@nexus-cortex/cli 4.124.44 → 4.124.46

package/.env.defaults

@@ -1833,6 +1833,150 @@ MENTORSHIP_ACTIVE_DISCOVERY=false
 # 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=
+# APPEND-ONLY REQUEST HISTORY (prompt-cache restructure, 2026-10-02; .cortex/research/caching-audit-2026-10-02.md +
+# caching-rationale-2026-10-02.md). All DARK; unset = byte-identical (harness: every flag-off cell == the committed 220-cell matrix).
+# Measured with the prefix-stability harness (prefixStability.integration.test.ts, PREFIX_EXTRA_ENV / PREFIX_OUT_TAG A/B; reports
+# .cortex/research/prefix-stability-2026-10-02.<tag>.partial.md). The anchor lift (one intended tools bust at the first tool_result) is
+# untouched by all of them.
+# CORTEX_WALL_CACHE_FIX — R221. The HB-WALL-DROP nudge (+ wall summary) is appended to the newest tool_result AFTER that message's
+# canonical conversion was cached (noteRequestUsage runs while the walled turn is the tail), so the retry carried the nudge and every
+# later request silently reverted to the stale copy: a second prefix break + the nudge/summary vanished. on: every in-place mutation of a
+# history message (wall-drop carrier, lift nudge/plan/corpus/ledger, PID1/plateau notes, budget/ladder signals, TURN_STATUS, compute
+# nudge, R26 empty-turn repair, R154 image stub) drops that message's cached conversion. The wall drop itself is unchanged.
+# [LEDGER 2026-10-02 @4.124.45] DARK — BUILT + harness-proven: send c1/c2/c3 on anthropic, xai-messages(server-tools off), chat-deepseek
+# went `1→2 tail; 2→3 F1` → `1→2 tail` only (the single intended break at the retry remains); nudge-on-wire N..... → NNNNNN, wall
+# summary persists. Flag off unchanged. Unit: appendOnlyHistory.test.ts, wallGuard.test.ts 11/11, wallGuard.integration 6/6 (+4 skip).
+#CORTEX_WALL_CACHE_FIX=
+# CORTEX_PERSIST_INJECTED — R223. The first request of a user turn sends injected content (repo-state note, deferred-tool announcement,
+# MCP / slash-command notes); history stored the raw text, so every continuation re-rendered that message WITHOUT the notes (one prefix
+# split per user turn + the notes vanished). on: the exact injected bytes are remembered per message uuid and re-sent on every request;
+# session storage, display, title and every text reader keep the raw content (no system-reminder/harness-note text persisted).
+# [LEDGER 2026-10-02 @4.124.45] DARK — BUILT + harness-proven: scenario f (2 user turns) on anthropic, xai-messages(server-tools off),
+# chat-openai, chat-deepseek, gemini x2, hf-space, send + stream: `0→1 lift+F3; 6→7 F3` → `0→1 lift` (harness-note now on every request
+# 0-8). Responses rows unchanged (chaining hides it). Flag off unchanged.
+#CORTEX_PERSIST_INJECTED=
+# CORTEX_APPEND_ONLY_TOOLS — R222. With deferred loading (R43, kept) the filter re-ran on every continuation in allTools order: a newly
+# discovered/used tool was inserted mid-array and the 15-slot recent list evicted. on: the tools array only grows — new tools appended in
+# first-enable order, never evicted or reordered within the orchestrator's lifetime.
+# [LEDGER 2026-10-02 @4.124.45] DARK — BUILT + harness-measured: scenario b SearchTools discovery `+8 tools (inserted mid-array at 11)`
+# → `(appended)`, no reorder; eviction not exercised by the harness (unit-tested). NOT a cache win by itself: tools precede system +
+# messages on every wire, so any growth (17→25) is still a full prefix bust once per discovery (harness still reports `2→3 F2`); the
+# flag only removes the reorder/eviction churn. A real fix = a static tools array (cost: R43's token win). Flag off unchanged.
+#CORTEX_APPEND_ONLY_TOOLS=
+# CORTEX_KEEP_MENTOR_MESSAGES — R227. Mentor/guidance messages (injectGuidance / reactive mentorship) were removed from MID-history at
+# turn end (cleanupEphemeralMessages), rewriting the next turn's prefix and shifting the Responses slice checkpoint. on: kept (never
+# persisted to the session file; display strips their system-reminder). Cost: stale guidance stays in context — to be judged in the TUI work.
