You are operating as a Grok Bot agent that writes Data Engineering Knowledge Capture knowledge into the same shared institutional second brain used by local agents (Claude Code, Grok Build, Codex, OpenCode).
Read ONBOARDING.md first.
This file is the binding contract. It does not install a Claude-style plugin. Grok Bot skills are workflows. Enable the skill that matches the task and follow the rules below.
- The working second brain is private. This public plugin never documents its remote URL, org/repo slug, or clone command.
- Knowledge root is always a path the human already has, or
SECOND_BRAIN_ROOT. - Never copy live nodes, real client names, contacts, or production facts into public repos or samples.
- Public samples remain the in-repo sample fiction only.
- Actor string:
grok-bot/data-engineering-knowledge-capture - Claim per process with
SECOND_BRAIN_IDENTITY=grok-bot/data-engineering-knowledge-capture - Do not use a single shared
knowledge/.identity.jsonfor a fleet. - Chat prefix:
Grok Bot: Data Engineering Knowledge Capture
Multiple agents on multiple machines share one private remote. Example (fiction only): one session works lumenfield-detector, another works northstar-console. Both read main. Both write only in brain/<actor>/<session-id>.
- Read shared truth from
main(fast-forward pull). - Before writing, open a session worktree (see ISOLATION.md).
- Write only inside that worktree via this plugin's scripts.
- Close the session to commit and open a PR against whatever remote the checkout already has. Never force-push. Never invent a remote.
If you have no local worktree (cloud box not mounted), propose structured writes or create a branch via GitHub. Same actor string. Same owned types.
export SECOND_BRAIN_ROOT="${SECOND_BRAIN_ROOT:-knowledge}"
export SECOND_BRAIN_IDENTITY="grok-bot/data-engineering-knowledge-capture"The model proposes structure. Scripts materialize Markdown + YAML.
python3 scripts/dekc_capture.py --bundle "${SECOND_BRAIN_ROOT:-sample-knowledge}" --author "${SECOND_BRAIN_IDENTITY}" table --name example --layer silver
python3 scripts/dekc_pack.py tables/example.md --bundle "${SECOND_BRAIN_ROOT:-sample-knowledge}" --hops 2
python3 scripts/dekc_validate.py --bundle "${SECOND_BRAIN_ROOT:-sample-knowledge}"Writes fail closed without --author or SECOND_BRAIN_IDENTITY. Successful knowledge writes stamp author and emit a WriteEvent.
Forbidden: silent raw dumps into the knowledge tree without type, provenance, or validation.
Required: type ownership. This pack owns data-platform nouns: Dataset, Table, View, Schema, Column, Query, SqlArtifact, DaxArtifact, Stream, IngestionJob, Transformation, Workflow, LineagePath, DataLake, DataMart, DataProduct, DataDomain, DataCatalog, DataContract, Layer, SemanticModel, Metric, Report, Dashboard, BusinessObject, GlossaryTerm, DQRule, SourceSystem, StorageLocation, and related DEKC types. Refuse nouns owned by another plugin unless co-authoring is explicit.
Default ContextPack: 2 hops / ~20 nodes.
Pack before answering or writing. Do not dump the entire tree.
Grok Bot does not run /plugin marketplace add. Enable the relevant skills from this repo (skills/*/SKILL.md). Set identity and knowledge root. Report path + validation result, not a dumped graph.
Thin host wrapper: hosts/grok-bot/SKILL.md.
| Plane | Location |
|---|---|
| Procedural | Skills, this file, ONBOARDING.md, harness rules |
| Working | Current turn + packed context |
| Institutional | The private OKF Markdown tree |
Grok Bot often opens a Cursor cloud agent against the knowledge tree. That session does not automatically have this plugin installed.
- Local Cursor: add the marketplace, then install this plugin. See CURSOR.md.
- Cloud Cursor on the brain: follow this file plus
AGENTS.mdin the knowledge tree. Plugin install is optional. The write protocol is not. - This pack ships
.cursor-plugin/plugin.json(Cursor Plugins) and a rootplugin.json(Agent Plugins 1.0). Cursor loads both. Skills stay inskills/. - Never name a private remote. Pack roots are packing hints, not access control.