peaks-loop 4.0.0-beta.2 → 4.0.0-beta.4

package/dist/src/services/crystallization/crystallization-service.d.ts

@@ -0,0 +1,239 @@
+import type Database from "better-sqlite3";
+import { z } from "zod";
+import { type CrystallizationEvent, type CrystallizationEventInput, type CrystallizationEventStatus, type EvidenceBrief } from "./crystallization-types.js";
+import { listCrystallizationEvents } from "./crystallization-store.js";
+import { type LoopReleaseInput } from "../loop/loop-release-types.js";
+/**
+ * CrystallizationService — spec §5 (post-run crystallization)
+ * + §4.5 + §4.7 + §10 RL-2 / RL-3 / RL-7.
+ *
+ * Hard rules enforced at THIS layer (defense in depth; the Zod
+ * schemas above and the brief refine guard add the second wall):
+ *
+ *   - AC-4 / RL-2: a crystallization_event is written only when the
+ *     candidate task has `task_status === 'completed'` AND
+ *     `gates_passed === true` AND `evidence_collected === true`.
+ *     Any pre-run attempt (status !== 'completed') is rejected
+ *     with `CRYSTALLIZATION_PRE_RUN`. This is the pre-run block
+ *     enforced by AC-4, AC-5, AC-6, AC-15, AC-16, AC-17 of the M5
+ *     slice.
+ *   - AC-6 / RL-3: writing a (loop_release, main_bee_release,
+ *     loop_bee_relation) pair happens in a SINGLE TRANSACTION. The
+ *     brief persist on crystallization_event also lives inside
+ *     the same transaction (the brief is the durable evidence;
+ *     dropping either side would break consistency).
+ *   - AC-7 / RL-1: this layer never asks the user to hand-author
+ *     JSON / manifest / CLI verb; all inputs are typed payloads
+ *     the LLM submits on the user's behalf.
+ *   - AC-15 / RL-7: the brief is mandatory; the
+ *     `EvidenceBriefSchema.refine` guard rejects partial briefs
+ *     at the parse boundary.
+ *   - AC-16 / RL-7: the brief has all 4 sections
+ *     (what_happened / why_it_matters / what_learned /
+ *     what_action). The service throws `BriefSectionError` if the
+ *     payload fails the guard.
+ *
+ * Sub-domain error codes:
+ *
+ *   - CRYSTALLIZATION_PRE_RUN: task_status / gates_passed /
+ *     evidence_collected precondition failed.
+ *   - CRYSTALLIZATION_INVALID_TASK_STATE: a task input field is
+ *     malformed (e.g. unknown task_status).
+ *   - MISSING_BRIEF_SECTION: a brief is missing at least one of
+ *     its 4 sections (raised from the evidence-brief-builder
+ *     refine; surfaced as-is by the service).
+ *   - CRYSTALLIZATION_TX_FAILED: a transaction-side effect failed
+ *     (FK constraint, duplicate key, …). The service re-raises
+ *     richer domain errors first; this is the catch-all for
+ *     unexpected SQL errors.
+ */
+export declare const CRYSTALLIZATION_TASK_STATUSES: readonly ["completed", "in_progress", "scratch", "blocked", "failed"];
+export type CrystallizationTaskStatus = (typeof CRYSTALLIZATION_TASK_STATUSES)[number];
+/**
+ * The pre-run task state the service requires. Crystallization is
+ * GATED on these three flags. The validator rejects any input that
+ * does not match. Adding a new flag means a schema-version bump; for
+ * now the three are required.
