@principles/core 1.284.17 → 1.284.19

package/dist/runtime-v2/internalization/evaluator-runner.js

@@ -8,9 +8,7 @@ import { HUMAN_REVIEW_REASON, planOwnerVerdictOverrideResume, markOwnerResolutio
 import { EvaluatorPromptBuilder, deriveRequirementLedger } from './evaluator-prompt-builder.js';
 import { reconcileLineageEcho } from './peer-runner-contracts.js';
 import { BasePeerRunner } from '../runner/base-peer-runner.js';
-import { EVALUATOR_STAGE1_MANIFEST, EVALUATOR_STAGE2_MANIFEST } from './context-manifests.js';
 import { extractIntentContract } from './intent-contract.js';
-import { evaluateFlaggedCriteria, isForcedStage2 } from './progressive-evaluator.js';
 // PRI-426: single-round adversarial sandbox replay in succeedTask.
 import { evaluateRefinerRuleHostGate } from './refiner-rulehost-gate.js';
 import { attributionFromLayer, partitionV2OutOfScopeFailures, resolveRequiresContextVersionFromArtifact } from './rule-reliability-validation.js';
@@ -121,53 +119,6 @@ function extractScribeArtifactId(artificerContentJson) {
         return direct;
     return null;
 }
-/**
- * Layer 0 (design §6.1, F17): evaluator's edge predecessor is `artificer` —
- * NOT scribe, even though buildContext loads both. The scribe artifact is
- * still consumed by invokeRuntime for code-review intent consistency; only the
- * artificer goes into `predecessorSummary`. Reusing the already-loaded string
- * keeps the writer path at zero extra store reads (F3). Returns null when no
- * artificer artifact was resolved — the writer then emits
- * `artifact_summary_predecessor_absent` and writes only the self `summary`.
- */
-function toArtificerPredecessor(context) {
-    if (!context.artificerArtifact || !context.sourceArtificerArtifactId)
-        return null;
-    let contentJson;
-    try {
-        contentJson = JSON.parse(context.artificerArtifact);
-    }
-    catch {
-        contentJson = context.artificerArtifact;
-    }
-    return {
-        artifactId: context.sourceArtificerArtifactId,
-        runnerKind: 'artificer',
-        contentJson,
-    };
-}
-/**
- * Check Stage 1 output for contract violations (design §6.5.4, Req 7.18).
- * Returns a reason string when the output is malformed, or null when valid.
- * Covers the failure modes observed in Phase 0 testing:
- *   - non-object (null, string, array)
- *   - missing `evaluation` sub-object (the core evaluator output structure)
- *   - `evaluation.decision` absent or not a string
- */
-function checkStage1ContractOutput(output) {
-    if (output === null || typeof output !== 'object' || Array.isArray(output)) {
-        return 'output_not_object';
-    }
-    const rec = output;
-    if (!Object.hasOwn(rec, 'evaluation') || rec.evaluation === null || typeof rec.evaluation !== 'object' || Array.isArray(rec.evaluation)) {
-        return 'evaluation_missing';
-    }
-    const ev = rec.evaluation;
-    if (!Object.hasOwn(ev, 'decision') || typeof ev.decision !== 'string') {
-        return 'decision_missing';
-    }
-    return null;
-}
 export const DEFAULT_EVALUATOR_RUNNER_OPTIONS = {
     pollIntervalMs: 5_000,
     timeoutMs: 300_000,
@@ -320,112 +271,20 @@ export class EvaluatorRunner extends BasePeerRunner {
         this.emitEvent('no_artificer_artifact', taskId, {});
         throw new PDRuntimeError('input_invalid', 'Artificer dependency artifact not found');
     }
-    /** Synthetic RunHandle prefix for progressive-evaluator pre-resolved output. */
-    static PROGRESSIVE_RUN_ID_PREFIX = '__progressive_evaluator_';
-    progressiveFinalOutput = undefined;
-    progressiveRunActive = false;
     async invokeRuntime(taskId, context) {
-        // Layer 2 (design §6.5, task 9.11): when progressive_evaluator flag is on,
-        // run two-stage evaluation (Stage 1 summary → flagged? → Stage 2 tier2).
-        // When off, the existing single-stage path runs unchanged.
-        if (!this.isProgressiveEvaluatorEnabled()) {
-            return this.invokeRuntimeSingleStage(taskId, context);
-        }
-        // ── Two-stage progressive evaluation ──
-        // Stage 1: summary-level evaluation (same prompt as single-stage, but
-        // uses EVALUATOR_STAGE1_MANIFEST for focused context).
