feat: add AutoGen integration for Hindsight

[DK09876]

Summary

• Adds hindsight-autogen package providing AutoGen FunctionTool instances for persistent long-term memory
• Factory function create_hindsight_tools() returns retain/recall/reflect tools compatible with AssistantAgent(tools=[...])
• Async-native — uses aretain/arecall/areflect directly in AutoGen's async runtime
• Full parameter set: tags, metadata, document_id, recall types, reflect context/schema
• Global config support via configure() with per-call overrides
• 31 unit tests, docs page, integrations.json entry

Test plan

• 31 unit tests pass (uv run pytest tests/ -v)
• Ruff lint and format clean
• Manual E2E test with running Hindsight server
• Cookbook notebook (separate PR in hindsight-cookbook)

🤖 Generated with Claude Code

[nicoloboschi]

Review

Overall: Clean, well-structured integration with good test coverage. A few issues worth addressing before merge:

Must Fix

1. release-integration.sh needs updating
The release script must be made aware of this new integration so it gets published alongside the others.

2. verbose config field is unused
HindsightAutoGenConfig.verbose is declared but never read by any tool function. Either wire it into the tool logging or remove it.

3. Hardcoded defaults scattered across multiple locations
"mid", 4096, "any" appear as fallback defaults in both config.py and tools.py. If the defaults drift, behavior becomes inconsistent. Centralize them as constants or always require config.

4. No validation on enum-like fields
budget accepts any string — no check that it's low/mid/high. Same for recall_tags_match (any/all/any_strict/all_strict). A typo like "medium" would silently pass through and fail at the API level.

5. asyncio.sleep(3) in docs/README
The Quick Start shows await asyncio.sleep(3) after retain with the comment "Wait for Hindsight to finish processing." This is a fragile magic number — it'll confuse users and fail for larger documents. Either document this limitation clearly or provide a polling helper.

6. Missing ruff config in pyproject.toml
The other Python packages in the repo use ruff with line-length = 120. This package has no [tool.ruff] section, so it'll use ruff defaults (88 chars). Should align with the rest of the monorepo.

Minor / Nits

• Optional from typing is used alongside native list[str] syntax — pick one style. Since requires-python = ">=3.10", native str | None is fine everywhere.
• errors.py defines HindsightError but tool functions catch broad Exception and re-wrap — consider catching httpx / client-specific errors instead so programming errors (e.g., TypeError) aren't swallowed.
• _client.py hardcodes timeout=30.0 — should this be configurable or at least match the default in hindsight-client?
• No py.typed marker for PEP 561 type checking support.

Looks Good

• Async-native design matches AutoGen's runtime correctly
• FunctionTool usage is idiomatic for AutoGen
• Test coverage is thorough (31 tests covering tools, config fallbacks, error paths)
• Tool descriptions are clear and will guide LLM tool selection well
• Selective tool inclusion (include_retain/recall/reflect) is a nice touch
• Docs page and integrations.json entry included

[nicoloboschi]

LGTM