pinecall 0.0.6 → 0.0.8

checksums.yaml

@@ -1,7 +1,7 @@
 ---
 SHA256:
-  metadata.gz: e3afca7edf6bc33bb5b6fc9ebed23d76eecac2bb6039fb243d252d379e509e50
-  data.tar.gz: f3a39bb373fa2565bc6f49f04429d8f560dcc5cd8d01f9cc8d874f3ae452808d
+  metadata.gz: d385aad2e86c89ef8639d736478845f8eb2299d6eda64fb0dfbfa14ffd6ea15a
+  data.tar.gz: 04d630be1150261ec88676d98b120744f0c5b6ae5242145071e94d470b9da580
 SHA512:
-  metadata.gz: bd6ea499f8cdbe11a9faf6e25715848c470755e8318705beee191db517446096866e79bd830c32ba77ed4c38c7da134ccdacf90ae973c7169d18baa5a58abd93
-  data.tar.gz: 9ee25075169e9581cafd1c58a67eee8b47705cc82bbd3d6aa8568b1a35db6ac20039089bb72de8004bbd973772679725208c3fdca5d469a8b292b2de53badd63
+  metadata.gz: 5f84b1700e9b536be0b573bc7f6a41ae5e63fb8803e5c586adec62c7928958b82e54e154dfaf9a23a6d2f47591782040557f57f9f521d13415d99fae204ab9c8
+  data.tar.gz: b9b958f3c8ee08bc72f02c5f9fb1edd04a6e6e20520db20a3d85fa01d69471c0ddfe2589cbfe2aefd0126bb769918a3d512ec15344112cfe838ef5a848899082

data/CHANGELOG.md

@@ -6,6 +6,22 @@ version number is the human's call.
 
 ## [Unreleased]
 
+## [0.0.8] — The wire reads evals.spent, call.started.expects and call.summary.run (2026-10-11)
+
+- The wire reads `evals.spent` (what the harness asked a model — a judge tried, personas
+  written, a case group diagnosed — its evals counted), `call.started.expects` (what a simulated
+  caller expects, replacing `accepts_when`/`declines_when`), a
+  `call.summary`'s `run`, and `credits.exhausted` for every quota the runtime refuses for —
+  `hosted_apps`, `evals`, `simulations` and `budget_usd` were missing. Needs runtime 0.1.46.
+
+## [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 = "0.0.8"
 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/registry.rb

@@ -34,6 +34,7 @@ module Pinecall
         "confirm.granted" => "ConfirmGranted",
         "confirm.request" => "ConfirmRequest",
         "credits.exhausted" => "CreditsExhausted",
+        "evals.spent" => "EvalsSpent",
         "custom" => "Custom",
         "dev.request" => "DevRequest",
         "docs.sources" => "DocsSources",
@@ -118,7 +119,7 @@ module Pinecall
       }.freeze
 
       # Every event this protocol declares, in the schema's own order.
-      EVENT_TYPES = %w[agent.configured agent.detached agent.draining agent.registered agent.state agent.transcript attention.answered attention.requested call.attached call.claimed call.dialing call.ended call.line call.ringing call.score call.started call.summary call.transferred callback.requested code.claimed code.issued confirm.declined confirm.granted confirm.request credits.exhausted custom dev.request docs.sources dtmf.received error event.received fleet.full log.caught_up log.gap memory.ops message.taken message.waiting metrics.avatar metrics.eot metrics.eou metrics.interruption metrics.llm metrics.realtime metrics.stt metrics.tts metrics.vad participant.joined participant.left participant.speaking pong prompt.changed room.opened room.sent spend.unusual state.changed supervisor.ended supervisor.released supervisor.said supervisor.took_over supervisor.transferred supervisor.whispered tool.call tool.result tools.changed track.published track.unpublished turn.agent turn.user user.state user.transcript vendor.switched].freeze
+      EVENT_TYPES = %w[agent.configured agent.detached agent.draining agent.registered agent.state agent.transcript attention.answered attention.requested call.attached call.claimed call.dialing call.ended call.line call.ringing call.score call.started call.summary call.transferred callback.requested code.claimed code.issued confirm.declined confirm.granted confirm.request credits.exhausted custom dev.request docs.sources dtmf.received error evals.spent event.received fleet.full log.caught_up log.gap memory.ops message.taken message.waiting metrics.avatar metrics.eot metrics.eou metrics.interruption metrics.llm metrics.realtime metrics.stt metrics.tts metrics.vad participant.joined participant.left participant.speaking pong prompt.changed room.opened room.sent spend.unusual state.changed supervisor.ended supervisor.released supervisor.said supervisor.took_over supervisor.transferred supervisor.whispered tool.call tool.result tools.changed track.published track.unpublished turn.agent turn.user user.state user.transcript vendor.switched].freeze
 
       # Every command an app may send.
       COMMAND_TYPES = %w[agent.configure agent.drain agent.register agent.reply agent.say call.attention call.callback call.claim call.dial call.dtmf call.event call.hangup call.hold call.log call.mute call.opt_out call.transfer call.unhold call.unmute dev.answer participant.mute participant.remove ping prompt.set room.invite room.send session.configure state.set supervisor.verb tool.result tools.set].freeze

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,
@@ -91,8 +93,7 @@ module Pinecall
           to: { kind: :str, required: true },
           run: { kind: :str, null: true },
           persona: { kind: :str, null: true },
-          accepts_when: { kind: :str, null: true },
-          declines_when: { kind: :str, null: true },
+          expects: { kind: :list, items: { kind: :str }, null: true },
           caller: { kind: :ref, null: true, ref: "Contact", required: true },
           started_at: { kind: :float, required: true },
           env: { kind: :ref, ref: "Env" },
@@ -108,7 +109,9 @@ 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 },
+          run: { kind: :str, null: true }
         }.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_events.rb

@@ -213,10 +213,20 @@ module Pinecall
         }.freeze,
         "CreditsExhausted" => {
           org: { kind: :str, required: true },
-          quota: { kind: :enum, values: %w[minutes messages agents concurrent_calls memory_facts knowledge_chunks numbers seats llm_tokens], required: true },
+          quota: { kind: :enum, values: %w[minutes messages agents concurrent_calls memory_facts knowledge_chunks numbers seats llm_tokens hosted_apps evals simulations budget_usd], required: true },
           used: { kind: :float, required: true },
           limit: { kind: :int, required: true }
         }.freeze,
+        "EvalsSpent" => {
+          kind: { kind: :enum, values: %w[try generate propose], required: true },
+          name: { kind: :str, required: true },
+          env: { kind: :ref, ref: "Env", required: true },
+          calls: { kind: :list, items: { kind: :str }, required: true },
+          evals: { kind: :int, required: true },
+          judge_calls: { kind: :int, required: true },
+          judge_cost_usd: { kind: :float, null: true },
+          own_key: { kind: :bool }
+        }.freeze,
         "Custom" => {
           name: { kind: :str, required: true },
           data: { kind: :json, required: true }

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: 0.0.8
 platform: ruby
 authors:
 - Pinecall