+ */
+export declare const CrystallizationTaskStateSchema: z.ZodObject<{
+    task_id: z.ZodString;
+    task_status: z.ZodLiteral<"completed">;
+    gates_passed: z.ZodLiteral<true>;
+    evidence_collected: z.ZodLiteral<true>;
+}, "strip", z.ZodTypeAny, {
+    task_id: string;
+    task_status: "completed";
+    gates_passed: true;
+    evidence_collected: true;
+}, {
+    task_id: string;
+    task_status: "completed";
+    gates_passed: true;
+    evidence_collected: true;
+}>;
+export type CrystallizationTaskState = z.infer<typeof CrystallizationTaskStateSchema>;
+export type CrystallizationIntegrityErrorCode = "CRYSTALLIZATION_PRE_RUN" | "CRYSTALLIZATION_INVALID_TASK_STATE" | "CRYSTALLIZATION_TX_FAILED" | "MISSING_BRIEF_SECTION";
+export declare class CrystallizationIntegrityError extends Error {
+    readonly code: CrystallizationIntegrityErrorCode | "MISSING_BRIEF_SECTION";
+    readonly findings: ReadonlyArray<{
+        path: string;
+        message: string;
+    }>;
+    constructor(code: CrystallizationIntegrityErrorCode, message: string, findings?: ReadonlyArray<{
+        path: string;
+        message: string;
+    }>);
+}
+export interface CrystallizePayload {
+    /** The candidate task state. Must satisfy the pre-run gate. */
+    task: CrystallizationTaskState;
+    /** The loop_release to create (new). Must include main bee via relation. */
+    loop_input: LoopReleaseInput;
+    /**
+     * The main bee_release to create. Note: bee_release requires a
+     * manifest row to keep the existing 4.x schema consistent;
+     * `bee_input` here carries ONLY the manifest-level fields the
+     * service uses. A future slice will swap in the full
+     * retainRelease() flow.
+     */
+    bee_input: BeeCrystallizeInput;
+    /**
+     * Optional NL reason the LLM authored; surfaced on the
+     * loop_bee_relation row (spec §4.6 `reason`).
+     */
+    bee_relation_reason: string;
+    /**
+     * The 4-section evidence brief (spec §4.7).
+     */
+    evidence_brief: EvidenceBrief;
+    /** Optional structured bullets (RL-7: counts may support the brief, not replace it). */
+    evidence_bullets?: string[];
+    /** Source trace pointers for the brief (spec §4.5). */
+    source_trace_pointers?: string[];
+    /** Optional evaluator summary (independent scorers — spec §4.5). */
+    evaluator_summary?: string;
+    /** Optional user decision summary (spec §4.5). */
+    user_decision_summary?: string;
+    /** Trigger classification (spec §5.4). */
+    trigger: CrystallizationEventInput["trigger"];
+}
+/**
+ * Minimum bee_release payload the service needs to create a main bee
+ * alongside a loop. The full bee_release schema (see
+ * `src/services/skillhub/release-retain.ts`) carries manifest rows,
+ * segment refs, file rows, etc.; M5 only writes the bee_release header
+ * + a single bee_manifest row sufficient to keep the 4.x FK
+ * constraints green. Full bee retaining lands in a future slice.
+ *
+ * Schema definition is hoisted above the interface so the typed
+ * interface can reference the inferred Zod type without forward-
+ * reference gymnastics.
+ */
+declare const BeeNameSchema: z.ZodString;
+export interface BeeCrystallizeInput {
+    bee_name: z.infer<typeof BeeNameSchema>;
+    version: string;
+    description: string;
+    user_intent_raw?: string;
+    parent_version?: string;
+    changelog?: string;
+}
+/**
+ * The crystallization result. The brief is returned AS PERSISTED
+ * (re-parsed through the schema) so callers receive the canonical
+ * shape, not the loose payload.
+ */
+export interface CrystallizationResult {
+    loop_release_id: string;
+    bee_release_id: number;
+    loop_bee_relation_id: number;
+    crystallization_event_id: string;
+    loop_release_lifecycle_status: CrystallizationEventStatus;
+}
+/**
+ * Service interface for the post-run crystallization flow.
+ */
+export declare class CrystallizationService {
+    private readonly db;
+    constructor(db: Database.Database);
+    /**
+     * Read a CrystallizationEvent by id.
+     */
+    read(id: string): CrystallizationEvent | undefined;
+    /**
+     * List events. Filter by lifecycle_status and/or the FK pointer
+     * columns. When no filters are supplied, lists ALL events
+     * (newest-first).
