ace-llm 0.39.2 → 0.41.0

checksums.yaml

@@ -1,7 +1,7 @@
 ---
 SHA256:
-  metadata.gz: cb59cc8e89a88f8c35c88148d6364c109cff56aa65393a8c6998636cfe3217f8
-  data.tar.gz: 1f0850c65fcafd1bdda99e7e6c61f99267cd32a8414d7842ce315078a1ed5983
+  metadata.gz: 2fae0fd9e7709654101b452e68a39a32412c1788948f60ff6b3ea619ca4877d8
+  data.tar.gz: c5ecb76fe5e258e27e5c71ddd8d0f760637ca7a0f67766a7239fe2e60acf3dbf
 SHA512:
-  metadata.gz: 9c71aed42c8a25be6d78cb5da76d61b833632767165b65c51249509667fd9d1bd7a5fd3be46e8d16f3f7b1e11715dd0b2654bbe74565dfefcda13f4d5bddc8f2
-  data.tar.gz: 00bf1c6a53d6268a90613953c3434768ea82421b73e0bdf9814e4ab00cc885a59e7749c1efaa104bcde825bda577356fcc8dec2e0c073a87c010c5c34ac745f0
+  metadata.gz: a885b5933135a969eb4f91a4e013af1e9a35b1e79450c8f47b8ce2b9afa372226c482c1023bab1033438a04b2fc95fbc6ccf4f8592e5b5ce6b763d0820bfb429
+  data.tar.gz: b378d04c4280ede3e36d60b01fbe7ee821f28597ac0c9074745961126d642db7e822af19c44f0362d15eaf5067e3ed6b4ee9cf9a2779abf15652e979f9313114

data/.ace-defaults/llm/presets/codex/ro.yml

@@ -1,5 +1,4 @@
 timeout: 600
 cli_args:
-  - --full-auto
   - --sandbox
   - read-only

data/.ace-defaults/llm/presets/codex/rw.yml

@@ -1,3 +1,4 @@
 timeout: 600
 cli_args:
-  - --full-auto
+  - --sandbox
+  - workspace-write

data/.ace-defaults/llm/presets/pi/ro.yml

@@ -1 +1,5 @@
 timeout: 600
+cli_args:
+  - --tools
+  - read,grep,find,ls
+  - --no-extensions

data/.ace-defaults/llm/thinking/pi/high.yml

@@ -0,0 +1,3 @@
+cli_args:
+  - --thinking
+  - high

data/.ace-defaults/llm/thinking/pi/low.yml

@@ -0,0 +1,3 @@
+cli_args:
+  - --thinking
+  - low

data/.ace-defaults/llm/thinking/pi/max.yml

@@ -0,0 +1,3 @@
+cli_args:
+  - --thinking
+  - max

data/.ace-defaults/llm/thinking/pi/medium.yml

@@ -0,0 +1,3 @@
+cli_args:
+  - --thinking
+  - medium

data/.ace-defaults/llm/thinking/pi/xhigh.yml

@@ -0,0 +1,3 @@
+cli_args:
+  - --thinking
+  - xhigh

data/CHANGELOG.md

@@ -7,6 +7,23 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 ## [Unreleased]
 
+## [0.41.0] - 2026-09-27
+
+### Changed
+
+- Query results now derive the top-level `usage` record from provider metadata token counts (model, input/output/cached/total). Absent counts stay absent; a measured zero cache read is kept so cost models can distinguish "no cache" from "unknown". Sessions no longer report "usage unavailable" when the provider supplied tokens.
+
+## [0.40.0] - 2026-09-24
+
+### Fixed
+
+- Update Codex CLI presets for the new `codex exec` interface: replace the
+  removed `--full-auto` flag with explicit `--sandbox read-only` (ro) and
+  `--sandbox workspace-write` (rw); `yolo` keeps
+  `--dangerously-bypass-approvals-and-sandbox`.
+- Preserve resolved provider, model, preset and reasoning metadata for the executed request, including fallback targets.
+- Accept `max` reasoning in selectors and Pi reasoning configuration.
+
 ## [0.39.2] - 2026-09-13
 