-        const { message: stage1Message, systemPrompt: stage1SystemPrompt } = await this.buildEvaluatorPrompt(taskId, context, EVALUATOR_STAGE1_MANIFEST);
-        const stage1Output = await this.runSingleEvaluation(taskId, stage1Message, stage1SystemPrompt);
-        // 9.4c (design §6.5.4): Stage 1 output contract violation check.
-        // Detect malformed Stage 1 output shapes that Phase 0 testing identified:
-        // non-object, missing required evaluator fields, or the output being a
-        // string (markdown-fenced / truncated). In all cases, emit a structured
-        // degradation event and force Stage 2 (rc-3 + rc-9: never silently pass).
-        //
-        // NOTE (Req 7.18): The full attemptStructuredOutputRepair channel is not
-        // wired here because it requires an llmCaller callback that connects to
-        // runtimeAdapter's internal LLM invocation path — a complex integration
-        // deferred to a follow-up PR. This simplified check covers the detected
-        // failure modes (non-object, missing evaluation field) and ensures they
-        // are never silently treated as "pass".
-        const stage1ContractViolated = checkStage1ContractOutput(stage1Output);
-        if (stage1ContractViolated !== null) {
-            this.emitEvent('stage1_output_contract_violation', taskId, {
-                reason: stage1ContractViolated,
-                detail: `Stage 1 output contract violated (${stage1ContractViolated}) — forcing Stage 2.`,
-            });
-        }
-        // Evaluate flagged criteria on Stage 1 output.
-        // When contract is violated, ALL criteria fields are undetermined → forces Stage 2.
-        const d1 = stage1ContractViolated !== null
-            ? { flagged: false, reasons: [], undetermined: ['stage1_output_contract_violation'] }
-            : evaluateFlaggedCriteria(stage1Output);
-        const forced = isForcedStage2(taskId);
-        if (!d1.flagged && !forced && d1.undetermined.length === 0) {
-            // Stage 1 is sufficient — return its output as the final result.
-            this.progressiveFinalOutput = stage1Output;
-            this.progressiveRunActive = true;
-            return { runId: `${EvaluatorRunner.PROGRESSIVE_RUN_ID_PREFIX}stage1`, runtimeKind: this.getRuntimeKind(), startedAt: new Date().toISOString() };
-        }
-        // Stage 2 triggered: independent re-evaluation with tier2 context.
-        // rc-7 / ERR-015 / ERR-018 / ERR-019: Stage 2 does NOT receive Stage 1
-        // output, concerns, or the FlaggedDecision. It is a fully independent call.
-        const { message: stage2Message, systemPrompt: stage2SystemPrompt, stage2Evidence } = await this.buildEvaluatorPrompt(taskId, context, EVALUATOR_STAGE2_MANIFEST);
-        if (stage2Evidence.state === 'required_unavailable') {
-            // Review round (information floor): Stage 2 is the deep-evidence stage;
-            // its REQUIRED evidence is unavailable, so there is NO safe fallback
-            // that still carries it — the full artificer artifact is the
-            // implementer's view, not the pain/dreamer deep evidence. Telemetry
-            // alone is not correctness: issuing an authoritative verdict from a
-            // Stage-2 prompt that lacks its declared evidence would silently
-            // proceed. Refuse the LLM round entirely (0 extra calls) and fail
-            // loud (input_invalid is permanent — no blind retry), mirroring the
-            // PR-A repair-evidence-unavailable contract.
-            this.emitEvent('stage2_required_evidence_unavailable', taskId, {
-                manifestId: EVALUATOR_STAGE2_MANIFEST.manifestId,
-                requiredPaths: stage2Evidence.unresolvedRequired,
-                nextAction: 'verify_durable_diagnosis_and_dreamer_ancestry_before_reevaluation',
-            });
-            throw new PDRuntimeError('input_invalid', `evaluator stage2: required tier2 evidence unavailable (${stage2Evidence.unresolvedRequired.join(', ')}) — refusing to issue a deep-evidence verdict without it.`);
-        }
-        const stage2Output = await this.runSingleEvaluation(taskId, stage2Message, stage2SystemPrompt);
-        this.progressiveFinalOutput = stage2Output;
-        this.progressiveRunActive = true;
-        return { runId: `${EvaluatorRunner.PROGRESSIVE_RUN_ID_PREFIX}stage2`, runtimeKind: this.getRuntimeKind(), startedAt: new Date().toISOString() };
-    }
-    /**
-     * Override pollUntilTerminal: for synthetic progressive-evaluator RunHandles,
-     * the LLM call already completed inside invokeRuntime — return succeeded
-     * immediately without hitting the real runtime adapter (which would fail on
-     * the synthetic runId). For normal handles, delegate to the base class.
