@bjoernrochel/sevro 1.0.0 → 1.1.0

package/dist/src/evaluation/trial.js

@@ -0,0 +1,593 @@
+import { rm } from "node:fs/promises";
+import { hashJson } from "../identity.js";
+import { allocateCandidateRuntime, candidateArtifactRoots, releaseCandidateRuntime, } from "../candidate-artifacts.js";
+import { join } from "node:path";
+import { advisoryBaseRevision } from "../fixtures/advisory.js";
+import { prepareTrialWorkspace } from "./resources.js";
+import { hostArtifacts, hostObservations } from "../evidence/host-results.js";
+import { TrialEvidenceWriter, } from "../evidence/trial-writer.js";
+import { usage } from "../evidence/usage.js";
+import { runAdvisoryReview } from "./advisory.js";
+import { gradeCandidateOutput, gradeShellChecks, gradeGitHeadChecks, gradeExtensionChecks, } from "./grading.js";
+import { gradeSemanticChecks } from "./semantic.js";
+import { HostExecutionFailure, } from "./types.js";
+import { clearFixtureContents, removeDisposableDirectory, } from "../fixtures/cleanup.js";
+import { retainWorkspace } from "../fixtures/workspace-state.js";
+import { gitHeadRevision } from "../graders/git-head.js";
+import { InstrumentationEvidenceError, snapshotInstrumentation, verifyAppliedInstrumentation, } from "../instrumentation.js";
+import { materializeNativeTranscriptView } from "../native-transcript-view.js";
+import { applyTaskVerdictPolicy, assessTrial, } from "../results.js";
+import { scheduleTrials } from "./scheduler.js";
+import { rethrowDiagnosticPersistenceError } from "../host-diagnostics.js";
+import { evaluationProgress } from "./progress.js";
+const MAX_FINAL_MESSAGE_BYTES = 8 * 1024 * 1024;
+function ensureNotCancelled(signal) {
+    if (signal?.aborted)
+        throw new Error("cancelled");
+}
+function hostFailureDiagnostic(execution, error) {
+    if (execution === "cancelled")
+        return { code: "sevro.run.cancelled", message: "run cancelled" };
+    if (error instanceof InstrumentationEvidenceError)
+        return { code: "sevro.instrumentation.mismatch", message: error.message };
+    if (error instanceof HostExecutionFailure)
+        return { code: "sevro.host.failed", message: error.message };
+    return {
+        code: "sevro.host.failed",
+        message: "host execution did not complete",
+    };
+}
+function extensionFinalMessageData(result) {
+    return result?.finalMessage == null ? {} : { text: result.finalMessage };
+}
+function compareDeclaredChecks(left, right, order) {
+    const leftIndex = order.get(left.id);
+    const rightIndex = order.get(right.id);
+    // Match the stable NaN comparator fallback; assessment reports undeclared IDs.
+    if (leftIndex === undefined || rightIndex === undefined)
+        return 0;
+    return leftIndex - rightIndex;
+}
+function selectedTrialTaskPolicy(options, recommendation) {
+    const id = options.extension?.session.identity.selectedTaskVerdictPolicy;
+    return id ? { id, recommendation } : null;
+}
+function initialTrialState(dry) {
+    return {
+        diagnostic: undefined,
+        persisted: false,
+        gitHeadBase: null,
+        advisoryRevision: null,
+        hostResult: null,
+        appliedInstrumentation: [],
+        additionalObservations: [],
+        producedArtifacts: [],
+        execution: dry ? "not_run" : "completed",
+        candidateDurationMs: null,
+        completeness: "unavailable",
+        observation: {
+            id: "sevro.observation.final-message",
+            source: "",
+            completeness: "unavailable",
+            data: {},
+        },
+        gradingObservations: [],
+        checks: [],
+        extensionMetrics: [],
+        domainOutcomes: [],
+        taskPolicyRecommendation: null,
+        graderError: false,
+        assessment: {
+            execution: { status: "not_run" },
+            grading: { status: "not_requested" },
+            task: { verdict: "not_assessed" },
+        },
+        advisoryReview: null,
+        retainedAuxiliaryWorkspaces: [],
+    };
+}
+class EvaluationTrial {
+    options;
+    context;
+    trial;
+    stopAdmission;
+    fixture;
+    state;
+    evidence;
+    constructor(options, context, trial, stopAdmission, fixture) {
+        this.options = options;
+        this.context = context;
+        this.trial = trial;
+        this.stopAdmission = stopAdmission;
+        this.fixture = fixture;
+        this.state = initialTrialState(options.dry);