 ### Fixed

data/README.md

@@ -35,7 +35,7 @@ Deterministic coverage lives in `test/fast/` and `test/feat/`. Scenario assets s
 
 ## How It Works
 
-1. Select a model -- by alias, `provider:model`, with a thinking level suffix (`:low`/`:medium`/`:high`), or an `@preset` -- and submit a prompt.
+1. Select a model -- by alias, `provider:model`, with a configured thinking level suffix (`:low`/`:medium`/`:high`/`:xhigh`/`:max`), or an `@preset` -- and submit a prompt.
 2. The provider router resolves the target through [ace-llm-providers-cli](../ace-llm-providers-cli) adapters, applying fallback and retry rules from the [config cascade](../ace-support-config).
 3. The response is returned as text, markdown, or JSON with optional token usage metadata.
 
@@ -43,7 +43,7 @@ Deterministic coverage lives in `test/fast/` and `test/feat/`. Scenario assets s
 
 **Switch providers with aliases** - use short names like `gflash`, `sonnet`, `opus` instead of full `provider:model` notation. Aliases resolve through versioned YAML in [`.ace-defaults/`](docs/usage.md).
 
-**Control reasoning depth** - append a thinking level (`codex:gpt:high`, `claude:sonnet:low`) to tune reasoning budgets. Supported CLI providers: `claude`, `codex` (levels: `low`, `medium`, `high`, `xhigh`).
+**Control reasoning depth** - append a thinking level (`codex:gpt:high`, `claude:sonnet:low`, `pi:glm5:max`) to tune reasoning budgets. Supported CLI providers: `claude`, `codex` (`low`, `medium`, `high`, `xhigh`) and `pi` (including `max`). Each provider/level pair requires a matching thinking configuration.
 
 **Run preset-driven prompts** - apply execution profiles with `@preset` or `--preset`. Built-in presets for CLI providers: `@ro` (read-only), `@rw` (read-write), `@yolo` (full autonomy). Supported by: `claude`, `codex`, `gemini`, `opencode`, `pi`, `agy`.
 

data/lib/ace/llm/molecules/fallback_orchestrator.rb

@@ -33,6 +33,7 @@ module Ace
         # @raise [Error] If all providers and retries exhausted
         def execute(primary_provider:, registry:)
           @start_time = Time.now
+          @primary_provider = primary_provider
           @visited_providers.clear
 
           # If fallback disabled, just execute with primary
@@ -84,6 +85,7 @@ module Ace
             client = get_client(provider_name, registry)
             return yield client, provider_name
           rescue => error
+            raise if error.is_a?(Ace::LLM::ConfigurationError) && provider_name != @primary_provider
             last_error = error
 
             # Handle the error - returns :retry or :stop_and_fallback

data/lib/ace/llm/molecules/provider_model_parser.rb

@@ -10,7 +10,7 @@ module Ace
       # ProviderModelParser handles parsing and validation of provider:model syntax
       # for the unified LLM query interface.
       class ProviderModelParser
-        THINKING_LEVELS = %w[low medium high xhigh].freeze
+        THINKING_LEVELS = %w[low medium high xhigh max].freeze
 
