Read and reuse AI coding session context across Claude Code, Codex, Antigravity/agent hooks, Cursor adapters, and compatible AI tools. Use this skill whenever the user asks to read, inspect, resume, compare, or continue another session by name or session ID, asks what happened in the previous session, or starts a new session that should inherit the nearest prior context. The paired provider adapter automatically loads the nearest previous project session at startup and resolves explicit session-read requests.
中文简介正在准备中。当前显示原文说明。
查看安装方法安装前可以查看交给 AI 代理的指令内容。
The executable implementation is bundled under this skill's hooks/ directory. Copy the complete skill directory when installing session memory; copying only SKILL.md installs the instructions but not the runtime hook. Provider entrypoints prefer this local bundle and accept SESSION_MEMORY_HOOK as an explicit override.
Use the provider adapter to recover focused context from earlier work without pasting an entire transcript into the current context.
On a new Claude Code or Codex startup, the hook reads the newest completed transcript for the current project, excluding the current session. It injects a bounded summary containing the session title, session ID, recent user messages, and recent assistant text. Resume, clear, compact, and subagent starts do not receive this automatic prior-session load.
Claude titles come from transcript title events such as ai-title or custom-title. Codex titles come from ~/.codex/session_index.jsonl (thread_name) and the matching rollout file.
Users can request a specific session with one of these forms:
/session read "session title"
/session-read session-id-prefix
read session "session title"
đọc session "tên session"
Use quotes when a title contains spaces. The hook resolves Claude /rename titles, configured aliases in session-aliases.json, exact session IDs, and unique session ID prefixes.
To inspect available sessions:
/session list
The same lookup is available from a terminal for each adapter:
python .claude/hooks/session_memory.py list
python .claude/hooks/session_memory.py read "session title"
python .codex/hooks/session_memory.py list
python .codex/hooks/session_memory.py read "session title"
python .cursor/hooks/session_memory.py list
Codex and compatible hooks emit plain text. Claude emits its native JSON hook envelope. The shared implementation also supports --provider generic for an AI integration that accepts plain hook stdout and supplies compatible session transcript paths.
.claude/settings.json..codex/hooks.json; its adapter reads ~/.codex/session_index.jsonl and rollout files..agents/hooks.json through the UTF-8 wrapper..cursor/hooks/session_memory.py; this repository has no Cursor hook manifest, so invoke it from the host's hook configuration or CLI.--provider generic and provide a compatible JSONL backend through the host integration. Plain stdout is the expected hook output./session list rather than guessing.还没有评价。欢迎分享你的使用感受。
无需登录即可发表。欢迎分享使用感受、提问，或说明在哪些环境下无法运行。
还没有评论。来写第一条吧。
Build robust backend systems — APIs (REST, GraphQL, gRPC), authentication (OAuth 2.1, JWT), databases, performance optimization, security (OWASP Top 10), scalability patterns (microservices, caching, sharding), and testing. Use when designing APIs, implementing auth, optimizing queries, handling security vulnerabilities, building microservices, or developing production-ready backend systems.
中文简介正在准备中。当前显示原文说明。
Terse output mode — strips filler/articles/hedging, ~75% token reduction. Activate whenever: context is filling up, user says be brief/terse/caveman/less tokens, or CAVEMAN_TRIGGERED appears in context. Persists every response until CAVEMAN_RELEASED or user says stop/normal mode.
中文简介正在准备中。当前显示原文说明。
Explore and debate solutions before writing code. Use when the user poses a design question, asks "how should I approach X", wants to think through options before committing, or says anything like "let's brainstorm", "what's the best way to...", "I'm not sure how to tackle this", or "should I use X or Y". Always runs before /ck:plan for novel or ambiguous features. No code is written — only: explore → clarify → spec → report.
中文简介正在准备中。当前显示原文说明。
Implement a feature phase by phase from a phased JSON master or Markdown plan. Before each phase, confirm whether to create/run unit tests and whether to run ck:quality, unless explicit flags supply those choices. Supports token-budgeted execution, resumable state, and TDD handoff to ck:test.
中文简介正在准备中。当前显示原文说明。
Fix a bug using Scout → Diagnose → Fix → Review → Finalize. Use when the user pastes an error message, stack trace, or test failure, or says "fix this bug", "something's broken", "tests are failing", "I'm getting an error". Also accepts `--from-quality <report>` or `--from-test <report>` to fix exactly the findings/failures a report lists. Modes (pick one): --fast (trivial errors — lint, type, build — skip scout and review), --hard (mandatory review, no auto-approve).
中文简介正在准备中。当前显示原文说明。
Plan a feature or system before implementation. Use when the user says "plan this", "I want to build X", "how do I implement Y", or when /ck:brainstorm produces a spec.md. Always run before /ck:cook. Modes (pick one): --fast (simple, single-file), --hard (research + red-team + validate). Composable flag (combine with any mode): --tdd — propagates into the cook pipeline. Every phase gets Design Constraints and Quality/Testing State; ck:cook confirms optional test and quality checks per phase.
中文简介正在准备中。当前显示原文说明。