@nexus-cortex/server 4.45.0 → 4.46.0

package/.cortex/agents/autoresearch-agent.md

@@ -17,6 +17,8 @@ model: inherit
 
 You are ONE of several auto-research agents the PM spawned on the **same** backlog deficiency. Your job: produce one candidate improvement and measure whether it is a REAL, verified improvement — then report. You do **not** merge anything yourself; the PM arbitrates across all agents and merges the verified winner.
 
+The same is true when the LOOP spawned you: `cortex autoresearch loop --width N` (4.45+) fans each round into N parallel Fixer arms — different models, same goal, own worktree each. If you are the Fixer in such a round, other arms are attacking the same deficiency concurrently; the gate compares every arm at FWER-adjusted significance and only the best accepted candidate merges. Same rules apply: smallest real fix, no eval-gaming — you win by merit, not by volume.
+
 ## Read your task first
 Your prompt contains:
 - The **deficiency** (what to improve) and how it's measured (task-set / metric).

package/.cortex/skills/autoresearch/SKILL.md

@@ -44,9 +44,11 @@ Identical agents on identical prompts waste the parallelism — they trace the s
 
 **Diversify the SEARCH; keep the EVALUATION identical** — every arm is judged by the *same* metric + the *same* gate (one shared judge). Letting arms pick their own metric is reward-hacking. Keep N small with **sharp** distinctions (4–5 genuinely different approaches beat many near-duplicates) — N arms ≈ N× the spend, so buy breadth, not duplicates.
 
+**The loop AUTOMATES this doctrine (4.45+).** `cortex autoresearch loop --width N` fans each round into N parallel arms natively — one worktree + one Fixer model per arm, per-arm `fixer:<model>` strategy labels recorded into the matrix, the gate FWER-adjusted to the true family width, best accepted arm merges (judge reviews accepted arms in effect order). Pick the arm pool with `--arm-models a,b,c` (explicit) or `--providers deepseek,anthropic,…` (flagship model per provider; auto-detected from funded keys when omitted), and set `--missing-provider-key-policy platform_fallback|omit|redistribute` for arms whose provider key isn't funded. Manual Task-dispatch arms (above) remain the right tool when arms need distinct PERSONAS/prompts, not just distinct models — the loop's arms share one goal prompt.
+
 ## 3. DELEGATE (pick the execution path by how you're accessed)
-- **Local cortex harness** (you're driving cortex, or inside it): set `AUTORESEARCH_AGENTS=native` and delegate via the **Task tool** (`subagent_type: autoresearch-agent`), one per strategy, each prompt = the plan + that arm's persona/strategy + `EXECUTION MODE: native`. Or drive the CLI directly: `cortex autoresearch fix` / `experiment` / `loop`.
-- **Hosted at scale** (external agent): the hosted **`nexus-autoresearch` MCP** is LIVE. Relay the plan to its tools (`start_autoresearch_campaign` / `autoresearch_experiment` / `autoresearch_fix`) — `EXECUTION MODE: mcp`. Hosted campaigns run the **LLM judge gate ON by default** (see §4); pass `requireJudge:false` to disable or `judgeRubric` to customize.
+- **Local cortex harness** (you're driving cortex, or inside it): set `AUTORESEARCH_AGENTS=native` and delegate via the **Task tool** (`subagent_type: autoresearch-agent`), one per strategy, each prompt = the plan + that arm's persona/strategy + `EXECUTION MODE: native`. Or drive the CLI directly: `cortex autoresearch fix` / `experiment` / `loop` — for a multi-provider swarm campaign, `loop --width N --arm-models/--providers …` runs the fan-out natively (§2).
+- **Hosted at scale** (external agent): the hosted **`nexus-autoresearch` MCP** is LIVE. Relay the plan to its tools (`start_autoresearch_campaign` / `autoresearch_experiment` / `autoresearch_fix`) — `EXECUTION MODE: mcp`. Campaigns take `width` (1–8), `armModels`, `providers`, and `missingProviderKeyPolicy` for the multi-provider swarm fan-out. Hosted campaigns run the **LLM judge gate ON by default** (see §4); pass `requireJudge:false` to disable or `judgeRubric` to customize.
 
 The agents EXPLORE; they do not merge. They each return a candidate + its verdict.
 

package/.cortex/skills/cortex-bench/SKILL.md

@@ -290,6 +290,8 @@ cortex autoresearch experiment --experiment-tag <id> \
 ```
 `--json` → `{verdict, holdoutVerdict, regressedTasks, mergeEligible, benchSummaries, cortexDir, jsonlPaths}`. Owns the "two builds not one relabel" correctness (each arm served from its OWN built code; refuses `baseRef===candidateRef`). `--cortex-dir` is the per-experiment **artifact** (all 3 `.cortex/*.jsonl`). `mergeEligible` = keep ∧ fwerAdjusted ∧ holdout-keep (FALSE without holdout). Validated live end-to-end (both a real **discard** and a real **keep** with `mergeEligible:true`).
 
+**The multi-arm campaign runner — `cortex autoresearch loop --width N` (4.45+).** The swarm-width family is now AUTOMATED: `loop --width N` fans each round into N parallel Fixer arms (own worktree + own model each, pool via `--arm-models`/`--providers`, `--missing-provider-key-policy` for unfunded arms), passes `--n-family <width>` to every arm's experiment so the FWER bar is honest, records per-arm `fixer:<model>` strategy labels into the matrix, and merges only the best gate-accepted (and judge-approved, if required) arm. This IS the N-of-this-experiment machinery — don't hand-orchestrate N `experiment` calls when a width'd loop round does it with correct family statistics.
+
 ### Hard-won benchmarking gotchas (from the first live runs, 2026-06-08)
 - **Server returns `model` as an OBJECT** `{id, provider}`, not a string — any recorder must extract `.id` (fixed in serverRunner/estimateCost, `a4097aaf8`).
 - **`PROJECT_PATH`, not `PROJECT_ROOT`**, controls the orchestrator's project context (system messages, CORTEX.md, agents) — set it when spawning a harness for a specific checkout (fixed in startServer, `3f549daa5`).

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@nexus-cortex/server",
-  "version": "4.45.0",
+  "version": "4.46.0",
   "description": "Thin Express server wrapper for Nexus Cortex core library",
   "type": "module",
   "main": "dist/index.js",
@@ -20,8 +20,8 @@
     "prepack": "node ../../scripts/copy-pkg-cortex-scaffold.mjs"
   },
   "dependencies": {
-    "@nexus-cortex/core": "4.45.0",
-    "@nexus-cortex/executors": "4.45.0",
+    "@nexus-cortex/core": "4.46.0",
+    "@nexus-cortex/executors": "4.46.0",
     "chalk": "^5.3.0",
     "cors": "^2.8.5",
     "dotenv": "^16.4.5",