         # Result object for parsed provider:model combinations.
         ParseResult = Struct.new(:provider, :model, :preset, :thinking_level, :valid, :error,

data/lib/ace/llm/molecules/role_resolver.rb

@@ -8,7 +8,7 @@ module Ace
     module Molecules
       # RoleResolver maps role names to an available concrete selector.
       class RoleResolver
-        THINKING_LEVELS = %w[low medium high xhigh].freeze
+        THINKING_LEVELS = %w[low medium high xhigh max].freeze
 
         def initialize(registry: nil, configuration: nil)
           @registry = registry || ClientRegistry.new

data/lib/ace/llm/molecules/thinking_level_loader.rb

@@ -8,7 +8,7 @@ module Ace
     module Molecules
       # Loads provider-scoped thinking-level overrides from llm/thinking/*.yml.
       class ThinkingLevelLoader
-        ALLOWED_LEVELS = %w[low medium high xhigh].freeze
+        ALLOWED_LEVELS = %w[low medium high xhigh max].freeze
 
         class << self
           def load_for_provider(provider, level)

data/lib/ace/llm/query_interface.rb

@@ -1,5 +1,7 @@
 # frozen_string_literal: true
 
+require "shellwords"
+
 require_relative "molecules/client_registry"
 require_relative "molecules/provider_model_parser"
 require_relative "molecules/preset_loader"
@@ -145,11 +147,13 @@ module Ace
 
         result = {
           text: text_content,
-          model: final_model,
-          provider: parse_result.provider,
-          preset: resolved_preset,
-          thinking_level: parse_result.thinking_level,
-          usage: response[:usage],
+          model: response.dig(:execution, :model) || final_model,
+          provider: response.dig(:execution, :provider) || parse_result.provider,
+          preset: response.fetch(:execution, {}).fetch(:preset, resolved_preset),
+          thinking_level: response.fetch(:execution, {}).fetch(:thinking_level, parse_result.thinking_level),
+          execution: response[:execution],
+          requested_selector: provider_model,
+          usage: build_usage(response[:metadata]),
           metadata: response[:metadata]
         }
 
@@ -178,6 +182,27 @@ module Ace
 
       private
 
+      # Derive the measured usage record from client metadata. Providers hand
+      # token counts to session recording through metadata keys; absent counts
+      # stay absent (never zero) so cost models can tell "no cache" from
+      # "unknown". Returns nil when the client reported no token counts at all.
+      def self.build_usage(metadata)
+        meta = metadata.is_a?(Hash) ? metadata : {}
+        input = meta[:input_tokens] || meta["input_tokens"]
+        output = meta[:output_tokens] || meta["output_tokens"]
+        return nil if input.nil? && output.nil?
+
+        usage = {model: meta[:model] || meta["model"]}
+        usage[:input_tokens] = input unless input.nil?
+        usage[:output_tokens] = output unless output.nil?
+        cached = meta[:cached_tokens] || meta["cached_tokens"]
+        usage[:cached_tokens] = cached unless cached.nil?
+        total = meta[:total_tokens] || meta["total_tokens"]
+        usage[:total_tokens] = total.nil? ? (input.to_i + output.to_i) : total
+        usage
+      end
+      private_class_method :build_usage
+
       def self.resolve_preset_name(suffix_preset, explicit_preset)
         suffix = suffix_preset&.to_s&.strip
         explicit = explicit_preset&.to_s&.strip
@@ -364,7 +389,8 @@ module Ace
         registry:, fallback_config:, timeout:, debug:, role_fallbacks: nil, preset: nil, thinking_level: nil, option_builder:)
         if fallback_config.disabled?
           client = registry.get_client(provider, model: model, timeout: timeout)
-          return client.generate(messages, **generation_opts)
+          response = client.generate(messages, **generation_opts)
+          return response.merge(execution: execution_identity(provider, model, preset, thinking_level, response, generation_opts))
         end
 
