@nexus-cortex/cli 4.124.42 → 4.124.44

package/.env.defaults

@@ -147,6 +147,18 @@ DEFAULT_MODEL_ID=deepseek-flash
 # [LEDGER 2026-09-10 @4.102.0] deepseek-flash — DeepSeek catalog collapse: `deepseek-flash` IS V4.1 Flash (vision native, same
 #   price as the retired v4-flash / vision-exp names, which now route here anyway). One helper card, no alias risk.
 HELPER_MODEL_ID=deepseek-flash
+# CORTEX_HELPER_MATCH_PROVIDER — R217 (4.124.44, 2026-10-02). HELPER_MODEL_ID and MENTORSHIP_HELPER_MODEL both default to deepseek-flash, so
+# every action model got DeepSeek as its judge / planner / mentor / wall-summarizer / title writer. on = when a role model is NOT explicitly
+# set, it resolves from the ACTION model's provider via the selectHelperModel map (anthropic→claude-haiku-4-5, openai→gpt-4.1-mini,
+# google→gemini-2.5-flash-lite, xai→grok-4.3, deepseek→deepseek-flash; unmapped providers keep deepseek-flash). Explicit settings always
+# win. "Explicit" is decided by VALUE (.env.defaults loads into process.env, so presence cannot tell): a value equal to the shipped
+# default (deepseek-flash) counts as defaulted — to pin deepseek-flash for a non-DeepSeek action model, leave this off. Routed: every
+# HelperModelMiddleware helper/mentor call (generateGuidance = all mentor surfaces + web summary + wall summary + error guidance; doctrine
+# curation; session title; turn summary/prediction; frame-candidate author). Not routed: compaction / context rejection (already
+# provider-mapped), vision hand-off (VISION_HELPER_MODEL), sub-agent model, orchestrator-side labels. off/unset (default) = byte-identical.
+# [LEDGER 2026-10-02 @4.124.44] DARK — BUILT + unit-tested (helperProviderMatch.test.ts: explicit wins, provider matrix, live model switch,
+# off unchanged). NO bench arm yet. evidence: transport map 10-02 (R217).
+#CORTEX_HELPER_MATCH_PROVIDER=
 
 # Model that backs WebSearch / WebFetch tools. Provider auto-detected from ID prefix.
 # When unset, auto-picks based on available API keys (Gemini > Anthropic > XAI),
@@ -601,6 +613,31 @@ CORTEX_LIFT_PLAN=true
 #CORTEX_LIFT_PLAN_TIMEOUT_MS=90000
 #CORTEX_LIFT_PLAN_INTERACTIVE=false
 #CORTEX_LIFT_PLAN_RECON_TIMEOUT_MS=8000
+# CORTEX_INTERACTIVE_PROFILE — R219 HB-INTERACTIVE-PROFILE (DARK, 2026-10-02). The headless levers behind the TB4.0 gains live only in
+#   the bench cell env, so an interactive TUI (neoncortex / fuzzycortex / `cortex` on a terminal) runs without every DARK one. A named
+#   profile, resolved in ONE place (packages/core/src/config/interactiveProfile.ts, applied by bootstrapEnv), turns a curated set on for
+#   INTERACTIVE sessions only (stdin AND stdout are TTYs — the orchestrator's own test) and only for levers you did not set: it sits
+#   ABOVE this shipped file and BELOW the real env, ./.env, <pkg>/.env and ~/.cortex/.env (.env.local still overrides). Filled keys are
+#   recorded in CORTEX_INTERACTIVE_PROFILE_APPLIED (harness-set) and show source `profile` in GET /health/config. Applied at launch.
+#   off (default; unset = byte-identical) | bench-lite =
+#     CORTEX_LOOP_TOOL_BLOCK=true  CORTEX_EMPTY_TURN_CONTINUE=true  CORTEX_WALL_DROP=on  CORTEX_WALL_SUMMARY=on
+#     CORTEX_REASONING_EXHAUST_BACKOFF=false  CORTEX_STEER_INPUTS=full  CORTEX_ORIENT_V2=1
+#   (failure-path-only or steering-input levers: no deadline assumption, no task-tree writes, no new pause before the first action).
+#   Already ON for every session via this file (not the profile's job): ENDTURN_GATE/REQUIREMENTS/RESOLVER(+ABSTAIN), ORIENT_BOOTSTRAP,
+#   SLICE_NUDGE, MENTORSHIP_ENABLED; ENDTURN_TIER=essential and LIFT_PLAN_DOCTRINE=v1 are code defaults.
