@nexus-cortex/cli 4.124.61 → 4.124.63

package/.env.defaults

@@ -1050,6 +1050,9 @@ CORTEX_ENDTURN_GATE=true
 #   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).
+# [LEDGER 2026-10-06 cell 2] "caps alone" control t4e (bundle, new caps, planner fully fixed): 3/14 vs c23 8/14 — 12 walls, 31 accept-with-gap.
+#   The new ceiling does NOT lift the plain bundle (it feeds loops); only Arm B (rules + scope + Flash judge) holds 7–8/14 on it, and only the
+#   ladder makes it affordable. Pin confound unresolved (no old-cap control at the new pin). r-tb4-caps2-2026-10-06.md.
 #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
@@ -1064,8 +1067,37 @@ CORTEX_ENDTURN_GATE=true
 #   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).
+# [LEDGER 2026-10-06 @4.124.60 first arm] t4g (Arm B + clock + ladder 65536,16384 / force at 2) vs t4f (same without ladder): 6/14 vs 7/14
+#   (undetermined), walls 9 → 4, model $14.23 → $5.39 (−62%); 4 ladder events (2 cap-only, 2 forced thinking-off tool calls). Mechanism proven
+#   on the bench; pass cost not established at n=1. Ledger r-tb4-caps2-2026-10-06.md.
 #CORTEX_WALL_RETRY_LADDER=
 #CORTEX_WALL_FORCE_ACTION_AT=2
+# CORTEX_SECOND_ATTEMPT / _MAX / _MIN_REMAINING / _TRIGGER / _RESERVE_MS / _FLOOR_MS / _OUT_DIR — HB-SECOND-ATTEMPT (R239 / unification plan
+# P4, 2026-10-07, DARK). The R194 second-attempt chain, moved from the bench adapter into the LIBRARY (operator 10-06: "the library IS the
+# harness; the adapter must never decide"). After a finish whose LAST endturn_resolver action is in TRIGGER (default accept-with-gap,
+# accept-low-confidence; add accept to retry confident finishes; add none to retry a give-up = a turn with no verdict), while >= MIN_REMAINING
+# (0.5) of the ORIGINAL deadline remains on the wall clock and remaining − RESERVE (5 min) >= FLOOR (15 min), and fewer than MAX (1) extra
+# attempts ran: the workspace is restored to its task-start snapshot (shadow git outside the tree, vendor dirs + files > 20 MB excluded) and a
+# FRESH orchestrator (own decisions.jsonl via OrchestratorConfig.stateDir, own sessions dir, deadline = remaining − reserve) runs the same task.
+# Selection over every attempt's tree: the workspace's shipped check (make test / test.sh / run_tests.sh / pytest) when it separates them >
+# verdict (accept > gap/low-confidence > no finish) > FEWEST tool-call turns (r-selector 09-28; promoted above the judge's count 10-07) > more
+# judge checks passed > earliest. The pick's tree is restored; the response is the pick with usage.session SUMMED over every attempt (deep:
+# helper + sub-agent buckets too — the adapter summed top-level numbers only) + usage/metadata.secondAttempt. Artifacts in OUT_DIR (the bench
+# passes /logs/agent): a<k>.json, response/decisions/session.attempt<k>.*, check<k>, second-attempt.json. Events: second_attempt
+# {phase: decision|selection}. Non-streaming /v1/messages only (streaming = P2/P5). Every failure leaves attempt 1 in place.
+# Evidence (adapter-side chain, same algorithm): r-tb4-judge-replay-2026-09-24 (75% of failures leave >= 50% budget unused), s2a/s2b, t4r/t4s
+# (chain opened 3), t4h (widened trigger, IN FLIGHT 10-07). Library A/B vs the adapter chain on the 14 c23 tasks is owed BEFORE the adapter
+# copy is deleted (plan :40). Adapter: forwards these keys; its own chain stands down when CORTEX_SECOND_ATTEMPT is set.
+# [LEDGER 2026-10-07 @4.124.63] DARK — BUILT + unit-tested (attemptController.test.ts 21/21 = the adapter's test tables ported 1:1,
+#   workspaceSnapshot.test.ts, usageAccounting sumSessionUsage, lever registration test) + loop-integration test (two scripted attempts).
+#   No arm yet.
+#CORTEX_SECOND_ATTEMPT=
+#CORTEX_SECOND_ATTEMPT_MAX=1
+#CORTEX_SECOND_ATTEMPT_MIN_REMAINING=0.5
+#CORTEX_SECOND_ATTEMPT_TRIGGER=accept-with-gap,accept-low-confidence
+#CORTEX_SECOND_ATTEMPT_RESERVE_MS=300000
+#CORTEX_SECOND_ATTEMPT_FLOOR_MS=900000
+#CORTEX_SECOND_ATTEMPT_OUT_DIR=
 # 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.
@@ -2100,6 +2132,9 @@ MENTORSHIP_ACTIVE_DISCOVERY=false
 # [LEDGER 2026-10-02 @4.124.45] DARK — BUILT + unit-tested (request shape + usage carried). Live probe: DeepSeek streaming
 # 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.
+# [LEDGER 2026-10-06 @4.124.62 R236] DEFAULT FLIPPED, provider-aware: unset = ON for openai | deepseek (live-probed 10-02) | groq, OFF for every
+#   other provider (they may reject the field); explicit on forces any provider; explicit off disables everywhere. Streaming usage had been
+#   zeros by default — the session ledger, lastUsageAnchor and the exhaustion token counts were wrong on every stream (parity map row 4).
 #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,

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@nexus-cortex/cli",
-  "version": "4.124.61",
+  "version": "4.124.63",
   "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": "4.124.61",
+    "@nexus-cortex/core": "4.124.63",
     "chalk": "^4.1.2",
     "cli-spinners": "^2.9.0",
     "commander": "^11.0.0",