chore(release): 2.3.0 version bump, changelog and status line (Refs #177)

CHANGELOG.md

@@ -1,5 +1,55 @@
 # Changelog
 
+## 2.3.0 — 2026-10-11
+
+Phase 9 recursive self-improvement & model distillation. The factory can now
+propose better prompts, distill its own verified work into local-model
+training data, and hot-swap its own binaries — every loop fail-closed and
+human-gated by construction: nothing rewrites prompts, weights or shipped
+code without a recorded human approval, and the canary stays OFF until a
+human explicitly enables it. The published trio (core/CLI/plugin) releases
+in lockstep as 2.3.0.
+
+**Behaviour changes.**
+
+- **#174: Evolutionary prompt & cognitive strategy optimizer.**
+  `generatePromptCandidates` (`packages/core/src/improve/optimizer.ts`)
+  turns telemetry failure clusters into targeted prompt variations
+  (append-guidance, trim-bloat, multi-generation combine); `evolvePrompts`
+  grades each against the benchmark suite and keeps only strict Pareto
+  winners (zero regressions plus a strictly better pass rate or lower
+  spend/steps). Winners become review-only proposal dirs with before/after
+  benchmark receipts; `applyPromptPatch` is the sole applier and refuses
+  without a matching recorded human approval. Signal-free telemetry yields
+  no candidates and vacuous baselines are refused.
+- **#175: Synthetic trajectory harvester for distillation.**
+  `harvestTrajectories` captures multi-turn trajectories of passed, verified
+  phase worktrees only (green gate summaries plus a non-vacuous audit chain,
+  derived by `es improve export-dataset` from on-disk evidence); every
+  string is secret-scrubbed, diffs are path- and existence-verified, and
+  unverified/missing/escaping inputs are skipped with reasons. Exports
+  Unsloth/Axolotl/HF-compatible chat JSONL (`sft`) and prompt/chosen/rejected
+  pairs (`dpo`, never inventing negatives) with a provenance manifest.
+- **#176: Canary self-compilation and hot-swap harness.**
+  `CanaryRunner` (`packages/fleet/src/canary.ts`) builds candidates in an
+  isolated staging sandbox, shadow-tests them beside the live worker, and
+  atomically re-points the active symlink only when green — rolling back
+  instantly when red. OFF by default: construction without `enabled: true`
+  and a recorded human approval throws. The rollback target is frozen at
+  construction, guarded writes deny the rollback tree and the active link,
+  and candidates built inside the rollback tree are refused.
+
+**Security.**
+
+- No autonomous path rewrites prompts, weights or shipped code: the
+  optimizer and canary apply only under recorded human approvals, and the
+  canary pipeline seam has no silent defaults.
+- Distillation datasets are secret-scrubbed (token families, credentials,
+  private-key blocks) with redaction counts; DPO export refuses to invent
+  rejected responses.
+- Test fixtures build fake secrets by interpolation (repo precedent) so
+  static scanners never see credential-like literals.
+
 ## 2.2.0 — 2026-10-11
 
 Phase 8 anti-Goodhart verification & mutation oracle. Tautological tests,

README.md

@@ -5,11 +5,12 @@
 > phases under dual QA, and open a PR a human merges. Mechanical gates and a
 > human-only approval channel keep autonomous runs honest.
 
-> **Status: 2.2.0.** Phase 8 anti-Goodhart verification & mutation oracle:
-> tautological suites fail the mutation gate below an 80% kill rate,
-> programmer and QA seats refuse shared-weight providers (`models.programmer`
-> vs `models.qa`), and factory fitness is tracked on 54 oracle-graded SWE
-> benchmark instances with pass@1, spend and regression deltas (#169–#171).
+> **Status: 2.3.0.** Phase 9 recursive self-improvement & model distillation:
+> prompt candidates evolve against telemetry failure clusters and ship only
+> behind a strict Pareto gate with benchmark receipts (`improve/optimizer`),
+> passed phase worktrees export scrubbed SFT/DPO datasets
+> (`es improve export-dataset`), and candidate builds promote through an
+> explicitly human-enabled canary with instant rollback (#174–#176).
 > Install from npm (`@heretek-ai/epistemic-swarm`, `@heretek-ai/es-core`,
 > `@heretek-ai/es-cli`); run `es key seal` once.
 

bun.lock

@@ -14,7 +14,7 @@
     },
     "packages/cli": {
       "name": "@heretek-ai/es-cli",
-      "version": "2.2.0",
+      "version": "2.3.0",
       "bin": {
         "es": "./bin/es.js",
         "epistemic-swarm": "./bin/es.js",
@@ -26,7 +26,7 @@
     },
     "packages/core": {
       "name": "@heretek-ai/es-core",
-      "version": "2.2.0",
+      "version": "2.3.0",
       "dependencies": {
         "vscode-jsonrpc": "9.0.3",
         "vscode-languageserver-protocol": "3.18.4",
@@ -50,7 +50,7 @@
     },
     "packages/opencode": {
       "name": "@heretek-ai/epistemic-swarm",
-      "version": "2.2.0",
+      "version": "2.3.0",
       "dependencies": {
         "@heretek-ai/es-core": "workspace:*",
         "@opencode/plugin": "2.0.24",

packages/cli/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@heretek-ai/es-cli",
-  "version": "2.2.0",
+  "version": "2.3.0",
   "esRelease": {
     "order": 2
   },

packages/cli/src/version.ts

@@ -1,2 +1,2 @@
 /** The `es` CLI version. Single source: kept equal to the package version by test/cli.test.ts. */
-export const VERSION = "2.2.0"
+export const VERSION = "2.3.0"

packages/core/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@heretek-ai/es-core",
-  "version": "2.2.0",
+  "version": "2.3.0",
   "esRelease": {
     "order": 1
   },

packages/opencode/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@heretek-ai/epistemic-swarm",
-  "version": "2.2.0",
+  "version": "2.3.0",
   "esRelease": {
     "order": 3
   },