Add content_class to provider contract and require it in Hermes

.claude-plugin/marketplace.json

@@ -12,7 +12,7 @@
       "name": "claude-code",
       "source": "./plugins/claude-code",
       "description": "Persistent semantic memory for Claude Code - user preferences, project context, prior decisions, and codebase facts that survive across sessions.",
-      "version": "0.1.18",
+      "version": "0.2.0",
       "category": "productivity",
       "homepage": "https://docs.atomicstrata.ai/integrations/coding-agents/claude-code",
       "license": "Apache-2.0"

packages/sdk/schema/v1/provider-contract.schema.json

@@ -33,6 +33,12 @@
       "type": "string",
       "enum": ["fact", "episode", "summary", "procedure", "document"]
     },
+    "ContentClass": {
+      "title": "ContentClass",
+      "description": "Sensitivity class of supplied content: 'summary' (distilled, hosted-safe), 'redacted' (sensitive spans removed by the caller), or 'raw' (verbatim prompt/response/diff/source). This contract exposes it on verbatim ingest only; core's HTTP API also consults it on extraction paths, where it decides whether the raw transcript is withheld from the durable audit episode. A core running RAW_CONTENT_POLICY=reject (the default) refuses a verbatim write of 'raw' or unclassified content. Never inferred by a provider — the caller chooses it, so omitting it fails closed rather than mislabeling raw content as safe.",
+      "type": "string",
+      "enum": ["summary", "redacted", "raw"]
+    },
     "Message": {
       "title": "Message",
       "type": "object",
@@ -104,6 +110,7 @@
         "mode": { "const": "verbatim" },
         "content": { "type": "string" },
         "kind": { "$ref": "#/$defs/MemoryKind" },
+        "content_class": { "$ref": "#/$defs/ContentClass" },
         "scope": { "$ref": "#/$defs/Scope" },
         "provenance": { "$ref": "#/$defs/Provenance" },
         "metadata": { "type": "object", "additionalProperties": true }

packages/sdk/src/memory/__tests__/provider-contract-schema.test.ts

@@ -86,6 +86,31 @@ describe('provider-contract v1 schemas', () => {
     expect(ingest({ ...GOOD_INGEST, mode: 'binary' })).toBe(false);
   });
 
+  // The SDK forwards content_class to core on the verbatim path, and core
+  // refuses an unclassified verbatim write under RAW_CONTENT_POLICY=reject.
+  // VerbatimIngest is additionalProperties:false, so the schema has to carry
+  // the field or the very payload the SDK emits is contract-invalid.
+  // Verbatim only, deliberately: core also consults content_class on
+  // extraction paths for audit-transcript redaction, but exposing it there
+  // changes what is durably retained — tracked as tech debt, not done here.
+  it.each(['summary', 'redacted', 'raw'])('accepts content_class %s on verbatim', (contentClass) => {
+    const { ingest } = buildValidators();
+    expect(ingest({ ...GOOD_INGEST, content_class: contentClass })).toBe(true);
+  });
+
+  it('rejects an unknown content_class', () => {
+    const { ingest } = buildValidators();
+    expect(ingest({ ...GOOD_INGEST, content_class: 'public' })).toBe(false);
+  });
+
+  it('still rejects an unknown ingest field', () => {
+    // Guards the guard: proves additionalProperties:false is doing work, so
+    // the acceptance above means the field was added rather than the schema
+    // having stopped constraining anything.
+    const { ingest } = buildValidators();
+    expect(ingest({ ...GOOD_INGEST, not_a_real_field: 'x' })).toBe(false);
+  });
+
   it('rejects a retrieval receipt missing trace_id', () => {
     const { searchPage } = buildValidators();
     const receipt: Record<string, unknown> = { ...GOOD_SEARCH_PAGE.retrieval };

plugins/claude-code/.claude-plugin/plugin.json

@@ -1,6 +1,6 @@
 {
   "name": "atomicmemory",
-  "version": "0.1.18",
+  "version": "0.2.0",
   "description": "Persistent semantic memory for Claude Code — user preferences, project context, prior decisions, and codebase facts that survive across sessions.",
   "author": {
     "name": "AtomicMemory",

