[codex] add loop engineering runtime

README.md

@@ -1,30 +1,57 @@
-# DESIGNOSFORGE v2.0.0 Source for Codex
+# DESIGNOSFORGE v2.1.0 Source for Codex
 
 DESIGNOSFORGE is an open-source Codex agent and skill system for turning AI design from prompt guessing into a governed, inspectable design workflow.
 
-v2.0.0 moves the project from prompt governance into a mathematical design kernel. It now models design decisions as vectors, probabilities, constraints, ranked candidate directions, critic scores, and failure-memory retrieval.
+v2.1.0 adds **Loop Engineering**: a system-runtime layer for scheduled, event-driven, multi-agent, tool-connected loops that can keep project context, validate results, and write durable memory.
 
-## What Changed In v2.0
+## What Changed In v2.1
 
-- `DesignKernel`: a new end-to-end orchestration core for intent parsing, routing, memory, constraints, candidates, critics, tool planning, and PromptPacketV2.
-- `DesignMathEngine`: Chinese/Latin vectorization, cosine and jaccard similarity, softmax route probability, entropy, confidence margin, Pareto front, TOPSIS, weighted utility, and residual-risk penalty.
-- `PromptPacketV2`: a richer design contract with route math, memory math, constraint math, candidate optimization, critic aggregation, and failure memory.
-- Stronger city identity behavior: routes public-cultural logo tasks toward modular identity systems and actively penalizes generic landmark stacking.
-- Stronger photography behavior: locks face identity, body anatomy, expression, clothing, light direction, skin texture, and documentary scene truth.
-- Stronger EnvArt behavior: CADMCP remains the source-fidelity route for CAD/DWG/DXF, construction drawings, layer semantics, wall/opening topology, and plan-to-board workflows.
+- `DesignKernel`: keeps the v2 mathematical orchestration core for intent parsing, routing, memory, constraints, candidates, critics, tool planning, and PromptPacketV2.
+- `LoopPromptPack`: remains the independent prompt-iteration companion pack for self-refine, failed-result recovery, branch search, visual-result repair, and seamless video loops.
+- `LoopEngineeringBlueprint`: new runtime blueprint for scheduler, worktree isolation, skill context, external connectors, validation gates, and persistent memory.
+- `LoopEngineeringOS`: new registry/capability route for long-running agent loops, GitHub/CI/PR workflows, and complete project-following workflows.
+- GitHub release planning now defaults to `v2.1.0` and includes compile, source validator, CLI smoke, and pytest gates.
+- Skill packaging now exposes Loop Engineering as a separate layer rather than replacing PromptPacketV2 or LoopPromptPack.
 
 ## Why It Matters
 
-Most AI design workflows fail after generation starts. DESIGNOSFORGE moves quality control before generation:
+Most AI design workflows fail because the agent loses the project thread. DESIGNOSFORGE v2.1 makes the workflow answer six engineering questions before it runs:
 
