Release 1.0.0-rc.1: unified memory workspace and server-independent local installation

.gitattributes

@@ -2,6 +2,7 @@
 *.sh text eol=lf
 02-tmcra-memory-api/deploy/tmcra text eol=lf
 *.py text eol=lf
+*.ps1 text eol=lf
 *.js text eol=lf
 *.mjs text eol=lf
 *.ts text eol=lf

.github/workflows/release-gate.yml

@@ -44,6 +44,9 @@ jobs:
       - name: Verify portable service and model substitution
         working-directory: 02-tmcra-memory-api
         run: python -m unittest test_portable_release.py test_tmcra_service_model_flexibility.py
+      - name: Verify feedback and local-runtime safety contracts
+        working-directory: 02-tmcra-memory-api
+        run: python -m unittest discover -s tmcra_service/tests -q
       - name: Run frozen benchmark contracts
         working-directory: 09-tmcra-benchmarks
         env:

.gitignore

@@ -15,3 +15,5 @@ tmcra_service_state/
 03-tmcra-web-console/public/downloads/tmcra-codex-*.zip
 03-tmcra-web-console/desktop/tmcra-memory/resources/tmcra-codex-*
 04-tmcra-desktop/resources/tmcra-codex-*
+release/
+docs/LOCAL_DEPLOYMENT_PRODUCTION_AUDIT.zh-CN.md

