@nexus-cortex/cli 4.124.57 → 4.124.58

package/.env.defaults

@@ -368,6 +368,10 @@ MENTORSHIP_HELPER_MODEL=deepseek-flash
 # [LEDGER 2026-09-10 @4.106.0] BUILT from the cell-m pilot (0/10 pro lift plans, 1/8 verdicts at the 4000 cap; direct API:
 #   completion_tokens 2045 = 1641 reasoning + ~400 content). UNTESTED on the bench — the flash-none vs flash-high re-pilot
 #   (K=2) is the efficacy test; until then CORTEX_MENTOR_REASONING=none stays the k=5 COMMON.
+# [LEDGER 2026-10-05] the 16000 pinned in every cell env since cell-n-r1 (Flash-on-MiMo, 09-11) rode into c21–c23 + the 10-04 TB4 arms under
+#   a PRO planner on TB4: planner max_tokens 20,000 (+4,000 plan budget), clamped to the 65,536 card; first-call clock 180 s ≈ 16.8K tokens
+#   at Pro's 94 tok/s → 16–26% of TB4 first plans aborted → thinking-OFF truncated retry → those tasks pass 30% vs 51% (r-tb4-valA
+#   addendum 4). 10-05 cap-raise cell: 60000 with CORTEX_LIFT_PLAN_BUDGET_TOKENS=6000 + CORTEX_LIFT_PLAN_TIMEOUT_MS=1000000 (first call 600 s).
 #CORTEX_MENTOR_REASONING_ALLOWANCE=
 # [LEDGER 2026-09-11 @4.107.0] cell-n-r1 at 16000 (cap 20K): flash@high/max still truncated ~30% (retry rescued, ~100 s); reasoningTok p90 16.8K/15.3K. Moot while thinking-on is DARK.
 # [LEDGER 2026-09-11 @4.106.1] cell-m-r1: flash@high thinking-on calls used 10.2K/10.6K reasoning tokens (med/p90) when they delivered and hit the 16K cap on 7/18 → 12K is ~50% short for flash@high; set 16000 on the next run (cap 20K). flash@max delivered at 3.8K/13.4K under 28K.
@@ -1020,6 +1024,10 @@ CORTEX_ENDTURN_GATE=true
 # walled turn + nudge on the tool_result tail; off = pre-existing; streamMessage wall cases skipped — scripted stream has no reasoning deltas).
 # [LEDGER 2026-10-01 c23 → written back 2026-10-04] ON in c23 d/w: wall_drop engaged in both; d +3.7, w +4.7 vs g +3.9 (undetermined vs g, w the
 #   cheapest arm); v39a (w) 5/8 +1.8, v41b 5/8 +1.7. IN THE STANDARD BUNDLE (w stack, operator-ratified for the full TB4.0 run).
+# [LEDGER 2026-10-05 hazard] photonic-waveguide-routing in t4r/t4s/t4j: triple 65,536-token reasoning exhaustion on a ~200 K-char Read of
+#   the task's checker → WALL_DROP dropped the model's OWN walled reasoning ×3 (summary ~700 chars kept; the tool result is never dropped —
+#   corrected 10-05 from wallGuard.ts) → model gave up with hours left (abnormal-exit-bypass). c23 g (no wall levers) passed 2/2, c23 w 1/2.
+#   r-tb4-valA addendum M2 + correction.
 #CORTEX_WALL_DROP=
 # CORTEX_OUTPUT_CAP_TOKENS — HB-OUTPUT-CAP (DARK, 2026-10-01). Per-call output cap for the action model (fills maxTokens only when the request
 # did not set one; helper/mentor calls unaffected). c22 acting calls' reasoning: p99 14K, p99.9 40K, 7 of 19,337 above 48K → a wall costs
@@ -1030,7 +1038,17 @@ CORTEX_ENDTURN_GATE=true
 # completed calls produced > 48K output (0.06%) and 7 of them sat in PASSING runs (6 x photonic-waveguide-routing at 48-62K: its routing plan
 # is one long think; + distributed-dedup r3 58.5K). Any cap below ~63K reaches passing calls; the saving is ~0.5 walls/session x ~17K output
 # tokens (~1 cent). Not armed in c23; wall drop + wall summary carry the wall savings. Revisit only with a passing-call-safe value (>= 63K).
+# [LEDGER 2026-10-05 @4.124.58] THE CEILING WAS OURS: the DeepSeek cards said outputTokens 65,536 and the gateway sends the card's limit as
+#   max_tokens on EVERY action call (GatewayTranslationLayer ~767) — every reasoning wall of the arc hit our own 64K. DeepSeek's API
+#   accepts max_tokens up to 393,216 (384K; defaults 64K thinking / 128K at effort max when unset); DeepSeek recommends ≥ 256K; AA runs at
+#   the disclosed max. Cards now 393,216 → an unset lever sends 393,216. Use this lever to pin a LOWER cap (e.g. 65536 to reproduce the
+#   old behaviour as a control, 131072 = the provider's max-effort default). Pairs with CORTEX_HTTP_TIMEOUT_MS (below).
 #CORTEX_OUTPUT_CAP_TOKENS=