+     */
+    list(opts?: Parameters<typeof listCrystallizationEvents>[1]): {
+        schema_version: "peaks.crystallization/1";
+        id: string;
+        lifecycle_status: "candidate" | "stable" | "retired";
+        trigger: "user_explicit" | "llm_suggested" | "success_default_prompt" | "similar_task_recurrence";
+        evidence_brief: {
+            what_happened: string;
+            why_it_matters: string;
+            what_learned: string;
+            what_action: string;
+        };
+        evidence_bullets: string[];
+        source_trace_pointers: string[];
+        evaluator_summary: string;
+        user_decision_summary: string;
+        created_at: string;
+        created_loop_release_id?: string | undefined;
+        updated_loop_release_id?: string | undefined;
+        created_bee_release_id?: number | undefined;
+        updated_bee_release_id?: number | undefined;
+    }[];
+    /**
+     * Update a crystallization event's lifecycle_status. Always allowed
+     * (no schema-version gate). Returns the updated row or undefined if
+     * the id is absent.
+     *
+     * A `stable` transition is the natural post-crystallization state;
+     * the service does NOT enforce user-confirmation here (per spec
+     * §5.6 promotion happens on a separate `peaks loop promote` path).
+     * `retired` is the dispose path (`peaks asset dispose`).
+     */
+    updateStatus(id: string, next: CrystallizationEventStatus): CrystallizationEvent | undefined;
+    /**
+     * Pre-run gate: throws if any of `task_status`, `gates_passed`,
+     * `evidence_collected` is not as required. Called from
+     * `crystallize` BEFORE the transaction begins.
+     *
+     * Reachable as a public method so the CLI / tests can probe the
+     * gate in isolation.
+     */
+    assertReady(task: CrystallizationTaskState): CrystallizationTaskState;
+    /**
+     * Run the post-run crystallization flow.
+     *
+     * Steps (all in ONE better-sqlite3 transaction, atomic):
+     *
+     *   1. Pre-run gate (task must be completed / gates_passed /
+     *      evidence_collected).
+     *   2. Re-validate the 4-section brief (EvidenceBriefSchema.parse,
+     *      EvidenceBriefSchema.refine).
+     *   3. Insert loop_release.
+     *   4. Insert bee_release header (+ a single bee_manifest row to
+     *      keep the 4.x FK contract).
+     *   5. Insert loop_bee_relation row with role='main'.
+     *   6. Insert crystallization_event with the brief persisted.
+     *
+     * Returns the four ids + the new loop's lifecycle_status. On any
+     * failure the transaction is rolled back and the partial rows
+     * disappear.
+     *
+     * Throws:
+     *   - CrystallizationIntegrityError(CRYSTALLIZATION_PRE_RUN) on
+     *     the pre-run gate.
+     *   - BriefSectionError or ZodError if the brief is malformed;
+     *     surfaced as
+     *     CrystallizationIntegrityError(MISSING_BRIEF_SECTION).
+     *   - CrystallizationIntegrityError(CRYSTALLIZATION_TX_FAILED)
+     *     on any SQL transaction failure.
+     */
+    crystallize(payload: CrystallizePayload): CrystallizationResult;
+}
+export {};

package/dist/src/services/crystallization/crystallization-service.js

@@ -0,0 +1,397 @@
+import { z, ZodError } from "zod";
+import { CrystallizationEventSchema, EvidenceBriefSchema, } from "./crystallization-types.js";
+import { BriefSectionError } from "./evidence-brief-builder.js";
+import { CRYSTALLIZATION_SCHEMA_VERSION, ensureCrystallizationEventTable, getCrystallizationEvent, listCrystallizationEvents, newCrystallizationId, updateCrystallizationEventStatus, } from "./crystallization-store.js";
+import { insertLoopRelease } from "../loop/loop-release-store.js";
+import { LoopReleaseSchema } from "../loop/loop-release-types.js";
+import { insertLoopBeeRelation } from "../loop/loop-bee-relation-store.js";
+import { LoopBeeRelationSchema, } from "../loop/loop-bee-relation-types.js";
+/**
+ * CrystallizationService — spec §5 (post-run crystallization)
+ * + §4.5 + §4.7 + §10 RL-2 / RL-3 / RL-7.
+ *
+ * Hard rules enforced at THIS layer (defense in depth; the Zod
+ * schemas above and the brief refine guard add the second wall):
+ *
+ *   - AC-4 / RL-2: a crystallization_event is written only when the
+ *     candidate task has `task_status === 'completed'` AND
+ *     `gates_passed === true` AND `evidence_collected === true`.
+ *     Any pre-run attempt (status !== 'completed') is rejected
+ *     with `CRYSTALLIZATION_PRE_RUN`. This is the pre-run block
+ *     enforced by AC-4, AC-5, AC-6, AC-15, AC-16, AC-17 of the M5
+ *     slice.
