deepclause-sdk 0.0.87 → 0.0.89

package/dist/prolog-src/deepclause_mi.pl

@@ -24,6 +24,7 @@
     post_exec_result/3,
     post_sample_token_result/3,
     post_llm_result/3,
+    post_judgment_result/3,
     provide_input/2,
     mi/3,
     % Tool engine predicates for isolated tool execution
@@ -50,6 +51,7 @@
 :- dynamic session_exec_result/2.  % session_exec_result(SessionId, Result)
 :- dynamic session_sample_token_result/2. % session_sample_token_result(SessionId, Result)
 :- dynamic session_llm_result/2. % session_llm_result(SessionId, Result)
+:- dynamic session_judgment_result/2. % session_judgment_result(SessionId, Result)
 :- dynamic session_trace_log/2.    % session_trace_log(SessionId, TraceLog) - accumulated trace entries
 :- dynamic session_trace_enabled/2. % session_trace_enabled(SessionId, true/false)
 :- dynamic session_pending_signal/2. % session_pending_signal(SessionId, Signal) - fallback for engine_fetch
@@ -576,6 +578,7 @@ process_engine_result(request_sample_token(Prompt, Allowed), request_sample_toke
     payload{prompt: Prompt, allowedTokens: Allowed}) :- !.
 process_engine_result(request_llm(Messages), request_llm, '',
     payload{messages: Messages}) :- !.
+process_engine_result(request_judgment(Payload), request_judgment, '', Payload) :- !.
 %% Tool result from inline tool execution
 process_engine_result(tool_result(Result), tool_result, '',
     payload{result: Result}) :- !.
@@ -599,6 +602,7 @@ destroy_engine(SessionId) :-
     retractall(session_exec_result(SessionId, _)),
     retractall(session_sample_token_result(SessionId, _)),
     retractall(session_llm_result(SessionId, _)),
+    retractall(session_judgment_result(SessionId, _)),
     retractall(session_trace_log(SessionId, _)),
     retractall(session_trace_enabled(SessionId, _)),
     (   current_module(SessionId)
@@ -645,6 +649,17 @@ post_sample_token_result(SessionId, Status, TokenOrError) :-
     assertz(session_sample_token_result(SessionId, result{status: Status, token: TokenOrError})),
     post_signal_to_engine(SessionId, sample_token_done).
 
+%% post_judgment_result(+SessionId, +Status, +ValuesOrError)
+%% On success Values is the list of answer values in question order.
+post_judgment_result(SessionId, success, Values) :-
+    retractall(session_judgment_result(SessionId, _)),
+    assertz(session_judgment_result(SessionId, result{status: success, values: Values})),
+    post_signal_to_engine(SessionId, judgment_done).
+post_judgment_result(SessionId, failure, Error) :-
+    retractall(session_judgment_result(SessionId, _)),
+    assertz(session_judgment_result(SessionId, result{status: failure, error: Error})),
+    post_signal_to_engine(SessionId, judgment_done).
+
 %% post_llm_result(+SessionId, +Status, +TextOrError)
 post_llm_result(SessionId, Status, TextOrError) :-
     retractall(session_llm_result(SessionId, _)),
@@ -1042,6 +1057,7 @@ process_tool_engine_result(request_sample_token(Prompt, Allowed), request_sample
     payload{prompt: Prompt, allowedTokens: Allowed}) :- !.
 process_tool_engine_result(request_llm(Messages), request_llm, '',
     payload{messages: Messages}) :- !.
+process_tool_engine_result(request_judgment(Payload), request_judgment, '', Payload) :- !.
 process_tool_engine_result(request_agent_loop(Desc, Vars, Tools, Memory, ToolScope, ReasoningEffort, RecipeContext), request_agent_loop, '',
     payload{taskDescription: Desc, outputVars: Vars, userTools: Tools, memory: Memory, toolScope: ToolScope, reasoningEffort: ReasoningEffort, recipeContext: RecipeContext}) :- !.
 %% Legacy 5-arg format (no reasoning/recipe)
@@ -1313,6 +1329,175 @@ wait_sample_token_signal(SessionId) :-
     ;   wait_sample_token_signal(SessionId)
     ).
 
+%% ============================================================
+%% Semantic Judgment Predicates
+%% ============================================================
+%% judge/2 asks one batch of typed questions about an explicit state and binds
+%% the answers into output variables carried by the spec terms. The one-off
+%% predicates are sugar over the same request.
+%%
+%%   judge(Message, [
+%%       choose("Which team?", [billing, orders, account]) - Team,
+%%       rate("How frustrated?", [calm, frustrated, angry]) - Frustration,
+%%       verify("Does it ask for a refund?") - Refund
+%%   ]).
+%%
+%% Backends (llm, jev, mock) are selected with with_judgment/2, the workspace
+%% configuration, or a per-run override. require_judgment/2 gates a goal on a
+%% backend capability such as calibrated.
