions → policies}/permissions.rb

@@ -2,7 +2,7 @@
 
 module Ask
   module Agent
-    module Extensions
+    module Policies
       class Permissions
         DEFAULT_TOOLS = %i[write edit bash destroy].freeze
 

data/lib/ask/agent/{extensions → policies}/rate_limiter.rb

@@ -2,7 +2,7 @@
 
 module Ask
   module Agent
-    module Extensions
+    module Policies
       class RateLimiter
         def initialize(max_calls_per_minute: 20, max_tool_calls_per_turn: 5)
           @max_calls_per_minute = max_calls_per_minute

data/lib/ask/agent/session.rb

@@ -421,7 +421,7 @@ module Ask
       def build_audit_log(config)
           config ||= Ask::Agent.configuration.audit_log
           return nil unless config
-          Ask::Agent::Extensions::AuditLog.new(self, adapter: config)
+          Ask::Agent::Policies::AuditLog.new(self, adapter: config)
         end
 
       # Build the approval queue + policy when approval is enabled.
@@ -451,7 +451,7 @@ module Ask
           )
         end
 
-        policy = Ask::Agent::Extensions::ApprovalPolicy.new(
+        policy = Ask::Agent::Policies::ApprovalPolicy.new(
           queue: queue,
           require_approval: policy_opts[:require_approval],
           tools: @tools

data/lib/ask/agent/version.rb

@@ -2,6 +2,6 @@
 
 module Ask
   module Agent
-    VERSION = "0.27.0"
+    VERSION = "0.28.0"
   end
 end

data/lib/ask/agent.rb

@@ -21,11 +21,20 @@ module Ask
 
     class UnknownAgent < Error; end
 
-    module Extensions
-      autoload :Permissions, "ask/agent/extensions/permissions"
-      autoload :RateLimiter, "ask/agent/extensions/rate_limiter"
-      autoload :AuditLog, "ask/agent/extensions/audit_log"
-      autoload :ApprovalPolicy, "ask/agent/extensions/approval_policy"
+    # Policies are opt-in, replaceable implementations of the tool-lifecycle
+    # hook seam (before_tool / after_tool). The agent loop runs without them
+    # and their semantics are unchanged by their absence; users compose and
+    # swap them freely (see Ask::Agent::Hooks for the seam).
+    #
+    # A policy is NOT core machinery. Core mechanisms live on Session — the
+    # approval queue, the :pending result status, and the `approval: true`
+    # option are core; Policies::ApprovalPolicy is the reference
+    # classification policy wired on top of them.
+    module Policies
+      autoload :Permissions, "ask/agent/policies/permissions"
+      autoload :RateLimiter, "ask/agent/policies/rate_limiter"
+      autoload :AuditLog, "ask/agent/policies/audit_log"
+      autoload :ApprovalPolicy, "ask/agent/policies/approval_policy"
     end
 
     module Middleware
@@ -236,8 +245,8 @@ module Ask
       yield configuration
     end
 
-    def self.load_extensions
-      Dir[File.expand_path("agent/extensions/*.rb", __dir__)].each { |f| require f }
+    def self.load_policies
+      Dir[File.expand_path("agent/policies/*.rb", __dir__)].each { |f| require f }
     rescue Errno::ENOENT
     end
   end

metadata

@@ -1,7 +1,7 @@
 --- !ruby/object:Gem::Specification
 name: ask-agent
 version: !ruby/object:Gem::Version
-  version: 0.27.0
+  version: 0.28.0
 platform: ruby
 authors:
 - Kaka Ruto
@@ -174,11 +174,6 @@ files:
 - lib/ask/agent/definition.rb
 - lib/ask/agent/evaluator.rb
 - lib/ask/agent/events.rb
-- lib/ask/agent/extensions/approval_policy.rb
-- lib/ask/agent/extensions/audit_log.rb
-- lib/ask/agent/extensions/audit_log/active_record_writer.rb
-- lib/ask/agent/extensions/permissions.rb
-- lib/ask/agent/extensions/rate_limiter.rb
 - lib/ask/agent/hooks.rb
 - lib/ask/agent/loop.rb
 - lib/ask/agent/meta_agent.rb
@@ -190,6 +185,11 @@ files:
 - lib/ask/agent/middleware/retry_on_failure.rb
 - lib/ask/agent/persistence/base.rb
 - lib/ask/agent/persistence/in_memory.rb
+- lib/ask/agent/policies/approval_policy.rb
+- lib/ask/agent/policies/audit_log.rb
+- lib/ask/agent/policies/audit_log/active_record_writer.rb
+- lib/ask/agent/policies/permissions.rb
+- lib/ask/agent/policies/rate_limiter.rb
 - lib/ask/agent/reflector.rb
 - lib/ask/agent/scheduler.rb
 - lib/ask/agent/session.rb