@nexus-cortex/cli 4.124.28 → 4.124.31
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- package/.env.defaults +24 -0
- package/package.json +2 -2
package/.env.defaults
CHANGED
|
@@ -1124,6 +1124,30 @@ CORTEX_ENDTURN_GATE=true
|
|
|
1124
1124
|
# P<0.0001 vs per-task control rates; 4 always-pass tasks lost. The judge quotes a failing blind check (specFrac 0.17-0.25, below the veto
|
|
1125
1125
|
# fraction) as its own named check → ordinary veto → the agent reshapes correct output to the invented check. KEEP DARK; redesign so check
|
|
1126
1126
|
# text never reaches the judge/writer (r-mimo64-val-2026-09-27.md §2).
|
|
1127
|
+
# CORTEX_JUDGE_SPEC_TEXT — R174e (2026-09-27, in tree): shown (default) | hidden. Hidden removes the spec checks' TEXT from the judge prompt and
|
|
1128
|
+
# the writer's veto message: spec evidence acts only through the averaged fraction gate (R174d); a judge CHECK line that echoes a spec
|
|
1129
|
+
# check is not run as a named check (closes the namedFailed bypass the sp arm hit); a fraction hold over a MEETS gives the writer a neutral
|
|
1130
|
+
# "re-verify each stated requirement against the task's own entry points; change nothing the task did not ask for" lead.
|
|
1131
|
+
# [LEDGER 2026-09-27 in tree] DARK — unit-tested; efficacy UNMEASURED (needs a MiMo cell: spec tests on + hidden vs the 4.124.27 base).
|
|
1132
|
+
# [LEDGER 2026-09-27 @4.124.30] MiMo verification cell sh (K=2, spec tests + hidden): 67/128 = 52.3%, −8.3 pts vs baseline per-task rates, P=0.0005 —
|
|
1133
|
+
# HARMFUL (half the sp arm's −16; the fraction gate still holds correct MEETS: 26 holds). R174 line CLOSED as a finish gate on this family. DARK.
|
|
1134
|
+
#CORTEX_JUDGE_SPEC_TEXT=
|
|
1135
|
+
# CORTEX_ENDTURN_REQ_COVERAGE — EndTurn Stage 4f requirement coverage (2026-09-27, in tree). Evidence: MiMo tb64 sweep, 186 failing hidden tests
|
|
1136
|
+
# in 111 failed sessions (.bench/mimo/fail-criteria-sweep-v2.jsonl, reasoning-first labels): 101 check a requirement the task states (72) or
|
|
1137
|
+
# implies (29), and the agent's own verification covered it only partially; 85 assert details the task never gives (AST shape, exact source
|
|
1138
|
+
# lines, hashes) — out of reach for any verification lever. Stage 4f extracts the task's OWN requirement clauses deterministically (bullets +
|
|
1139
|
+
# must/exactly/do-not/preserve/remain/unchanged/only sentences; 3–12 per MiMo task) and, once per turn, rejects a finish whose `requirements`
|
|
1140
|
+
# attestation leaves clauses uncovered, quoting them verbatim — no model, no invented check (the R174 lesson). Needs CORTEX_ENDTURN_REQUIREMENTS.
|
|
1141
|
+
# Banks `req_coverage` {clauses, uncovered, nudged}. [LEDGER 2026-09-27 in tree] DARK — unit + gate tests; efficacy UNMEASURED (MiMo cell).
|
|
1142
|
+
# [LEDGER 2026-09-27 @4.124.30] MiMo verification cell rc (K=2): 72/128 = 56.2%, −4.4 pts, one-sided P=0.048 — UNDETERMINED leaning negative; fires on
|
|
1143
|
+
# 63% of sessions (the 40% token-overlap matcher counts paraphrased rows as uncovered). DARK; retest only with a stricter matcher (r-mimo64-verif-2026-09-27.md).
|
|
1144
|
+
#CORTEX_ENDTURN_REQ_COVERAGE=
|
|
1145
|
+
# CORTEX_TOOL_DOCTRINE — tool-description doctrine variant (2026-09-28, in tree). mimo-v1 = the three edits BOTH MiMo doctrine mines proposed
|
|
1146
|
+
# (sweep 320 sessions + verification cell 255): Read over bash file viewing (Read + Bash descriptions), no reference mining in parent dirs /
|
|
1147
|
+
# sibling projects / whole filesystem, ad hoc installs count as manifest changes only if the final manifest includes them. Anchored appends
|
|
1148
|
+
# (tested: every anchor exists once; unset = byte-identical). Wording keeps the install directive and project-test regression checks legitimate.
|
|
1149
|
+
# [LEDGER 2026-09-28 in tree] DARK — efficacy UNMEASURED (MiMo cell, b config + mimo-v1, K=2, vs sweep pooled 194/320).
|
|
1150
|
+
#CORTEX_TOOL_DOCTRINE=
|
|
1127
1151
|
#CORTEX_JUDGE_SPEC_TESTS_SAMPLES=
|
|
1128
1152
|
#CORTEX_JUDGE_SPEC_VETO_FRAC=
|
|
1129
1153
|
# CORTEX_JUDGE_VETO_MIN_REMAINING / _GAP_HOLD_MIN_INTERVAL_MS / _GAP_HOLD_PLAN_MAX_SIMILARITY — R173c (4.121.0, 2026-09-20, cell g1
|
package/package.json
CHANGED
|
@@ -1,6 +1,6 @@
|
|
|
1
1
|
{
|
|
2
2
|
"name": "@nexus-cortex/cli",
|
|
3
|
-
"version": "4.124.
|
|
3
|
+
"version": "4.124.31",
|
|
4
4
|
"description": "Nexus Cortex CLI - Terminal interface for multi-provider LLM orchestration",
|
|
5
5
|
"main": "dist/index.js",
|
|
6
6
|
"type": "module",
|
|
@@ -19,7 +19,7 @@
|
|
|
19
19
|
"prepack": "node ../../scripts/copy-pkg-cortex-scaffold.mjs"
|
|
20
20
|
},
|
|
21
21
|
"dependencies": {
|
|
22
|
-
"@nexus-cortex/core": "4.124.
|
|
22
|
+
"@nexus-cortex/core": "4.124.31",
|
|
23
23
|
"chalk": "^4.1.2",
|
|
24
24
|
"cli-spinners": "^2.9.0",
|
|
25
25
|
"commander": "^11.0.0",
|