+ *   - AC-6 / RL-3: writing a (loop_release, main_bee_release,
+ *     loop_bee_relation) pair happens in a SINGLE TRANSACTION. The
+ *     brief persist on crystallization_event also lives inside
+ *     the same transaction (the brief is the durable evidence;
+ *     dropping either side would break consistency).
+ *   - AC-7 / RL-1: this layer never asks the user to hand-author
+ *     JSON / manifest / CLI verb; all inputs are typed payloads
+ *     the LLM submits on the user's behalf.
+ *   - AC-15 / RL-7: the brief is mandatory; the
+ *     `EvidenceBriefSchema.refine` guard rejects partial briefs
+ *     at the parse boundary.
+ *   - AC-16 / RL-7: the brief has all 4 sections
+ *     (what_happened / why_it_matters / what_learned /
+ *     what_action). The service throws `BriefSectionError` if the
+ *     payload fails the guard.
+ *
+ * Sub-domain error codes:
+ *
+ *   - CRYSTALLIZATION_PRE_RUN: task_status / gates_passed /
+ *     evidence_collected precondition failed.
+ *   - CRYSTALLIZATION_INVALID_TASK_STATE: a task input field is
+ *     malformed (e.g. unknown task_status).
+ *   - MISSING_BRIEF_SECTION: a brief is missing at least one of
+ *     its 4 sections (raised from the evidence-brief-builder
+ *     refine; surfaced as-is by the service).
+ *   - CRYSTALLIZATION_TX_FAILED: a transaction-side effect failed
+ *     (FK constraint, duplicate key, …). The service re-raises
+ *     richer domain errors first; this is the catch-all for
+ *     unexpected SQL errors.
+ */
+/* ---------------------------------------------------------------------- */
+/* Pre-run task gate — AC-4 / RL-2                                        */
+/* ---------------------------------------------------------------------- */
+export const CRYSTALLIZATION_TASK_STATUSES = [
+    "completed",
+    "in_progress",
+    "scratch",
+    "blocked",
+    "failed",
+];
+/**
+ * The pre-run task state the service requires. Crystallization is
+ * GATED on these three flags. The validator rejects any input that
+ * does not match. Adding a new flag means a schema-version bump; for
+ * now the three are required.
+ */
+export const CrystallizationTaskStateSchema = z.object({
+    task_id: z.string().trim().min(1).max(256),
+    task_status: z.literal("completed", {
+        errorMap: () => ({
+            message: "task_status must be 'completed' for crystallization (AC-4 / RL-2)",
+        }),
+    }),
+    gates_passed: z.literal(true, {
+        errorMap: () => ({
+            message: "gates_passed must be true for crystallization (AC-4 / RL-2)",
+        }),
+    }),
+    evidence_collected: z.literal(true, {
+        errorMap: () => ({
+            message: "evidence_collected must be true for crystallization (AC-4 / RL-2)",
+        }),
+    }),
+});
+export class CrystallizationIntegrityError extends Error {
+    code;
+    findings;
+    constructor(code, message, findings = []) {
+        super(message);
+        this.name = "CrystallizationIntegrityError";
+        this.code = code;
+        this.findings = findings;
+    }
+}
+/**
+ * Minimum bee_release payload the service needs to create a main bee
+ * alongside a loop. The full bee_release schema (see
+ * `src/services/skillhub/release-retain.ts`) carries manifest rows,
+ * segment refs, file rows, etc.; M5 only writes the bee_release header
+ * + a single bee_manifest row sufficient to keep the 4.x FK
+ * constraints green. Full bee retaining lands in a future slice.
+ *
+ * Schema definition is hoisted above the interface so the typed
+ * interface can reference the inferred Zod type without forward-
+ * reference gymnastics.
+ */
+const BeeNameSchema = z
+    .string()
+    .trim()
+    .min(1, "bee_name is required")
+    .max(200)
+    .regex(/^[a-z0-9][a-z0-9._-]*$/, {
+    message: "bee_name must be kebab-/snake-case starting with a lowercase letter or digit",
+});
+const BeeCrystallizeInputSchema = z.object({
+    bee_name: BeeNameSchema,
+    version: z
+        .string()
+        .regex(/^\d+\.\d+\.\d+(?:[-+][0-9A-Za-z.-]+)?$/, {
+        message: "version must be a semver string (e.g. 0.1.0)",
+    }),
+    description: z.string().trim().min(1).max(4000),
+    user_intent_raw: z.string().trim().max(4000).optional(),
+    parent_version: z.string().trim().max(64).optional(),
+    changelog: z.string().trim().max(8000).optional(),
+});
+/**
+ * Service interface for the post-run crystallization flow.