plugins/claude-code/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@atomicmemory/claude-code-plugin",
-  "version": "0.1.18",
+  "version": "0.2.0",
   "description": "AtomicMemory plugin for Claude Code — persistent semantic memory across sessions.",
   "private": false,
   "license": "Apache-2.0",

plugins/codex/.codex-plugin/plugin.json

@@ -1,6 +1,6 @@
 {
   "name": "atomicmemory",
-  "version": "0.1.18",
+  "version": "0.2.0",
   "description": "AtomicMemory memory layer for Codex. Pluggable semantic memory — swap backends through the SDK's MemoryProvider model by config, not code change.",
   "author": {
     "name": "AtomicMemory",

plugins/codex/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@atomicmemory/codex-plugin",
-  "version": "0.1.18",
+  "version": "0.2.0",
   "description": "AtomicMemory plugin for OpenAI Codex — plugin manifest, MCP server config, and memory protocol skill.",
   "private": true,
   "license": "Apache-2.0",

plugins/codex/skills/atomicmemory/SKILL.md

@@ -10,7 +10,7 @@ description: >
 license: Apache-2.0
 metadata:
   author: AtomicMemory
-  version: "0.1.18"
+  version: "0.2.0"
   category: ai-memory
   tags: "memory, semantic-search, codex, pluggable"
 ---

plugins/cursor/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@atomicmemory/cursor-plugin",
-  "version": "0.1.18",
+  "version": "0.2.0",
   "description": "AtomicMemory integration for Cursor - MCP configuration and project rules for persistent semantic memory.",
   "private": true,
   "license": "Apache-2.0",

plugins/hermes/CHANGELOG.md

@@ -1,5 +1,15 @@
 # Changelog
 
+## 0.2.0 - 2026-07-29
+
+### Changed
+
+- **Breaking (tool contract):** `atomicmemory_conclude` now requires a `content_class` parameter (`summary`, `redacted`, or `raw`) describing the sensitivity of the text being stored. A call that omits it, or passes an unrecognized value, is rejected at the tool boundary with a message naming the parameter.
+
+  This is required to store facts against AtomicMemory Core 1.2.0 and later. Core defaults to `RAW_CONTENT_POLICY=reject`, which refuses a verbatim write carrying no content class — previously `conclude` sent an unclassified write and received `422 raw_content_rejected`, with nothing indicating which tool call caused it. The class is never chosen on the caller's behalf: guessing it could label a raw transcript as hosted-safe.
+
+  Existing deployments need no configuration change, but a model that calls `conclude` without the new parameter will get a tool error instead of a stored fact until it adapts.
+
 ## 0.1.13 - 2026-05-15
 
 ### Fixed

plugins/hermes/client.py

@@ -158,4 +158,5 @@ def ingest_verbatim(
         scope: Scope,
         provenance: Provenance,
         metadata: dict[str, Any] | None = None,
+        content_class: str | None = None,
     ) -> IngestResult: ...

plugins/hermes/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@atomicmemory/hermes-plugin",
-  "version": "0.1.18",
+  "version": "0.2.0",
   "description": "AtomicMemory native Hermes memory provider — Python SDK-backed, cross-tool memory by default.",
   "publishConfig": {
     "access": "public",

plugins/hermes/plugin.yaml

@@ -1,5 +1,5 @@
 name: atomicmemory
-version: 0.1.18
+version: 0.2.0
 description: "AtomicMemory native Hermes memory provider — Python SDK-backed, cross-tool memory by default."
 pip_dependencies:
   - "atomicmemory>=1.1.2,<2.0.0"

plugins/hermes/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "atomicmemory-hermes"
-version = "0.1.18"
+version = "0.2.0"
 description = "AtomicMemory native Hermes memory provider."
 readme = "README.md"
 requires-python = ">=3.10"

