akm-cli 0.9.17-alpha.8 → 0.9.17

package/dist/core/config/schema/index-config.js

@@ -10,19 +10,6 @@ import { z } from "zod";
 import { warnOnce } from "../../warn.js";
 import { engineName, LlmInvocationOverridesSchema, nonEmptyString, positiveInt } from "./primitives.js";
 // ── Index / per-pass ────────────────────────────────────────────────────────
-//
-// WI-9.6c: `graphExtractionIncludeTypes` is no longer validated against a
-// hardcoded allowlist (the prior GRAPH_EXTRACTION_INCLUDE_TYPES_ALLOWED,
-// which included a stale `wiki` entry and was already missing `fact` from the
-// runtime consumer's own list — the schema-level allowlist had drifted from
-// reality). Accept-any until Chunk 2 sources a real type list from adapter
-// metadata: the field is now an array of arbitrary non-empty strings.
-// Runtime consumers already handle unknown/unsupported type strings
-// gracefully — src/indexer/graph/graph-extraction.ts's
-// `SUPPORTED_GRAPH_EXTRACTION_INCLUDE_TYPES` set (and `collectEligibleFiles`)
-// silently skips any type it doesn't recognize (no placement entry ⇒ zero
-// eligible files for that type; no crash). This is a permissive-direction
-// behavior change: configs with a previously-rejected type string now parse.
 const INDEX_PASS_RETIRED_KEYS = new Set([
     "endpoint",
     "provider",
@@ -60,12 +47,8 @@ export const IndexPassConfigSchema = z.preprocess((raw, ctx) => {
     timeoutMs: z.union([positiveInt, z.null()]).optional(),
     enabled: z.boolean().optional(),
     llm: LlmInvocationOverridesSchema.optional(),
-    graphExtractionBatchSize: positiveInt.optional(),
-    // Accept-any until Chunk 2 (WI-9.6c) — no longer enum-restricted.
-    graphExtractionIncludeTypes: z.array(z.string().min(1)).nonempty().optional(),
 })
     .passthrough());
-const MetadataEnhanceSchema = z.object({ enabled: z.boolean().optional() }).passthrough();
 const IndexDefaultsSchema = z
     .object({
     engine: engineName.optional(),
@@ -76,16 +59,14 @@ const IndexDefaultsSchema = z
     .passthrough();
 /**
  * Index config is a union of reserved feature sections and per-pass entries.
- * Passthrough so per-pass entries (keyed by arbitrary pass names like `graph`,
- * `enrichment`) can live next to the reserved keys.
+ * Passthrough so per-pass entries (keyed by arbitrary pass names like
+ * `memory`, or a retired one like `graph`) can live next to the reserved keys.
  * The outer preprocess emits the legacy parser's actionable error messages
  * for the two most common type-shape mistakes:
  *   - An array at the `index` block.
  *   - A non-object at `index.<passName>`.
- * Inner field validation (graphExtractionIncludeTypes shape, invocation
- * overrides, provider-key rejection) is delegated to {@link IndexPassConfigSchema}.
- * `graphExtractionIncludeTypes` accepts arbitrary non-empty strings
- * (WI-9.6c — no hardcoded type allowlist; accept-any until Chunk 2).
+ * Inner field validation (invocation overrides, provider-key rejection) is
+ * delegated to {@link IndexPassConfigSchema}.
  */
 const IndexConfigRuntimeSchema = z.preprocess((raw, ctx) => {
     if (raw === undefined || raw === null)
@@ -93,7 +74,7 @@ const IndexConfigRuntimeSchema = z.preprocess((raw, ctx) => {
     if (Array.isArray(raw)) {
         ctx.addIssue({
             code: z.ZodIssueCode.custom,
-            message: 'Invalid `index` config: expected an object keyed by pass name (e.g. `{ "enrichment": { "enabled": false } }`).',
+            message: 'Invalid `index` config: expected an object keyed by pass name (e.g. `{ "memory": { "enabled": false } }`).',
         });
         return raw;
     }
@@ -119,7 +100,6 @@ const IndexConfigRuntimeSchema = z.preprocess((raw, ctx) => {
 }, z
     .object({
     defaults: IndexDefaultsSchema.optional(),
-    metadataEnhance: MetadataEnhanceSchema.optional(),
 })
     .catchall(IndexPassConfigSchema));
 // The runtime catchall correctly validates arbitrary pass objects, but its

