feat(confidence-engine): add completed episode reasoning seam

2026-09-01 09:28:02 +01:00
parent ab655e2222
commit c89cc51ae6
2 changed files with 264 additions and 0 deletions
@@ -31,6 +31,8 @@ import {
  validateGraphReferences,
 } from "./utils.js";
 import { validateFindings, produceFindingInformedSummary } from "./finding-helpers.js";
+import { prepareCompletedEpisode } from "./episode-preparation.js";
+import { buildEpisodeAwareGraphPrompt } from "./prompt-builder-episode.js";

 function toValidationErrors(error) {
  return (
@@ -1106,3 +1108,107 @@ async function updateCaseWithDependencies(body, dependencies = {}) {
    }),
  };
 }
+
+// ── Episode-aware reconsideration seam ─────────────────────
+
+/**
+ * Reason over a completed focused episode and return a parsed GraphUpdateProposal.
+ *
+ * This is the smallest sibling to updateCase that reuses existing provider,
+ * model, and parsing mechanics — only the prompt path differs.
+ */
+export async function reconsiderCompletedEpisode(epiParams, deps = {}) {
+  const { situationGraph: epiGraph, targetNodeId, contributions, findings, episode: rawEpisode } = epiParams;
+
+  // Accept prepared episode or produce it deterministically
+
+  let episode;
+
+  if (rawEpisode) {
+    episode = rawEpisode;
+  } else if (epiParams?.turns || epiParams?.eligibleCanonicalFindings) {
+    // Already a prepared episode passed as first positional arg
+    episode = epiParams;
+  } else {
+    episode = prepareCompletedEpisode(epiParams);
+  }
+
+  if (!episode?.turns && !episode?.eligibleCanonicalFindings) {
+    return {
+      success: false,
+      stage: "input_validation",
+      error: "Invalid prepared episode — requires turns or findings",
+      statusCode: 400,
+    };
+  }
+
+  const getModelName = deps.getModelName;
+  const buildEpisodePrompt = deps.buildEpisodePrompt;
+  const parseProposal = deps.parseProposal;
+  const provider = epiParams.provider ?? deps.provider;
+
+  let modelName = null;
+  let rawResponse;
+  const startedAt = Date.now();
+
+  try {
+    modelName = deps.modelName ?? (deps.config ? deps.config.OLLAMA_MODEL : null) ?? getModelName?.() ?? null;
+
+    const promptBuilder = buildEpisodePrompt ?? buildEpisodeAwareGraphPrompt;
+    const prompt = promptBuilder({ episode });
+
+    const prov = provider ?? getProvider();
+    rawResponse = await prov.generateReconstruction(prompt, modelName);
+  } catch (error) {
+    return {
+      success: false,
+      stage: "provider",
+      error: "Graph update proposal generation failed",
+      providerErrors: [
+        sanitiseErrorMessage(
+          error,
+          "Provider failed to generate graph update proposal from episode",
+        ),
+      ],
+      diagnostics: {
+        promptVersion: null,
+        modelName,
+        responseDurationMs: Date.now() - startedAt,
+        normalisationsApplied: [],
+      },
+      statusCode: 502,
+    };
+  }
+
+  const parser = parseProposal ?? parseGraphUpdateProposal;
+  const parsedProposal = parser(rawResponse);
+  const responseDurationMs = Date.now() - startedAt;
+
+  if (!parsedProposal.success) {
+    return {
+      success: false,
+      stage: "proposal_validation",
+      error: "Invalid graph update proposal",
+      proposalErrors: parsedProposal.errors,
+      diagnostics: {
+        promptVersion: null,
+        modelName,
+        responseDurationMs,
+        normalisationsApplied: parsedProposal.normalisationsApplied,
+      },
+      statusCode: 502,
+    };
+  }
+
+  return {
+    success: true,
+    stage: "proposal_ready",
+    proposal: parsedProposal.proposal,
+    diagnostics: {
+      promptVersion: null,
+      modelName,
+      responseDurationMs,
+      normalisationsApplied: parsedProposal.normalisationsApplied,
+    },
+  };
+}
@@ -0,0 +1,158 @@
+import { beforeEach, describe, expect, it, vi } from "vitest";
+
+// ── Fixtures ────────────────────────────────────────────────
+
+function makePreparedEpisode(overrides = {}) {
+  return {
+    situationGraph: {
+      centralStatement: "Complaint counts increased.",
+      nodes: [
+        { id: "n-unknown", label: "Rate denominator", kind: "unknown", status: "unknown" },
+      ],
+      edges: [],
+      activeUnknownNodeId: "n-unknown",
+      resolvedNodeIds: [],
+    },
+    targetNodeId: "n-unknown",
+    turns: [
+      { contributionId: "contrib-a", sequence: 1, question: "What denominator?", answer: "100 units" },
+      { contributionId: "contrib-b", sequence: 2, question: "Is it consistent?", answer: "Yes" },
+    ],
+    eligibleCanonicalFindings: [
+      { findingId: "f-1", contributionId: "contrib-a", proposition: "Denominator is fixed at 100.", endorsement: null },
+    ],
+    excludedFindingProvenance: [],
+    ...overrides,
+  };
+}
+
+function makeProposalResponse(overrides = {}) {
+  return {
