feat(integrations): add AutoGen adapter + integration CI; fix broken adapters (#170)

Roadmap Phase 1, Week 3-4: Framework Adapters.

New: AutoGen (AG2 / autogen-core v0.4+) adapter
- mimir_autogen.MimirMemory implements the autogen_core.memory.Memory protocol
  (add/query/update_context/clear/close), maps MemoryContent onto Mimir entities,
  injects mimir_context as a SystemMessage before each turn
- Persistent stdio session, pyproject, README, 7 passing tests (framework stubbed)

Fixes uncovered while adding CI (no CI ran the Python tests before):
- crewai + langgraph: init handshake read the wrong JSON-RPC id (req_id-1=0 vs
  init id 1), causing a 30s stall on every real session start. Now reads init_id
- crewai + langgraph tests: FakePopen still used the old communicate() API after
  the adapters moved to persistent write/readline stdio — every test errored with
  'FakePopen has no attribute stdin'. Rewrote both fakes to drive the stdio session
- langgraph adapter passed nonexistent --ollama-url and misused --embedding-model
  (expects an ONNX path, not a model name). Now maps to --llm-endpoint/--llm-model

New CI: .github/workflows/integrations.yml — matrix-tests all three adapters
(autogen, crewai, langgraph) on every integrations/** change. langgraph installs
the real package since it imports langgraph.store.base at module load.

Verified: 7 (autogen) + 3 (crewai) + 6 (langgraph) = 16 tests passing locally.

Author: tcconnally

.github/workflows/integrations.yml

@@ -0,0 +1,44 @@
+name: Integrations
+
+on:
+  push:
+    branches: [main, master]
+    paths:
+      - "integrations/**"
+      - ".github/workflows/integrations.yml"
+  pull_request:
+    branches: [main, master]
+    paths:
+      - "integrations/**"
+      - ".github/workflows/integrations.yml"
+
+jobs:
+  test:
+    runs-on: ubuntu-latest
+    strategy:
+      fail-fast: false
+      matrix:
+        include:
+          # autogen & crewai tests stub their framework deps — no install needed.
+          - adapter: autogen
+            extra_deps: ""
+          - adapter: crewai
+            extra_deps: ""
+          # langgraph adapter imports langgraph.store.base at module load,
+          # so the real package must be present to collect the tests.
+          - adapter: langgraph
+            extra_deps: "langgraph"
+    steps:
+      - uses: actions/checkout@v4
+
+      - name: Set up Python
+        uses: actions/setup-python@v5
+        with:
+          python-version: "3.11"
+
+      - name: Install test deps
+        run: pip install pytest pytest-asyncio ${{ matrix.extra_deps }}
+
+      - name: Run ${{ matrix.adapter }} adapter tests
+        working-directory: integrations/${{ matrix.adapter }}
+        run: python -m pytest tests/ -v

integrations/README.md

@@ -7,9 +7,10 @@ Ready-to-use adapters that connect Mimir to popular AI agent frameworks.
 | Framework | Type | Directory |
 |---|---|---|
 | **LangGraph** (LangChain) | `BaseStore` implementation | [`langgraph/`](langgraph/) |
+| **CrewAI** | Agent Tool | [`crewai/`](crewai/) |
+| **AutoGen** (AG2 / autogen-core) | `Memory` implementation | [`autogen/`](autogen/) |
 | **Claude Code** (Anthropic) | MCP server config | [`../docs/integration/claude-code.md`](../docs/integration/claude-code.md) |
 | **Cursor** | MCP server config | [`../docs/integration/cursor.md`](../docs/integration/cursor.md) |
-| **CrewAI** | Agent Tool | [`crewai/`](crewai/) |
 
 ## Adding a New Integration
 

integrations/autogen/README.md

@@ -0,0 +1,89 @@
+# Mimir for AutoGen
+
+Persistent long-term memory for [AutoGen](https://github.com/microsoft/autogen)
+(AG2 / `autogen-core` v0.4+) agents, backed by [Mimir](https://github.com/Perseus-Computing-LLC/mimir).
+
+`MimirMemory` implements the `autogen_core.memory.Memory` protocol, so it drops
+straight into an `AssistantAgent(memory=[...])`. Stored knowledge is injected
+into the model context before each inference, giving your agents memory that
+survives across sessions, processes, and crews.
+
+## Install
+
+```bash
+# Install Mimir (the binary)
+curl -sSL https://raw.githubusercontent.com/Perseus-Computing-LLC/mimir/main/scripts/bootstrap.sh | bash
+
+# Install the adapter
+pip install -e integrations/autogen
+```
+
+## Usage
+
+```python
+import asyncio
+from autogen_agentchat.agents import AssistantAgent
+from autogen_ext.models.openai import OpenAIChatCompletionClient
+from mimir_autogen import MimirMemory
+
+
+async def main():
+    memory = MimirMemory(db_path="~/.mimir/data/agent.db")
+
+    # Seed a fact
+    from autogen_core.memory import MemoryContent, MemoryMimeType
+    await memory.add(MemoryContent(
+        content="The user prefers TypeScript over JavaScript.",
+        mime_type=MemoryMimeType.TEXT,
+        metadata={"category": "preferences", "key": "language"},
+    ))
+
+    agent = AssistantAgent(
+        name="assistant",
+        model_client=OpenAIChatCompletionClient(model="gpt-4o"),
+        memory=[memory],
+    )
+
+    result = await agent.run(task="What language should I use for this project?")
+    print(result.messages[-1].content)
+
+    await memory.close()
+
+
+asyncio.run(main())
+```
+
+## How it maps to Mimir
+
+| AutoGen `Memory` method | Mimir tool | Behavior |
+|---|---|---|
+| `add(MemoryContent)` | `mimir_remember` | Content → `body_json`; `metadata.category`/`metadata.key` route the entity |
+| `query(text)` | `mimir_recall` | FTS5 keyword search → list of `MemoryContent` |
+| `update_context(ctx)` | `mimir_context` | Prepends the rendered memory block as a `SystemMessage` |
+| `clear()` | `mimir_prune` | Soft-deletes (archives) this memory's category |
+| `close()` | — | Shuts down the persistent Mimir stdio process |
+
+## Configuration
+
+```python
+MimirMemory(
+    binary="mimir",                       # or absolute path: /usr/local/bin/mimir
+    db_path="~/.mimir/data/mimir.db",
+    category="autogen",                   # default category for add()
+    context_limit=10,                     # entities injected by update_context()
+    encryption_key="~/.mimir/secret.key", # optional AES-256-GCM at rest
+    llm_endpoint="http://localhost:11434/api/generate",  # optional, for hybrid search
+    llm_model="nomic-embed-text",         # optional embedding/RAG model
+)
+```
+
+## Notes
+
+- The adapter keeps a **persistent** Mimir stdio session — the process is
+  spawned once and reused across all calls (no per-call cold start). Call
+  `await memory.close()` when done, or let `__del__` reap it.
+- `add()` accepts `metadata={"category": ..., "key": ...}` to control where the
+  entity lands. Without an explicit key, a timestamped key is generated so
+  repeated `add()` calls never collide.
+- Use a **project-specific** `db_path` to isolate memories per agent or per
+  workspace.