akm-cli 0.9.27-alpha.2 → 0.9.27

package/CHANGELOG.md

@@ -6,6 +6,52 @@ The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.1.0/).
 
 ## [Unreleased]
 
+## [0.9.27] - 2026-10-07
+
+The stable release of the 0.9.27 line: 0.9.27-alpha.1, alpha.2 and alpha.3.
+When upgrading from 0.9.26:
+
+- **The nightly queue stops refilling with duplicates.** Consolidate no longer
+  re-offers a memory whose body a reviewer rejected, and the drain's judge sees
+  the five knowledge notes nearest a promotion and rejects one they already
+  cover. A confident `subsumed` or `supersedes` retirement that the second look
+  confirms resolves unattended, like a duplicate (alpha.2).
+- **The judge rejects a promotion that is not durable:** a status, rollout,
+  version or test count, a plan not yet carried out, or something retired or
+  replaced. On 65 labelled promotions it accepted 13 of 61 bad ones before and
+  3 to 6 after, keeping 3 of the 4 good ones (alpha.3).
+- **Distill writes a lesson only when its memory holds one.** The writer may
+  answer that there is none; its judge rejects a lesson that only records what
+  was done, repeats an asset the library holds or claims what the memory does
+  not say; both see the related lessons, knowledge notes and skills (alpha.2).
+- **Reflect skips feedback an accepted exact fix already resolved**, and no
+  longer writes a second copy of a skill's reference files (alpha.1, alpha.2).
+- **Fixed:** an old (0.9.1-era) index is recreated instead of crashing every
+  `akm index`; an updated index scores like a fresh one; strict APIs that reject
+  `chat_template_kwargs` get one retry without it, and every response schema is
+  strict-valid; extract reports a lesson it refuses; a codex dispatch with a
+  schema leaves no temp folder (alpha.1).
+- **Removed:** `scripts/akm-eval`, now in itlackey/akm-eval (alpha.2).
+
+## [0.9.27-alpha.3] - 2026-10-07
+
+### Fixed
+
+- **The drain's judge rejects a promotion that is a status snapshot, a plan or
+  something retired, not only a duplicate.** Its rubric for a promotion (a memory
+  proposed as a new knowledge note) said to reject a note that is wrong, a
+  duplicate or contradicts the live asset, and to accept a correct, valuable one,
+  so it judged only whether the note was new: the reasons it gave for accepting a
+  note pinned to commits and versions, a rollout status or a retired host were
+  "distinct from the existing notes". The reject line now also names a note that
+  reports the state of something that changes (a status, rollout, branch, commit,
+  version, test count, "as of <date>"), a plan not yet carried out, or something
+  already retired, replaced or superseded, and says a lesson drawn from an
+  incident is durable. On the owner's 65 labelled promotions (chat/qwen3.8-27b,
+  two runs each) the judge accepted 7 and 8 of the 14 stale ones and 4 and 5 of
+  the 7 ephemeral ones before, and 1 and 3, 0 and 1 after. Other proposals keep
+  the old rubric. No new settings.
+
 ## [0.9.27-alpha.2] - 2026-10-07
 
 ### Removed

package/dist/commands/proposal/drain.js

@@ -144,6 +144,8 @@ export function fencedBlock(text) {
 }
 /** The judgment prompt: the proposal, the live asset it would overwrite, same-ref siblings, and for a promotion the nearest knowledge notes. */
 export function buildJudgmentPrompt(proposal, reason, ctx) {
+    // A promotion (a memory proposed as a new knowledge note) must also be durable.
+    const promotion = proposal.source === "consolidate" && ctx.liveAsset === undefined;
     const sections = [
         "You are adjudicating a pending knowledge-base proposal no quality judge has",
         "passed yet. Decide whether to accept, reject, or defer it.",
@@ -174,7 +176,9 @@ export function buildJudgmentPrompt(proposal, reason, ctx) {
         }
         sections.push("", "Reject the promotion if these notes already cover what it says, even in other words.");
     }
-    sections.push("", "## Your task", 'Return ONLY a JSON object: {"decision": "accept" | "reject" | "defer", "reason": "<short reason>"}.', "- accept: the proposed content is a correct, valuable update worth committing.", "- reject: the proposal is wrong, a duplicate, or contradicts the live asset.", "- defer: you cannot decide from the provided context (leave it pending).", "Output the JSON object and nothing else.");
+    sections.push("", "## Your task", 'Return ONLY a JSON object: {"decision": "accept" | "reject" | "defer", "reason": "<short reason>"}.', "- accept: the proposed content is a correct, valuable update worth committing.", promotion
+        ? '- reject: the proposal is wrong, a duplicate, contradicts the live asset, or is not durable: it reports the state of something that changes (a status, rollout, branch, commit, version, test count, "as of <date>"), is a plan not yet carried out, or describes something already retired, replaced or superseded. A lesson drawn from an incident is durable.'
+        : "- reject: the proposal is wrong, a duplicate, or contradicts the live asset.", "- defer: you cannot decide from the provided context (leave it pending).", "Output the JSON object and nothing else.");
     return sections.join("\n");
 }
 /** A verdict from raw runner output (the first JSON object), or null. */

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "akm-cli",
-  "version": "0.9.27-alpha.2",
+  "version": "0.9.27",
   "type": "module",
   "description": "akm (Agent Knowledge Manager) — a portable, local-first capability library for AI agents. Discover, load, share, and improve reusable skills, scripts, workflows, and knowledge across any shell-capable coding agent, including Claude Code, OpenCode, and Cursor.",
   "keywords": [