-     */
-    async pollUntilTerminal(runHandle) {
-        if (this.progressiveRunActive && runHandle.runId.startsWith(EvaluatorRunner.PROGRESSIVE_RUN_ID_PREFIX)) {
-            return { status: 'succeeded', runId: runHandle.runId };
-        }
-        return super.pollUntilTerminal(runHandle);
-    }
-    /**
-     * Intercept fetchAndParseOutput for progressive-evaluator runs: when the
-     * runId is a synthetic progressive handle, return the pre-resolved output
-     * directly (the actual LLM call already happened inside invokeRuntime).
-     */
-    async fetchAndParseOutput(runId, taskId) {
-        if (this.progressiveRunActive && runId.startsWith(EvaluatorRunner.PROGRESSIVE_RUN_ID_PREFIX)) {
-            return this.progressiveFinalOutput;
-        }
-        return super.fetchAndParseOutput(runId, taskId);
+        // Single-stage evaluation is the only path — the two-stage
+        // progressive_evaluator branch was retired with PRI-819 R-06.
+        const { message, systemPrompt } = await this.buildEvaluatorPrompt(taskId, context);
+        return this.runtimeAdapter.startRun({
+            agentSpec: { agentId: this.resolvedOptions.agentId, schemaVersion: 'v1' },
+            taskRef: { taskId },
+            inputPayload: message,
+            contextItems: [],
+            outputSchemaRef: 'evaluator-output-v1',
+            timeoutMs: this.resolvedOptions.timeoutMs,
+            systemPrompt,
+        });
     }
-    /**
-     * Check Stage 1 output for contract violations (design §6.5.4, Req 7.18).
-     * Returns a reason string when the output is malformed, or null when valid.
-     */
-    // eslint-disable-next-line @typescript-eslint/class-methods-use-this -- pure function, no instance state needed
-    checkStage1Contract(_output) { return null; }
-    /**
-     * Build the evaluator prompt with the given manifest's resolved context.
-     * Shared by single-stage and two-stage paths.
-     */
     /**
      * PRI-630 收敛契约 (SPEC §18.1): 解析上轮评估上下文 — 从 dependency
      * artificer 的 repairPayload.sourceEvaluatorTaskId 找到上轮 evaluator,
@@ -537,19 +396,9 @@ export class EvaluatorRunner extends BasePeerRunner {
         };
     }
     /**
-     * Stage 2 required-evidence outcome (review round / information floor):
-     *  - `focused`            — required tier2 (`diagnostician.raw.evidence`,
-     *                           `dreamer.raw.candidates`) resolved from the
-     *                           durable CandidateLineage.
-     *  - `fallback_other`     — non-required fallback (envelope sparsity, budget
-     *                           flag off): the full artificer artifact is used,
-     *                           which is the pre-PR-B legacy assembly.
-     *  - `required_unavailable` — REQUIRED deep evidence is absent/truncated or
-     *                           the lineage is corrupt. No safe legacy fallback
-     *                           exists for a deep-evidence stage: the caller MUST
-     *                           refuse to send the Stage 2 prompt.
+     * Build the evaluator prompt from the full predecessor artifacts.
      */
-    async buildEvaluatorPrompt(taskId, context, manifest) {
+    async buildEvaluatorPrompt(taskId, context) {
         let parsedArtificerArtifact = null;
         if (context.artificerArtifact) {
             try {
@@ -568,46 +417,6 @@ export class EvaluatorRunner extends BasePeerRunner {
                 parsedScribeArtifact = context.scribeArtifact;
             }
         }
-        // Resolve manifest-injected focused fields (Layer 1 + PR B tier2).
-        const artificerPred = toArtificerPredecessor(context);
-        const isStage2 = manifest.manifestId === EVALUATOR_STAGE2_MANIFEST.manifestId;
-        let resolutionOutcome = isStage2
-            ? { state: 'fallback_other' }
-            : { state: 'not_stage2' };
-        if (artificerPred !== null) {
-            // Stage 2 is by definition the deep-evidence stage, so its tier2 raw
-            // fields (`diagnostician.raw.evidence`, `dreamer.raw.candidates`) are
-            // REQUIRED. When they cannot be resolved from the durable CandidateLineage
-            // — or the budget cannot carry them — resolution falls back to the
-            // authoritative full artificer artifact rather than injecting a silently
-            // thinner context (design §34/§35: no silent required-field loss).