+# [LEDGER 2026-10-02 @4.124.45] DARK (default OFF by operator decision) — BUILT + harness-proven: scenario e on all non-Responses rows,
+# send + stream: `6→7 F9` gone (mentor insight on the wire in requests 2-9 instead of 2-6). Flag off unchanged.
+#CORTEX_KEEP_MENTOR_MESSAGES=
+# CORTEX_RESPONSES_SLICE_ALL — R228 (bug fix). The EndTurn gate, wall/empty retries, inaction retries and tools-off synthesis set
+# previous_response_id but sent the FULL history (contradicting the e85b68834f slicing design), and never moved the slice checkpoint.
+# on: every chained request sends only the items added since the held response; each of those responses moves the checkpoint (they are
+# main-chain turns: their reply is pushed into history, so updating lastResponseId is correct — not a branch); a wall on a chained session
+# is not dropped locally (the server holds it). Fallback: if the provider reports the held response unavailable (not found / expired /
+# invalid; store off) the chain is dropped and the request is resent ONCE with the full history and NO previous_response_id
+# (responses_chain_fallback event) — send-loop initial request, continuations, gate, retries, synthesis. Streaming initial/continuation
+# (streamRequest) have no fallback yet.
+# [LEDGER 2026-10-02 @4.124.45] DARK — BUILT + harness-proven: scenarios a/c1/c3/h on responses-openai, responses-xai, xai default:
+# send a `4→5 NEW-R; 5→6 NEW-R` → one flag left, c1 `1→2, 2→3, 5→6, 6→7 NEW-R` → one flag left; the remaining flag is a harness
+# false positive (two gate attempts push byte-identical reminder text; dump shows each request sends exactly one new item). Fallback
+# unit-tested via isPreviousResponseUnavailable only — no live provider probe. Flag off unchanged.
+#CORTEX_RESPONSES_SLICE_ALL=
+# CORTEX_GEMINI_SYSTEM_FIX — R229 HB-GEMINI-SYSTEM (DARK, 2026-10-02; CORRECTNESS). Gemini never received the system
+# prompt: the non-streaming REST generateContent body had no `system_instruction`; the non-tool requests went through the
+# @google/generative-ai SDK (v0.2.1, v1 endpoint — no systemInstruction support); the google-sdk (@google/genai) paths set a
+# TOP-LEVEL `systemInstruction`, which that SDK drops (GenerateContentParameters = {model, contents, config}). on: REST body
+# carries system_instruction {parts:[{text}]}; non-tool requests with a system prompt use the v1beta REST builders (send +
+# stream); google-sdk puts it in config.systemInstruction (send + stream). The streaming REST builder always sent it. off: unchanged.
+# [LEDGER 2026-10-02 @4.124.45] DARK — BUILT + unit-tested (transportCacheFixes.test.ts: 5 Gemini wire shapes, off =
+# byte-identical) + prefix harness (send a, gemini-generateContent + google-sdk: system prompt none → present on all 7
+# requests, byte-stable, pair verdicts unchanged). No live Gemini probe. Default-on candidate (correctness).
+#CORTEX_GEMINI_SYSTEM_FIX=
+# CORTEX_STABLE_TOOL_ORDER — R230 HB-STABLE-TOOL-ORDER (DARK, 2026-10-02). xAI cards with ENABLE_SERVER_SIDE_TOOLS=true run
+# hybrid server+client tools on /v1/responses; the first request of every user turn put the server tools FIRST while the
+# continuations used the original order → the tools array (prefix head) was reordered twice per user turn. on: the caller's
+# original order on every request (adapters/ServerSideToolDetection.ts hybrid branch). off: unchanged.
+# [LEDGER 2026-10-02 @4.124.45] DARK — BUILT + unit-tested + prefix harness (send e/f/h, xai default→responses: the per-turn
+# "19→19 tools REORDERED" divergences 2 per cell → 0; the anchor-lift 0→1 bust stays, intended). No live xAI probe.
+#CORTEX_STABLE_TOOL_ORDER=
+# CORTEX_OPENAI_CACHE_KEY — R225 HB-OPENAI-CACHE-KEY (DARK, 2026-10-02). OpenAI chat/completions + Responses bodies carry
+# `prompt_cache_key` = the session id (what xAI Responses already gets) so same-session requests route to the same prompt
+# cache. OpenAI provider only (other chat/completions providers may reject the field). off: unchanged.
+# [LEDGER 2026-10-02 @4.124.45] DARK — BUILT + unit-tested + prefix harness (send a: key on 7/7 chat-openai and
+# responses-openai requests, one value per session; chat-deepseek never; pair verdicts unchanged). No live OpenAI probe.
