@nexus-cortex/cli 4.124.59 → 4.124.61

package/.env.defaults

@@ -626,6 +626,13 @@ CORTEX_LIFT_PLAN=true
 # [LEDGER 2026-09-27] MiMo K=3 wire (banked `mentor`, 192 lift plans): pro thinking-on @max median 108 s, p90 166 s, 15/192 thinking-off retries — 12–22% of a 900 s MiMo agent budget. e6 (09-18) found no effort level separating (max 10/28, high 12/29, none 11/29 at 10× faster lift); max buys nothing measured. Candidate: CORTEX_LIFT_PLAN_REASONING=none (operator decision).
 #CORTEX_LIFT_PLAN_EFFORT=max
 #CORTEX_LIFT_PLAN_BUDGET_TOKENS=4000
+# [LEDGER 2026-10-06] TRAP: this lever governs ONLY the orchestrator's outer wait (withTimeout around generateTaskPlan); the helper's own
+#   first-call abort (the one that fires at 180 s and triggers the thinking-OFF truncated retry) follows the per-effort thinking table
+#   (mentorRole THINKING_TIMEOUT_MS max=300 s × FIRST_CALL_SHARE 0.6) because the lift-plan mentor config is resolved WITHOUT a timeoutMs
+#   input (HelperModelMiddleware ~1617). The 10-05 cap cell set this to 1,000,000 and every abort still recorded firstCallTimeoutMs 180,000.
+#   Workaround = CORTEX_MENTOR_THINKING_TIMEOUT_MS (all thinking-on surfaces); proper fix = pass this lever into the lift-plan mentor resolve.
+# [LEDGER 2026-10-06 @4.124.60] FIXED — the lift-plan mentor resolve now receives this lever as timeoutMs, so the first-call abort = 0.6 ×
+#   max(this, thinking table): 1,000,000 → 600 s (mentorRole.test.ts). Cells may drop the CORTEX_MENTOR_THINKING_TIMEOUT_MS workaround.
 #CORTEX_LIFT_PLAN_TIMEOUT_MS=90000
 #CORTEX_LIFT_PLAN_INTERACTIVE=false
 #CORTEX_LIFT_PLAN_RECON_TIMEOUT_MS=8000
@@ -1044,6 +1051,21 @@ CORTEX_ENDTURN_GATE=true
 #   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_WALL_RETRY_LADDER / CORTEX_WALL_FORCE_ACTION_AT — HB-WALL-RETRY-LADDER (R241/R240, 2026-10-06, DARK). Under the 393,216 cap every
+# wall of the 10-05 TB4 cell was a degenerate repetition loop (40/40) and the retry re-looped from its first tokens at full price; DeepSeek has
+# no server-side cycle breaker in thinking mode (probed 10-06: penalties deprecated, temperature inert, stop sequences never reach
+# reasoning_content). LADDER = max_tokens for the retry after the n-th exhaustion wall of a turn ("65536,16384": wall 1 → 65,536, wall 2+ →
+# 16,384; last rung repeats; never above the request's own cap). FORCE_ACTION_AT (default 2) = from that wall on the retry carries
+# tool_choice:required, which the gateway turns into thinking OFF on DeepSeek → the model must emit a tool call (probed 10-06: works
+# mid-conversation, thinking resumes next call; costs ONE uncached prefix read). Exhaustion kinds only, never a no-tools nudge. R240 rides
+# along: the retry's usage is now counted in the session ledger. Events: wall_retry_ladder {wallIndex, cap, forceAction, iteration}.
+# [LEDGER 2026-10-06 @4.124.60] DARK — BUILT + unit/integration-tested (wallRetryLadder.test.ts, wallGuard.integration.test.ts); first arm pending.
+# [LEDGER 2026-10-06 @4.124.60 WIRE-PROVEN] local headless session through the capture proxy (OUTPUT_CAP 1200, LADDER 1150,1100, FORCE_AT 2):
+#   wall 1 (iter 5) → retry max_tokens 1150, no forced choice; wall 2 (iter 7) → retry max_tokens 1100 + tool_choice "required" +
+#   thinking {type: disabled} on the wire → the model acted and finished the task. Events wall_retry_ladder {wallIndex 1 cap 1150}, {wallIndex 2
+#   cap 1100 forceAction true}. Note: with CORTEX_REASONING_EXHAUST_BACKOFF unset the retry effort also stepped down (bundle sets it false).
+#CORTEX_WALL_RETRY_LADDER=
+#CORTEX_WALL_FORCE_ACTION_AT=2
 # 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.
@@ -2079,6 +2101,10 @@ MENTORSHIP_ACTIVE_DISCOVERY=false
 # accepted stream_options (200; usage on the final chunk WITH finish_reason, not a separate choices:[] chunk). The prefix
 # harness cannot reach the streaming builders.
 #CORTEX_STREAM_USAGE=
+# [LEDGER 2026-10-06 @4.124.61 R238] The chat-completions stream's finalMessage now SETTLES on any early exit (consumer break / abort /
+#   throw) with the partial message: finish_reason 'aborted', content + reasoning_content + tool_calls accumulated so far, x_partial: true,
+#   x_estimated_output_tokens (chars/4 — the provider usage chunk never arrives on an abort), x_error on a throw. Natural end byte-identical.
+#   Unblocks an on-stream loop detector with abort (plan P1/P3). Messages and Responses readers untouched.
 # CORTEX_OPENAI_TOOLS_REASONING — HB-OPENAI-TOOLS-REASONING (DARK, 2026-10-02; stack-portability N1). OpenAI reasoning cards never declare
 # reasoning.toggleable, so the gateway withheld the request/action effort (CORTEX_ACTION_EFFORT was a NOOP on every OpenAI path).
 # on: the gateway forwards the effort for OpenAI reasoning cards → Responses `reasoning.effort` (+ summary auto). Chat/completions is

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@nexus-cortex/cli",
-  "version": "4.124.59",
+  "version": "4.124.61",
   "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": "4.124.59",
+    "@nexus-cortex/core": "4.124.61",
     "chalk": "^4.1.2",
     "cli-spinners": "^2.9.0",
     "commander": "^11.0.0",