release: 5.8.10 — live model-behavior verification; fix a false-failing memory-clear test (five packages lockstep)

Ran a representative LlmVerification PlayMode subset against a live OpenAI-compatible
endpoint (LM Studio, reference model qwen3.5-4b-mtp) to audit tool-call/skill/model behavior.

Fixed:
- AllToolCalls_MemoryTool_WriteAppendClear asserted the clear tool REMOVED the store entry
  (!store.TryLoad), but the memory tool's clear action EMPTIES the document (Change("")) and
  keeps the record by design ("clear empties memory"). The run failed even though
  HasCompletedMemoryAction("clear") already proved the model emitted and completed the real
  memory(action=clear) call and the document was empty. It now asserts "no entry OR empty
  content", matching the documented semantics; the misleading "responded with text instead"
  warning is corrected.

Verified (live, qwen3.5-4b-mtp): tool-calling, custom agents (ToolsOnly/ToolsAndChat/ChatOnly/
WithAction), skill self-service (read-then-use), skill-tool proxy, skill tool discovery, memory
write/append/clear, and the execute_lua Lua-authoring pipeline (model writes correct
sandbox-scoped Lua) — all pass. Verdict: the tool-call/skill design is sound; a 4B model handles
the whole surface, and the tool contract explicitly guards narration-instead-of-action. (Z.AI had
no balance and no spark/opencode OpenAI endpoint was available, so the documented local reference
model was used.)

Author: NeoXider

Assets/CoreAI/CHANGELOG.md

@@ -2,6 +2,28 @@
 
 ## [Unreleased]
 
+## 5.8.10 - Live model-behavior verification; fix a false-failing memory-clear test (2026-07-13)
+
+### Fixed
+
+- **`AllToolCalls_MemoryTool_WriteAppendClear` no longer fails when the model clears memory correctly.** The
+  test asserted the `clear` tool REMOVED the store entry (`!store.TryLoad(...)`), but the memory tool's
+  `clear` action EMPTIES the document (`MemoryMutationPlan.Change("")`) and keeps the record by design
+  ("clear empties memory"). So the run failed even though `HasCompletedMemoryAction("clear")` already proved
+  the model emitted and completed the real `memory(action=clear)` call and the document was empty. It now
+  asserts "no entry OR empty content", matching the documented clear semantics; the misleading "model
+  responded with text instead" warning is corrected (the model DID call the tool).
+
+### Verified (live model behavior — LM Studio, reference model `qwen3.5-4b-mtp`)
+
+- Ran a representative subset of the LlmVerification PlayMode suite against a live OpenAI-compatible endpoint:
+  tool-calling, custom agents (ToolsOnly / ToolsAndChat / ChatOnly / WithAction), skill self-service
+  (read-then-use), skill-tool proxy, skill tool discovery, memory write/append/clear, and the `execute_lua`
+  Lua-authoring pipeline (the model writes correct sandbox-scoped Lua) — all pass. The tool-call/skill design
+  is sound: a 4B model handles the full surface correctly, and the tool contract explicitly guards against
+  narration-instead-of-action. (Z.AI had no balance and no spark/opencode OpenAI endpoint was available, so
+  the documented local benchmark reference model was used.)
+
 ## 5.8.9 - Demo/benchmark review wave: per-scene gameplay-binding seam + honest fixes (2026-07-13)
 
 ### Added

Assets/CoreAI/package.json

@@ -1,6 +1,6 @@
 {
   "name": "com.neoxider.coreai",
-  "version": "5.8.9",
+  "version": "5.8.10",
   "displayName": "CoreAI",
   "unity": "6000.0",
   "description": "Portable C# core for CoreAI \u2014 LLM agents that call your game code. Orchestration, function-calling tools, agent memory, AgentBuilder, SkillSet, and production resilience (retries, timeouts, rate limits), with no UnityEngine dependency. LLM and Lua are optional modules (COREAI_NO_LLM / COREAI_NO_LUA), so the core compiles without them \u2014 install only what your game uses. Unity integration ships in com.neoxider.coreaiunity.",