+# CORTEX_HTTP_TIMEOUT_MS — HTTP timeout for the NON-STREAMING chat call (2026-10-05). The openai SDK default (600 s, 2 retries) bounds the
+# whole response body; a 384K-token thinking turn decodes 20–70 min, so the default would abort it and retry at double cost. Unset =
+# derived from the request's max_tokens at 50 tok/s with a 600-s floor (65,536 → 1,311 s; 393,216 → 7,864 s); integer ms overrides.
+# [LEDGER 2026-10-05 @4.124.58] BUILT + unit-tested (httpTimeout.test.ts); first bench use = the 10-05 cap-raise control cell.
+#CORTEX_HTTP_TIMEOUT_MS=
 # CORTEX_WALL_SUMMARY / _TIMEOUT_MS — HB-WALL-SUMMARY (DARK, 2026-10-01). On a reasoning-exhaustion wall, before HB-WALL-DROP removes the
 # turn, one helper call (the EndTurn judge's mentor wire, ~500-token budget) condenses the walled reasoning (head 4K + tail 16K chars) into
 # CONCLUDED / STUCK ON / NEXT; the result rides INSIDE the wall nudge (same cache-safe tail), attributed as written by the harness — item-11c
@@ -1165,6 +1183,12 @@ CORTEX_ENDTURN_GATE=true
 # budgeted cap and the R173 gap rules keep their fraction logic. all: the v1 behaviour (every fraction rule sees below-budget) — measured
 # 10-04 to over-correct (cwbfl 20/40 vs bundle−MAX_VETOES 25/40: accept-with-gap doubled). Events carry budgetFloorMode.
 # [LEDGER 2026-10-04 @4.124.57] DARK — BUILT + unit-tested; first arm cwbfl2 = MiMo bundle + FLOOR_S=1200 (mode evidence), expect ≈ 25/40.
+# [LEDGER 2026-10-05 @4.124.57 first arm] POSITIVE: cwbfl2 (full MiMo bundle incl. MAX_VETOES=3 + FLOOR_S=1200, mode evidence) = 27/40 —
+#   best Flash MiMo arm of the arc; +10 vs the bundle alone (17), +4 vs no-bundle (23), +2 vs hand-removed cap (25); judge FA 27%; floor
+#   fired 82/82 with the R160 cap intact. Replicate K=2 before claiming; proposed for STANDARD_BUNDLE. Ledger r-mimo64-bundle-2026-10-04.md §8.
+# [LEDGER 2026-10-05 @4.124.57 K=2] REPLICATED: cwbfl3 26/40 → per-run 27 / 26 (mean 26.5), union 31, both 22; +9.5 vs the bundle
+#   alone, +3.5 vs no-bundle 23, +1.5 vs hand-removed cap 25; both repeats above the Flash bar 22.4. WIN over the bundle-without-floor;
+#   first MiMo config above the bar on two repeats. Proposed STANDARD_BUNDLE addition (FLOOR_S=1200, MODE=evidence).
 #CORTEX_BUDGET_FLOOR_MODE=
 # (no lever) R161 HB-JUDGE-TASKSHAPE (4.112.0, 2026-09-16): `isTaskShaped()` gated the EndTurn finish judge and the deadline-exit mentor
 # with a fixed coding-verb list — prompts like "Eliminate cumulative layout shift…", "To determine the beta activity…", "You are given …
@@ -1252,6 +1276,9 @@ CORTEX_ENDTURN_GATE=true
 #   undetermined; engaged (task_rule_plan 40 / task_rule_judge 74); protected rejects 2 vs 3. TB4 validation run A is still the real test.
 # [LEDGER 2026-10-04 TB4 arm] R211–R214 with R208–R210 on TB4 (t4r): 3/11 vs c23 6/11 — no lift, negative-leaning (uefi-bootkit, risk-scorer-replay
 #   lost). Stays dark. Ledger r-tb4-valA-2026-10-04.md.
+# [LEDGER 2026-10-05 mechanism] t4r trajectories: rule judge meets=false 32/36 (89%) → 14 vetoes + 11 accept-with-gap; sessions ended
+#   MID-VETO at remFrac ≥ 0.6 (cls 24,936 s). A never-satisfied gate on long budgets; harmless on 900-s MiMo only because the budget rules
+#   cut it off. Needs a fit gate + rule-veto cap before any revival. r-tb4-valA-2026-10-04.md addendum M1.
 #CORTEX_RULE_INPUTS=
 # R212 INTERP = L2 interpretation enumeration: named alternatives each computed; keep the one consistent with every explicit hint and all data.
 # [LEDGER 2026-10-02 @4.124.43] DARK — BUILT + unit-tested (taskRules.test.ts). NO bench arm yet. evidence: tb4-never-passed-2026-10-02 (foodstuff-beta-activity, glycan-ms2-elucidation, hof-topology-interpenetration, ontology-kg-querying).

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@nexus-cortex/cli",
-  "version": "4.124.57",
+  "version": "4.124.58",
   "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": "4.124.57",
+    "@nexus-cortex/core": "4.124.58",
     "chalk": "^4.1.2",
     "cli-spinners": "^2.9.0",
     "commander": "^11.0.0",