-- one dominant focal anchor instead of scattered fragments
-- grid, density, and negative-space rules instead of visual noise
-- exact visible text instead of pseudo-text and mojibake
-- structured PromptPacketV2 output instead of loose prompt paragraphs
-- project-context locks instead of mixing commercial and academic competition logic
-- memory-case recommendations instead of vague style recall
-- mathematical route and candidate audits instead of black-box taste claims
-- GitHub-ready release workflows instead of one-off local experiments
+- Who wakes the loop?
+- How do parallel agents avoid collisions?
+- How does the agent know project habits?
+- What external systems can it touch?
+- Who validates the result?
+- How does it remember yesterday?
+
+Principle: the model may forget, but the repository must not.
+
+## Capability Call Table
+
+| User intent | Primary route | Output | Validation |
+| --- | --- | --- | --- |
+| General design strategy | `DesignKernel` | `PromptPacketV2` | route math, critic aggregation, constraints |
+| Iterative prompt refinement | `LoopPromptEngine` | `LoopPromptPack` | one-axis iteration, stop conditions, failure memory |
+| Durable agent loops | `LoopEngineeringOS` | `LoopEngineeringBlueprint` | scheduler, worktree isolation, verifier split, memory |
+| Brand, VI, city identity | `brandVIos` | mark system direction and prompt packet | grid derivation, dynamic system, no random landmarks |
+| Academic infovis boards | `InfoVisOS` / `LayeredBoardComposer` | board strategy and modules | thesis hierarchy, evidence modules, text accuracy |
+| EnvArt/CAD production | `EnvArtCADMCPBridge` | CAD-aware plan or prompt packet | CAD health, DXF audit, geometry locks |
+| Photography/retouching | `PhotographyOS` | edit or shoot plan | identity preservation, light direction, artifact check |
+| GitHub/release work | `GitHubManager` / GitHub plugin | release plan or PR body | compileall, tests, source validator, CI/PR checks |
+
+## Quick Check
+
+```powershell
+$env:PYTHONPATH='.'
+$env:PYTHONUTF8='1'
+py -m app.cli capabilities
+py -m app.cli kernel loop-engineering "Loop Engineering 调度 issue CI，worktree 并行隔离，validator 验收，写入长期记忆"
+py -m app.cli kernel loop-prompt "create a design prompt with three loop iterations and self critique"
+py -m app.cli github release-plan --version v2.1.0
+py -m compileall app tests tools
+py -m pytest -q
+py tools\validate_source_skill.py
+```
 
 ## Open Source
 
@@ -36,35 +63,21 @@ Use it to study, adapt, and extend:
 - design-agent orchestration
 - visual prompt governance
 - mathematical design routing
+- Loop Engineering runtime design
 - aesthetic quality gates
 - LoRA aesthetic corpus planning
-- photography and retouching aesthetic-memory planning
+- photography and retouching workflows
 - environmental-art CAD/DWG/DXF source-fidelity workflows
-- training-aware aesthetic memory indexing
-- project-context routing for commercial, academic competition, and public cultural work
 - GitHub-ready release workflows
 
 See `docs/CODEX_INSTALL.md` for local Codex skill installation.
 
-## Quick Check
-
-```bash
-PYTHONPATH=. python -m app.cli capabilities
-PYTHONPATH=. python -m app.cli kernel plan "为安徽省钢城马鞍山市设计城市标识系统logo，要求现代、公共文化传播、不要堆砌地标"
-PYTHONPATH=. python -m app.cli kernel math-audit "拯救课堂纪实照片，不要改变人物本来的面貌形象"
-PYTHONPATH=. python -m app.cli kernel prompt-packet "用CADMCP审核环艺DWG平面并生成展板分析图提示词"
-PYTHONPATH=. python -m app.cli lora audit-corpus
-PYTHONPATH=. python -m app.cli lora build-memory-index
-PYTHONPATH=. python -m app.cli envart-cad plan "用CADMCP审核环艺DWG平面并生成展板分析图提示词"
-PYTHONPATH=. python -m app.cli github release-plan --version v2.0.0
-PYTHONPATH=. pytest -q
-```
-
 ## GitHub
 
-This source package includes a GitHub Actions workflow, PR body, release notes, and a source skill validator.
-After binding a target remote, push the main branch and tag `v2.0.0`, then open a draft PR using `docs/PR_BODY_v2.0.0.md`.
+This source package includes GitHub workflow docs, PR bodies, release notes, and a source skill validator.
+
+For v2.1 release work, use `docs/PR_BODY_v2.1.0.md` and `docs/RELEASE_NOTES_v2.1.0.md`. Do not publish releases or notify external systems without explicit user intent.
 