+#   NOT in bench-lite — NEEDS-ADAPTATION (follow-ups): the lift plan in interactive sessions (CORTEX_LIFT_PLAN_INTERACTIVE stays the only
+#   opener: silent 30-110 s mentor call after the first tool result with no planning-status event to the TUIs, one-shot per orchestrator
+#   so only the first request of a chat is planned, grader-oriented persona); LIFT_PLAN_SCOPE (repair-task edit scope); INVENTORY (whole-tree
+#   walk through execSync blocks the TUI event loop); TURN_STATUS (deadline clock; auto is a no-op without a deadline); VERIFY_SCALE and
+#   RULE_INPUTS/INTERP/HOLDOUT/INDEP (more EndTurn-judge vetoes against a human who wants to stop); ACTION_EFFORT (the TUIs have their own
+#   effort selector). BENCH-ONLY: WORKSPACE_CLEAN (moves project memory out of the project), ORIENT_BASELINE (repair-task revert baseline),
+#   PID1_GUARD (container PID 1 assumption), PLATEAU_STOP (optimisation-task measurement).
+# [LEDGER 2026-10-02 @4.124.44] DARK — BUILT + unit-tested (interactiveProfile.test.ts 16/16: name resolution, interactive vs headless,
+#   explicit env wins, off byte-identical, bootstrapEnv layering, effective-config `profile` source, lift-plan gate; interactiveLiftPlanGate
+#   .integration.test.ts 6/6 both loops: headless plans, interactive + bench-lite does not, explicit CORTEX_LIFT_PLAN_INTERACTIVE=true
+#   plans). NO interactive field read yet. evidence: R219.
+#CORTEX_INTERACTIVE_PROFILE=
+#CORTEX_INTERACTIVE_PROFILE_APPLIED=
 #   CORTEX_ENDTURN_RESOLVER — the finish-side twin of the lift planner: AFTER the mechanical EndTurn gate
 #   accepts, a bounded max-reasoning mentor adjudicates "does the work product meet the task's real
 #   requirements?" MEETS finishes; GAP vetoes with a fix plan (the narrow-door model executes it instead of
@@ -1025,6 +1062,10 @@ CORTEX_ENDTURN_GATE=true
 # [LEDGER 2026-10-02 @4.124.42] wording fix after review: instruction limited to continuous objectives (never pass/fail or test counts); reminder adds 'if the task's required checks are not yet passing, this is not a stop signal — keep fixing them' (a flat test-count series would otherwise have pushed an early finish).
 # noisy-line parsing, firing cap, lever-off path, STATUS segment; plateauStop.integration.test.ts: both loops, off = no instruction/
 # reminder/segment, on = instruction + 2 reminders + PLATEAU segment). NO bench arm yet. evidence: R210 v39a layout-config-recreation2 (483 iterations, 5.3 h).
+# [LEDGER 2026-10-02 @4.124.43] DARK — BUILT + unit-tested: new mode `status` = boot instruction + metric tracking + STATUS score segment ONLY
+# (segment without the ` — PLATEAU` label); the detector runs on the same cadence (same cap/spacing) but steers NO reminder — each would-be
+# firing is banked as plateau_detected {name, dir, best, improvementPct, points, firing, iteration, streaming} so a run measures when `on`
+# would have fired. on / off unchanged (plateauStop.test.ts mode cases). NO bench arm yet. Values: off | on | status.
 #CORTEX_PLATEAU_STOP=
 # [LEDGER 2026-10-02 @4.124.41] DARK — default 0.002 (0.2%) relative; untested on a bench. evidence: R210 v39a layout (840,636 → 851,547 over 5.3 h; per-window gains not yet measured).