package/dist/core/file-change.js

@@ -4,13 +4,21 @@
 /**
  * The content of a proposal's PRIMARY change (WI-6.2 envelope read path).
  *
- * By construction `changes[0].after === payload.content`; consumers read
- * through this accessor so the single-content assumption lives in ONE place
- * once multi-file proposals (consolidate ops) ride the envelope. Typed
- * structurally (not against `Proposal`) so this module stays dependency-free.
+ * By construction `changes[0].after === payload.content` for a `create` or
+ * `update` change; consumers read through this accessor so the
+ * single-content assumption lives in ONE place. A `delete` primary change (a
+ * consolidate retire proposal, alpha.9) writes no content by definition, so
+ * it reads as `""` here rather than throwing — every generic proposal scan
+ * (dedup-by-body-hash, the triage drain's diff-size filter, the quality
+ * validators) can call this on ANY pending proposal, retire included,
+ * without special-casing the op first. Typed structurally (not against
+ * `Proposal`) so this module stays dependency-free.
  */
 export function proposalContent(p) {
-    const content = p.changes[0]?.after;
+    const primary = p.changes[0];
+    if (primary?.op === "delete")
+        return "";
+    const content = primary?.after;
     if (content === undefined)
         throw new Error("Proposal primary change has no content.");
     return content;

package/dist/core/improve-result.js

@@ -37,6 +37,9 @@ const COMMON_FIELDS = [
     "deadUrlCoverage",
     "reflectsWithErrorContext",
     "memoryInference",
+    // 0.9.17-alpha.9: no longer written (the LLM entity-graph extraction that
+    // wrote these was retired), kept readable for the same reason as
+    // `evalCasesWritten` above.
     "graphExtraction",
     "memoryInferenceDurationMs",
     "graphExtractionDurationMs",
@@ -119,7 +122,7 @@ function validateConsolidationPlan(value) {
     ]));
     if (!isRecord(value.configured))
         fail("plan.consolidation.configured must be an object");
-    requireExactFields(value.configured, new Set(["enabled", "minPoolSize", "limit", "maxChunkSize", "incrementalSince"]));
+    requireExactFields(value.configured, new Set(["enabled", "minPoolSize", "limit", "maxChunkSize"]));
     if (value.configured.enabled !== undefined && typeof value.configured.enabled !== "boolean") {
         fail("plan.consolidation.configured.enabled must be a boolean");
     }
@@ -128,9 +131,6 @@ function validateConsolidationPlan(value) {
             fail(`plan.consolidation.configured.${field} must be a number`);
         }
     }
-    if (value.configured.incrementalSince !== undefined && typeof value.configured.incrementalSince !== "string") {
-        fail("plan.consolidation.configured.incrementalSince must be a string");
-    }
     if (!isRecord(value.effective))
         fail("plan.consolidation.effective must be an object");
     requireExactFields(value.effective, new Set(["enabled", "minPoolSize", "limit", "chunkSize"]));
@@ -200,7 +200,15 @@ function validateProcessRoutingRows(value) {
             fail("plan.processes entries must be objects");
         requireExactFields(row, new Set(["process", "enabled", "engine", "model", "engineKind", "notices", "unavailable", "eligibleRefs"]));
         if (typeof row.process !== "string" ||
-            !(canonicalNames.includes(row.process) || row.process === "triage.judgment")) {
+            !(canonicalNames.includes(row.process) ||
+                row.process === "triage.judgment" ||
+                // 0.9.17-alpha.9: "graphExtraction" is no longer a process a fresh
+                // plan can route (the LLM entity-graph extraction it named was
+                // retired), but every pre-alpha.9 run recorded one here — kept
+                // readable, same as `stageNames`' "graph-extraction" above and
+                // `graphExtraction` in COMMON_FIELDS, so a historical run still
+                // decodes for `akm health` instead of failing this row outright.
+                row.process === "graphExtraction")) {
             fail("plan.processes.process is invalid");
         }
         if (seen.has(row.process))
@@ -385,10 +393,18 @@ function validateImprovePlan(value, dryRun, plannedRefNames) {
     if (value.limits.totalCeiling !== undefined && value.effectiveRefs.length > value.limits.totalCeiling) {
         fail("plan.effectiveRefs cannot exceed plan.limits.totalCeiling");
     }
-    validateProcessRoutingRows(value.processes);
+    // #947 added plan.processes (2026-09-09T09:03:19Z); every run recorded
+    // before it legitimately stored `plan` with no `processes` key at all —
+    // validate the rows only when present, same as `proactive` below, so a
+    // pre-#947 envelope still decodes (AGENTS.md "Reading persisted data").
+    if (value.processes !== undefined)
+        validateProcessRoutingRows(value.processes);
     if (value.proactive !== undefined)
         validateProactivePlan(value.proactive);
     validateConsolidationPlan(value.consolidation);
+    // "graph-extraction" is no longer a stage a fresh plan can name — retired
+    // 0.9.17-alpha.9 — but stays here so a historical envelope that captured a
+    // dry-run plan before the retirement still decodes.
     const stageNames = new Set(["consolidation", "extract", "graph-extraction", "memory-inference"]);
     if (!Array.isArray(value.stages))
         fail("plan.stages must be an array");