plugins/hermes/python_sdk.py

@@ -175,16 +175,21 @@ def ingest_verbatim(
         scope: Scope,
         provenance: Provenance,
         metadata: dict[str, Any] | None = None,
+        content_class: str | None = None,
     ) -> IngestResult:
-        raw = self._require_client().ingest(
-            {
-                "mode": "verbatim",
-                "content": content,
-                "scope": scope,
-                "provenance": _provenance_dict(provenance),
-                "metadata": metadata,
-            }
-        )
+        payload: dict[str, Any] = {
+            "mode": "verbatim",
+            "content": content,
+            "scope": scope,
+            "provenance": _provenance_dict(provenance),
+            "metadata": metadata,
+        }
+        # Only sent when the caller chose one. Never defaulted here: an
+        # unclassified write must fail closed at core rather than be relabeled
+        # safe by the bridge.
+        if content_class is not None:
+            payload["content_class"] = content_class
+        raw = self._require_client().ingest(payload)
         return _ingest_result(raw)
 
     def _require_client(self) -> Any:

plugins/hermes/tests/fakes.py

@@ -120,13 +120,15 @@ def ingest_verbatim(
         scope: Scope,
         provenance: Provenance,
         metadata: dict[str, Any] | None = None,
+        content_class: str | None = None,
     ) -> IngestResult:
         self._record(
             "ingest_verbatim",
             content=content,
             scope=scope,
             provenance=provenance,
             metadata=metadata,
+            content_class=content_class,
         )
         self._maybe_raise("ingest_verbatim")
         return self.ingest_response

plugins/hermes/tests/test_provider.py

@@ -130,7 +130,10 @@ def test_conclude_stamps_provenance_source_hermes(self) -> None:
         fake = FakeAtomicMemoryClient(ingest_response=IngestResult(created=["m1"]))
         provider, _ = _make_provider(fake=fake)
 
-        provider.handle_tool_call("atomicmemory_conclude", {"conclusion": "user prefers terse"})
+        provider.handle_tool_call(
+            "atomicmemory_conclude",
+            {"conclusion": "user prefers terse", "content_class": "summary"},
+        )
         provider.shutdown()
 
         call = fake.calls_to("ingest_verbatim")[0]
@@ -139,6 +142,47 @@ def test_conclude_stamps_provenance_source_hermes(self) -> None:
         self.assertTrue(call.kwargs["provenance"].source_url.startswith("hermes://session/"))
         self.assertEqual(call.kwargs["metadata"], {"kind": "fact"})
 
+    def test_conclude_forwards_the_caller_chosen_content_class(self) -> None:
+        for chosen in ("summary", "redacted", "raw"):
+            with self.subTest(content_class=chosen):
+                fake = FakeAtomicMemoryClient(ingest_response=IngestResult(created=["m1"]))
+                provider, _ = _make_provider(fake=fake)
+
+                provider.handle_tool_call(
+                    "atomicmemory_conclude",
+                    {"conclusion": "a fact", "content_class": chosen},
+                )
+                provider.shutdown()
+
+                call = fake.calls_to("ingest_verbatim")[0]
+                self.assertEqual(call.kwargs["content_class"], chosen)
+
+    def test_conclude_rejects_a_missing_content_class_at_the_tool_boundary(self) -> None:
+        # Fails here with a message naming the parameter, rather than reaching
+        # core and returning 422 raw_content_rejected with no indication of
+        # which tool call caused it.
+        fake = FakeAtomicMemoryClient(ingest_response=IngestResult(created=["m1"]))
+        provider, _ = _make_provider(fake=fake)
+
+        raw = provider.handle_tool_call("atomicmemory_conclude", {"conclusion": "a fact"})
+        provider.shutdown()
+
+        self.assertIn("content_class", raw)
+        self.assertEqual(fake.calls_to("ingest_verbatim"), [])
+
+    def test_conclude_rejects_an_unknown_content_class(self) -> None:
+        fake = FakeAtomicMemoryClient(ingest_response=IngestResult(created=["m1"]))
+        provider, _ = _make_provider(fake=fake)
+
+        raw = provider.handle_tool_call(
+            "atomicmemory_conclude",
+            {"conclusion": "a fact", "content_class": "public"},
+        )
+        provider.shutdown()
+
+        self.assertIn("Invalid content_class", raw)
+        self.assertEqual(fake.calls_to("ingest_verbatim"), [])
+
 