 #CORTEX_PLATEAU_REL=
@@ -1136,6 +1177,31 @@ CORTEX_ENDTURN_GATE=true
 # endturn_resolver gains {verifyScale}. Needs CORTEX_ENDTURN_RESOLVER=true.
 # [LEDGER 2026-10-01 @4.124.40] DARK — BUILT + unit-tested (endTurnResolver.test.ts R208 cases). NO bench arm yet.
 #CORTEX_VERIFY_SCALE=
+# CORTEX_RULE_INPUTS / CORTEX_RULE_INTERP / CORTEX_RULE_HOLDOUT / CORTEX_RULE_INDEP — R211–R214 HB-TASK-RULES (DARK, 2026-10-02). TB4.0
+# requirement-miss pass over the 16 never-passed tasks that fail ONE identical hidden test every run: an input field ignored or a given
+# constant altered to make the problem feasible; an ambiguous convention settled by heuristic while the right alternative sat in the
+# agent's own caveats; passing the visible sample / own oracle while hidden variants fail; producer = verifier with measured discrepancies
+# left unreconciled. Four general rules, each its own flag (off | on) so they can be A/B'd, each with TWO carriers (training/taskRules.ts):
+# A planner — one doctrine bullet per active rule in the lift-plan persona (plannerSystem; one combined block just before the output
+# instruction, after the CORTEX_LIFT_PLAN_SCOPE bullet), so the plan carries the step; B judge — one rubric clause per active rule appended to
+# the EndTurn judge persona after the R208 scale item (resolverSystemPrompt): evidence that does not establish the rule is a GAP with a
+# concrete CHECK. All off → both personas byte-identical. The rules only ask for MORE verification. Events: task_rule_plan {rules,
+# blockChars, planChars, retire} once per session when the delivered plan was written under them; task_rule_judge {surface, rules,
+# clauseChars, roundsUsed, meets, parsed, confidence} per adjudication; endturn_resolver gains {taskRules}. Carrier A needs
+# CORTEX_LIFT_PLAN on; carrier B needs CORTEX_ENDTURN_RESOLVER=true.
+# R211 INPUTS = L1 input-coverage audit: every field/column/parameter of the provided inputs accounted for; given constants never changed.
+# [LEDGER 2026-10-02 @4.124.43] DARK — BUILT + unit-tested (taskRules.test.ts: off = byte-identical planner + judge personas). NO bench arm yet. evidence: tb4-never-passed-2026-10-02 (cargo-flight-dispatch, data-anonymization, live-database-cutover).
+#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).
+#CORTEX_RULE_INTERP=
+# R213 HOLDOUT = L3 held-out / variant validation: when hidden inputs or generalisation are implied, validate beyond the visible sample.
+# [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 (bun-sourcemap-leak, sound-change-cascade, session-window-debug).
+#CORTEX_RULE_HOLDOUT=
+# R214 INDEP = L4 independent verifier: final check must not reuse the producing logic; unreconciled measured discrepancy blocks the finish;
+# run the task's own verification commands/limits.
+# [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 (layout-config-recreation, retro-console-soc, satb-audio-transcription).
+#CORTEX_RULE_INDEP=
 # CORTEX_TURN_CONTRACT — HB-TURN-CONTRACT (4.115.0, 2026-09-17; Terminus-2 disparity "suppress vs channel"). The boot-minimal door SUPPRESSES
 # deliberation ("Prefer acting over deliberating"; tuned at effort low). At effort high the model fought it: tb4-flash-v3 had 58 turns in
 # 23 sessions reason to the 65536 cap and emit nothing (R153). Terminus 2 CHANNELS deliberation: think freely, then every response must
@@ -1147,6 +1213,16 @@ CORTEX_ENDTURN_GATE=true
 # reasoning counts inside max_tokens and starves the answer) — both harnesses hit the same wall; this lever changes the RECOVERY and the FRAME.
 # [LEDGER 2026-09-17 @4.115.0] BUILT + unit-tested (turnContract.test.ts). Field: PENDING (tb4 pilot-12, channel arm at effort max).
 #CORTEX_TURN_CONTRACT=
+# CORTEX_FAMILY_PRESETS — R217 (4.124.44, 2026-10-02). promptPreset 'boot-minimal' is set ONLY on the 5 DeepSeek cards, and it is the only
+# prompt that points the model at the `.cortex/orient` script — Claude/GPT/Grok/Gemini got the full corpus and never the orient pointer, so
+# the ORIENT_* / baseline levers were unreachable for them. on = a card WITHOUT its own promptPreset gets its family preset; today ONE
+# generic preset for every family, 'boot-minimal-generic' = the boot-minimal text minus the DeepSeek-tuned "Prefer acting over
+# deliberating." clause (tuned at DeepSeek effort low; fought at high effort, R153), same minimal mass, same orient pointer, same
+# delegation-hint / turn-contract levers. DeepSeek cards unchanged; env CORTEX_PROMPT_MASS / CORTEX_SYSTEM_PROMPT_FILE still win.