 ## Codex Placement
 
-Use `codex_skill/designos-forge/SKILL.md` as the Codex skill entry. The included Codex entry has been upgraded to `designos-forge` v2.0.0 with a mathematical DesignKernel, PromptPacketV2, aesthetic memory, failure memory, photography, EnvArt CADMCP fusion, prompt precision, text/encoding health, environment-aware routing, and Git/GitHub release planning.
+Use `codex_skill/designos-forge/SKILL.md` as the Codex skill entry. The included skill has been upgraded to `designos-forge` v2.1.0 with DesignKernel, PromptPacketV2, LoopPromptPack, LoopEngineeringOS, aesthetic memory, failure memory, photography, EnvArt CADMCP fusion, prompt precision, text/encoding health, project-context routing, GitOps, and GitHub release planning.

app/__init__.py

@@ -1 +1 @@
-__version__ = "2.0.0"
+__version__ = "2.1.0"

app/cli.py

@@ -39,15 +39,15 @@ def main():
     github_parser = sub.add_parser("github")
     github_parser.add_argument("action", choices=("status", "release-plan", "pr-template"))
     github_parser.add_argument("--repo", default=".")
-    github_parser.add_argument("--version", default="v2.0.0")
+    github_parser.add_argument("--version", default="v2.1.0")
     quality_parser = sub.add_parser("quality")
     quality_parser.add_argument("action", choices=("audit", "guardrails"))
     quality_parser.add_argument("text")
     envart_cad_parser = sub.add_parser("envart-cad")
     envart_cad_parser.add_argument("action", choices=("plan",))
     envart_cad_parser.add_argument("text")
     kernel_parser = sub.add_parser("kernel")
-    kernel_parser.add_argument("action", choices=("plan", "prompt-packet", "loop-prompt", "math-audit", "record-failure"))
+    kernel_parser.add_argument("action", choices=("plan", "prompt-packet", "loop-prompt", "loop-engineering", "math-audit", "record-failure"))
     kernel_parser.add_argument("text")
     kernel_parser.add_argument("--domain", default="general-design")
     kernel_parser.add_argument("--failure-mode", default="unspecified_failure")
@@ -94,6 +94,8 @@ def main():
                 print(json_dumps(plan["prompt_packet_v2"]))
             elif args.action == "loop-prompt":
                 print(json_dumps(plan["loop_prompt_pack"]))
+            elif args.action == "loop-engineering":
+                print(json_dumps(plan["loop_engineering"]))
             elif args.action == "math-audit":
                 print(json_dumps(plan["math_trace"]))
             else:

app/core/capabilities.py

@@ -15,7 +15,18 @@
     "ConstraintPenaltyModel", "DesignStateGraph", "HybridRouter", "AestheticGenomeExtractor",
     "DesignMemoryVectorIndex", "FailureMemoryBank", "MultiCandidateGenerator",
     "CriticEnsemble", "ConstraintSolver", "ToolExecutionPlanner", "PromptPacketV2Builder",
-    "LoopPromptEngine", "LoopPromptPackBuilder",
+    "LoopPromptEngine", "LoopPromptPackBuilder", "LoopEngineeringBlueprintBuilder", "LoopEngineeringOS",
+)
+
+CALL_TABLE = (
+    ("general_design_strategy", "DesignKernel", "PromptPacketV2", "route math, critic aggregation, constraints"),
+    ("iterative_prompt_refinement", "LoopPromptEngine", "LoopPromptPack", "one-axis iteration, stop conditions, failure memory"),
+    ("durable_agent_loop", "LoopEngineeringOS", "LoopEngineeringBlueprint", "scheduler, worktree isolation, verifier split, persistent memory"),
+    ("brand_vi_city_identity", "brandVIos", "PromptPacketV2", "grid derivation, dynamic system, no random landmark stacking"),
+    ("academic_infovis_board", "InfoVisOS/LayeredBoardComposer", "PromptPacketV2", "thesis hierarchy, evidence modules, text accuracy"),
+    ("envart_cad_production", "EnvArtCADMCPBridge", "CAD-aware plan", "CAD health, DXF audit, geometry locks"),
+    ("photography_retouching", "PhotographyOS", "edit or shoot plan", "identity preservation, light direction, artifact check"),
+    ("github_release_work", "GitHubManager/GitOpsManager", "release plan or PR body", "compileall, tests, source validator, CI/PR checks"),
 )
 
