v0.2.22

# Release Notes for v0.2.22

This release moves conflict detection onto Moorcheh's cloud `agent/run` pipeline, with live progress streamed to the CLI and web UI and support for cancelling a run. `memanto memory sync` now writes high-confidence memories straight into each connected agent's instruction file instead of `MEMORY.md`. The Claude Code integration is folded into `memanto connect`, and connect hooks are more robust. It also adds three integration packages: Vapi voice agents, AG2 (AutoGen) and Amazon Bedrock AgentCore Runtime.

## New Features
- **Cloud conflict detection via Moorcheh `agent/run`** (`memanto/app/clients/agent_conflict.py`, `memanto/app/services/daily_analysis_service.py`)
  - On the cloud backend, conflict reports now call `POST /agent/run` with `task=detect_conflicts` and compare `fact` and `preference` memories for the chosen date. They no longer depend on that day's session summary files existing.
  - If the SSE stream ends before the report arrives, Memanto polls `GET /agent/runs/{run_id}` until the report is ready (300s default timeout).
  - Results are converted to the existing conflict schema: compatible and duplicate pairs and self-referencing pairs are dropped, and `keep_both` becomes `merge`. `created_at` and `source` are then added from document metadata.
  - Non-cloud backends keep the previous LLM-based path, which reads session summaries.
- **Streaming conflict-report endpoint** (`memanto/app/ui/routes/ui_router.py`)
  - New `POST /api/ui/conflicts/generate/stream` returns SSE events: `progress` (with `message`, `step` and `detail`), `result`, `error` and `done`. It sends `: keepalive` comments during long waits.
  - When the client disconnects, the endpoint cancels the upstream stream and polling through a `cancel_event`.
  - Added to the TypeScript SDK's `sdks/typescript/openapi.json`.
- **Live conflict-detection progress in the UI and CLI** (`memanto/app/ui/static/index.html`, `memanto/cli/commands/memory.py`)
  - The web UI shows a step tracker (Starting → Loading & similarity → Analyzing contradictions → Complete) along with the Moorcheh run ID.
  - The CLI conflict-detection command shows live status messages and prints the run ID when it finishes.
- **Dynamic memory injection for `memanto memory sync`** (`memanto/cli/commands/memory_mgmt.py`, `memanto/cli/connect/updater.py`)
  - `memanto memory sync` now recalls the most relevant active `instruction`, `preference` and `goal` memories (`min_confidence` 0.8). It writes them into the dynamic sentinel section of each connected agent's instruction file and `SKILL.md`, instead of exporting `MEMORY.md`.
  - The target connection is detected automatically. New `--connection` and `--scope local|global` options set the target explicitly when there's more than one candidate.
  - `--limit` now defaults to `10` (previously `25`). For `--okf` it is still the maximum per type.
  - If no memories match, the dynamic sections are cleared.
- **Vapi integration: `memanto-vapi`** (`integrations/vapi/`)
  - A webhook server (`memanto-vapi serve`) that gives Vapi voice assistants a lasting memory:
    - At call start, context is injected through `{{memanto_context}}`.
    - During the call, the assistant has `memanto_recall` and `memanto_remember` tools.
    - At call end, the transcript is turned into lessons.
  - `MEMANTO_VAPI_SCOPE=shared` (the default) keeps one memory per assistant. `caller` adds private memories per caller, tagged with an HMAC of the phone number or `customer.externalId` keyed by `MEMANTO_VAPI_CALLER_SALT`.
  - In caller scope, anything the assistant saves mid-call stays private to that caller. Shared lessons come only from end-of-call extraction, so a caller can't push their details into other callers' context.
  - Requests are authenticated with `MEMANTO_VAPI_SECRET`. A malformed JSON body returns `400`, not `500`.
- **AG2 (AutoGen) integration: `memanto-ag2` (preview)** (`integrations/ag2/`)
  - `register_memanto_tools()` registers remember, recall and answer as AG2 `register_for_llm`/`register_for_execution` tools. `include_remember`, `include_recall` and `include_answer` select which tools to register.
  - Binds the `SdkClient` session in a thread-safe way, handles agents that already exist, and supports AG2 0.9.x schema/`LLMConfig`.
- **Amazon Bedrock AgentCore Runtime integration: `memanto-agentcore` (preview)** (`integrations/agentcore/`)
  - `MemantoRuntimeAdapter` runs recall → execute → retain on each handler turn, using `before_turn`, `after_turn` and `run_turn`. Retain runs asynchronously by default.
  - Memory is keyed by a stable `tenant-{tenant_id}-user-{user_id}-agent-{agent_name}` ID, not `runtimeSessionId`, so it survives when AgentCore sessions are recycled. A missing `user_id` raises `AgentResolutionError`.

