pinecall 0.0.6 → 0.0.7

checksums.yaml

@@ -1,7 +1,7 @@
 ---
 SHA256:
-  metadata.gz: e3afca7edf6bc33bb5b6fc9ebed23d76eecac2bb6039fb243d252d379e509e50
-  data.tar.gz: f3a39bb373fa2565bc6f49f04429d8f560dcc5cd8d01f9cc8d874f3ae452808d
+  metadata.gz: f33416ca47b6dac53cbee96b457cd9d8961c27439b7f5568834ca64c561c5e89
+  data.tar.gz: deaf2b2a704a6dbda1a8b88f21f46fabb327ee7dc22eab9f858cf2e5dca7d694
 SHA512:
-  metadata.gz: bd6ea499f8cdbe11a9faf6e25715848c470755e8318705beee191db517446096866e79bd830c32ba77ed4c38c7da134ccdacf90ae973c7169d18baa5a58abd93
-  data.tar.gz: 9ee25075169e9581cafd1c58a67eee8b47705cc82bbd3d6aa8568b1a35db6ac20039089bb72de8004bbd973772679725208c3fdca5d469a8b292b2de53badd63
+  metadata.gz: 671735c18ef00423d7e4c7a1fbefbd87c0f5087b23a0d9fa556bc151a19df928fa0c7bbebcb9977858bdf3cb54645a7d570d0e3e0fe6ad5bb5fb8977575d45ea
+  data.tar.gz: 5ed8532ef612a31608e11e70016045e8da92d35dcfe078e6dd976680e729695b4e8b6eee2448cb4567991592294fe67cb4519ad53ed05bbf27878c81ed1d66e0

data/CHANGELOG.md

@@ -6,6 +6,14 @@ version number is the human's call.
 
 ## [Unreleased]
 
+## [0.0.7] — Every judge is a model's answer, and the class names the judge's model (2026-10-10)
+
+- `judge "<vendor>/<model>", builds:, options:`: the model the agent's calls are judged on, over
+  the org's choice and Pinecall's. On a key of the org's own — a local model through
+  `options: { base_url: "…" }` among them — its evals are never billed.
+- The wire reads what the runtime now writes: a judge's `na` and `classified` answers with their
+  `choice` or `score`, `call.score`'s `evals` and `own_key`, and `call.summary`'s `simulated`.
+
 ## [0.0.6] — A greeting is words or :improvise, a hangup is words or true, and the ears choose who ends the turn (2026-10-10)
 
 - `greeting "…"` (said as written), `greeting :improvise` (the model opens on its prompt) and

data/lib/pinecall/agent/config.rb

@@ -13,7 +13,7 @@ module Pinecall
       AS_WRITTEN = %i[phone whatsapp web channel_rules].freeze
 
       # The settings a class may declare. Must match the TypeScript package's `ENVIRONMENT`.
-      ENVIRONMENT = %i[language voice llm stt greeting hangup turn says hears knowledge docs memory record].freeze
+      ENVIRONMENT = %i[language voice llm stt judge greeting hangup turn says hears knowledge docs memory record].freeze
 
       # The ones written as the wire's own fields, keyword by keyword: `greeting say: "…"`.
       AS_KEYWORDS = %i[turn knowledge docs memory].freeze
@@ -90,6 +90,16 @@ module Pinecall
           environment[:stt] = { **Config.model_of(model), builds:, options:, end_of_turn: ends }.compact
         end
 
+        # The model the agent's calls are judged on, over the org's choice and the platform's:
+        # `judge "openai/qwen3-32b", options: { base_url: "http://gpu:8000/v1" }`. On a key of the
+        # org's own, a local model's server among them, its evals are never billed; `builds` and
+        # `options` run on the org's own key alone.
+        def judge(model = NOTHING, builds: nil, options: nil)
+          return environment[:judge] if model.equal?(NOTHING)
+
+          environment[:judge] = { **Config.model_of(model), builds:, options: }.compact
+        end
+
         # How a call opens: the words, said as written, or :improvise for the model's own. The
         # caller cannot cut it short unless `interruptible: true`.
         def greeting(opening = NOTHING, interruptible: nil)

data/lib/pinecall/version.rb

@@ -2,5 +2,5 @@
 
 module Pinecall
   # Set by the maintainer at release time.
-  VERSION = "0.0.6"
+  VERSION = "0.0.7"
 end

