peaks-loop 4.0.0-beta.8 → 4.0.0-beta.9

package/dist/services/context/threshold.js

@@ -0,0 +1,76 @@
+/**
+ * G9.3 — forced compression gate threshold constants + tier evaluation.
+ *
+ * 256K default context capacity is a conservative proxy. The LLM's real
+ * capacity is its private business (R-1 / R-8 / R-10 / R-13 boundary
+ * inherited from slice #009); we use prompt size as the gate signal.
+ *
+ * See: `.peaks/memory/sub-agent-headroom-forced-compression-gate.md`
+ * for the full G9 rule (RL-27..RL-32, AC-50..AC-65).
+ */
+export const CONTEXT_CAPACITY_DEFAULT_BYTES = 256 * 1024; // 256K
+export const THRESHOLD_SOFT_WARN_RATIO = 0.5; // 50%
+export const THRESHOLD_NEAR_LIMIT_RATIO = 0.75; // 75% — user red line
+export const THRESHOLD_HARD_REJECT_RATIO = 0.80; // 80%
+export const THRESHOLD_EMERGENCY_RATIO = 0.90; // 90%
+/**
+ * Compute the threshold tier for a given prompt size. Pure function
+ * (no IO, no side effects). The `capacityBytes` parameter lets callers
+ * override the default 256K proxy (e.g. for testing or for a future
+ * per-IDE capacity override).
+ */
+export function evaluateThresholdTier(promptSize, capacityBytes = CONTEXT_CAPACITY_DEFAULT_BYTES) {
+    if (!Number.isFinite(promptSize) || promptSize < 0) {
+        throw new Error(`evaluateThresholdTier: promptSize must be ≥ 0 (got ${promptSize})`);
+    }
+    if (!Number.isFinite(capacityBytes) || capacityBytes <= 0) {
+        throw new Error(`evaluateThresholdTier: capacityBytes must be > 0 (got ${capacityBytes})`);
+    }
+    const ratio = promptSize / capacityBytes;
+    const warnings = [];
+    let tier;
+    if (ratio >= THRESHOLD_EMERGENCY_RATIO) {
+        tier = 'emergency';
+        warnings.push('PROMPT_EMERGENCY');
+    }
+    else if (ratio >= THRESHOLD_HARD_REJECT_RATIO) {
+        tier = 'hard-reject';
+        warnings.push('PROMPT_TOO_LARGE');
+    }
+    else if (ratio >= THRESHOLD_NEAR_LIMIT_RATIO) {
+        tier = 'near-limit';
+        warnings.push('CONTEXT_NEAR_LIMIT');
+    }
+    else if (ratio >= THRESHOLD_SOFT_WARN_RATIO) {
+        tier = 'soft-warn';
+        warnings.push('CONTEXT_SOFT_WARN');
+    }
+    else {
+        tier = 'ok';
+    }
+    return {
+        tier,
+        ratio,
+        bytesUsed: promptSize,
+        capacityBytes,
+        warnings
+    };
+}
+/**
+ * Convert a threshold tier to a machine-readable code suitable for the
+ * CLI envelope's `code` field. Used by both the CLI and the hook layer.
+ */
+export function tierToCode(tier) {
+    switch (tier) {
+        case 'ok':
+            return 'OK';
+        case 'soft-warn':
+            return 'CONTEXT_SOFT_WARN';
+        case 'near-limit':
+            return 'CONTEXT_NEAR_LIMIT';
+        case 'hard-reject':
+            return 'PROMPT_TOO_LARGE';
+        case 'emergency':
+            return 'PROMPT_EMERGENCY';
+    }
+}

package/dist/services/context/tokenizer.d.ts

@@ -0,0 +1,12 @@
+/**
+ * Per spec §4.1 Step 3 — Tokenizer (non-mutating).
+ *
+ * Hard constraint H3: structured metadata > bare strings. Each collector
+ * artifact and each fetched doc gets a tokenized metadata record so
+ * peaks-rd/qa can do relative-anomaly detection later.
+ *
+ * Immutability H (common/coding-style): the output is frozen so callers
+ * cannot mutate it. Use spread to derive new outputs in future slices.
+ */
+import type { CollectorOutput, DocRetrieverOutput, TokenizerOutput } from './types.js';
+export declare function tokenize(collector: CollectorOutput, docRetriever: DocRetrieverOutput, now?: Date): TokenizerOutput;