01-tmcra-agent-memory-algorithm/deploy/local-model-profiles.json

@@ -0,0 +1,132 @@
+{
+  "schema_version": "tmcra.local-model-profiles.1",
+  "verified_upstream_on": "2026-09-06",
+  "status": "windows_local_preview_partial_validation",
+  "installation_enabled": true,
+  "scope": "full_local_runtime_preview",
+  "full_memory_system_ready": false,
+  "hardware_requirements_are": "conservative_planning_estimates_not_benchmark_results",
+  "profiles": [
+    {
+      "id": "lite-cpu",
+      "name_zh": "轻量版",
+      "recommendation_zh": "普通笔记本、无独显电脑；中英文短记忆，低并发",
+      "validation": "cpu_ingest_and_raw_recall_passed_complex_compile_timed_out",
+      "system_ram_gib_min": 8,
+      "system_ram_gib_recommended_for_full_memory": 16,
+      "retrieval_vram_gib_recommended": 0,
+      "weights_bytes": 941234298,
+      "embedding": {
+        "repo_id": "intfloat/multilingual-e5-small",
+        "revision": "614241f622f53c4eeff9890bdc4f31cfecc418b3",
+        "license": "MIT",
+        "upstream": "https://huggingface.co/intfloat/multilingual-e5-small",
+        "weights": [{"file": "model.safetensors", "bytes": 470641600, "sha256": "1a55775f53449dac10a2bcbc312469fac40b96d53198c407081a831f81c98477"}],
+        "dimensions": 384,
+        "model_max_tokens": 512,
+        "pooling": "mean",
+        "normalize": true,
+        "query_prefix": "query: ",
+        "document_prefix": "passage: ",
+        "padding_side": "right"
+      },
+      "reranker": {
+        "repo_id": "cross-encoder/mmarco-mMiniLMv2-L12-H384-v1",
+        "revision": "1427fd652930e4ba29e8149678df786c240d8825",
+        "license": "Apache-2.0",
+        "upstream": "https://huggingface.co/cross-encoder/mmarco-mMiniLMv2-L12-H384-v1",
+        "weights": [{"file": "model.safetensors", "bytes": 470592698, "sha256": "5daeca2481a76b5976a2bdc32f0a78532b6716da4f8cd3ff59460ef8d2f359b4"}],
+        "adapter": "sequence-classification",
+        "model_max_tokens": 512,
+        "tmcra_fusion_checkpoint_compatible": false
+      },
+      "required_work": ["memory_pressure_retest", "complex_compile_latency", "organizer_validation", "full_service_restart_validation"]
+    },
+    {
+      "id": "balanced-bge",
+      "name_zh": "均衡版",
+      "recommendation_zh": "优先复现现有生产检索模型；16GB 以上内存，建议有独显",
+      "validation": "production_model_stack_verified_consumer_installer_pending",
+      "system_ram_gib_min": 16,
+      "system_ram_gib_recommended_for_full_memory": 32,
+      "retrieval_vram_gib_recommended": 6,
+      "weights_bytes": 4542217682,
+      "embedding": {
+        "repo_id": "BAAI/bge-m3",
+        "revision": "5617a9f61b028005a4858fdac845db406aefb181",
+        "license": "MIT",
+        "upstream": "https://huggingface.co/BAAI/bge-m3",
+        "weights": [{"file": "pytorch_model.bin", "bytes": 2271145830, "sha256": "b5e0ce3470abf5ef3831aa1bd5553b486803e83251590ab7ff35a117cf6aad38"}],
+        "dimensions": 1024,
+        "model_max_tokens": 8192,
+        "pooling": "cls",
+        "normalize": true,
+        "query_prefix": "",
+        "document_prefix": "",
+        "padding_side": "right"
+      },
+      "reranker": {
+        "repo_id": "BAAI/bge-reranker-v2-m3",
+        "revision": "953dc6f6f85a1b2dbfca4c34a2796e7dde08d41e",
+        "license": "Apache-2.0",