+
+mi_call(judge(State, Specs), StateIn, StateOut) :-
+    !,
+    mi_call_judgment(State, Specs, StateIn, StateOut).
+
+mi_call(choose(State, Question, Options, Choice), StateIn, StateOut) :-
+    !,
+    mi_call_judgment(State, [choose(Question, Options) - Choice], StateIn, StateOut).
+
+mi_call(rate(State, Question, Levels, Level), StateIn, StateOut) :-
+    !,
+    mi_call_judgment(State, [rate(Question, Levels) - Level], StateIn, StateOut).
+
+mi_call(verify(State, Question, Truth), StateIn, StateOut) :-
+    !,
+    mi_call_judgment(State, [verify(Question) - Truth], StateIn, StateOut).
+
+mi_call(probability(State, Question, Probability), StateIn, StateOut) :-
+    !,
+    mi_call_judgment(State, [probability(Question) - Probability], StateIn, StateOut).
+
+mi_call(holds(State, Question), StateIn, StateOut) :-
+    !,
+    mi_call_judgment(State, [verify(Question) - Truth], StateIn, StateOut),
+    (   Truth == yes
+    ->  true
+    ;   fail
+    ).
+
+mi_call(holds(State, Question, Threshold), StateIn, StateOut) :-
+    !,
+    mi_call_judgment(State, [probability(Question) - Probability], StateIn, StateOut),
+    (   number(Probability), Probability >= Threshold
+    ->  true
+    ;   fail
+    ).
+
+mi_call(with_judgment(Backend, Goal), StateIn, StateOut) :-
+    !,
+    set_judge_backend(StateIn, Backend, ScopedState),
+    mi_call(Goal, ScopedState, StateWithScope),
+    clear_judge_backend(StateWithScope, StateOut).
+
+mi_call(require_judgment(Capability, Goal), StateIn, StateOut) :-
+    !,
+    push_required_judgment(StateIn, Capability, ScopedState),
+    mi_call(Goal, ScopedState, StateWithScope),
+    pop_required_judgment(StateWithScope, StateOut).
+
+%% mi_call_judgment(+State, +Specs, +StateIn, -StateOut)
+mi_call_judgment(State, Specs, StateIn, StateOut) :-
+    consume_gas(StateIn, State1),
+    get_session_id(SessionId),
+    get_depth(State1, Depth),
+    split_specs(Specs, Questions, Outputs),
+    length(Questions, Count),
+    add_trace_entry(SessionId, llm_call, judgment, [Count], Depth),
+    get_judge_backend(State1, Backend),
+    get_required_judgments(State1, Requires),
+    engine_yield(request_judgment(payload{
+        state: State,
+        questions: Questions,
+        backend: Backend,
+        requires: Requires
+    })),
+    wait_judgment_signal(SessionId),
+    (   session_judgment_result(SessionId, Result)
+    ->  retract(session_judgment_result(SessionId, _)),
+        (   Result.status == success
+        ->  add_trace_with_result(SessionId, exit, judgment, [Count], Result.values, Depth),
+            bind_judgment_outputs(Outputs, Result.values),
+            StateOut = State1
+        ;   (   get_dict(error, Result, ErrorMsg), ErrorMsg \= ""
+            ->  format(atom(WarnMsg), 'Warning: judgment failed: ~w', [ErrorMsg])
+            ;   format(atom(WarnMsg), 'Warning: judgment failed (status=~w)', [Result.status])
+            ),
+            add_trace_entry(SessionId, fail, judgment, [Count], Depth),
+            engine_yield(output(WarnMsg)),
+            fail
+        )
+    ;   add_trace_entry(SessionId, fail, judgment, [Count], Depth),
+        engine_yield(output('Warning: judgment failed (no result returned)')),
+        fail
+    ).
+
+wait_judgment_signal(SessionId) :-
+    get_next_signal(SessionId, Signal),
+    (   Signal == judgment_done
+    ->  true
+    ;   Signal = error_signal(_)
+    ->  true
+    ;   wait_judgment_signal(SessionId)
+    ).
+
+%% split_specs(+Specs, -Questions, -Outputs)
+%% Each spec is `Question - Output`. Outputs are bound after the batch.
+split_specs([], [], []).
+split_specs([Spec-Output|Rest], [Question|Questions], [Output|Outputs]) :-
+    spec_question(Spec, Question),
+    split_specs(Rest, Questions, Outputs).
+
+spec_question(choose(Instruction, Options),
+    question{kind: choose, instruction: Instruction, options: Options}).
+spec_question(rate(Instruction, Levels),
+    question{kind: rate, instruction: Instruction, levels: Levels}).
+spec_question(verify(Instruction),
+    question{kind: verify, instruction: Instruction}).