+    addedNodes: [],
+    updatedNodes: [
+      {
+        nodeId: "n-unknown",
+        previousStatus: "unknown",
+        newStatus: "resolved",
+        previousValue: null,
+        newValue: "100 units",
+        reason: "Denominator confirmed.",
+      },
+    ],
+    addedEdges: [],
+    removedEdgeIds: [],
+    resolvedUnknownNodeIds: ["n-unknown"],
+    affectedNodeIds: [],
+    selectedQuestion: null,
+    answerMeaning: null,
+    structuralActionRequired: false,
+    ...overrides,
+  };
+}
+
+const MOCK_CONFIG = { OLLAMA_MODEL: "configured" };
+const TEST_EPISODE = makePreparedEpisode();
+
+// ── Tests ───────────────────────────────────────────────────
+
+describe("episode reasoning seam (reconsiderCompletedEpisode)", () => {
+  beforeEach(() => {
+    vi.resetModules();
+    vi.clearAllMocks();
+  });
+
+  it("uses the episode-aware prompt builder, not legacy buildGraphUpdatePrompt", async () => {
+    const buildEpisodeAwareGraphPrompt = vi.fn().mockReturnValue("EPISODE_PROMPT");
+    const buildGraphUpdatePrompt = vi.fn();
+    const provider = {
+      generateReconstruction: vi.fn().mockResolvedValue(makeProposalResponse()),
+    };
+
+    // epiParams carries provider via its own key (as in the API contract)
+    const result = await (await import("@/lib/graph/orchestrator.js")).reconsiderCompletedEpisode(
+      { ...TEST_EPISODE, provider },
+      { getModelName: () => "configured", buildEpisodePrompt: buildEpisodeAwareGraphPrompt },
+    );
+
+    expect(result.success).toBe(true);
+    expect(buildEpisodeAwareGraphPrompt).toHaveBeenCalledWith({ episode: expect.any(Object) });
+    expect(provider.generateReconstruction).toHaveBeenCalledWith("EPISODE_PROMPT", "configured");
+    expect(buildGraphUpdatePrompt).not.toHaveBeenCalled();
+  });
+
+  it("calls provider.generateReconstruction exactly once with the built episode prompt", async () => {
+    const buildEpisodeAwareGraphPrompt = vi.fn().mockReturnValue("EPISODE_PROMPT_V2");
+    const provider = {
+      generateReconstruction: vi.fn().mockResolvedValue(makeProposalResponse()),
+    };
+
+    const { reconsiderCompletedEpisode } = await import("@/lib/graph/orchestrator.js");
+    const result = await reconsiderCompletedEpisode(
+      { ...TEST_EPISODE, provider },
+      { getModelName: () => "configured", buildEpisodePrompt: buildEpisodeAwareGraphPrompt },
+    );
+
+    expect(result.success).toBe(true);
+    expect(provider.generateReconstruction.mock.calls.length).toBe(1);
+    expect(provider.generateReconstruction).toHaveBeenCalledWith("EPISODE_PROMPT_V2", "configured");
+  });
+
+  it("reuses existing parser — returns parsed GraphUpdateProposal shape for valid JSON", async () => {
+    const provider = {
+      generateReconstruction: vi.fn().mockResolvedValue(JSON.stringify(makeProposalResponse())),
+    };
+
+    const { reconsiderCompletedEpisode } = await import("@/lib/graph/orchestrator.js");
+    const result = await reconsiderCompletedEpisode(
+      TEST_EPISODE,
+      { getModelName: () => "configured", provider },
+    );
+
+    expect(result.success).toBe(true);
+    expect(result.stage).toBe("proposal_ready");
+    expect(result.proposal).toMatchObject({
+      updatedNodes: [expect.objectContaining({ nodeId: "n-unknown" })],
+      resolvedUnknownNodeIds: ["n-unknown"],
+      answerMeaning: null,
+      structuralActionRequired: false,
+    });
+  });
+
+  it("preserves existing parser error behaviour for invalid provider result", async () => {
+    const provider = {
+      generateReconstruction: vi.fn().mockResolvedValue("{ not valid json }"),
+    };
+
+    const { reconsiderCompletedEpisode } = await import("@/lib/graph/orchestrator.js");
+    const result = await reconsiderCompletedEpisode(
+      TEST_EPISODE,
+      { getModelName: () => "configured", provider },
+    );
+
+    expect(result.success).toBe(false);
+    expect(result.stage).toBe("proposal_validation");
+    expect(result.error).toBe("Invalid graph update proposal");
+    expect(result.statusCode).toBe(502);
+    expect(result.proposalErrors).toBeInstanceOf(Array);
+  });
+
+  it("does not perform any graph mutation or application", async () => {
+    const applyValidatedProposal = vi.fn();
+    const provider = {
+      generateReconstruction: vi.fn().mockResolvedValue(makeProposalResponse()),
+    };
+
+    const { reconsiderCompletedEpisode } = await import("@/lib/graph/orchestrator.js");
+    const result = await reconsiderCompletedEpisode(
+      TEST_EPISODE,
+      { getModelName: () => "configured", provider, applyValidatedProposal },
+    );
+
+    expect(result.success).toBe(true);
+    expect(applyValidatedProposal).not.toHaveBeenCalled();
+
+    // No mutation of the input episode's graph
+    expect(TEST_EPISODE.situationGraph.nodes.length).toBe(1);
+  });
+
+});