+        "upstream": "https://huggingface.co/BAAI/bge-reranker-v2-m3",
+        "weights": [{"file": "model.safetensors", "bytes": 2271071852, "sha256": "d9e3e081faff1eefb84019509b2f5558fd74c1a05a2c7db22f74174fcedb5286"}],
+        "adapter": "sequence-classification",
+        "model_max_tokens": 8192,
+        "production_runtime_max_tokens": 1280,
+        "tmcra_fusion_checkpoint_compatible": true
+      },
+      "required_work": ["single_recall_lane", "desktop_worker_and_context_budgets", "desktop_installer", "offline_end_to_end_validation"]
+    },
+    {
+      "id": "quality-qwen",
+      "name_zh": "增强版候选",
+      "recommendation_zh": "32GB 以上内存、高显存电脑；长文本和跨语言检索候选，需与均衡版实测比较",
+      "validation": "upstream_artifacts_verified_runtime_pending",
+      "system_ram_gib_min": 32,
+      "system_ram_gib_recommended_for_full_memory": 64,
+      "retrieval_vram_gib_recommended": 16,
+      "weights_bytes": 9235180368,
+      "embedding": {
+        "repo_id": "Qwen/Qwen3-Embedding-4B",
+        "revision": "5cf2132abc99cad020ac570b19d031efec650f2b",
+        "license": "Apache-2.0",
+        "upstream": "https://huggingface.co/Qwen/Qwen3-Embedding-4B",
+        "weights": [
+          {"file": "model-00001-of-00002.safetensors", "bytes": 4965826464, "sha256": "e70bfe3c970523fb7ef4eddffed2254ce3f1e7150c3de2af4342de129dd756f8"},
+          {"file": "model-00002-of-00002.safetensors", "bytes": 3077765624, "sha256": "ed1b87c8e9eb7e535a1a155e4fd00d9f4dba80e58a6db48a4c9f82cede7079c1"}
+        ],
+        "dimensions": 2560,
+        "model_max_tokens": 32768,
+        "pooling": "last_token",
+        "normalize": true,
+        "query_prefix": "Instruct: Given a query about previous conversations, retrieve the relevant source passages.\nQuery: ",
+        "document_prefix": "",
+        "padding_side": "left"
+      },
+      "reranker": {
+        "repo_id": "Qwen/Qwen3-Reranker-0.6B",
+        "revision": "e61197ed45024b0ed8a2d74b80b4d909f1255473",
+        "license": "Apache-2.0",
+        "upstream": "https://huggingface.co/Qwen/Qwen3-Reranker-0.6B",
+        "weights": [{"file": "model.safetensors", "bytes": 1191588280, "sha256": "27cd75a405b9c1b46b59abfd88aaa209e6fed2a1972cde9b70e7659537c5e65b"}],
+        "adapter": "causal-lm-yes-no",
+        "model_max_tokens": 32768,
+        "tmcra_fusion_checkpoint_compatible": false
+      },
+      "required_work": ["qwen_yes_no_reranker_adapter", "service_variable_embedding_dimensions", "bounded_gpu_batches", "full_index_rebuild", "quality_and_latency_comparison"]
+    }
+  ],
+  "deployment_contract": {
+    "automatic_profile_switch_on_existing_index": false,
+    "cloud_fallback_in_full_local_mode": false,
+    "cloud_account_required_in_full_local_mode": false,
+    "generation_model_included_in_weights_bytes": false,
+    "model_cache_and_python_runtime_included_in_weights_bytes": false,
+    "required_index_identity": ["repo_id", "revision", "dimensions", "pooling", "query_prefix", "document_prefix", "normalization", "precision", "chunking_policy"],
+    "existing_sources_must_be_preserved": true,
+    "existing_indexes_must_be_rebuilt_before_activation": true,
+    "runtime_downloads_after_setup": false,
+    "readiness_requires_actual_inference": true
+  }
+}