+        this.state.candidateRuntimeRoot = fixture.candidateRuntimeRoot;
+        this.evidence = new TrialEvidenceWriter({
+            runDir: context.resources.runDir,
+            trial,
+            workspace: fixture.workspace,
+            preparationArtifacts: context.preparationArtifacts,
+        });
+    }
+    async run() {
+        try {
+            if (this.fixture.preparationCancelled)
+                return await this.cancelPreparation();
+            await this.initializeGradingBases();
+            await this.runCandidate();
+            await this.retainCandidateArtifacts();
+            this.observeCandidate();
+            await this.prepareCandidateArtifacts();
+            this.progress("grading", this.state.execution === "completed" ? "started" : "skipped");
+            this.gradeOutput();
+            await this.gradeShell();
+            await this.gradeGitHead();
+            await this.gradeSemantic();
+            await this.gradeExtension();
+            this.assess();
+            if (this.state.execution === "completed")
+                this.progress("grading", "completed");
+            await this.runAdvisory();
+            await this.retainRawCandidate();
+            return await this.persistTrial();
+        }
+        finally {
+            await this.cleanup();
+        }
+    }
+    async cancelPreparation() {
+        this.failCandidate(undefined);
+        await this.retainCandidateArtifacts();
+        this.observeCandidate();
+        this.assess();
+        return this.persistTrial();
+    }
+    async initializeGradingBases() {
+        this.state.gitHeadBase = this.context.prepared.checks.preparedGitHead.length
+            ? await gitHeadRevision(this.fixture.workspace, this.options.signal)
+            : null;
+        this.state.advisoryRevision = this.options.advisoryHost
+            ? await advisoryBaseRevision(this.fixture.workspace)
+            : null;
+    }
+    candidateRequest() {
+        return {
+            retainHostDiagnostics: (bytes) => this.evidence.retainArtifacts([{ id: "sevro.host.diagnostics", bytes: Buffer.from(bytes) }], "host-"),
+            candidateRuntimeRoot: this.state.candidateRuntimeRoot,
+            prompt: this.fixture.trialPrompt,
+            ...(this.fixture.trialFollowUpPrompt !== undefined
+                ? { followUpPrompt: this.fixture.trialFollowUpPrompt }
+                : {}),
+            workspace: this.fixture.workspace,
+            condition: this.options.condition,
+            fixtureBinDir: this.fixture.fixtureBinDir,
+            runtimePolicy: this.candidateRuntimePolicy(),
+            instrumentation: snapshotInstrumentation(this.context.prepared.requestedInstrumentation),
+            ...structuredClone(this.context.prepared.invocation.hostRequest),
+            signal: this.options.signal,
+        };
+    }
+    candidateRuntimePolicy() {
+        if (this.options.runtimePolicy)
+            return this.options.runtimePolicy;
+        if (!this.context.prepared.fixture.fixtureSetup)
+            return undefined;
+        return {
+            format: "sevro.runtime.v1",
+            environment: {},
+            readOnlyRoots: [],
+        };
+    }
+    candidateReservedIds() {
+        return new Set([
+            "sevro.observation.final-message",
+            "sevro.observation.workspaces",
+            ...this.context.preparationArtifacts.map((item) => item.id),
+            ...this.state.additionalObservations.map((item) => item.id),
+            ...this.options.case.checks.map((item) => item.id),
+            ...this.context.prepared.checks.preparedShell.map((item) => `sevro.observation.shell.${item.id}`),
+            ...(this.context.prepared.checks.preparedGitHead.length
+                ? ["sevro.observation.git-head"]
+                : []),
+            ...(this.context.prepared.checks.preparedSemantic.length
+                ? ["sevro.semantic.verdicts"]
+                : []),
+        ]);
+    }
+    acceptCandidateResult(response) {
+        this.state.appliedInstrumentation = verifyAppliedInstrumentation(this.context.prepared.requestedInstrumentation, response.appliedInstrumentation, this.options.condition, response.actualCondition);
+        if (response.finalMessage !== null &&
+            Buffer.byteLength(response.finalMessage, "utf8") > MAX_FINAL_MESSAGE_BYTES)