+spec_question(verify(Instruction, Criteria),
+    question{kind: verify, instruction: Instruction, criteria: Criteria}).
+spec_question(probability(Instruction),
+    question{kind: probability, instruction: Instruction}).
+
+bind_judgment_outputs([], []).
+bind_judgment_outputs([Output|Outputs], [Value|Values]) :-
+    Output = Value,
+    bind_judgment_outputs(Outputs, Values).
+
+%% get_judge_backend(+State, -Backend)
+get_judge_backend(State, Backend) :-
+    (   get_dict(judge_backend, State, Scoped)
+    ->  Backend = Scoped
+    ;   Backend = default
+    ).
+
+set_judge_backend(StateIn, Backend, StateOut) :-
+    StateOut = StateIn.put(judge_backend, Backend).
+
+clear_judge_backend(StateIn, StateOut) :-
+    (   get_dict(judge_backend, StateIn, _)
+    ->  del_dict(judge_backend, StateIn, _, StateOut)
+    ;   StateOut = StateIn
+    ).
+
+get_required_judgments(State, Requires) :-
+    (   get_dict(judgment_requires, State, Requires)
+    ->  true
+    ;   Requires = []
+    ).
+
+push_required_judgment(StateIn, Capability, StateOut) :-
+    get_required_judgments(StateIn, Existing),
+    (   member(Capability, Existing)
+    ->  StateOut = StateIn
+    ;   append(Existing, [Capability], Updated),
+        StateOut = StateIn.put(judgment_requires, Updated)
+    ).
+
+pop_required_judgment(StateIn, StateOut) :-
+    (   get_dict(judgment_requires, StateIn, _)
+    ->  del_dict(judgment_requires, StateIn, _, StateOut)
+    ;   StateOut = StateIn
+    ).
+
 mi_call(llm(Messages, Reply), StateIn, StateOut) :-
     !,
     consume_gas(StateIn, State1),
@@ -1687,6 +1872,28 @@ mi_call(fail, _StateIn, _StateOut) :-
 mi_call(false, _StateIn, _StateOut) :-
     fail.
 
+%% mi_call(assertz(Head), +StateIn, -StateOut)
+%% mi_call(asserta(Head), +StateIn, -StateOut)
+%% mi_call(retract(Head), +StateIn, -StateOut)
+%% mi_call(retractall(Head), +StateIn, -StateOut)
+%% Database predicates must operate on the session module so DML-side facts
+%% (draft/4, claim/3, ...) are asserted/retracted in the right module.
+mi_call(assertz(Head), StateIn, StateIn) :-
+    get_session_id(SessionId),
+    assertz(SessionId:Head).
+
+mi_call(asserta(Head), StateIn, StateIn) :-
+    get_session_id(SessionId),
+    asserta(SessionId:Head).
+
+mi_call(retract(Head), StateIn, StateIn) :-
+    get_session_id(SessionId),
+    retract(SessionId:Head).
+
+mi_call(retractall(Head), StateIn, StateIn) :-
+    get_session_id(SessionId),
+    retractall(SessionId:Head).
+
 %% ============================================================
 %% List Predicates - Interpreted for proper state threading
 %% ============================================================
@@ -2183,6 +2390,16 @@ is_mi_special_predicate(without_tools(_,_)).
 is_mi_special_predicate(with_reasoning(_,_)).
 %% Recipe context scoping
 is_mi_special_predicate(with_recipe(_,_)).
+%% Semantic judgment predicates
+is_mi_special_predicate(judge(_,_)).
+is_mi_special_predicate(choose(_,_,_,_)).
+is_mi_special_predicate(rate(_,_,_,_)).
+is_mi_special_predicate(verify(_,_,_)).
+is_mi_special_predicate(probability(_,_,_)).
+is_mi_special_predicate(holds(_,_)).
+is_mi_special_predicate(holds(_,_,_)).
+is_mi_special_predicate(with_judgment(_,_)).
+is_mi_special_predicate(require_judgment(_,_)).
 %% Context stack management
 is_mi_special_predicate(get_memory(_)).
 is_mi_special_predicate(push_context).
@@ -2201,6 +2418,10 @@ is_mi_special_predicate(!).
 is_mi_special_predicate(true).
 is_mi_special_predicate(fail).
 is_mi_special_predicate(false).
+is_mi_special_predicate(assertz(_)).
+is_mi_special_predicate(asserta(_)).
+is_mi_special_predicate(retract(_)).
+is_mi_special_predicate(retractall(_)).
 is_mi_special_predicate(_:_).
 %% List predicates
 is_mi_special_predicate(member(_,_)).
@@ -2305,8 +2526,10 @@ mi_call_dispatch(Goal, StateIn, StateOut) :-
         clause(SessionId:Goal, Body),
         NewDepth is Depth + 1,
         set_depth(StateIn, NewDepth, State1),
+        % Soft cut: execute the body for every solution (so findall/backtracking
+        % works), but still record a fail trace when the body has none.