package/dist/core/improve-types.js

@@ -32,7 +32,6 @@ export function classifyImproveAction(mode) {
         case "reflect":
         case "distill":
         case "memory-inference":
-        case "graph-extraction":
             return "accepted";
         case "reflect-cooldown":
         case "reflect-skipped":

package/dist/core/loopback.js

@@ -20,10 +20,11 @@
  *
  * The ONE local-vs-remote answer for concurrency defaults — the workflow
  * engine's frozen engine concurrency (`workflows/concurrency-policy.ts`) and
- * the indexer's LLM pool (`indexer/indexer.ts` `getDefaultLlmConcurrency`)
- * both classify through here. NOT shared with the website snapshot fetcher's
- * SSRF policy (`sources/snapshot-fetchers/host-guard.ts`), which deliberately
- * draws different lines for a different threat model.
+ * the embedding pool (`resolveEmbeddingConcurrency`,
+ * `src/llm/embedders/remote.ts`) both classify through here. NOT shared with
+ * the website snapshot fetcher's SSRF policy
+ * (`sources/snapshot-fetchers/host-guard.ts`), which deliberately draws
+ * different lines for a different threat model.
  *
  * @module core/loopback
  */
@@ -92,14 +93,8 @@ export function isLoopbackEndpoint(endpoint) {
  * endpoint: 1 for a loopback endpoint (a local model server serves one
  * inference at a time; concurrent requests thrash it — reload thrash, HTTP
  * 500 "Model reloaded"), 2 for a remote one (enough to overlap request
- * latency without hammering a rate-limited API). A leaf helper so both
- * callers — the indexer's LLM enrichment pool (`getDefaultLlmConcurrency`,
- * `src/indexer/indexer.ts`) and the embedding pool
- * (`resolveEmbeddingConcurrency`, `src/llm/embedders/remote.ts`) — share one
- * definition instead of mirroring it; `src/llm/embedders/remote.ts` cannot
- * import `getDefaultLlmConcurrency` directly (`src/indexer/indexer.ts`
- * already depends on this module transitively through
- * materialize-embeddings.ts).
+ * latency without hammering a rate-limited API). Used by the embedding pool
+ * (`resolveEmbeddingConcurrency`, `src/llm/embedders/remote.ts`).
  */
 export function defaultConcurrencyForEndpoint(endpoint) {
     return isLoopbackEndpoint(endpoint) ? 1 : 2;