+            throw new Error("oversized host result");
+        this.state.additionalObservations = hostObservations(response, this.options.host.id, new Set([
+            ...this.context.preparationArtifacts.map((item) => item.id),
+            ...this.options.case.checks.map((item) => item.id),
+        ]));
+        this.state.producedArtifacts = hostArtifacts(response, this.candidateReservedIds());
+        if (response.executionFailed) {
+            this.state.execution = "failed";
+            this.stopAdmission();
+            this.state.diagnostic = {
+                code: "sevro.host.failed",
+                message: response.failureMessage ?? "host execution did not complete",
+            };
+        }
+    }
+    discardCandidateResult() {
+        this.state.hostResult = null;
+        this.state.appliedInstrumentation = [];
+        this.state.additionalObservations = [];
+        this.state.producedArtifacts = [];
+    }
+    failCandidate(error) {
+        this.stopAdmission();
+        this.discardCandidateResult();
+        try {
+            if (error instanceof HostExecutionFailure) {
+                this.acceptCandidateResult(error.result);
+                this.state.hostResult = error.result;
+            }
+        }
+        catch (validationError) {
+            this.discardCandidateResult();
+            error = validationError;
+        }
+        this.state.execution = this.options.signal?.aborted
+            ? "cancelled"
+            : "failed";
+        this.state.diagnostic = hostFailureDiagnostic(this.state.execution, error);
+    }
+    async runCandidate() {
+        if (this.options.dry) {
+            this.progress("candidate", "skipped");
+            return;
+        }
+        this.progress("candidate", "started");
+        const started = performance.now();
+        try {
+            ensureNotCancelled(this.options.signal);
+            this.state.hostResult = await this.options.host.run(this.candidateRequest());
+            this.acceptCandidateResult(this.state.hostResult);
+            if (this.options.signal?.aborted) {
+                this.state.execution = "cancelled";
+                this.stopAdmission();
+                this.state.diagnostic = hostFailureDiagnostic("cancelled", undefined);
+            }
+        }
+        catch (error) {
+            rethrowDiagnosticPersistenceError(error);
+            this.failCandidate(error);
+        }
+        finally {
+            this.state.candidateDurationMs = Math.max(0, performance.now() - started);
+            this.progress("candidate", "completed");
+        }
+    }
+    async retainCandidateArtifacts() {
+        await this.evidence.retainArtifacts(this.state.producedArtifacts, "host-");
+        await this.evidence.verifyArtifacts();
+        const bundle = this.state.producedArtifacts.find((item) => item.id === "sevro.native-transcripts.bundle");
+        if (this.options.runtimePolicy?.nativeTranscripts && bundle)
+            this.state.candidateTranscriptRoot =
+                await materializeNativeTranscriptView(bundle.bytes);
+    }
+    observeCandidate() {
+        this.state.observation = this.evidence.observeCandidate(this.options.host.id, this.state.hostResult);
+        this.state.completeness = this.state.observation.completeness;
+    }
+    failGrading(message, cancelable = true) {
+        if (cancelable && this.options.signal?.aborted) {
+            this.state.execution = "cancelled";
+            this.state.diagnostic = {
+                code: "sevro.run.cancelled",
+                message: "run cancelled",
+            };
+            return;
+        }
+        this.state.graderError = true;
+        this.state.diagnostic = { code: "sevro.grader.error", message };
+        this.stopAdmission();
+    }
+    canGrade() {
+        return this.state.execution === "completed" && !this.state.graderError;
+    }
+    async prepareCandidateArtifacts() {
+        if (this.state.execution !== "completed")
+            return;
+        try {
+            this.state.candidateArtifacts = await candidateArtifactRoots(this.state.candidateRuntimeRoot);
+        }
+        catch {
+            this.failGrading("candidate artifact directories are missing or unsafe");
+        }
+    }
+    acceptGradingFailure(failure) {
+        if (failure)
+            this.failGrading(failure.message, failure.cancelable);