+# off/unset (default) = byte-identical (non-DeepSeek cards keep the full corpus).
+# [LEDGER 2026-10-02 @4.124.44] DARK — BUILT + unit-tested (familyPresets.test.ts: per-family selection on/off, DeepSeek identical on/off,
+# off byte-identical). NO bench arm yet. evidence: transport map 10-02 (R217).
+#CORTEX_FAMILY_PRESETS=
 # CORTEX_FINISH_CONFIRM / CORTEX_FINISH_CONFIRM_MIN_REMAINING / CORTEX_FINISH_CONFIRM_MAX — R167 HB-FINISH-CONFIRM (4.116.0, 2026-09-17,
 # pilot-12 hand-curation). The early-finish class is the whole game on the confident-wrong tail: 8/12 first finish attempts per arm
 # had >= 70% of the wall budget left (Terminus 2 too: 8/12 confident-wrong finishes in 15–105 min, its bare "are you sure" changed
@@ -1724,6 +1800,40 @@ MENTORSHIP_ACTIVE_DISCOVERY=false
 # Set to false to opt out (restores the pre-R63 promptless behavior).
 #CORTEX_DELIVER_SYSTEM_PROMPT=false
 
+# CORTEX_RESPONSES_INLINE_REMINDERS — R215 HB-RESPONSES-REMINDERS (DARK, 2026-10-02). The Responses path cut every
+# `<system-reminder>` block out of the conversation and merged it into `instructions`; xAI Responses never receives
+# `instructions` → those reminders were DROPPED there, and on OpenAI they lost their place after the tool output. Levers
+# delivered this way: lift plan, requirement ledger, PID1 note, plateau instruction, lift nudge, deferred-corpus header.
+# on: reminders stay IN PLACE as user input_text items (tags unchanged); a user item the adapter emitted before its
+# function_call_output items is moved after them (tool output first, then the text). `instructions` then carries only
+# the R63 static system prompt (xAI: system-role item at chain start, unchanged). OpenAI + xAI Responses. off: unchanged.
+# [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).
+#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
+# (CORTEX_WALL_DROP / CORTEX_WALL_SUMMARY / CORTEX_REASONING_EXHAUST_BACKOFF) and the truncated-continue nudge never fired
+# there. on: Responses status 'incomplete' + incomplete_details.reason 'max_output_tokens' → 'max_output_tokens' (other
+# incomplete reasons passed through; else the status, e.g. 'completed'); the streaming finalMessage carries the terminal
+# event's status + incomplete_details; hf-space → its finish_reason (synthesized 'stop' / 'tool_calls' — the Space exposes
+# no truncation signal, so hf-space truncation stays invisible). off: unchanged.
+# [LEDGER 2026-10-02 @4.124.44] DARK — BUILT + unit-tested (transportFixes.test.ts: live shapes → classifier truncated +
+# 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).
+#CORTEX_RESPONSES_STOP_REASON=
+# 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
+# text) of the previous user message (stable prefix) and the newest user message, within Anthropic's 4-breakpoint cap
+# (system + tool + existing markers counted; one free slot → the stable prefix only). Anthropic provider only — the xAI
+# Messages branch is untouched. off: unchanged.
+# [LEDGER 2026-10-02 @4.124.44] DARK — BUILT + unit-tested (transportFixes.test.ts: placement, cap <= 4, no mutation,
+# send + stream, xAI untouched). Live probe: DeepSeek /anthropic accepted message-level cache_control (200, cache
+# fields 0 — it ignores it). No live Claude probe. evidence: transport map 10-02 (R215/R216/R218).
+#CORTEX_ANTHROPIC_HISTORY_CACHE=
+
 # Prompt-mass composition (BASH_PLUS_SPEC P6): full (default) | minimal (only
 # the core system prompt; guides/CLAUDE.md/memory dropped) | defer (minimal on
 # turn 1, full corpus appended once at the first tool-result boundary).

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@nexus-cortex/cli",
-  "version": "4.124.42",
+  "version": "4.124.44",
   "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": "4.124.42",
+    "@nexus-cortex/core": "4.124.44",
     "chalk": "^4.1.2",
     "cli-spinners": "^2.9.0",
     "commander": "^11.0.0",