package/dist/services/context/tokenizer.js

@@ -0,0 +1,43 @@
+function freshDecayScore(fetchedAt, now) {
+    const ageMs = now.getTime() - new Date(fetchedAt).getTime();
+    const ageDays = ageMs / (1000 * 60 * 60 * 24);
+    // Linear decay: 1.0 at day 0, 0.5 at day 30, 0.0 at day 60+
+    if (ageDays >= 60)
+        return 0;
+    return Math.max(0, 1 - ageDays / 60);
+}
+export function tokenize(collector, docRetriever, now = new Date()) {
+    const metadata = [];
+    for (const doc of docRetriever.fetchedDocs) {
+        metadata.push({
+            id: `doc:${doc.dep}@${doc.version}`,
+            kind: 'doc',
+            version: doc.version,
+            blastRadius: doc.sections.map((s) => s.title),
+            conflictScore: 0, // v1: no cross-source conflict detection yet
+            timeDecayScore: freshDecayScore(doc.fetchedAt, now),
+            tags: ['fetched', doc.source, doc.stale ? 'stale' : 'fresh'],
+        });
+    }
+    for (const mem of collector.memoryEntries) {
+        metadata.push({
+            id: `memory:${mem.path}`,
+            kind: 'memory',
+            blastRadius: [mem.title],
+            conflictScore: 0,
+            timeDecayScore: mem.relevanceScore,
+            tags: ['memory'],
+        });
+    }
+    for (const file of collector.files) {
+        metadata.push({
+            id: `code:${file.path}`,
+            kind: 'code',
+            blastRadius: [file.path],
+            conflictScore: 0,
+            timeDecayScore: 1,
+            tags: [file.kind],
+        });
+    }
+    return Object.freeze({ metadata: Object.freeze(metadata) });
+}

package/dist/services/context/types.d.ts

@@ -0,0 +1,93 @@
+/**
+ * Per spec §4.1 — the single source of truth for context.json shape.
+ * The Zod schema in context-schema.ts must mirror this interface exactly;
+ * the schema is the runtime validator, the interface is the compile-time contract.
+ *
+ * Hard constraint H8: Audit trail must be hashable. `sha256` is the field
+ * that lets peaks-state-lock chain signatures between stages.
+ */
+export type ContextVersion = '1.0';
+export type Audience = 'peaks-rd' | 'peaks-qa' | 'peaks-mut' | 'all';
+export type DepsMode = 'locked' | 'latest';
+export type FileKind = 'source' | 'test' | 'config' | 'doc';
+export interface CollectedFile {
+    readonly path: string;
+    readonly kind: FileKind;
+    readonly lines: number;
+    readonly hash: string;
+}
+export interface GitStatus {
+    readonly branch: string;
+    readonly lastCommit: string;
+    readonly dirty: boolean;
+}
+export interface MemoryEntry {
+    readonly path: string;
+    readonly title: string;
+    readonly relevanceScore: number;
+    readonly excerptHash: string;
+}
+export interface DepInfo {
+    readonly version: string;
+    readonly source: 'package.json' | 'pnpm-lock.yaml' | 'yarn.lock';
+    readonly resolved: string;
+}
+export interface DocSection {
+    readonly title: string;
+    readonly tokenEstimate: number;
+    readonly excerpt: string;
+}
+export interface FetchedDoc {
+    readonly dep: string;
+    readonly version: string;
+    readonly source: 'local-cache' | 'remote-fetch';
+    readonly url?: string;
+    readonly fetchedAt: string;
+    readonly contentHash: string;
+    readonly sections: ReadonlyArray<DocSection>;
+    readonly stale: boolean;
+}
+export interface SkippedDoc {
+    readonly dep: string;
+    readonly reason: 'unconfigured' | 'network_error' | 'version_unknown';
+}
+export type MetaKind = 'doc' | 'code' | 'memory' | 'git';
+export interface TokenizedItem {
+    readonly id: string;
+    readonly kind: MetaKind;
+    readonly version?: string;
+    readonly blastRadius: ReadonlyArray<string>;
+    readonly conflictScore: number;
+    readonly timeDecayScore: number;
+    readonly tags: ReadonlyArray<string>;
+}
+export interface CollectorOutput {
+    readonly files: ReadonlyArray<CollectedFile>;
+    readonly gitStatus: GitStatus;
+    readonly memoryEntries: ReadonlyArray<MemoryEntry>;
+    readonly deps: Readonly<Record<string, DepInfo>>;
+}
+export interface DocRetrieverOutput {
+    readonly fetchedDocs: ReadonlyArray<FetchedDoc>;
+    readonly skipped: ReadonlyArray<SkippedDoc>;
+}
+export interface TokenizerOutput {
+    readonly metadata: ReadonlyArray<TokenizedItem>;
+}
+export interface RendererOutput {
+    readonly audience: Audience;
+    readonly renderedAt: string;
+    readonly sizeBytes: number;
+    readonly truncated: boolean;
+    readonly truncatedReason?: 'doc_budget_exceeded' | 'section_count_exceeded';
+}
+export interface ContextJson {
+    readonly version: ContextVersion;
+    readonly goal: string;
+    readonly generatedAt: string;
+    readonly sha256: string;
+    readonly collector: CollectorOutput;
+    readonly docRetriever: DocRetrieverOutput;
+    readonly tokenizer: TokenizerOutput;
+    readonly renderer: RendererOutput;
+}

package/dist/services/context/types.js

@@ -0,0 +1,9 @@
+/**
+ * Per spec §4.1 — the single source of truth for context.json shape.
+ * The Zod schema in context-schema.ts must mirror this interface exactly;
+ * the schema is the runtime validator, the interface is the compile-time contract.
+ *
+ * Hard constraint H8: Audit trail must be hashable. `sha256` is the field
+ * that lets peaks-state-lock chain signatures between stages.
+ */
+export {};

package/dist/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 {};