## Improvements
- **Claude Code integration merged into `memanto connect`** (`integrations/claudecode/`, `memanto/cli/connect/engine.py`)
  - `claudecode-memanto` 0.1.1 is now a thin `install`/`uninstall` wrapper around `memanto connect`'s `claude-code` install. It requires `memanto>=0.2.21`.
  - Removed the separate hook scripts, LLM extractor, profile and skill-map modules, which the core integration now covers.
- **More robust connect hooks** (`memanto/cli/connect/assets/hooks/`, `memanto/cli/connect/engine.py`)
  - Hooks now call the absolute Python interpreter (`"<sys.executable>" -m memanto`) and scripts from the installed package. They no longer rely on `python`/`python3`/`memanto` being on `PATH`, or on copying scripts into the agent's config directory.
  - Memanto-owned hooks are marked `"_managed_by": "memanto"`. Install and uninstall remove only Memanto's own commands and keep other tools' hooks, even in a shared hook group. Reinstalling replaces Memanto's hooks without adding duplicates.
  - Removed the hardcoded fallback `SessionStart` payloads from the agent registry.
- **Agent instruction templates v1.0.1** (`memanto/cli/connect/templates.py`)
  - Merged the recall and remember triggers into one 10-point `<thinking>` checklist, with a mandatory `memanto memory sync` and recall at the start of each session.
  - Added schema rules for valid types, provenance values and the confidence range, plus worked examples, to cut down on made-up metadata.
  - Existing installs are flagged as outdated so they can be updated with `memanto connect update`.
- **UI agent resolution** (`memanto/app/ui/routes/ui_router.py`)
  - Conflict and daily-summary UI routes now fall back to the browser session cookie when there's no explicit `agent_id` and no CLI session marker.
- **Learning-type parsing heuristic** (`memanto/app/services/memory_parsing_service.py`)
  - Learning detection now matches "in retrospect", "looking back" and "going forward" in place of "in hindsight".

## Bug Fixes
- **Copilot `applyTo` frontmatter duplicated on template updates** (`memanto/cli/connect/engine.py`)
  - Updating a `*.instructions.md` file that already had `applyTo` frontmatter added a second frontmatter block. The update now replaces only the sentinel block and skips frontmatter that's already present.

## Tests
- `tests/test_agent_conflict_client.py`: parsing SSE events, progress descriptions, finalizing and polling runs, and cancellation during streaming and polling.
- `tests/test_dynamic_memory_injection.py`: selecting the connection and scope for sync, and making sure global scope never changes local instruction files.
- `tests/test_daily_analysis_query_length.py`: switched to checking the agent-based conflict detection path.
- `tests/test_connect_engine.py`, `tests/test_cli.py`, `tests/test_unit.py`: updated for managed-hook filtering, conflict-report progress callbacks and the new `memory sync` behavior.
- `integrations/vapi/tests/`, `integrations/ag2/tests/`, `integrations/agentcore/tests/`: unit and live test suites for the new integrations.

## Full Changelog
Full Changelog: https://github.com/moorcheh-ai/memanto/compare/v0.2.21...v0.2.22