+    }
+    gradeOutput() {
+        if (this.state.execution !== "completed")
+            return;
+        const graded = gradeCandidateOutput({
+            result: this.state.hostResult,
+            checks: this.context.prepared.checks.prepared,
+            evidenceRefs: this.evidence.hasCandidateMessage
+                ? [this.state.observation.id]
+                : [],
+        });
+        this.state.checks = graded.checks;
+        this.acceptGradingFailure(graded.failure);
+    }
+    shellRootRequest() {
+        const isolation = this.options.shellIsolation;
+        if (!isolation)
+            return null;
+        return {
+            workspace: this.fixture.workspace,
+            projectRoot: this.context.prepared.fixture.projectRoot,
+            resultsRoot: this.options.resultsRoot,
+            additionalRoots: [
+                ...isolation.protectedRoots,
+                ...(this.options.preparationSources
+                    ? [this.options.preparationSources.root]
+                    : []),
+                ...(this.context.prepared.fixture.source.kind === "repository"
+                    ? [this.context.prepared.fixture.source.repository.path]
+                    : []),
+                this.context.resources.runStateRoot,
+            ],
+        };
+    }
+    async gradeShell() {
+        if (!this.canGrade() ||
+            this.context.prepared.checks.preparedShell.length === 0)
+            return;
+        const graded = await gradeShellChecks({
+            checks: this.context.prepared.checks.preparedShell,
+            protectedRoots: this.shellRootRequest(),
+            execution: {
+                candidateTranscriptRoot: this.state.candidateTranscriptRoot,
+                candidateArtifactRoots: this.state.candidateArtifacts,
+                workspace: this.fixture.workspace,
+                fixtureBinDir: this.fixture.fixtureBinDir,
+                toolchainBinDir: this.options.shellIsolation?.toolchainBinDir,
+                runtimePolicy: this.options.runtimePolicy,
+                uvRuntimeCache: this.options.shellIsolation?.uvRuntimeCache,
+                privateStateRoot: join(this.context.resources.stateDir, "shell-sandbox"),
+                signal: this.options.signal,
+            },
+        });
+        this.acceptGradingFailure(graded.failure);
+        this.state.gradingObservations.push(...graded.observations);
+        this.state.checks.push(...graded.checks);
+    }
+    async gradeGitHead() {
+        if (!this.canGrade() ||
+            this.context.prepared.checks.preparedGitHead.length === 0)
+            return;
+        const graded = await gradeGitHeadChecks({
+            checks: this.context.prepared.checks.preparedGitHead,
+            base: this.state.gitHeadBase,
+            workspace: this.fixture.workspace,
+            signal: this.options.signal,
+        });
+        this.acceptGradingFailure(graded.failure);
+        this.state.gradingObservations.push(...graded.observations);
+        this.state.checks.push(...graded.checks);
+    }
+    completeCandidateMessage() {
+        const result = this.state.hostResult;
+        if (result?.finalMessage == null || !result.complete)
+            return null;
+        return result.finalMessage;
+    }
+    async gradeSemantic() {
+        if (!this.canGrade())
+            return;
+        const graded = await gradeSemanticChecks({
+            checks: this.context.prepared.checks.preparedSemantic,
+            declaredChecks: this.options.case.checks,
+            host: this.options.semanticHost,
+            message: this.completeCandidateMessage(),
+            candidateWorkspace: this.fixture.workspace,
+            candidateTranscriptRoot: this.state.candidateTranscriptRoot,
+            candidateArtifactRoots: this.state.candidateArtifacts,
+            projectRoot: this.context.prepared.fixture.projectRoot,
+            reservedWorkspace: this.context.resources.workspaces.semantic(this.trial),
+            runtimePolicy: this.options.runtimePolicy,
+            signal: this.options.signal,
+        }, {
+            evidence: this.evidence,
+            onFailure: (failure) => {
+                this.acceptGradingFailure(failure);
+            },
+            retainWorkspace: (workspace) => this.retainAuxiliaryWorkspace(workspace),
+        });
+        this.state.gradingObservations.push(...graded.observations);
