跳至正文
OpenAgentSkill
浏览已开放 · 目录记录日期 2026-11-21 · 自动更新和账户服务暂时关闭。
搜索与研究

nacl-init

Inspect or initialize a NaCl project with a per-project Neo4j Community graph and project-local MCP. Use for first setup, bootstrap, and repair planning.

ITSalt · 已有静态检查记录

安装与使用

运行命令前先阅读来源与依赖说明。复制命令不会自动安装或执行。

npx skills add ITSalt/NaCl --skill nacl-init

技能说明


name: nacl-init description: Inspect or initialize a NaCl project with a per-project Neo4j Community graph and project-local MCP. Use for first setup, bootstrap, and repair planning.

NaCl Init

Read the Skills-only runtime contract.

This entry must work before any project MCP exists. Never call an installation doctor, a package gateway, or a checkout-relative/global skill path.

Resolve one explicit absolute project root. For an empty new project without config.yaml, invoke the bundle-relative project-creation planner with the root, project name, optional one-line description, and optional stack. It performs no writes. Present its exact config.yaml, AGENTS.md, Git action, planHash, and CREATE_NACL_PROJECT:<sha256> confirmation, then stop.

After the user repeats that exact confirmation, invoke the bundle-relative project-creation applier with the same inputs, plan-hash, and confirmation. It recomputes the plan under a create lock before writing. For a genuinely empty non-Git directory it creates config.yaml and concise repository-specific AGENTS.md, initializes Git and commits only those initial artifacts. It preserves an existing AGENTS.md and never auto-adopts a non-Git directory containing files or a linked worktree. Stop on any non-success result.

AGENTS.md records durable project context, constraints, and verified commands; keep it concise and use a closer nested AGENTS.md for directory-specific rules. Never put credentials in it. For an existing config.yaml without a stable project.id, present the exact add-only config change and stop for confirmation before writing it; never derive identity during a read.

Before bootstrap, invoke the bundle-relative plan runner with the explicit root, ID, database, and chosen loopback Bolt/HTTP ports. This command is read-only: it performs no Docker call, network request, or mutation. Show its exact plan fields, planHash, and fresh INIT_LOCAL_GRAPH:<project-id>:<sha256> token, then stop. Never accept the old static token or reconstruct a token manually.

After the user repeats that exact token, invoke the bundled POSIX or PowerShell runner, respectively the POSIX runner or PowerShell runner, with the same root, ID, database, ports, and token. Never pass a password. The runner recomputes the plan immediately before its first mutation; stale or mismatched state is BLOCKED/PLAN_TOKEN_STALE with zero mutation.

Treat the exact token as an execution request, not as a reason to re-plan, diagnose, or ask for another confirmation. Run the selected runner exactly once, capture its complete stdout and stderr, and do not infer project state while it is running. The runner's terminal machine-readable line is the sole authority for the outcome:

  • NACL_SKILLS_ONLY_BOOTSTRAP: status=... code=... is the successful bootstrap receipt.
  • NACL_GRAPH_RESULT: status=... code=... is a blocked, failed, or partially verified bootstrap receipt.

Report those fields verbatim. Never claim PARTIAL_BOOTSTRAP_STATE, missing files, or an incomplete bootstrap from an LLM inference, a directory name, or a prior message. If the runner reports FAILED with rollback=VERIFIED, report that failure and rollback; the graph bootstrap is not partial and no recovery plan is needed. A retry requires a newly generated plan and fresh user confirmation. Only if the runner reports PARTIALLY_VERIFIED with rollback=INCOMPLETE, invoke plan-project-graph.mjs --diagnose-only once, then report the returned JSON evidence without embellishment. If no terminal receipt is produced, report RUNNER_RECEIPT_MISSING and run that same read-only diagnosis; do not assert a state before its result.

Preserve the runner's status/code. A successful runner returns PARTIALLY_VERIFIED/RESTART_REQUIRED with bootstrap=VERIFIED and initialization=NOT_RUN. Then stop and ask the user to open a new task in this project so Codex loads the newly created project .codex/config.toml. The current task must never report overall initialization VERIFIED.

In the new task, overall initialization is VERIFIED only after all of these same-task gates succeed through the actual project nacl_neo4j MCP:

  1. Record real MCP initialize and tools/list discovery, requiring the exact read-cypher and write-cypher tool names. Missing discovery is closed non-success.
  2. Run graph connectivity health, read the complete nacl-graph-gateway schema ledger, require versions 1–3 with their packaged checksums, and read back every required constraint.
  3. Execute the bundled named read sa_statistics_extensions unchanged and record its result. A generic RETURN 1 alone is insufficient.
  4. Invoke the plan runner with --verification-plan, show its random idempotency key, parameterized write/read-back statements, planHash, and exact VERIFY_NACL_INITIALIZATION:<project-id>:<sha256> token, then stop.
  5. After the user freshly repeats that token, make exactly one parameterized write-canary call with the plan's statement/parameters, followed by a separate read call with the read-back statement. Require matching project ID, idempotency key, and integer revision. Never accept a static, old, or reconstructed verification token. Never automatically retry the write; any failure requires a new verification plan, new idempotency key, and fresh user confirmation.

Return these exact result fields: status, code, initializationState, mcpServerKey, mcpInitialize, mcpToolsList, readTool, writeTool, graphHealth, schemaVersion, schemaChecksum, namedRead, writeCanary, and writeReadback. Set `s

阅读完整来源 ↗