Assets/CoreAIBenchmark/package.json

@@ -1,6 +1,6 @@
 {
   "name": "com.neoxider.coreaibenchmark",
-  "version": "5.8.9",
+  "version": "5.8.10",
   "displayName": "CoreAI Benchmark",
   "unity": "6000.0",
   "description": "LLM game-creation benchmark harness and scenarios for CoreAI: drives execute_lua and world_command through real PlayMode scenarios (G1-G8), grades tool correctness, reasoning, instruction adherence, and determinism, and reports suite/role fitness scores. Depends on com.neoxider.coreai / com.neoxider.coreaiunity / com.neoxider.coreaimods.",

Assets/CoreAIHub/package.json

@@ -1,6 +1,6 @@
 {
   "name": "com.neoxider.coreaihub",
-  "version": "5.8.9",
+  "version": "5.8.10",
   "displayName": "CoreAI Hub (UI Toolkit)",
   "unity": "6000.0",
   "description": "Optional UI Toolkit window for CoreAI: a tabbed Hub that renders pages registered into CoreAI's HubPageRegistry (Chat, Settings, Statistics, Mods, and C#/Lua-authored pages). Lazy per-tab content, live tab rebuild on register/unregister, semi-transparent runtime panel, and a future path to Unity 6.5 world-space PanelRenderer. Install this package only when you want the built-in Hub window; without it CoreAI still exposes the registry so you can render pages on your own uGUI/UITK canvas via the API. Depends on com.neoxider.coreai.",

Assets/CoreAIMods/package.json

@@ -1,6 +1,6 @@
 {
   "name": "com.neoxider.coreaimods",
-  "version": "5.8.9",
+  "version": "5.8.10",
   "displayName": "CoreAI Mods (Lua)",
   "unity": "6000.0",
   "description": "Lua modding layer for CoreAI: the Lua-CSharp sandbox (managed, AOT-safe, WebGL/IL2CPP-ready — bundled, no external package), execute_lua and manage_mods tools, the Lua Modding skill, the mod runtime, persistent mod stores, gameplay bindings, and mod UI. Depends on com.neoxider.coreai / com.neoxider.coreaiunity. Install this package only when you want AI/player-authored Lua mods; without it CoreAI compiles Lua-free.",

Assets/CoreAiUnity/Tests/PlayMode/LlmVerification/AllToolCallsPlayModeTests.cs

@@ -306,24 +306,27 @@ public IEnumerator AllToolCalls_MemoryTool_WriteAppendClear()
                             "Runtime clear/tool-result behavior is covered by deterministic tests; this is model/backend compliance.");
                     }
 
-                    //  :      tool call'
-                    bool memoryCleared = !store.TryLoad(BuiltInAgentRoleIds.CoreMechanic, out _);
+                    // WHY: the memory tool's `clear` action EMPTIES the document (MemoryMutationPlan.Change(""))
+                    // — it deliberately keeps the store record with Memory == "", it does not delete the entry.
+                    // So "cleared" means "no entry OR empty content", not "entry absent" (the old check wrongly
+                    // failed the run even though HasCompletedMemoryAction("clear") above already proved the model
+                    // emitted and completed the real tool call).
+                    bool memoryCleared =
+                        !store.TryLoad(BuiltInAgentRoleIds.CoreMechanic, out AgentMemoryState clearedState)
+                        || string.IsNullOrEmpty(clearedState?.Memory);
 
                     if (!memoryCleared)
                     {
-                        string currentMemory = store.TryLoad(BuiltInAgentRoleIds.CoreMechanic, out AgentMemoryState s)
-                            ? s.Memory
-                            : "(none)";
-                        Debug.LogWarning($"[AllToolCalls]  CLEAR FAILED: Memory NOT cleared by tool call. " +
-                                         $"Current memory: '{currentMemory}'. Model responded with text instead.");
+                        Debug.LogWarning($"[AllToolCalls]  CLEAR FAILED: memory(action=clear) completed but the " +
+                                         $"document is not empty. Current memory: '{clearedState?.Memory}'.");
                     }
                     else
                     {
                         Debug.Log($"[AllToolCalls]  Memory cleared by tool call");
                     }
 
