@nexus-cortex/cli 4.124.51 → 4.124.54

package/.env.defaults

@@ -1822,6 +1822,13 @@ MENTORSHIP_ACTIVE_DISCOVERY=false
 # [LEDGER 2026-10-02 @4.124.44] DARK — BUILT + unit-tested (transportFixes.test.ts: OpenAI/xAI/streaming request shapes,
 # off = byte-identical). Live probe: DeepSeek /responses accepted function_call → function_call_output → user reminder
 # item (200, model obeyed the reminder). NO bench arm. evidence: transport map 10-02 (R215/R216/R218).
+# [LEDGER 2026-10-03 @4.124.52] LIVE on OpenAI, full harness (cell-smoke, gpt-6-luna + deepseek-v4-pro mentor, payload log): the one
+# reminder-carrying tool turn went as [function_call_output, message] — inline after the tool output; instructions carried only the static
+# prompt. First bench arm = STDB run g6rs2 (vs g6pro2 without the lever), 2026-10-03.
+# [LEDGER 2026-10-03 @4.124.51 cell] NEGATIVE SIGNAL on its first bench arm: STDB MiMo-40 g6rs2 (gpt-6-luna action, deepseek-v4-pro mentor,
+#   R215+R216 on) = 15/40 vs the otherwise-identical g6pro2 = 24/40 (−9; 10 PASS→FAIL / 1 FAIL→PASS; protected rejects 5 vs 3). R216 never
+#   fired (0 truncations) so the drop is R215's. Engagement: effcfg gate active + smoke payload-log shape; no wire log in the cell. Stays DARK.
+#   Ledger omniclaude-v4/.cortex/bench/r-mimo64-g6pro2-g6rs2-2026-10-03.md.
 #CORTEX_RESPONSES_INLINE_REMINDERS=
 # CORTEX_RESPONSES_STOP_REASON — R216 HB-RESPONSES-STOPREASON (DARK, 2026-10-02). extractStopReason returned undefined for
 # the responses and hf-space patterns, so a max-output-tokens cutoff was never classified `truncated` and the wall levers
@@ -1834,7 +1841,21 @@ MENTORSHIP_ACTIVE_DISCOVERY=false
 # reasoning exhaustion; streaming response.incomplete). Live probe: DeepSeek /responses max_output_tokens 8 → status
 # incomplete, incomplete_details {reason: max_output_tokens}, same in the streaming response.incomplete event.
 # evidence: transport map 10-02 (R215/R216/R218).
+# [LEDGER 2026-10-03 @4.124.52] on in STDB run g6rs2 (first bench arm); the 10-03 OpenAI smoke produced no truncation, so the classifier
+# path is still unit-test + DeepSeek-probe evidence only (it is response parsing — cannot 400).
+# [LEDGER 2026-10-03 @4.124.51 cell] UNEXERCISED in g6rs2 (40 tasks, 0 wall/truncation events) — no evidence either way; stays DARK.
 #CORTEX_RESPONSES_STOP_REASON=
+# CORTEX_CHAT_REMINDERS_TO_SYSTEM — R231 HB-CHAT-REMINDERS-SYSTEM (DARK, 2026-10-03). chat/completions mirror of the Responses-OFF
+# reminder delivery: every <system-reminder> block carried in a USER text block (lift plan, requirement ledger, PID1 note, plateau
+# instruction, lift nudge, deferred-corpus header, mentor insight) is lifted out of the conversation and appended to the system message
+# on every request; reminders prepended to tool results stay in place (same scope as the Responses extractor); the compaction marker
+# stays. off (default): unchanged — in place with the tags stripped (ChatCompletionsAPIAdapter.stripSystemReminderTags). Cache note:
+# the system message changes whenever a new reminder appears → DeepSeek prefix-cache miss on those turns (Responses-OFF pays the same).
+# Why: MiMo-40 g6rs2 (R215 on = in place) 15/40 vs g6pro2 (reminders in `instructions`) 24/40 on GPT-6 Luna; a DeepSeek cell has never
+# had the system-message delivery. A/B = deepseek-flash action + V4 Pro mentor, MiMo-40, off vs on; readout incl. judge false-accept.
+# [LEDGER 2026-10-03 @4.124.53] DARK — BUILT + unit-tested (transportFixes.test.ts: lift from user text blocks only, tool-result and
+#   compaction marker untouched, off = byte-identical; gateway scoped to ChatCompletionsAPIAdapter). NO bench arm yet.
+#CORTEX_CHAT_REMINDERS_TO_SYSTEM=
 # CORTEX_ANTHROPIC_HISTORY_CACHE — R218 HB-ANTHROPIC-CACHE (DARK, 2026-10-02). Anthropic Messages requests set cache_control
 # only on the system prompt and the last tool, so the conversation history was never cached on Claude. on (and
 # ANTHROPIC_PROMPT_CACHING not false): rolling message breakpoints on the last markable block (not thinking, not empty
@@ -2282,6 +2303,13 @@ CORTEX_ORIENT_BOOTSTRAP=1
 # CORTEX_LIFT_PLAN_INTERACTIVE=true. DARK; validate adaptive/auto on TB2.1 or TB4.0 (build-style tasks) before any default.
 # [LEDGER 2026-09-27 @4.124.27] MiMo sweep (r-mimo64-sweep-2026-09-27.md): with SCOPE=true + BASELINE=1 on the o2 config, protected-file rejects
 # fell to 13/320 (4.1%) vs 31/320 (9.7%) in the five prior runs — mechanism works; pass rate unchanged (5 arms pooled 60.6% vs control 59.9%). DARK.
+# [LEDGER 2026-10-03 @4.124.52] CORTEX_LIFT_PLAN_SCOPE=task — ONE task-conditioned scope rule for every grader/surface (operator: a
+# per-grader flag is a trap — production has no grader, and the strict MiMo rule rode the stack into TB4). Named files bind; forbidden
+# paths + vendored/third-party code, tests, runners, fixtures, data off-limits unless the task points there; otherwise every file the fix
+# needs (no file-count limit); revert INCIDENTAL changes only, re-running the task check without a change before removing it. Judge gets
+# SCOPE RULE (never vetoes on file count). Evidence for the change: v43b sglang-qwen-burst — judge vetoed a working 8-file fix twice under
+# SCOPE=true ("smallest set"), agent stashed 7 detector files, 1-file patch failed 10/10 hidden tests (only sglang run ever under SCOPE=true;
+# prior 12/20). DARK. Validate: MiMo regression (protected-reject rate must stay ~4%, not back to ~10%) + TB4 multi-file guards; then default.
 #CORTEX_ORIENT_BASELINE=
 #CORTEX_LIFT_PLAN_SCOPE=
 # [LEDGER 2026-09-02 @4.89.0] code default 30 (4.88.0) — 20 made the break rung (2N=24) unreachable; preserves 40%/80% density semantics.

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@nexus-cortex/cli",
-  "version": "4.124.51",
+  "version": "4.124.54",
   "description": "Nexus Cortex CLI - Terminal interface for multi-provider LLM orchestration",
   "main": "dist/index.js",
   "type": "module",
@@ -19,7 +19,7 @@
     "prepack": "node ../../scripts/copy-pkg-cortex-scaffold.mjs"
   },
   "dependencies": {
-    "@nexus-cortex/core": "4.124.51",
+    "@nexus-cortex/core": "4.124.54",
     "chalk": "^4.1.2",
     "cli-spinners": "^2.9.0",
     "commander": "^11.0.0",