         (   mi_call(Body, State1, State2)
-        ->  add_trace_entry(SessionId, exit, Functor, Args, Depth),
+        *-> add_trace_entry(SessionId, exit, Functor, Args, Depth),
             set_depth(State2, Depth, StateOut)
         ;   add_trace_entry(SessionId, fail, Functor, Args, Depth),
             fail

package/dist/runner.d.ts

@@ -3,6 +3,7 @@
  */
 import type { SWIPLModule } from './prolog/loader.js';
 import type { CompactionOptions, DMLEvent, MemoryMessage, LLMBackend, ToolDefinition, ToolPolicy, TypedVar } from './types.js';
+import type { JudgeSelection } from './judge/types.js';
 import { type ReasoningType } from './system/config/model-database.js';
 export declare function decodeAgentOutputVars(rawOutputVars: unknown): (string | TypedVar)[];
 export interface RunnerOptions {
@@ -20,6 +21,7 @@ export interface RunnerOptions {
     reasoningBudgetMap?: Record<string, number>;
     contextWindow?: number;
     llmBackend?: LLMBackend;
+    judgeSelection?: JudgeSelection;
 }
 export interface InternalRunOptions {
     args?: unknown[];
@@ -32,6 +34,8 @@ export interface InternalRunOptions {
     signal?: AbortSignal;
     initialMessages?: MemoryMessage[];
     compaction?: CompactionOptions;
+    /** Per-run default judgment backend name override. */
+    judgeBackend?: string;
 }
 /**
  * DML execution engine
@@ -48,6 +52,7 @@ export declare class DMLRunner {
     private sessionId;
     private currentMemory;
     private stepCounter;
+    private judgeCache;
     constructor(swipl: SWIPLModule, options: RunnerOptions);
     /**
      * Get the current conversation memory
@@ -98,6 +103,13 @@ export declare class DMLRunner {
      * Handle direct one-shot LLM request (llm/2)
      */
     private handleLlm;
+    /**
+     * Handle a semantic judgment request (judge/2 and the one-off predicates).
+     */
+    private handleJudgment;
+    private parseJudgmentPayload;
+    private completeJudgment;
+    private postJudgmentResult;
     /**
      * Handle exec request (exec() predicate)
      */

package/dist/runner.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"runner.d.ts","sourceRoot":"","sources":["../src/runner.ts"],"names":[],"mappings":"AAAA;;GAEG;AAIH,OAAO,KAAK,EAAE,WAAW,EAAE,MAAM,oBAAoB,CAAC;AACtD,OAAO,KAAK,EACV,iBAAiB,EACjB,QAAQ,EACR,aAAa,EACb,UAAU,EACV,cAAc,EACd,UAAU,EACV,QAAQ,EACT,MAAM,YAAY,CAAC;AAkBpB,OAAO,EAAiC,KAAK,aAAa,EAAE,MAAM,mCAAmC,CAAC;AAuGtG,wBAAgB,qBAAqB,CAAC,aAAa,EAAE,OAAO,GAAG,CAAC,MAAM,GAAG,QAAQ,CAAC,EAAE,CAmBnF;AAED,MAAM,WAAW,aAAa;IAC5B,KAAK,EAAE,MAAM,CAAC;IACd,QAAQ,EAAE,MAAM,CAAC;IACjB,WAAW,EAAE,MAAM,CAAC;IACpB,SAAS,EAAE,MAAM,CAAC;IAClB,OAAO,CAAC,EAAE,MAAM,CAAC;IACjB,KAAK,CAAC,EAAE,OAAO,CAAC;IAChB,SAAS,CAAC,EAAE,OAAO,CAAC;IACpB,KAAK,CAAC,EAAE,OAAO,CAAC;IAEhB,eAAe,CAAC,EAAE,MAAM,CAAC,MAAM,EAAE,MAAM,CAAC,MAAM,EAAE,GAAG,CAAC,CAAC,CAAC;IACtD,UAAU,CAAC,EAAE,iBAAiB,CAAC;IAC/B,aAAa,CAAC,EAAE,aAAa,CAAC;IAC9B,kBAAkB,CAAC,EAAE,MAAM,CAAC,MAAM,EAAE,MAAM,CAAC,CAAC;IAC5C,aAAa,CAAC,EAAE,MAAM,CAAC;IACvB,UAAU,CAAC,EAAE,UAAU,CAAC;CACzB;AAED,MAAM,WAAW,kBAAkB;IACjC,IAAI,CAAC,EAAE,OAAO,EAAE,CAAC;IACjB,MAAM,CAAC,EAAE,MAAM,CAAC,MAAM,EAAE,OAAO,CAAC,CAAC;IACjC,aAAa,CAAC,EAAE,MAAM,CAAC;IACvB,QAAQ,CAAC,EAAE,MAAM,CAAC;IAClB,KAAK,EAAE,GAAG,CAAC,MAAM,EAAE,cAAc,CAAC,CAAC;IACnC,UAAU,EAAE,UAAU,GAAG,IAAI,CAAC;IAC9B,eAAe,EAAE,CAAC,MAAM,EAAE,MAAM,KAAK,OAAO,CAAC,MAAM,CAAC,CAAC;IACrD,MAAM,CAAC,EAAE,WAAW,CAAC;IACrB,eAAe,CAAC,EAAE,aAAa,EAAE,CAAC;IAClC,UAAU,CAAC,EAAE,iBAAiB,CAAC;CAChC;AA2CD;;;;;;;GAOG;AACH,qBAAa,SAAS;IACpB,OAAO,CAAC,KAAK,CAAc;IAC3B,OAAO,CAAC,OAAO,CAAgB;IAC/B,OAAO,CAAC,MAAM,CAAiB;IAC/B,OAAO,CAAC,SAAS,CAAc;IAC/B,OAAO,CAAC,aAAa,CAAuB;IAC5C,OAAO,CAAC,WAAW,CAAK;gBAEZ,KAAK,EAAE,WAAW,EAAE,OAAO,EAAE,aAAa;IAKtD;;OAEG;IACH,SAAS,IAAI,aAAa,EAAE;IAI5B;;OAEG;IACI,GAAG,CAAC,IAAI,EAAE,MAAM,EAAE,OAAO,EAAE,kBAAkB,GAAG,cAAc,CAAC,QAAQ,CAAC;IAmJ/E;;OAEG;IACH,OAAO,CAAC,SAAS;IAsBjB;;OAEG;IACH,OAAO,CAAC,SAAS;IAMjB;;OAEG;IACH,OAAO,CAAC,WAAW;IAmBnB;;;OAGG;IACH,OAAO,CAAC,SAAS;IAQjB;;OAEG;IACH,OAAO,CAAC,YAAY;IA2BpB;;OAEG;IACH,OAAO,CAAC,YAAY;IAmBpB;;OAEG;IACH,OAAO,CAAC,UAAU;IA4BlB;;OAEG;YACY,eAAe;IAuS9B;;OAEG;YACY,iBAAiB;IA4ChC;;OAEG;YACY,SAAS;IAwCxB;;OAEG;YACY,UAAU;IAkDzB;;OAEG;IACH,OAAO,CAAC,YAAY;IA6HpB;;OAEG;IACH,OAAO,CAAC,iBAAiB;IAwCzB;;;;OAIG;IACH,OAAO,CAAC,eAAe;IAqDvB;;;;;;;;;;;;;;OAcG;YACW,mBAAmB;YAmbnB,uBAAuB;YAmDvB,iBAAiB;YA+DjB,sBAAsB;IAqBpC;;OAEG;IACH,OAAO,CAAC,wBAAwB;IAmBhC,OAAO,CAAC,wBAAwB;IAKhC,OAAO,CAAC,uBAAuB;IAM/B;;OAEG;IACH,OAAO,CAAC,eAAe;IAoBvB;;OAEG;IACH,OAAO,CAAC,cAAc;IAkBtB;;OAEG;IACH,OAAO,CAAC,qBAAqB;IAkB7B;;OAEG;IACH,OAAO,CAAC,aAAa;IAkBrB;;OAEG;IACH,OAAO,CAAC,mBAAmB;IAwB3B;;OAEG;IACH,OAAO,CAAC,YAAY;IAOpB;;OAEG;IACH,OAAO,CAAC,KAAK;IAWb;;OAEG;IACH,OAAO,CAAC,OAAO;IAWf;;OAEG;IACG,OAAO,IAAI,OAAO,CAAC,IAAI,CAAC;CAG/B"}
+{"version":3,"file":"runner.d.ts","sourceRoot":"","sources":["../src/runner.ts"],"names":[],"mappings":"AAAA;;GAEG;AAIH,OAAO,KAAK,EAAE,WAAW,EAAE,MAAM,oBAAoB,CAAC;AACtD,OAAO,KAAK,EACV,iBAAiB,EACjB,QAAQ,EACR,aAAa,EACb,UAAU,EACV,cAAc,EACd,UAAU,EACV,QAAQ,EACT,MAAM,YAAY,CAAC;AAIpB,OAAO,KAAK,EAAkC,cAAc,EAAwC,MAAM,kBAAkB,CAAC;AAgB7H,OAAO,EAAiC,KAAK,aAAa,EAAE,MAAM,mCAAmC,CAAC;AAuGtG,wBAAgB,qBAAqB,CAAC,aAAa,EAAE,OAAO,GAAG,CAAC,MAAM,GAAG,QAAQ,CAAC,EAAE,CAmBnF;AAED,MAAM,WAAW,aAAa;IAC5B,KAAK,EAAE,MAAM,CAAC;IACd,QAAQ,EAAE,MAAM,CAAC;IACjB,WAAW,EAAE,MAAM,CAAC;IACpB,SAAS,EAAE,MAAM,CAAC;IAClB,OAAO,CAAC,EAAE,MAAM,CAAC;IACjB,KAAK,CAAC,EAAE,OAAO,CAAC;IAChB,SAAS,CAAC,EAAE,OAAO,CAAC;IACpB,KAAK,CAAC,EAAE,OAAO,CAAC;IAEhB,eAAe,CAAC,EAAE,MAAM,CAAC,MAAM,EAAE,MAAM,CAAC,MAAM,EAAE,GAAG,CAAC,CAAC,CAAC;IACtD,UAAU,CAAC,EAAE,iBAAiB,CAAC;IAC/B,aAAa,CAAC,EAAE,aAAa,CAAC;IAC9B,kBAAkB,CAAC,EAAE,MAAM,CAAC,MAAM,EAAE,MAAM,CAAC,CAAC;IAC5C,aAAa,CAAC,EAAE,MAAM,CAAC;IACvB,UAAU,CAAC,EAAE,UAAU,CAAC;IACxB,cAAc,CAAC,EAAE,cAAc,CAAC;CACjC;AAED,MAAM,WAAW,kBAAkB;IACjC,IAAI,CAAC,EAAE,OAAO,EAAE,CAAC;IACjB,MAAM,CAAC,EAAE,MAAM,CAAC,MAAM,EAAE,OAAO,CAAC,CAAC;IACjC,aAAa,CAAC,EAAE,MAAM,CAAC;IACvB,QAAQ,CAAC,EAAE,MAAM,CAAC;IAClB,KAAK,EAAE,GAAG,CAAC,MAAM,EAAE,cAAc,CAAC,CAAC;IACnC,UAAU,EAAE,UAAU,GAAG,IAAI,CAAC;IAC9B,eAAe,EAAE,CAAC,MAAM,EAAE,MAAM,KAAK,OAAO,CAAC,MAAM,CAAC,CAAC;IACrD,MAAM,CAAC,EAAE,WAAW,CAAC;IACrB,eAAe,CAAC,EAAE,aAAa,EAAE,CAAC;IAClC,UAAU,CAAC,EAAE,iBAAiB,CAAC;IAC/B,sDAAsD;IACtD,YAAY,CAAC,EAAE,MAAM,CAAC;CACvB;AAkDD;;;;;;;GAOG;AACH,qBAAa,SAAS;IACpB,OAAO,CAAC,KAAK,CAAc;IAC3B,OAAO,CAAC,OAAO,CAAgB;IAC/B,OAAO,CAAC,MAAM,CAAiB;IAC/B,OAAO,CAAC,SAAS,CAAc;IAC/B,OAAO,CAAC,aAAa,CAAuB;IAC5C,OAAO,CAAC,WAAW,CAAK;IACxB,OAAO,CAAC,UAAU,CAAoC;gBAE1C,KAAK,EAAE,WAAW,EAAE,OAAO,EAAE,aAAa;IAKtD;;OAEG;IACH,SAAS,IAAI,aAAa,EAAE;IAI5B;;OAEG;IACI,GAAG,CAAC,IAAI,EAAE,MAAM,EAAE,OAAO,EAAE,kBAAkB,GAAG,cAAc,CAAC,QAAQ,CAAC;IAwJ/E;;OAEG;IACH,OAAO,CAAC,SAAS;IAsBjB;;OAEG;IACH,OAAO,CAAC,SAAS;IAMjB;;OAEG;IACH,OAAO,CAAC,WAAW;IAmBnB;;;OAGG;IACH,OAAO,CAAC,SAAS;IAQjB;;OAEG;IACH,OAAO,CAAC,YAAY;IA2BpB;;OAEG;IACH,OAAO,CAAC,YAAY;IAmBpB;;OAEG;IACH,OAAO,CAAC,UAAU;IA4BlB;;OAEG;YACY,eAAe;IAuS9B;;OAEG;YACY,iBAAiB;IA4ChC;;OAEG;YACY,SAAS;IAwCxB;;OAEG;YACY,cAAc;IAqD7B,OAAO,CAAC,oBAAoB;YAUd,gBAAgB;IAyD9B,OAAO,CAAC,kBAAkB;IAgB1B;;OAEG;YACY,UAAU;IAkDzB;;OAEG;IACH,OAAO,CAAC,YAAY;