+ */
+export class CrystallizationService {
+    db;
+    constructor(db) {
+        this.db = db;
+        // Idempotent — safe to call on every constructor invocation; the
+        // openStateDb pipeline already applied the SQL migration, but a
+        // caller-built DB (e.g. tests) needs this.
+        ensureCrystallizationEventTable(this.db);
+    }
+    /**
+     * Read a CrystallizationEvent by id.
+     */
+    read(id) {
+        return getCrystallizationEvent(this.db, id);
+    }
+    /**
+     * List events. Filter by lifecycle_status and/or the FK pointer
+     * columns. When no filters are supplied, lists ALL events
+     * (newest-first).
+     */
+    list(opts = {}) {
+        return listCrystallizationEvents(this.db, opts);
+    }
+    /**
+     * Update a crystallization event's lifecycle_status. Always allowed
+     * (no schema-version gate). Returns the updated row or undefined if
+     * the id is absent.
+     *
+     * A `stable` transition is the natural post-crystallization state;
+     * the service does NOT enforce user-confirmation here (per spec
+     * §5.6 promotion happens on a separate `peaks loop promote` path).
+     * `retired` is the dispose path (`peaks asset dispose`).
+     */
+    updateStatus(id, next) {
+        return updateCrystallizationEventStatus(this.db, id, next);
+    }
+    /**
+     * Pre-run gate: throws if any of `task_status`, `gates_passed`,
+     * `evidence_collected` is not as required. Called from
+     * `crystallize` BEFORE the transaction begins.
+     *
+     * Reachable as a public method so the CLI / tests can probe the
+     * gate in isolation.
+     */
+    assertReady(task) {
+        let parsed;
+        try {
+            parsed = CrystallizationTaskStateSchema.parse(task);
+        }
+        catch (err) {
+            if (err instanceof ZodError) {
+                const findings = err.issues.map((i) => ({
+                    path: i.path.join("."),
+                    message: i.message,
+                }));
+                const first = err.issues[0];
+                throw new CrystallizationIntegrityError("CRYSTALLIZATION_PRE_RUN", `pre-run gate failed: ${first?.message ?? "invalid task state"}`, findings);
+            }
+            throw err;
+        }
+        // Defense in depth — re-assert at the service boundary.
+        if (parsed.task_status !== "completed" ||
+            parsed.gates_passed !== true ||
+            parsed.evidence_collected !== true) {
+            throw new CrystallizationIntegrityError("CRYSTALLIZATION_PRE_RUN", "pre-run gate failed: task must be completed with gates_passed=true AND evidence_collected=true (AC-4 / RL-2)", [
+                {
+                    path: "task",
+                    message: "task_status must equal 'completed' and gates_passed and evidence_collected must both be true",
+                },
+            ]);
+        }
+        return parsed;
+    }
+    /**
+     * Run the post-run crystallization flow.
+     *
+     * Steps (all in ONE better-sqlite3 transaction, atomic):
+     *
+     *   1. Pre-run gate (task must be completed / gates_passed /
+     *      evidence_collected).
+     *   2. Re-validate the 4-section brief (EvidenceBriefSchema.parse,
+     *      EvidenceBriefSchema.refine).
+     *   3. Insert loop_release.
+     *   4. Insert bee_release header (+ a single bee_manifest row to
+     *      keep the 4.x FK contract).
+     *   5. Insert loop_bee_relation row with role='main'.
+     *   6. Insert crystallization_event with the brief persisted.
+     *
+     * Returns the four ids + the new loop's lifecycle_status. On any
+     * failure the transaction is rolled back and the partial rows
+     * disappear.
+     *
+     * Throws:
+     *   - CrystallizationIntegrityError(CRYSTALLIZATION_PRE_RUN) on
+     *     the pre-run gate.