 def capability_report():
@@ -25,4 +36,7 @@ def capability_report():
     lines.append("")
     lines.append("Skills:")
     lines.extend(f"- {skill.name}: {skill.description}" for skill in registry.list())
+    lines.append("")
+    lines.append("Capability Call Table:")
+    lines.extend(f"- {intent}: route={route}; output={output}; validation={validation}" for intent, route, output, validation in CALL_TABLE)
     return "\n".join(lines)

app/core/design_kernel.py

@@ -16,11 +16,11 @@
     weighted_sum,
 )
 from .envart_cadmcp import EnvArtCADMCPBridge
-from .loop_prompt import LoopPromptPackBuilder
+from .loop_prompt import LoopEngineeringBlueprintBuilder, LoopPromptPackBuilder
 from .task_router import ROUTES, TaskRouterAgent
 
 
-VERSION = "2.0.0"
+VERSION = "2.1.0"
 FAILURE_MEMORY_PATH = Path("lora_training_sandbox/aesthetic_corpus/failure_memory.jsonl")
 
 
@@ -182,6 +182,7 @@ class DesignKernelPlan:
     math_trace: dict
     prompt_packet_v2: dict
     loop_prompt_pack: dict
+    loop_engineering: dict
 
     def to_dict(self):
         return asdict(self)
@@ -759,6 +760,7 @@ def plan(self, intent, route, constraints):
         steps = ["parse_intent", "retrieve_memory", "solve_constraints", "rank_candidates", "run_critic_ensemble"]
         tools = []
         loop_detection = LoopPromptPackBuilder().detect(intent.raw_text, intent=intent, route=route)
+        loop_engineering_detection = LoopEngineeringBlueprintBuilder().detect(intent.raw_text, loop_detection=loop_detection)
         if constraints.get("cad_plan"):
             tools.extend(constraints["cad_plan"].get("preferred_tools", []))
         if "prompt_packet" in intent.delivery_modes:
@@ -768,6 +770,9 @@ def plan(self, intent, route, constraints):
             tools.append("LoopPromptPack")
             if loop_detection.loop_type == "seamless_video_loop":
                 tools.append("ShortDramaAIGC_OS")
+        if loop_engineering_detection.active:
+            steps.extend(["build_loop_engineering_blueprint", "define_scheduler_isolation_validation_memory"])
+            tools.append("LoopEngineeringBlueprint")
         if "git_release" in intent.delivery_modes:
             tools.extend(["GitOpsManager", "GitHubManager"])
         if route["skill_name"] == "PhotographyOS":
@@ -869,6 +874,7 @@ def __init__(self):
         self.failures = FailureMemoryBank()
         self.packet = PromptPacketV2Builder()
         self.loop_prompt = LoopPromptPackBuilder()
+        self.loop_engineering = LoopEngineeringBlueprintBuilder()
 
     def plan(self, text):
         intent = self.intent_parser.parse(text)
@@ -894,6 +900,14 @@ def plan(self, text):
             constraints=constraints,
             failure_memory=failure_memory,
         )
+        loop_engineering = self.loop_engineering.build(
+            intent.raw_text,
+            intent=intent,
+            route=route,
+            loop_detection=LoopPromptPackBuilder().detect(intent.raw_text, intent=intent, route=route),
+            constraints=constraints,
+            failure_memory=failure_memory,
+        )
         return DesignKernelPlan(
             schema_version=VERSION,
             intent=asdict(intent),
@@ -909,6 +923,7 @@ def plan(self, text):
             math_trace=math_trace,
             prompt_packet_v2=prompt_packet,
             loop_prompt_pack=loop_prompt_pack,
+            loop_engineering=loop_engineering,
         )
 
     def _math_trace(self, route, memory, constraints, candidates, critics, failure_memory):