         primary_provider_string = model ? "#{provider}:#{model}" : provider
@@ -391,9 +417,60 @@ module Ace
           else
             option_builder.call(target_selector)
           end
-          client.generate(messages, **opts)
+          parsed = Molecules::ProviderModelParser.new(registry: registry).parse(target_selector)
+          raise ConfigurationError, parsed.error unless parsed.valid?
+          primary = target_selector.to_s == primary_provider_string
+          response = client.generate(messages, **opts)
+          response.merge(execution: execution_identity(
+            parsed.provider, parsed.model,
+            primary ? preset : (parsed.preset || preset),
+            primary ? thinking_level : parsed.thinking_level,
+            response,
+            opts
+          ))
+        end
+      end
+
+      # Transport identity records what ACE executed, not a provider attestation.
+      def self.execution_identity(provider, model, preset, thinking_level, response, options)
+        model ||= response.dig(:metadata, :model) || response.dig(:metadata, "model")
+        finish_reason = response.dig(:metadata, :finish_reason) || response.dig(:metadata, "finish_reason") ||
+          response.dig(:metadata, :stop_reason) || response.dig(:metadata, "stop_reason")
+        status = if Ace::LLM::SUCCESSFUL_FINISH_REASONS.include?(finish_reason.to_s.downcase)
+          "succeeded"
+        else
+          "incomplete"
+        end
+        {provider: provider, model: model, preset: preset, thinking_level: effective_thinking_level(provider, thinking_level, options),
+         identity_source: "resolved_request", status: status}
+      end
+      private_class_method :execution_identity
+
+      def self.effective_thinking_level(provider, default, options)
+        return default unless %w[pi codex].include?(provider)
+
+        args = Array(options[:cli_args]).flat_map { |arg| Shellwords.split(arg.to_s) }
+        effective = default
+        args.each_with_index do |arg, index|
+          if provider == "pi"
+            candidate = if arg == "--thinking"
+              args[index + 1]
+            elsif arg.start_with?("--thinking=")
+              arg.delete_prefix("--thinking=")
+            end
+            effective = candidate if %w[off minimal low medium high xhigh max].include?(candidate)
+          else
+            config = if %w[-c --config].include?(arg)
+              args[index + 1]
+            elsif arg.match?(/\A(?:-c|--config)=?model_reasoning_effort=/)
+              arg.sub(/\A(?:-c|--config)=?/, "")
+            end
+            effective = config.split("=", 2).last if config&.start_with?("model_reasoning_effort=")
+          end
         end
+        effective
       end
+      private_class_method :effective_thinking_level
 
       def self.extract_text_content(response)
         if response[:text]

data/lib/ace/llm/version.rb

@@ -2,6 +2,6 @@
 
 module Ace
   module LLM
-    VERSION = "0.39.2"
+    VERSION = "0.41.0"
   end
 end

data/lib/ace/llm.rb

@@ -55,6 +55,8 @@ module Ace
     class ConfigurationError < Error; end
     class AuthenticationError < Error; end
 
+    SUCCESSFUL_FINISH_REASONS = %w[success stop end_turn complete completed stop_sequence].freeze
+
     # Define module namespaces
     module Atoms; end
     module Molecules; end

metadata

@@ -1,13 +1,13 @@
 --- !ruby/object:Gem::Specification
 name: ace-llm
 version: !ruby/object:Gem::Version
-  version: 0.39.2
+  version: 0.41.0
 platform: ruby
 authors:
 - Michal Czyz
 bindir: exe
 cert_chain: []
-date: 2026-09-13 00:00:00.000000000 Z
+date: 2026-09-27 00:00:00.000000000 Z
 dependencies:
 - !ruby/object:Gem::Dependency
   name: ace-support-cli
@@ -266,6 +266,11 @@ files:
 - ".ace-defaults/llm/thinking/codex/low.yml"
 - ".ace-defaults/llm/thinking/codex/medium.yml"
 - ".ace-defaults/llm/thinking/codex/xhigh.yml"
+- ".ace-defaults/llm/thinking/pi/high.yml"
+- ".ace-defaults/llm/thinking/pi/low.yml"
+- ".ace-defaults/llm/thinking/pi/max.yml"
+- ".ace-defaults/llm/thinking/pi/medium.yml"
+- ".ace-defaults/llm/thinking/pi/xhigh.yml"
 - ".ace-defaults/nav/protocols/guide-sources/ace-llm.yml"
 - CHANGELOG.md
 - LICENSE