IA6HpB;;OAEG;IACH,OAAO,CAAC,iBAAiB;IAwCzB;;;;OAIG;IACH,OAAO,CAAC,eAAe;IAqDvB;;;;;;;;;;;;;;OAcG;YACW,mBAAmB;YA+dnB,uBAAuB;YAmDvB,iBAAiB;YA+DjB,sBAAsB;IAqBpC;;OAEG;IACH,OAAO,CAAC,wBAAwB;IAmBhC,OAAO,CAAC,wBAAwB;IAKhC,OAAO,CAAC,uBAAuB;IAM/B;;OAEG;IACH,OAAO,CAAC,eAAe;IAoBvB;;OAEG;IACH,OAAO,CAAC,cAAc;IAkBtB;;OAEG;IACH,OAAO,CAAC,qBAAqB;IAkB7B;;OAEG;IACH,OAAO,CAAC,aAAa;IAkBrB;;OAEG;IACH,OAAO,CAAC,mBAAmB;IAwB3B;;OAEG;IACH,OAAO,CAAC,YAAY;IAOpB;;OAEG;IACH,OAAO,CAAC,KAAK;IAWb;;OAEG;IACH,OAAO,CAAC,OAAO;IAWf;;OAEG;IACG,OAAO,IAAI,OAAO,CAAC,IAAI,CAAC;CAG/B"}

package/dist/runner.js

@@ -5,6 +5,7 @@ import * as fs from 'fs/promises';
 import * as path from 'path';
 import { mountWorkspace } from './prolog/loader.js';
 import { runAgentLoop } from './agent.js';
+import { judgeAnswersToValues, judgeCacheKey, normalizeJudgeState, parseJudgeQuestions, validateJudgeAnswers, valuesToPrologList } from './judge/marshal.js';
 import { executeCompactor, getCompactionBindings, resolveCompactionOptions, } from './compaction.js';
 import { buildToolCompletionEvent, buildToolFailureEvent, buildToolStartEvent, } from './system/runtime/shell-tool-events.js';
 import { checkToolPolicy } from './tools.js';
@@ -145,6 +146,7 @@ export class DMLRunner {
     sessionId = '';
     currentMemory = [];
     stepCounter = 0;
+    judgeCache = new Map();
     constructor(swipl, options) {
         this.swipl = swipl;
         this.options = options;
@@ -162,6 +164,7 @@ export class DMLRunner {
         // Generate unique session ID
         this.sessionId = `sess_${Date.now()}_${Math.random().toString(36).slice(2)}`;
         const memoryId = `mem_${this.sessionId}`; // Kept for API compatibility
+        this.judgeCache.clear();
         try {
             // Mount workspace if path provided
             if (options.workspacePath) {
@@ -243,6 +246,9 @@ export class DMLRunner {
                     case 'request_llm':
                         yield* this.handleLlm(step.payload, options);
                         break;
+                    case 'request_judgment':
+                        yield* this.handleJudgment(step.payload, options);
+                        break;
                     case 'wait_input':
                         // Request user input
                         yield { type: 'input_required', prompt: step.prompt };
@@ -739,6 +745,115 @@ export class DMLRunner {
             yield { type: 'task_activity', taskState: succeeded ? 'completed' : 'failed', taskId: stepId };
         }
     }
+    /**
+     * Handle a semantic judgment request (judge/2 and the one-off predicates).
+     */
+    async *handleJudgment(payload, options) {
+        const parsed = this.parseJudgmentPayload(payload);
+        const ids = parsed.questions.map((question) => question.id);
+        const backendName = parsed.requestedName === 'default'
+            ? options.judgeBackend ?? this.options.judgeSelection?.defaultName ?? 'llm'
+            : parsed.requestedName;
+        const stepId = `judge_${this.stepCounter++}`;
+        yield {
+            type: 'task_activity',
+            taskState: 'started',
+            taskDescription: `judge[${backendName}]: ${ids.join(', ')}`,
+            taskId: stepId,
+            judgeBackend: backendName,
+            judgeQuestionIds: ids,
+            judgeState: 'started',
+        };
+        const result = await this.completeJudgment(parsed, { signal: options.signal, judgeBackend: options.judgeBackend });
+        if (result.error) {