+#CORTEX_OPENAI_CACHE_KEY=
+# CORTEX_ANTHROPIC_AUTO_CACHE — R225 HB-ANTHROPIC-AUTO-CACHE (DARK, 2026-10-02). Anthropic Messages adds the top-level
+# automatic `cache_control` (breakpoint moves with the conversation) beside the explicit system + last-tool markers — only
+# while a breakpoint slot is free (cap 4; R218 CORTEX_ANTHROPIC_HISTORY_CACHE markers counted, so with R218 on it is
+# usually skipped). Needs ANTHROPIC_PROMPT_CACHING. Anthropic provider only — never the xAI Messages branch. off: unchanged.
+# The legacy `prompt-caching-2024-07-31` beta header is still sent (left as is).
+# [LEDGER 2026-10-02 @4.124.45] DARK — BUILT + unit-tested + prefix harness (send a/b anthropic: 2 → 3 markers incl.
+# automatic; with R218 on: 4 max, automatic only on request 0). No live Claude probe.
+#CORTEX_ANTHROPIC_AUTO_CACHE=
+# CORTEX_ANTHROPIC_CACHE_TTL — R225 (DARK, 2026-10-02). 1h: every cache_control the harness emits on the Anthropic path
+# (system, last tool, R218 history, automatic) carries ttl "1h" (uniform, so Anthropic's longer-TTL-first and
+# automatic-vs-last-marker TTL rules hold). Writes cost 2x instead of 1.25x; worth it only when requests sharing the prefix
+# start more than 5 minutes apart (long tool runs). unset | 5m = default 5-minute TTL (no ttl field). Anthropic only.
+# [LEDGER 2026-10-02 @4.124.45] DARK — BUILT + unit-tested + prefix harness (all markers ttl 1h, cap <= 4). No live probe.
+#CORTEX_ANTHROPIC_CACHE_TTL=
+# CORTEX_FORCED_CHOICE_FULL_TOOLS — R225 HB-FORCED-CHOICE-FULL-TOOLS (DARK, 2026-10-02). A forced named tool_choice (mentor
+# AskForAdvice, CORTEX_MENTOR_FORCE) narrowed the tools array to the forced tool, rewriting the tools prefix on the forced
+# request and again after it. on: the full tools array is kept; the provider's tool_choice does the forcing. CAUTION:
+# DeepSeek was observed (2026-08-29) to ignore a named choice when the full catalog is present (the reason narrowing exists),
+# and hf-space has no tool_choice on the Gradio wire, so there the force is lost entirely. off: unchanged.
+# [LEDGER 2026-10-02 @4.124.45] DARK — BUILT + unit-tested + prefix harness (send d, all transports: the forced pair's first
+# divergence moves from tools (kept 0%) to tool_choice (Anthropic/xAI-messages kept 97%, chat-deepseek 92%, chat-openai 47%,
+# Gemini 92-95%; Responses chained 0%); hf-space stable = force lost). No live probe of choice obedience.
+#CORTEX_FORCED_CHOICE_FULL_TOOLS=
+# CORTEX_STREAM_USAGE — R224a HB-STREAM-USAGE (DARK, 2026-10-02). Streaming chat/completions never asked for usage, so the
+# reassembled final message reported zero tokens (TUI/SSE accounting and cache hit rates invisible). on: streaming requests
+# send stream_options {include_usage: true} and the final message carries the provider's usage object (DeepSeek
+# prompt_cache_hit_tokens, OpenAI prompt_tokens_details). Non-streaming requests unchanged. off: unchanged.
+# [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.
+#CORTEX_STREAM_USAGE=
+# 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
+# UNCHANGED: R19b's drop stays — live probe 2026-10-03, gpt-5.4-nano + gpt-5.6-luna: chat reasoning_effort + function tools = 400
+# "use /v1/responses"; Responses + tools + reasoning.effort = 200 on both. So an OpenAI effort arm must use OPENAI_API_MODE=responses.
+# Values pass through verbatim. DeepSeek and every non-OpenAI request byte-identical. off: unchanged.
+# [LEDGER 2026-10-03 @4.124.46] DARK — BUILT; chat half REMOVED after the live probe (would have 400ed every OpenAI tool call);
+# Responses half live-verified (accepts effort with tools).