                     Assert.IsTrue(memoryCleared,
-                        "Memory must be cleared by actual tool call. Memory store should be empty.");
+                        "Memory must be emptied by the clear tool call (empty document, entry may remain).");
                 }
 
                 Debug.Log("[AllToolCalls]  MEMORY TOOL TEST PASSED ");

Assets/CoreAiUnity/package.json

@@ -1,6 +1,6 @@
 {
   "name": "com.neoxider.coreaiunity",
-  "version": "5.8.9",
+  "version": "5.8.10",
   "displayName": "CoreAI Unity",
   "unity": "6000.0",
   "description": "Unity framework for LLM-powered NPCs and agents that call your game code \u2014 drop-in chat UI, function-calling tools, persistent memory, token-by-token streaming, and per-role LLM routing. Runs on a local GGUF model (via LLMUnity) or any OpenAI-compatible API; proven on a 4 GB local model. Unity layer for com.neoxider.coreai.",

TODO.md

@@ -2,10 +2,11 @@
 
 > Updated 2026-07-13. Tracks open work by priority. Shipped work is in `CHANGELOG.md` (both packages);
 > non-blocking future work in `Assets/CoreAiUnity/Docs/BACKLOG.md`.
-> Released: 5.8.9 (2026-07-13, all five packages in lockstep). Last full verification 2026-07-13 (batchmode):
-> mods EditMode 119 passed / 0 failed; full EditMode green; PlayMode `FastNoLlm` 56 passed / 0 failed (with
-> graphics — `-nographics` fails 5 GPU/RenderTexture camera tests, environment artifact). LLM-API PlayMode
-> tests run against a local LM Studio OpenAI endpoint (`qwen3.5-4b-mtp`) — see the LLM-API gate below.
+> Released: 5.8.10 (2026-07-13, all five packages in lockstep). Last full verification 2026-07-13 (batchmode):
+> full EditMode 1573 passed / 0 failed; PlayMode `FastNoLlm` 56 passed / 0 failed (with graphics —
+> `-nographics` fails 5 GPU/RenderTexture camera tests, an environment artifact); a live LlmVerification
+> subset is green on the LM Studio `qwen3.5-4b-mtp` OpenAI endpoint (tools / skills / agents / memory /
+> execute_lua) — see the LLM-API gate below.
 
 ## [A6] Deep audit wave (2026-07-13) — runtime / architecture / tests / security, all 22 findings fixed
 
@@ -109,6 +110,14 @@
       reflection that can no-op under IL2CPP stripping (degrades gracefully).
 - [x] Representative local 4B live checks: memory write and real `world_command` spawn pass on
       `qwen3.5-4b-mtp`.
+- [x] **Model-behavior verification (5.8.10, LM Studio `qwen3.5-4b-mtp` OpenAI endpoint).** Ran a
+      representative LlmVerification PlayMode subset live: tool-calling, custom agents (all 4 modes), skill
+      self-service (read-then-use), skill-tool proxy, skill tool discovery, memory write/append/clear, and the
+      `execute_lua` Lua-authoring pipeline (model writes correct sandbox-scoped Lua) — all pass. Verdict: the
+      tool-call/skill design is sound (a 4B model handles the whole surface); the tool contract explicitly
+      guards narration-instead-of-action. Found + fixed one false-failing test (memory-clear asserted entry
+      removal vs the documented empty-document semantics). Path A LLM tests need the asset's `qwopus3.5-9b`
+      model loaded; running 4B+9B + Unity PlayMode together OOMs this machine (env resource limit, not a bug).
 - [ ] Verify the AI writes mods through the Hub chat in each kept Hub-enabled demo with local 4B/9B/27B
       (LM Studio) and Opus 4.8 via the bundled preset (`Assets/Resources/CoreAIPresets/`,
       bridge: `agent.sh openai-server -e claude -m opus`). *(API models are already proven through the