+            this.postJudgmentResult({ success: false, error: result.error });
+            yield { type: 'log', content: `judgment failed: ${result.error}` };
+        }
+        else {
+            this.postJudgmentResult({ success: true, values: result.values ?? [] });
+            if (result.usage) {
+                yield {
+                    type: 'usage',
+                    usage: {
+                        inputTokens: result.usage.inputTokens,
+                        outputTokens: result.usage.outputTokens,
+                        totalTokens: result.usage.totalTokens ?? result.usage.inputTokens + result.usage.outputTokens,
+                    },
+                    usageSource: 'judge',
+                    usageModel: result.usageModel,
+                };
+            }
+        }
+        yield {
+            type: 'task_activity',
+            taskState: result.error ? 'failed' : 'completed',
+            taskId: stepId,
+            judgeBackend: result.backendName,
+            judgeQuestionIds: result.ids,
+            judgeState: result.error ? 'failed' : 'completed',
+        };
+    }
+    parseJudgmentPayload(payload) {
+        const data = (payload ?? {});
+        return {
+            questions: parseJudgeQuestions(data.questions),
+            state: normalizeJudgeState(data.state),
+            requiredCapabilities: Array.isArray(data.requires) ? data.requires.map(String) : [],
+            requestedName: typeof data.backend === 'string' ? data.backend : 'default',
+        };
+    }
+    async completeJudgment(parsed, context) {
+        const selection = this.options.judgeSelection;
+        const defaultName = context.judgeBackend ?? selection?.defaultName ?? 'llm';
+        const backendName = parsed.requestedName === 'default' ? defaultName : parsed.requestedName;
+        const ids = parsed.questions.map((question) => question.id);
+        try {
+            const backend = selection?.backends.get(backendName);
+            if (!backend)
+                throw new Error(`Unknown judgment backend: ${backendName}`);
+            const missing = parsed.requiredCapabilities.filter((capability) => !backend.capabilities[capability]);
+            if (missing.length > 0) {
+                throw new Error(`Judgment backend '${backendName}' lacks required capabilities: ${missing.join(', ')}`);
+            }
+            const model = this.options.model;
+            const key = judgeCacheKey(backend.id, model, parsed.state, parsed.questions);
+            const cached = this.judgeCache.get(key);
+            if (cached) {
+                return { backendName, ids, values: judgeAnswersToValues(parsed.questions, cached) };
+            }
+            const response = await backend.complete({
+                state: parsed.state,
+                questions: parsed.questions,
+                signal: context.signal,
+                model,
+                requiredCapabilities: parsed.requiredCapabilities,
+            });
+            const answers = validateJudgeAnswers(parsed.questions, response.answers);
+            this.judgeCache.set(key, answers);
+            return {
+                backendName,
+                ids,
+                values: judgeAnswersToValues(parsed.questions, answers),
+                usage: response.usage,
+                usageModel: response.model ?? model,
+            };
+        }
+        catch (error) {
+            return {
+                backendName,
+                ids,