+     *   - BriefSectionError or ZodError if the brief is malformed;
+     *     surfaced as
+     *     CrystallizationIntegrityError(MISSING_BRIEF_SECTION).
+     *   - CrystallizationIntegrityError(CRYSTALLIZATION_TX_FAILED)
+     *     on any SQL transaction failure.
+     */
+    crystallize(payload) {
+        const task = this.assertReady(payload.task);
+        // Re-validate the 4-section brief up-front; the schema refines
+        // guard rejects partial briefs. This is the AC-15 / AC-16 / AC-17
+        // gate.
+        let brief;
+        try {
+            brief = payload.evidence_brief;
+            // Defensive re-parse even though the input is typed; ensures
+            // `parse()` runs the .refine() guard.
+            EvidenceBriefSchema.parse(brief);
+        }
+        catch (err) {
+            if (err instanceof ZodError || err instanceof BriefSectionError) {
+                const findings = err instanceof ZodError
+                    ? err.issues.map((i) => ({
+                        path: i.path.join("."),
+                        message: i.message,
+                    }))
+                    : err.findings;
+                throw new CrystallizationIntegrityError("MISSING_BRIEF_SECTION", "evidence_brief is missing at least one of the 4 required sections (spec §4.7 / RL-7)", [...findings]);
+            }
+            throw err;
+        }
+        // Re-validate loop_input via Zod so a hand-crafted caller input
+        // cannot bypass the schema. (M8 dogfood follow-up.)
+        const loopRow = LoopReleaseSchema.parse(payload.loop_input);
+        // Re-validate bee_input.
+        const beeRow = BeeCrystallizeInputSchema.parse(payload.bee_input);
+        // Loop-bee relation input (the relation's `reason` is the NL
+        // account the LLM authored). We validate the shape WITHOUT
+        // bee_release_id (which is filled in by the bee insert below),
+        // then validate the final tuple after the SQL writes.
+        const relationInputShape = LoopBeeRelationSchema.omit({
+            id: true,
+            created_at: true,
+            bee_release_id: true,
+        }).parse({
+            loop_release_id: loopRow.id,
+            role: "main",
+            reason: payload.bee_relation_reason,
+        });
+        const eventId = newCrystallizationId();
+        const eventCreatedAt = new Date().toISOString();
+        let loopReleaseId = loopRow.id;
+        let beeReleaseId = -1;
+        let relationId = -1;
+        let eventPersistedId = eventId;
+        const tx = this.db.transaction(() => {
+            // 1. Insert loop_release.
+            insertLoopRelease(this.db, loopRow);
+            // 2. Insert bee_release header. The existing
+            // `retainRelease` adds segment_ref / file / change rows; the
+            // minimal crystallization write here only opens the bee_release
+            // header row + a bee_manifest row. Future slices can promote
+            // the bee retain path to a richer flow.
+            const beeInsert = this.db.prepare(`INSERT INTO bee_release (
+           bee_name, version, source, archived_at, archived_by,
+           user_intent_raw, description, parent_version, changelog,
+           shareable, desktop_visible
+         ) VALUES (?, ?, 'user', ?, ?, ?, ?, ?, ?, 1, 1)`);
+            const beeInfo = beeInsert.run(beeRow.bee_name, beeRow.version, eventCreatedAt, "llm", beeRow.user_intent_raw ?? null, beeRow.description, beeRow.parent_version ?? null, beeRow.changelog ?? null);
+            beeReleaseId = Number(beeInfo.lastInsertRowid);
+            // Update the bee_release_pointer row so the latest version is
+            // discoverable via list tooling.
+            this.db
+                .prepare(`INSERT OR REPLACE INTO bee_release_pointer (bee_name, latest_version, released_at) VALUES (?, ?, ?)`)
+                .run(beeRow.bee_name, beeRow.version, eventCreatedAt);
+            // 3. Insert bee_manifest stub. Mirrors `retainRelease` minimum
+            // surface: promotion = 'candidate', requires_human = 0,
+            // requires_smoke = 0. Min-cycles left NULL.
+            this.db
+                .prepare(`INSERT INTO bee_manifest (
+             release_id, schema_version, description, segments_json,
+             entrypoint_preamble, promotion, min_cycles,
+             requires_human, requires_smoke, retire_on_misses
+           ) VALUES (?, ?, ?, '[]', '', 'candidate', NULL, 0, 0, NULL)`)
+                .run(beeReleaseId, "peaks.bee/1", beeRow.description);
+            // 4. Insert the loop_bee_relation row (FKs now valid). Re-parse
+            // through the input schema so the bee_release_id is validated
+            // too — defense in depth. `id` is autoincrement (not in input).
