[good first issue-platform adapt-Claude Code] Add native Memory plugin for Claude Code (hooks + MCP, no proxy)

[plainwuatlig]

Description | 描述

Adds MemoryCore/claude-code-plugin/: a native Claude Code client for Memory Gateway /v3/*, in the same shape as the OpenClaw client adapter (MemoryCore/openclaw-plugin/): a pure client over @tencentdb-agent-memory/memory-sdk-ts-v2, no extraction or storage, and no LLM-traffic proxy. Claude Code keeps talking to Anthropic directly; memory arrives through Claude Code's own lifecycle hooks and a stdio MCP server. It sits beside the existing proxy route (agents/claude-code/) as the choice for teams that do not want to move model traffic, keys, or billing to the proxy.

新增 MemoryCore/claude-code-plugin/：Memory Gateway /v3/* 的 原生 Claude Code 客户端，形态与 OpenClaw 客户端适配器一致——基于 @tencentdb-agent-memory/memory-sdk-ts-v2 的纯客户端，不做抽取和存储，不经过 LLM 流量代理。Claude Code 仍直连 Anthropic；记忆通过 Claude Code 自身的生命周期 Hook 与一个 stdio MCP server 接入。与现有代理路线（agents/claude-code/）并列，供不希望把模型流量、密钥与计费转到代理的团队选用。

Hooks (one process per event; state shared via hashed files in a state dir)

• UserPromptSubmit: searchAtomic for the prompt on every turn; readCore + listScenarios once per session; injected as additionalContext (<user-persona>, <scene-navigation>, <relevant-memories>, tool guide). A total Gateway outage injects nothing.
• Stop: the finished turn is read from the transcript named by transcript_path (fallback: the Claude Code projects directory) and sent as one addConversation: prompt, tool calls, tool results, intermediate and final text. Thinking/images dropped; tool traffic quoted with size caps; 8192-char chunks; 100 messages per batch. Credential-shaped substrings are redacted before sending (tool traffic carries env dumps and config files). Falls back to prompt + reply when no transcript can be read. Skipped while background_tasks / session_crons are present.
• SessionEnd: whatever transcript remains unsent. A failed batch leaves the marker at the last batch that landed and the turn unclaimed, so a later hook sends the rest.

MCP server (stdio): tdai_memory_search, tdai_conversation_search, tdai_scenario_read, tdai_memory_capture (a milestone in the agent's own words), and, when TDAI_KNOWLEDGE_URL is set, tdai_wiki_list / search / pages / read / write against the Knowledge Service /v3/wiki API. Wiki ids are always tool parameters; identity comes from TDAI_* env; the plugin carries no organisation-specific content.

Docs: bilingual README.md / README_CN.md, integrations/hooks.json, integrations/mcp.json.example; one row added to the agent tables in INSTALL.md / INSTALL_CN.md.

Related Issue | 关联 Issue

#926 (Adapters Wanted, Claude Code native route). Complements #1126's proxy route for Pi and agents/claude-code/ for Claude Code rather than replacing them.

Change Type | 修改类型

• Bug fix | Bug 修复
• New feature | 新功能
• Documentation update | 文档更新
• Code optimization | 代码优化

Self-test Checklist | 自测清单

• Verified locally | 本地验证通过 — npm run typecheck, npm run build, npm test (17 vitest tests: transcript normalisation and redaction, hook flows incl. partial failure and retry, MCP round-trips, Knowledge client envelopes). Live against a running Gateway + Knowledge Service: first-prompt recall 178 ms with persona/guide/L1, second prompt L1 only; Stop sent the turn in 346 ms with the secret redacted; queryConversation read back all four messages; MCP server lists 9 tools and returns real wiki and memory hits.
• No existing features affected | 无影响现有功能 — new directory only; two table rows in the install guides.

Additional Notes | 其他说明

Design choices worth a reviewer's eye: (1) per-turn capture at Stop rather than only at SessionEnd, because a 100-message batch costs the Gateway a few seconds and a session's SessionEnd fires once, so end-only capture loses the tail of long sessions; (2) redaction on the client, because the Gateway records what it is sent; (3) the MCP wiki tools include wiki_write, which MemoryKnowledge's own MCP server does not yet expose — happy to move that there instead if preferred.

Commit is DCO signed-off.

🤖 Generated with Claude Code

https://claude.ai/code/session_01EkVfNqVBnTAHnc62NUieKj

[Maxwell-Code07]

Thank you so much for your attention and contribution! We will arrange an internal review for this PR shortly, and all feedback will be shared right here in the discussion.