01-tmcra-agent-memory-algorithm/run_tmcra_v4_build.py

@@ -47,6 +47,9 @@ def _stage(root: Path, name: str, **values: Any) -> None:
 
 
 def _load_shell_environment(path: Path) -> dict[str, str]:
+    from tmcra_local_only import enabled, read_environment
+    if enabled():
+        return read_environment(path)
     if not path.is_file():
         raise BuildError(f"writer environment file is missing: {path}")
     command = 'set -a; source "$1"; env -0'
@@ -82,6 +85,10 @@ def _rotated(keys: Sequence[str], worker_index: int) -> str:
 
 def _worker_environment(base: Mapping[str, str], keys: Sequence[str], worker_index: int) -> dict[str, str]:
     environment = dict(base)
+    from tmcra_local_only import enabled, validate_environment
+    if enabled(environment):
+        validate_environment(environment)
+        return environment
     pool = _rotated(keys, worker_index)
     base_url = environment.get("TMCRA_DEEPSEEK_WRITER_BASE_URL") or environment.get("TMCRA_WRITER_BASE_URL") or "https://api.deepseek.com/v1"
     max_tokens = environment.get("TMCRA_WRITER_MAX_TOKENS", "16384")

01-tmcra-agent-memory-algorithm/shared_core_manifest.json

@@ -1,31 +1,31 @@
 {
   "schema_version": "tmcra.service.shared-core-manifest.1",
-  "service_version": "0.3.0-rc2",
-  "generated_on": "2026-08-17",
+  "service_version": "1.0.0-rc.1",
+  "generated_on": "2026-09-06",
   "algorithm_files": [
     {
       "path": "prepare_tmcra_v4_e2e_data.py",
       "sha256": "68bdd34631a0897730926b5f8f68b87ee7352900446baba62d1615c961a53c96"
     },
     {
       "path": "run_tmcra_v4_build.py",
-      "sha256": "1ebc54751fcfea91bbd8ea019cac45ca79cf60ce229bf7249d0427955093f93e"
+      "sha256": "4172caaa4e5a9b75dc2dec7c8188e6f585376aecb41e4f08a2514b25193ee4b9"
     },
     {
       "path": "run_tmcra_v4_compile_evidence.py",
       "sha256": "e869c9069bb0c8ad66e7f9bf469f17526269450cd0f58981ec9676d524234166"
     },
     {
       "path": "tmp_tmcra_v2_lme_pipeline.py",
-      "sha256": "22768faa73f61050373791179f83a33595b46a3d9fb35476ef8054222a446cbd"
+      "sha256": "7e1ade3b1902c14c405be975c99764b6fac4843bd3b2a8ace5185593ec2fe365"
     },
     {
       "path": "tmcra_v2_lme_pipeline.py",
       "sha256": "b09afe18bcc888f02cad2a0985698210ff172ba9b5e77d9b4405e4146fa6bc60"
     },
     {
       "path": "tmcra_v3_online_runtime.py",
-      "sha256": "75186638496fe9271b711e0ba76cb93b232a5684e75b565dfe4721bf821c0a12"
+      "sha256": "18b55050ff04bc8235b6c5aa2d53adf891a793179b4478bf4cfdfd6038c7195d"
     },
     {
       "path": "tmcra_v3_product_writer.py",
@@ -41,7 +41,7 @@
     },
     {
       "path": "tmcra_v4_batch_writer.py",
-      "sha256": "352568d64d831dbb28a228e9cd898f688636330b13ac2ff99e5ed36226a4dbf2"
+      "sha256": "a46a3695b08a8a5bd205863a43c416340863976f274d851ed2db4d04f5c125f5"
     },
     {
       "path": "tmcra_v4_cost_report.py",
@@ -53,15 +53,15 @@
     },
     {
       "path": "tmcra_v4_evidence_planner.py",
-      "sha256": "99dd6cd40bd205a4e9f10bc748b05aff848e187c9ce5f5f6fca9237a9d32aa99"
+      "sha256": "dfe546149fc8a5596ed311032020f36737eab37290a9e69255d2fae0fd708c7a"
     },
     {
       "path": "tmcra_v4_online_runtime.py",
-      "sha256": "4fee0d591cd4140353d0adf57287e93534607e06db7c92b5ab42fbb0afe45db2"
+      "sha256": "7ef1bb4ec78094fa166750b78b567079b7048ca52787cfbefd7e3aaf8a9a8a2e"
     },
     {
       "path": "tmcra_v4_slow_graph.py",
-      "sha256": "985e2aab12eba2b96bd53cc05e7352412fb45bd3d9fee3a0a8213b26ff4dba0b"
+      "sha256": "7ebf6d258b85c9a463e55c0dcdc817e9fc27b42fc93af7abdf48e62ecf8acf42"
     },
     {
       "path": "tmcra_v4_recall_planner.py",
@@ -78,6 +78,14 @@
     {
       "path": "tmcra_v4_typed_semantics.py",
       "sha256": "367917528c80829a9b99ecebc1f9c533fb77f39753713b65abadbddb9c4e43e5"
+    },
+    {
+      "path": "tmcra_local_only.py",
+      "sha256": "16da1be2c9d8bf881481ea64a09de30b423e3c4c496017e0b693e0bd2ba45e05"
+    },
+    {
+      "path": "tmcra_local_models.py",
+      "sha256": "f464a241c7f7ca2f78a52f7c03d7157520795281ac5c7e01faeaea4606274fb4"
     }
   ]
 }