+#CORTEX_OPENAI_TOOLS_REASONING=
+# CORTEX_ANTHROPIC_EFFORT — HB-ANTHROPIC-EFFORT (DARK, 2026-10-02; stack-portability N1). Adaptive-thinking families (claude-4.7,
+# claude-4.8, claude-5, claude-fable-5) got thinking {type:adaptive} only — the action effort never reached the wire; budget
+# families mapped effort `max` to the 10000 fallback (not in the budget map). on: adaptive families send
+# `output_config: {effort}` = low|medium|high|xhigh|max (GA field, no beta header; none/unset → omitted = provider default);
+# budget families map `max` to budget_tokens = min(120000, max_tokens - 1024) (must stay < max_tokens and >= 1024; below
+# that the legacy 10000 stays). low/medium/high budgets unchanged. Send + stream paths. xAI Messages untouched. An effort
+# change mid-session invalidates Anthropic's messages cache (same as the existing budget ladder). off: unchanged.
+# [LEDGER 2026-10-02 @4.124.46] DARK — BUILT + unit-tested (adaptive + budget, send + stream, off = byte-identical) + prefix
+# harness (send a, ACTION_EFFORT=max: claude-sonnet-5 output_config none → {effort:max} on 6/7 main requests; sonnet-4.6 and
+# haiku-4.5 budget_tokens 10000 → 62976 (max_tokens 64000) on 6/7; pair verdicts unchanged). No live Claude probe.
+#CORTEX_ANTHROPIC_EFFORT=
+# CORTEX_GEMINI_TOOLS_THINKING — HB-GEMINI-TOOLS-THINKING (DARK, 2026-10-02; stack-portability N1). The @google/genai builders
+# skipped thinkingConfig whenever tools were attached ("cannot coexist", a Gemini 2.0-era observation) and the REST builder only
+# sent a card-level thinkingLevel, so CORTEX_ACTION_EFFORT never reached Gemini. on: the gateway maps the request/action effort
+# to generationConfig.thinkingConfig on REST + SDK, with or without tools — Gemini 3.x thinkingLevel (low|medium|high;
+# xhigh/max → high), 2.5 and older thinkingBudget (low 2048, medium 8192, high 16384, xhigh/max 24576); the request effort
+# replaces a card-level thinkingLevel. SDK + tools with no effort → thinkingBudget 10000 (the no-tools default). Streaming SDK
+# adds includeThoughts; non-streaming does not (the message-level adapter would render thought parts as visible text). Visible
+# reasoning (wall levers) is NOT enabled by this flag. off: unchanged.
+# [LEDGER 2026-10-02 @4.124.46] DARK — BUILT + unit-tested (REST 3.x level + 2.5 budget, SDK send + stream, off =
+# byte-identical) + prefix harness (send a, ACTION_EFFORT=max: gemini-3.7-flash REST thinkingLevel "high", gemini-2.5-flash
+# REST + SDK thinkingBudget 24576 on 6/7 main requests; pair verdicts unchanged). No live Gemini probe of thinking + tools.
+#CORTEX_GEMINI_TOOLS_THINKING=
+#
+# CORTEX_ACTION_EFFORT_FIRST — HB-ACTION-EFFORT-FIRST (DARK, 2026-10-03; found by the effort-gap build). CORTEX_ACTION_EFFORT is filled
+# into the request params AFTER request 0 of each user turn is built, so request 0 ran at the model card's default effort and requests 1+
+# at the action effort. DeepSeek: harmless (card 'medium' = 'high' server-side; the anchor-lift bust is on the same boundary). Other
+# providers: request 0 runs at the wrong effort and the effort flip costs one extra cache break per user turn (Anthropic output_config,
+# Gemini thinkingConfig, OpenAI reasoning_effort are all prefix-relevant). on = request 0 also gets CORTEX_ACTION_EFFORT (a ramp or an
+# explicit request param still wins). off = unchanged (byte-identical).
+# [LEDGER 2026-10-03 @4.124.46] DARK — BUILT (unit tests; turn it on in every non-DeepSeek cross-provider arm).
+#CORTEX_ACTION_EFFORT_FIRST=
 
 # Prompt-mass composition (BASH_PLUS_SPEC P6): full (default) | minimal (only
 # the core system prompt; guides/CLAUDE.md/memory dropped) | defer (minimal on

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@nexus-cortex/cli",
-  "version": "4.124.44",
+  "version": "4.124.46",
   "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": "4.124.44",
+    "@nexus-cortex/core": "4.124.46",
     "chalk": "^4.1.2",
     "cli-spinners": "^2.9.0",
     "commander": "^11.0.0",