-            const requiredPaths = isStage2 ? [...EVALUATOR_STAGE2_MANIFEST.tier2] : [];
-            const resolved = await this.resolveContextInjectionAsync({
-                taskId,
-                manifest,
-                predecessorContentJson: artificerPred.contentJson,
-                startArtifactId: context.sourceArtificerArtifactId ?? undefined,
-                requiredPaths,
-            });
-            if (resolved.mode === 'focused') {
-                parsedArtificerArtifact = resolved.fields;
-                if (isStage2)
-                    resolutionOutcome = { state: 'focused' };
-            }
-            else if (isStage2
-                && resolved.mode === 'fallback'
-                && (resolved.reason === 'required_evidence_unresolved' || resolved.reason === 'lineage_unavailable')) {
-                // Review round: telemetry alone is not correctness. A deep-evidence
-                // stage whose REQUIRED evidence is unavailable must NOT proceed —
-                // there is no safe legacy fallback that contains the missing evidence
-                // (the full artificer artifact is the implementer's view, not the
-                // pain/dreamer deep evidence). The caller aborts before any LLM call.
-                resolutionOutcome = {
-                    state: 'required_unavailable',
-                    unresolvedRequired: resolved.unresolvedRequired,
-                };
-            }
-        }
         const builder = new EvaluatorPromptBuilder();
         const { message, systemPrompt } = builder.buildPrompt({
             taskId,
@@ -626,20 +435,7 @@ export class EvaluatorRunner extends BasePeerRunner {
             // undefined → prompt unchanged (backward compatible).
             intentContract: extractIntentContract(parsedScribeArtifact) ?? undefined,
         });
-        return { message, systemPrompt, stage2Evidence: resolutionOutcome };
-    }
-    /** Original single-stage invokeRuntime (flag-off path). */
-    async invokeRuntimeSingleStage(taskId, context) {
-        const { message, systemPrompt } = await this.buildEvaluatorPrompt(taskId, context, EVALUATOR_STAGE1_MANIFEST);
-        return this.runtimeAdapter.startRun({
-            agentSpec: { agentId: this.resolvedOptions.agentId, schemaVersion: 'v1' },
-            taskRef: { taskId },
-            inputPayload: message,
-            contextItems: [],
-            outputSchemaRef: 'evaluator-output-v1',
-            timeoutMs: this.resolvedOptions.timeoutMs,
-            systemPrompt,
-        });
+        return { message, systemPrompt };
     }
     async validateOutput(output, taskId, context) {
         // PRI-630 P1 评审修复: 修复轮把上轮 requirement ids 传入做完整覆盖校验
@@ -730,11 +526,7 @@ export class EvaluatorRunner extends BasePeerRunner {
                 sourceTaskId: taskId,
                 lineageArtifactIds,
                 validationStatus: 'pending',
-                // Layer 0 (design §6.1, task 3.11): evaluator's edge predecessor is
-                // artificer (NOT scribe — scribe is loaded for code review but is not
-                // the edge predecessor, F17). The scribe artifact still flows through
-                // the separate extractScribeArtifactId path (F3).
-                contentJson: this.buildArtifactContentJson(taskId, 'evaluator', output, toArtificerPredecessor(context)),
+                contentJson: JSON.stringify(output),
                 createdAt: now,
                 updatedAt: now,
             });
@@ -828,7 +620,7 @@ export class EvaluatorRunner extends BasePeerRunner {
                                     sourceTaskId: taskId,
                                     lineageArtifactIds,
                                     validationStatus: 'pending',
-                                    contentJson: this.buildArtifactContentJson(taskId, 'evaluator', finalOutput, toArtificerPredecessor(context)),
+                                    contentJson: JSON.stringify(finalOutput),
                                     createdAt: now,
                                     updatedAt: new Date().toISOString(),
                                 });
@@ -926,11 +718,7 @@ export class EvaluatorRunner extends BasePeerRunner {
                         sourceTaskId: taskId,
                         lineageArtifactIds,
                         validationStatus: 'pending',
-                        // Layer 0 (design §6.1): re-build the envelope on the final output
-                        // so the replay re-persist keeps the same `summary` /
-                        // `predecessorSummary` fields as the initial write (the first write
-                        // would otherwise be overwritten with a bare JSON.stringify).
-                        contentJson: this.buildArtifactContentJson(taskId, 'evaluator', finalOutput, toArtificerPredecessor(context)),
+                        contentJson: JSON.stringify(finalOutput),
                         createdAt: now,
                         updatedAt: new Date().toISOString(),
                     });