Imported from BrightWayAI/claude-cortex (.agents/skills/start-nucleus/SKILL.md). Install upstream with npx skills add BrightWayAI/claude-cortex --skill start-nucleus. Copyright stays with the author.

Cortex: start-nucleus

This is a thin host adapter. Starting from this loaded SKILL.md file, walk upward to the nearest directory containing AGENTS.md, commands/start-nucleus.md, and scripts/cortex_cli.py; call it <cortex-repo-root>. Do not assume the current working directory is the Cortex repository. If no such root is readable, stop.

Read <cortex-repo-root>/AGENTS.md, <cortex-repo-root>/commands/start-nucleus.md, <cortex-repo-root>/references/core-contract.md, and the current host's column (Claude Code, Cowork, Codex, or ChatGPT Work) in <cortex-repo-root>/references/capability-matrix.md completely before acting. The command file is the workflow authority. If those sources conflict, stop and report the conflict instead of inventing behavior.

Invocation

On Claude, use the text following /start-nucleus; on Codex, use the text following $start-nucleus; in ChatGPT Work, use the user's text after selecting Cortex or this skill with @. Natural-language activation is also allowed when the request clearly matches the description.

Host binding

1. Resolve <config-root> through the shared scripts/lib/config_root.py precedence chain. On filesystem-capable hosts, <cortex-repo-root>/hooks/session_start.py --print-root is available. In cloud Work, call cortex_status; the MCP bridge resolves the same chain on the machine that owns the memory. Do not create a host-specific pointer.
2. Require filesystem.read for memory access. On Claude Code, Codex, or Local Work, the resolved root must be inside the host's readable/writable roots as appropriate; Cowork requests its directory grant. In cloud Work, use cortex_recall and cortex_search; never claim direct access to device files.
3. Bind every optional operation through references/capability-matrix.md. Skip unavailable connectors, web search, scheduling, artifacts, or delegation exactly as that capability's degrade rule says.
4. Treat model-tier names in the canonical workflow as cost/latency intent. Use a host-available low-cost/fast model when possible; otherwise run inline and disclose the degradation.

Mutation boundary

For every local write under <config-root>/memory/, shell out to <cortex-repo-root>/scripts/cortex_cli.py or import the existing implementation from <cortex-repo-root>/scripts/lib/. In cloud Work, use cortex_configure only for a confirmed first-run pointer setup; for memory mutations use only cortex_add_note, cortex_update_section, cortex_reindex, and cortex_refresh_hot. Show the exact proposed change and obtain confirmation before setting user_confirmed=true. Replacing an existing pointer requires a separate confirmation before force_pointer=true. The bridge intentionally has no move, delete, append-line, or whole-file-write tool; preview or skip those steps. Never hand-edit memory files and never reimplement locking, atomic replacement, index generation, or hot-cache generation.

If the canonical step has no deterministic CLI/library path, show a proposed change or skip that step; do not turn prose into an unreviewed write. Session-end hooks are best-effort reminders, never proof that /remember completed.

Adapter status: partial

This adapter is intentionally partial. Never imply that its unavailable steps completed. Source access is host-specific and canonical bulk writes are not fully CLI-wired. In ChatGPT Work, preview or skip any step for which the bounded MCP bridge has no mutation tool.

End with a concise account of what was read, what changed, which optional capabilities were unavailable, and which partial steps were left unapplied.