data/lib/pinecall/wire/enums.rb

@@ -34,9 +34,10 @@ module Pinecall
       # Whose action ended the call. platform covers timeouts, errors and a drained worker.
       ENDED_BY = %w[caller agent supervisor platform].freeze
       # What one judge answered about a finished call. Held: the rule held. Broken: it did not,
-      # and the reason names the evidence. Deferred: the judge was asked and could not settle it.
-      # Skipped: nobody asked it — no model was reachable inside the call's judging budget.
-      SCORE_VERDICT = %w[held broken deferred skipped].freeze
+      # and the reason names the evidence. Classified: a choice or a score was answered. N/A: the
+      # question did not apply. Deferred: the judge was asked and could not settle it. Skipped:
+      # nobody asked it — no model, the ceiling reached, or the org's evals used up.
+      SCORE_VERDICT = %w[held broken classified na deferred skipped].freeze
       # Cold: the caller is sent on with a REFER on their SIP leg and the call ends here. Warm:
       # the number is dialled into the call's own room, the agent stays on the line until the
       # other side answers and then falls silent; the call ends when either of them hangs up.

data/lib/pinecall/wire/shapes_call_events.rb

@@ -81,6 +81,8 @@ module Pinecall
           judges: { kind: :list, items: { kind: :ref, ref: "Judgment" }, required: true },
           panel: { kind: :list, items: { kind: :str } },
           judge_calls: { kind: :int, required: true },
+          evals: { kind: :int },
+          own_key: { kind: :bool },
           judge_cost_usd: { kind: :float },
           judged_by: { kind: :ref, ref: "JudgedBy" }
         }.freeze,
@@ -108,7 +110,8 @@ module Pinecall
           turns: { kind: :int, required: true },
           usage: { kind: :list, items: { kind: :ref, ref: "ModelUsage" }, required: true },
           cost: { kind: :ref, ref: "Cost", required: true },
-          recording: { kind: :str }
+          recording: { kind: :str },
+          simulated: { kind: :bool }
         }.freeze,
         "CallTransferred" => {
           to: { kind: :str, required: true },

data/lib/pinecall/wire/shapes_doors.rb

@@ -23,6 +23,7 @@ module Pinecall
           voice: { kind: :ref, ref: "VoiceConfig" },
           llm: { kind: :ref, ref: "ModelConfig" },
           stt: { kind: :ref, ref: "ModelConfig" },
+          judge: { kind: :ref, ref: "ModelConfig" },
           turn: { kind: :ref, ref: "TurnConfig" },
           says: { kind: :list, items: { kind: :ref, ref: "Pronunciation" } },
           hears: { kind: :list, items: { kind: :str } },

data/lib/pinecall/wire/shapes_parts.rb

@@ -318,7 +318,9 @@ module Pinecall
           verdict: { kind: :ref, ref: "ScoreVerdict", required: true },
           criteria: { kind: :str, required: true },
           reason: { kind: :str, required: true },
-          evidence: { kind: :ref, ref: "JudgmentEvidence", required: true }
+          evidence: { kind: :ref, ref: "JudgmentEvidence", required: true },
+          choice: { kind: :str, null: true },
+          score: { kind: :int, null: true }
         }.freeze,
         "JudgmentEvidence" => {
           seqs: { kind: :list, items: { kind: :int }, required: true },

data/sig/pinecall.rbs

@@ -88,6 +88,7 @@ module Pinecall
     def self.voice: (?String, ?String?, ?model: String?, ?builds: String?, ?options: Hash[untyped, untyped]?) -> untyped
     def self.llm: (?String, ?temperature: Float?, ?builds: String?, ?options: Hash[untyped, untyped]?) -> untyped
     def self.stt: (?String, ?builds: String?, ?options: Hash[untyped, untyped]?, ?end_of_turn: (String | Symbol)?) -> untyped
+    def self.judge: (?String, ?builds: String?, ?options: Hash[untyped, untyped]?) -> untyped
     def self.language: (?String) -> untyped
     def self.says: (?Array[Hash[Symbol, String]]) -> untyped
     def self.record: (?bool) -> untyped

metadata

@@ -1,7 +1,7 @@
 --- !ruby/object:Gem::Specification
 name: pinecall
 version: !ruby/object:Gem::Version
-  version: 0.0.6
+  version: 0.0.7
 platform: ruby
 authors:
 - Pinecall