+            const fullRelation = LoopBeeRelationSchema.omit({
+                id: true,
+                created_at: true,
+            }).parse({
+                loop_release_id: relationInputShape.loop_release_id,
+                bee_release_id: beeReleaseId,
+                role: relationInputShape.role,
+                reason: relationInputShape.reason,
+                schema_version: "peaks.loop-bee-relation/1",
+            });
+            const relationRow = insertLoopBeeRelation(this.db, fullRelation);
+            relationId = relationRow.id;
+            // 5. Persist the crystallization_event row with the brief inline
+            // (the brief is the durable evidence — AC-17).
+            const eventInput = {
+                trigger: payload.trigger,
+                evidence_brief: brief,
+                evidence_bullets: payload.evidence_bullets ?? [],
+                source_trace_pointers: payload.source_trace_pointers ?? [],
+                evaluator_summary: payload.evaluator_summary ?? "",
+                user_decision_summary: payload.user_decision_summary ?? "",
+                created_loop_release_id: loopReleaseId,
+                created_bee_release_id: beeReleaseId,
+                // updated_* are not set on a CREATE crystallization; they will
+                // be set on future UPDATE crystallizations.
+                lifecycle_status: "candidate",
+            };
+            const persisted = insertCrystallizationEventRaw(this.db, eventInput, eventId, eventCreatedAt);
+            eventPersistedId = persisted.id;
+        });
+        try {
+            tx();
+        }
+        catch (err) {
+            // Map known SQLite constraint errors to friendlier codes; raise
+            // CrystallizationIntegrityError so the CLI can render the JSON
+            // envelope.
+            if (err instanceof CrystallizationIntegrityError ||
+                err instanceof BriefSectionError) {
+                throw err;
+            }
+            const message = err instanceof Error ? err.message : String(err);
+            throw new CrystallizationIntegrityError("CRYSTALLIZATION_TX_FAILED", `crystallization transaction failed: ${message}`, [
+                {
+                    path: "tx",
+                    message: "transaction rolled back; verify loop/bee ids are unique and FK targets exist",
+                },
+            ]);
+        }
+        // Silence the unused-variable lint; the parsed task is the source
+        // of truth at the entry boundary.
+        void task;
+        return {
+            loop_release_id: loopReleaseId,
+            bee_release_id: beeReleaseId,
+            loop_bee_relation_id: relationId,
+            crystallization_event_id: eventPersistedId,
+            loop_release_lifecycle_status: "candidate",
+        };
+    }
+}
+/* ---------------------------------------------------------------------- */
+/* Internal: thin writer that injects the deterministic eventId +         */
+/*           created_at so the crystallization flow can land all rows in  */
+/*           one tx before any non-idempotent residue lands.              */
+/* ---------------------------------------------------------------------- */
+function insertCrystallizationEventRaw(db, row, id, createdAt) {
+    const persisted = CrystallizationEventSchema.parse({
+        ...row,
+        id,
+        schema_version: CRYSTALLIZATION_SCHEMA_VERSION,
+        created_at: createdAt,
+    });
+    const stmt = db.prepare(`INSERT INTO crystallization_event (
+       id, trigger, evidence_brief_json, evidence_bullets_json,
+       source_trace_pointers_json, evaluator_summary,
+       user_decision_summary,
+       created_loop_release_id, updated_loop_release_id,
+       created_bee_release_id, updated_bee_release_id,
+       lifecycle_status, schema_version, created_at
+     ) VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?)`);
+    stmt.run(persisted.id, persisted.trigger, JSON.stringify(persisted.evidence_brief), JSON.stringify(persisted.evidence_bullets), JSON.stringify(persisted.source_trace_pointers), persisted.evaluator_summary, persisted.user_decision_summary, persisted.created_loop_release_id ?? null, persisted.updated_loop_release_id ?? null, persisted.created_bee_release_id ?? null, persisted.updated_bee_release_id ?? null, persisted.lifecycle_status, persisted.schema_version, persisted.created_at);
+    return persisted;
+}