 class SyncTurnUsesIngestMessages(unittest.TestCase):
     def test_sync_turn_sends_structured_messages_via_worker(self) -> None:

plugins/hermes/tools.py

@@ -64,8 +64,19 @@ def tool_error(message: str) -> str:
         "type": "object",
         "properties": {
             "conclusion": {"type": "string", "description": "The fact to store verbatim."},
+            "content_class": {
+                "type": "string",
+                "enum": ["summary", "redacted", "raw"],
+                "description": (
+                    "Sensitivity of the text you are storing: 'summary' if you distilled it "
+                    "yourself, 'redacted' if you removed sensitive spans, 'raw' if it is a "
+                    "verbatim prompt, response, diff, or source excerpt. Required: a core "
+                    "running the default raw-content policy refuses raw or unclassified "
+                    "writes, and this tool will not guess on your behalf."
+                ),
+            },
         },
-        "required": ["conclusion"],
+        "required": ["conclusion", "content_class"],
     },
 }
 
@@ -127,15 +138,33 @@ def _handle_context(provider: AtomicMemoryMemoryProvider, args: dict[str, Any])
     )
 
 
+# Mirrors the v1 ContentClass enum. Kept local so a malformed tool call is
+# rejected at this boundary rather than surfacing later as a core 422 with no
+# indication of which tool produced it.
+CONTENT_CLASSES = ("summary", "redacted", "raw")
+
+
 def _handle_conclude(provider: AtomicMemoryMemoryProvider, args: dict[str, Any]) -> str:
     conclusion = _clean_str(args.get("conclusion"))
     if not conclusion:
         return tool_error("Missing required parameter: conclusion")
+    content_class = _clean_str(args.get("content_class"))
+    if not content_class:
+        return tool_error(
+            "Missing required parameter: content_class (one of "
+            f"{', '.join(CONTENT_CLASSES)}). A core running the default raw-content "
+            "policy rejects unclassified writes; this tool does not choose for you."
+        )
+    if content_class not in CONTENT_CLASSES:
+        return tool_error(
+            f"Invalid content_class {content_class!r}; expected one of {', '.join(CONTENT_CLASSES)}."
+        )
     provider._require_client().ingest_verbatim(
         content=conclusion,
         scope=ingest_scope_dict(user_id=provider._user_id),
         provenance=provider._ingest_provenance(session_id=provider._session_id),
         metadata={"kind": "fact"},
+        content_class=content_class,
     )
     return json.dumps({"result": "Fact stored."})
 

plugins/openclaw/openclaw.plugin.json

@@ -1,7 +1,7 @@
 {
   "id": "atomicmemory",
   "name": "AtomicMemory",
-  "version": "0.1.18",
+  "version": "0.2.0",
   "description": "Persistent semantic memory for OpenClaw agents — cross-channel user memory and deterministic session snapshots via the AtomicMemory SDK's pluggable MemoryProvider model.",
   "kind": "memory",
   "skills": ["./skills/atomicmemory"],

plugins/openclaw/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@atomicmemory/openclaw-plugin",
-  "version": "0.1.18",
+  "version": "0.2.0",
   "description": "AtomicMemory plugin for OpenClaw — persistent semantic memory and deterministic session snapshots across channels.",
   "type": "module",
   "main": "dist/index.js",

plugins/openclaw/skills/atomicmemory/skill.yaml

@@ -1,5 +1,5 @@
 name: atomicmemory
-version: 0.1.18
+version: 0.2.0
 author:
   name: AtomicMemory
   url: https://atomicmem.ai