01-tmcra-agent-memory-algorithm/tmcra_local_models.py

@@ -0,0 +1,101 @@
+"""Pinned retrieval profiles shared by the API and its real index workers."""
+from __future__ import annotations
+
+import hashlib
+import json
+import os
+from pathlib import Path
+
+PROFILE_FILE = Path(__file__).resolve().parent / "deploy" / "local-model-profiles.json"
+
+
+def profiles():
+    return json.loads(PROFILE_FILE.read_text(encoding="utf-8"))["profiles"]
+
+
+def profile_by_id(identity):
+    for profile in profiles():
+        if profile["id"] == identity:
+            return profile
+    raise ValueError("unknown local model profile")
+
+
+def signature(profile):
+    # A model change, pooling change or window-policy change requires reindexing.
+    contract = {"embedding": profile["embedding"], "source_windows": "token-char-covered-v1",
+                "long_slow_embedding": "normalized-window-mean-v1"}
+    return hashlib.sha256(json.dumps(contract, sort_keys=True, separators=(",", ":")).encode()).hexdigest()
+
+
+def apply_local_profile(args):
+    if os.getenv("TMCRA_DEPLOYMENT_MODE") != "local":
+        return args
+    profile = profile_by_id(os.environ["TMCRA_LOCAL_PROFILE"])
+    embed = profile["embedding"]
+    rerank = profile["reranker"]
+    args.embedding_profile_id = profile["id"]
+    args.embedding_index_signature = signature(profile)
+    args.embedding_model = os.environ["TMCRA_EMBEDDING_MODEL"]
+    args.cross_model = os.environ["TMCRA_CROSS_MODEL"]
+    args.text_dim = embed["dimensions"]
+    args.embedding_max_length = embed["model_max_tokens"]
+    args.embedding_pooling = embed["pooling"]
+    args.embedding_query_prefix = embed["query_prefix"]
+    args.embedding_document_prefix = embed["document_prefix"]
+    args.embedding_padding_side = embed["padding_side"]
+    args.embedding_strict_max_length = True
+    args.embedding_long_document_policy = "window_mean"
+    args.reranker_mode = "fusion" if rerank["tmcra_fusion_checkpoint_compatible"] else "semantic-only"
+    args.reranker_adapter = rerank["adapter"]
+    args.cross_max_length = min(rerank["model_max_tokens"], 1280)
+    args.cross_batch_size = 2 if args.device == "cpu" else 8
+    args.batch_size = 4 if args.device == "cpu" else 8
+    return args
+
+
+def verify_index_identity(payload, args):
+    expected = getattr(args, "embedding_index_signature", "")
+    if expected and (payload.get("embedding_index_signature") != expected
+                     or payload.get("text_dim") != args.text_dim):
+        raise RuntimeError("index embedding identity differs; rebuild into a new generation before recall")
+
+
+def sha256_file(path):
+    digest = hashlib.sha256()
+    with Path(path).open("rb") as handle:
+        for block in iter(lambda: handle.read(8 * 1024 * 1024), b""):
+            digest.update(block)
+    return digest.hexdigest()
+
+
+def verify_weights(model, directory):
+    directory = Path(directory)
+    for asset in model["weights"]:
+        path = directory / asset["file"]
+        if (not path.is_file() or path.stat().st_size != asset["bytes"]
+                or sha256_file(path) != asset["sha256"]):
+            raise RuntimeError(f"model asset is missing or failed SHA-256 verification: {asset['file']}")
+
+
+QWEN_SYSTEM = ("<|im_start|>system\nJudge whether the Document meets the requirements based on the Query "
+              "and the Instruct provided. Note that the answer can only be \"yes\" or \"no\"."
+              "<|im_end|>\n<|im_start|>user\n")
+QWEN_SUFFIX = "<|im_end|>\n<|im_start|>assistant\n<think>\n\n</think>\n\n"
+
+
+def qwen_rerank_windows(tokenizer, query, document, *, max_length):
+    prefix = QWEN_SYSTEM + ("<Instruct>: Given a query about previous conversations, retrieve the relevant "
+                            f"source passages.\n<Query>: {query}\n<Document>: ")
+    head = tokenizer.encode(prefix, add_special_tokens=False)
+    tail = tokenizer.encode(QWEN_SUFFIX, add_special_tokens=False)
+    body = tokenizer.encode(document, add_special_tokens=False)
+    capacity = max_length - len(head) - len(tail)
+    if capacity < 32:
+        raise ValueError("reranker query leaves insufficient document context; shorten the query")
+    step = max(1, capacity - min(192, capacity // 4))
+    windows = []
+    for start in range(0, max(1, len(body)), step):
+        windows.append(head + body[start:start + capacity] + tail)
+        if start + capacity >= len(body):
+            break
+    return windows