+        this.state.checks.push(...graded.checks);
+    }
+    trialObservations() {
+        return [
+            this.state.observation,
+            ...this.state.gradingObservations,
+            ...this.state.additionalObservations,
+            {
+                id: "sevro.observation.workspaces",
+                source: "sevro.runner",
+                completeness: "complete",
+                data: {
+                    candidate: this.fixture.workspace,
+                    retainedAuxiliary: this.state.retainedAuxiliaryWorkspaces,
+                },
+            },
+        ];
+    }
+    extensionEvaluationRequest() {
+        return {
+            caseId: this.options.case.id,
+            execution: { status: this.state.execution },
+            observations: this.trialObservations().map((observation) => observation === this.state.observation
+                ? {
+                    ...observation,
+                    completeness: this.state.completeness,
+                    data: {
+                        ...observation.data,
+                        ...extensionFinalMessageData(this.state.hostResult),
+                    },
+                }
+                : observation),
+            builtinChecks: this.state.checks.map((check) => ({
+                id: check.id,
+                status: check.status,
+                ...(check.detail ? { detail: check.detail } : {}),
+                evidenceRefs: check.evidenceRefs ?? [],
+            })),
+            artifacts: this.evidence.artifactRefs,
+            extensionData: this.context.prepared.extensionData,
+        };
+    }
+    async gradeExtension() {
+        const extension = this.options.extension;
+        if (!extension || !this.canGrade())
+            return;
+        const graded = await gradeExtensionChecks({
+            evaluate: (request) => extension.session.evaluate(request),
+            request: this.extensionEvaluationRequest(),
+            declarations: this.context.prepared.checks.extensionDeclarations,
+        });
+        this.acceptGradingFailure(graded.failure);
+        this.state.checks.push(...graded.checks);
+        this.state.extensionMetrics = graded.metrics;
+        this.state.domainOutcomes = graded.domainOutcomes;
+        this.state.taskPolicyRecommendation = graded.recommendation;
+    }
+    assess() {
+        const declaredOrder = new Map(this.options.case.checks.map((check, index) => [check.id, index]));
+        this.state.checks.sort((left, right) => compareDeclaredChecks(left, right, declaredOrder));
+        const defaultAssessment = assessTrial({
+            execution: this.state.execution,
+            declaredChecks: this.options.case.checks
+                .filter((check) => !this.context.prepared.checks.replaced.has(check.grader))
+                .map((check) => check.id),
+            checks: this.state.checks,
+            graderError: this.state.graderError,
+            requiredEvidenceUnavailable: this.options.case.requiredEvidence.some((id) => !this.state.producedArtifacts.some((item) => item.id === id) &&
+                !this.trialObservations().some((item) => item.id === id && item.completeness === "complete")),
+        });
+        this.state.assessment = applyTaskVerdictPolicy(defaultAssessment, this.state.taskPolicyRecommendation, Boolean(this.options.extension?.session.identity.selectedTaskVerdictPolicy));
+    }
+    retainAuxiliaryWorkspace(workspace) {
+        this.state.retainedAuxiliaryWorkspaces.push(workspace);
+        this.context.resources.workspaces.retain(workspace);
+        return retainWorkspace(workspace);
+    }
+    async runAdvisory() {
+        const host = this.options.advisoryHost;
+        if (!host)
+            return;
+        this.state.advisoryReview = await runAdvisoryReview({
+            host,
+            candidateCompleted: this.state.execution === "completed",
+            candidateWorkspace: this.fixture.workspace,
+            candidateTranscriptRoot: this.state.candidateTranscriptRoot,
+            candidateArtifactRoots: this.state.candidateArtifacts,
+            baseRevision: this.state.advisoryRevision,
+            excludedPaths: this.options.advisoryExcludedPaths,
+            task: this.fixture.trialPrompt,
+            checks: this.state.checks,
+            declaredChecks: this.options.case.checks,