package/dist/core/parse.js

@@ -101,24 +101,24 @@ export function parseJsonResponse(raw) {
  * balanced `{ }` or `[ ]` structure in the text and attempts to parse that
  * substring. Returns `undefined` if no valid JSON structure is found.
  *
- * Shape preference is controlled by {@link ParseEmbeddedJsonOptions.expect}:
- * - `"any"` (default): non-array results are preferred — a `{…}` object found
- *   first is returned immediately; arrays (`[…]`) are a fallback.
- * - `"array"`: only top-level arrays are returned. The direct parse is
- *   accepted only if it is an array, and the scanner returns the first
- *   balanced `[…]` while skipping `{…}` openers entirely.
+ * Object-preferring: a `{…}` object found first is returned immediately;
+ * a top-level array (`[…]`) is only a fallback, returned if no object is
+ * found. Preserves the historical behaviour every current caller (all in
+ * `src/`) relies on.
+ *
+ * An array-preferring mode (`expect: "array"`, for callers that specifically
+ * expect a JSON array) existed here for batched LLM entity-graph extraction
+ * (#635) and was removed with it in 0.9.17-alpha.9 — every remaining caller
+ * calls this with no options, so it was dead weight, not a live choice.
  */
-export function parseEmbeddedJsonResponse(raw, options) {
-    const expectArray = options?.expect === "array";
+export function parseEmbeddedJsonResponse(raw) {
     const direct = parseJsonResponse(raw);
-    if (direct !== undefined && (!expectArray || Array.isArray(direct)))
+    if (direct !== undefined)
         return direct;
     const text = escapeJsonStringControls(stripCodeFences(stripThinkBlocks(raw)));
     let arrayFallback;
     // Scan only *top-level* balanced structures: once a `{…}`/`[…]` is matched we
     // jump `start` past its closing bracket rather than re-scanning its interior.
-    // This keeps array mode from salvaging an array *nested inside* a leading
-    // object (e.g. the `entities` array of a bare `{entities,relations}` object).
     for (let start = 0; start < text.length; start++) {
         const opener = text[start];
         if (opener !== "{" && opener !== "[")
@@ -161,13 +161,10 @@ export function parseEmbeddedJsonResponse(raw, options) {
         try {
             const parsed = JSON.parse(text.slice(start, end + 1));
             if (Array.isArray(parsed)) {
-                // First valid array wins in array mode; in "any" mode it is the
-                // fallback returned only if no object is found.
-                if (expectArray)
-                    return parsed;
+                // Fallback, returned only if no object is found.
                 arrayFallback ??= parsed;
             }
-            else if (!expectArray) {
+            else {
                 return parsed;
             }
             // Skip past this balanced structure so we don't descend into it.

package/dist/core/state/migrations.js

@@ -47,6 +47,7 @@ export const STATE_MIGRATION_SAFETY_BY_ID = Object.freeze({
     "026-proposals-strip-legacy-fragment-refs": "data-preserving-rebuild",
     "027-extract-sessions-seen-harness-rename": "data-preserving-rebuild",
     "028-improve-ledger": "historical-destructive",
+    "029-improve-ledger-content-hash": "additive",
 });
 export const STATE_MIGRATIONS = [
     // ── Migration 001 — initial schema ──────────────────────────────────────────
@@ -1284,6 +1285,20 @@ export const STATE_MIGRATIONS = [
       DROP TABLE IF EXISTS canary_queries;
     `,
     },
+    // ── Migration 029 — improve_ledger.content_hash ───────────────────────────
+    //
+    // The consolidate pair pass (alpha.9) selects initiators by content change,
+    // not a time window: no row means eligible, a row whose content_hash
+    // differs from the asset's current body means eligible, otherwise not
+    // (src/commands/improve/consolidate/pair-pass.ts, selectInitiators). Every
+    // other source still uses next_eligible_at; this column is additional, not
+    // a replacement.
+    {
+        id: "029-improve-ledger-content-hash",
+        up: `
+      ALTER TABLE improve_ledger ADD COLUMN content_hash TEXT DEFAULT NULL;
+    `,
+    },
 ];
 assertMigrationRegistry(STATE_MIGRATIONS);
 function assertStateMigrationSafetyRegistry() {