+                error: error instanceof Error ? error.message : String(error),
+            };
+        }
+    }
+    postJudgmentResult(result) {
+        if (result.success) {
+            this.query(`deepclause_mi:post_judgment_result('${this.sessionId}', success, ${valuesToPrologList(result.values ?? [])})`);
+        }
+        else {
+            this.query(`deepclause_mi:post_judgment_result('${this.sessionId}', failure, ${this.toPrologTerm(result.error ?? 'Unknown error')})`);
+        }
+    }
     /**
      * Handle exec request (exec() predicate)
      */
@@ -1145,6 +1260,49 @@ export class DMLRunner {
                         }
                         break;
                     }
+                    case 'request_judgment': {
+                        const parsed = this.parseJudgmentPayload(payload);
+                        const nestedStepId = `judge_${this.stepCounter++}`;
+                        const nestedIds = parsed.questions.map((question) => question.id);
+                        onToolEvent?.({
+                            type: 'task_activity',
+                            taskState: 'started',
+                            taskDescription: `judge[${parsed.requestedName}]: ${nestedIds.join(', ')}`,
+                            taskId: nestedStepId,
+                            judgeBackend: parsed.requestedName,
+                            judgeQuestionIds: nestedIds,
+                            judgeState: 'started',
+                        });
+                        const judgeResult = await this.completeJudgment(parsed, { signal: runContext.signal, judgeBackend: undefined });
+                        if (judgeResult.error) {
+                            this.postJudgmentResult({ success: false, error: judgeResult.error });
+                        }
+                        else {
+                            this.postJudgmentResult({ success: true, values: judgeResult.values ?? [] });
+                            if (judgeResult.usage) {
+                                onToolEvent?.({
+                                    type: 'usage',
+                                    usage: {
+                                        inputTokens: judgeResult.usage.inputTokens,
+                                        outputTokens: judgeResult.usage.outputTokens,
+                                        totalTokens: judgeResult.usage.totalTokens ??
+                                            judgeResult.usage.inputTokens + judgeResult.usage.outputTokens,
+                                    },
+                                    usageSource: 'judge',
+                                    usageModel: judgeResult.usageModel,
+                                });
+                            }
+                        }
+                        onToolEvent?.({
+                            type: 'task_activity',
+                            taskState: judgeResult.error ? 'failed' : 'completed',
+                            taskId: nestedStepId,
+                            judgeBackend: judgeResult.backendName,
+                            judgeQuestionIds: judgeResult.ids,
+                            judgeState: judgeResult.error ? 'failed' : 'completed',
+                        });
+                        break;
+                    }
                     case 'request_agent_loop': {
                         // Tool is calling task() or prompt() - run nested agent loop
                         const taskDescription = String(payload?.taskDescription ?? '');