Proposal: prompt lineage in the knowledge graph

Record every prompt as a node in the knowledge graph, so a reviewer can trace a commit or plan back to the prompt that produced it.

Proposal, not built. No prompt record exists in the knowledge graph or in .robos/ today.

Proposal

A robos:Prompt node type with SHACL validation. Fields named in the spec include the prompt text, system prompt, origin, app, agent session, model, provider, token usage, duration and outcome.
An origin on every prompt, so a reviewer can tell how it started: typed in a RobOS app, run by an autonomous RobOS job, run in a terminal outside RobOS, delegated by another agent, or generated by a benchmark.
Three capture points: submissions from <robos-ai-textarea>, agent sessions started by RobOS, and a shell wrapper for claude, gemini and gh copilot run in a terminal.
Append-only JSON-LD files under .robos/prompts/YYYY-MM-DD/, committed to Git.
Edges from a prompt to the plan, task and commit it led to.
A “Prompt History” view in Dev Central and a replay button that loads an old prompt back into the text area.

All logging would stay local.

<robos-ai-textarea> is a shared element in packages/robos-ui and is used in many app windows.
AI Prompt keeps its last 50 runs (packages/ai-prompt/main.js).
packages/robos-graph defines robos:PromptStrategy, robos:PromptOptimizer and robos:PromptCompiler. These describe prompting techniques, not individual runs. There is no robos:Prompt type and no .robos/prompts/ directory.

Raw note: ideas/inbox/prompt-knowledge-graph-logging.txt. The earlier long spec is in the git history of this file.