package/dist/core/time.js

@@ -95,23 +95,3 @@ export function parseSinceToIso(since) {
     }
     return parsed.toISOString();
 }
-/**
- * LENIENT variant for config-driven since-values that may be a duration
- * shorthand (e.g. `"30m"`, `"24h"`, `"7d"` per {@link DURATION_UNITS}):
- * a matching duration resolves to the ISO timestamp of `now - duration`;
- * anything else is returned UNCHANGED (assumed to already be an ISO
- * timestamp) — this function never throws.
- *
- * Deliberately distinct from {@link parseSinceToIso}, which rejects
- * unparseable input with a UsageError: the identity fallback here is
- * behavior consumers rely on (consolidate's incremental narrowing compares
- * the returned string lexicographically against file-mtime ISO strings, so
- * garbage input silently selects nothing rather than failing the run —
- * pinned by tests/commands/goldens-duration-flags.test.ts).
- */
-export function parseSinceToIsoLenient(since) {
-    const ms = parseDuration(since, DURATION_UNITS);
-    if (ms === null)
-        return since;
-    return new Date(Date.now() - ms).toISOString();
-}

package/dist/indexer/db/llm-cache.js

@@ -27,10 +27,10 @@ import { computeBodyHash, getLlmCacheEntry, upsertLlmCacheEntry, } from "../../s
  *                          already hashed the body (e.g. to reuse it elsewhere)
  *                          should pass this to avoid the redundant work.
  * @param cacheVariant  - Namespace token for the cache row so different passes'
- *                        rows do not collide (e.g. memory-inference vs graph).
+ *                        rows do not collide (e.g. memory-inference vs metadata-enhance).
  * @param hooks         - Optional event sink for telemetry (see {@link WithLlmCacheHooks}).
  */
-export async function withLlmCache(db, cacheKey, body, reEnrich, llmFn, validate, precomputedHash, cacheVariant = "", hooks) {
+export async function withLlmCache(db, cacheKey, body, reEnrich, llmFn, validate, precomputedHash, cacheVariant, hooks) {
     const bodyHash = precomputedHash ?? computeBodyHash(body);
     if (!reEnrich) {
         const cacheHit = bestEffort(() => {

package/dist/indexer/ensure-index.js

@@ -115,8 +115,8 @@ function hasNewerIndexableFiles(stashDir, builtAt, indexedPaths, indexedHashes)
                 continue;
             // Newer than the last full build — only stale if its current content
             // actually differs from what is indexed. No stored hash means the row
-            // predates content-hash tracking (or a hash-less enrichment pass), so
-            // fall back to the conservative mtime-stale answer.
+            // predates content-hash tracking, so fall back to the conservative
+            // mtime-stale answer.
             const indexedHash = indexedHashes.get(file);
             if (indexedHash === undefined)
                 return true;

package/dist/indexer/index-written-assets.js

@@ -15,9 +15,8 @@
  * This is NOT a general reindex. It upserts exactly the files the caller just
  * wrote: frontmatter/metadata via the shared matcher pipeline, the canonical
  * row with its transactionally owned FTS projection, and vectors for changed
- * entry IDs when semantic search is enabled. Index-time LLM passes, graph
- * extraction, `builtAt`, and the per-dir walk cache remain full-index
- * responsibilities.
+ * entry IDs when semantic search is enabled. `builtAt` and the per-dir walk
+ * cache remain full-index responsibilities.
  */
 import fs from "node:fs";
 import path from "node:path";