+            runtimePolicy: this.options.runtimePolicy,
+            signal: this.options.signal,
+        }, {
+            evidence: this.evidence,
+            retainWorkspace: (workspace) => this.retainAuxiliaryWorkspace(workspace),
+        });
+    }
+    retainRawCandidate() {
+        return this.evidence.retainRawCandidate(this.state.hostResult?.finalMessage);
+    }
+    trialEvidenceRecord() {
+        return {
+            caseId: this.options.case.id,
+            trial: this.trial,
+            executionMode: this.options.dry ? "dry" : "executed",
+            candidateDurationMs: this.state.candidateDurationMs,
+            condition: {
+                requested: this.options.condition,
+                actual: this.state.hostResult?.actualCondition ?? "unknown",
+                appliedInstrumentation: this.state.appliedInstrumentation,
+            },
+            observationCompleteness: this.state.completeness,
+            observations: this.trialObservations(),
+            metrics: this.state.extensionMetrics,
+            domainOutcomes: this.state.domainOutcomes,
+            taskVerdictPolicy: selectedTrialTaskPolicy(this.options, this.state.taskPolicyRecommendation),
+            advisoryReview: this.state.advisoryReview,
+            routes: this.context.prepared.routes,
+            usage: usage(this.state.hostResult),
+            rawResult: this.evidence.candidateRawResult(this.options.host.id),
+            artifactRefs: this.evidence.artifactRefs,
+        };
+    }
+    async persistTrial() {
+        const evidence = this.trialEvidenceRecord();
+        this.progress("evidence", "started");
+        const trialSummary = await this.evidence.persistTrial(this.options.case.id, this.state.assessment, this.state.checks, this.state.domainOutcomes, evidence);
+        this.state.persisted = true;
+        this.state.diagnostic = this.evidence.diagnosticLocation(this.state.diagnostic);
+        this.context.results.complete({
+            summary: trialSummary,
+            evidence,
+            diagnostic: this.state.diagnostic,
+        });
+        this.progress("evidence", "completed");
+        return !((this.state.execution !== "completed" &&
+            this.state.execution !== "not_run") ||
+            this.state.graderError);
+    }
+    progress(stage, status) {
+        evaluationProgress(this.options, stage, status, this.trial);
+    }
+    async cleanup() {
+        await removeDisposableDirectory(join(this.context.resources.stateDir, "shell-sandbox/runtime", hashJson(this.fixture.workspace)));
+        if (this.state.candidateTranscriptRoot)
+            await rm(this.state.candidateTranscriptRoot, {
+                recursive: true,
+                force: true,
+            });
+        if (this.state.persisted && !this.retainFailedWorkspace()) {
+            try {
+                await clearFixtureContents(this.fixture.workspace);
+                this.context.resources.workspaces.completed(this.fixture.reservation);
+            }
+            catch {
+                console.warn(`fixture cleanup failed; retained at ${this.fixture.workspace}`);
+            }
+        }
+    }
+    retainFailedWorkspace() {
+        return (this.state.execution === "failed" ||
+            this.state.execution === "cancelled" ||
+            this.state.graderError ||
+            this.state.assessment.grading.status === "unavailable" ||
+            this.state.assessment.task.verdict === "failed");
+    }
+}
+export async function executeEvaluationTrials(options, context) {
+    await scheduleTrials({
+        count: options.trialCount,
+        jobs: context.prepared.jobs,
+        signal: options.signal,
+    }, async (trial, stopAdmission) => {
+        const candidateRuntimeRoot = await allocateCandidateRuntime(options.runtimePolicy, Boolean(context.prepared.fixture.fixtureSetup));
+        try {
+            const fixture = await prepareTrialWorkspace(options, context.prepared.fixture, context.prepared.invocation, context.resources.workspaces, trial, candidateRuntimeRoot);
+            evaluationProgress(options, "preparation", "completed", trial);
+            return await new EvaluationTrial(options, context, trial, stopAdmission, fixture).run();
+        }
+        finally {
+            await releaseCandidateRuntime(candidateRuntimeRoot);
+        }
+    });
+}
+//# sourceMappingURL=trial.js.map