This repository starts from a strict premise: ni is the Project Intent
Compiler for AI Agents, not a generic agent harness.
Don't run the agent yet. Compile the intent first.
The first implementation target is a planning kernel:
conversation -> docs contract -> readiness gate -> lockfile -> 4000-char goal prompt
Downstream harnesses and execution systems may consume locked output later, but they must not become kernel-owned runtime behavior.
Use available prompts in numeric order. The original bootstrap sequence is:
000-readonly-review.md
001-reframe-docs-to-ni-kernel.md
002-bootstrap-cli.md
003-ni-init-docs-template.md
004-contract-model.md
005-readiness-status.md
006-lockfile-end.md
007-prompt-compiler-run.md
008-codex-skills.md
009-work-graph-proposal.md
010-generated-harness-contract.md
011-dogfood-ni-project.md
012-codex-exec-experiment-later.md
Later follow-up prompts may use higher task numbers, such as
029-repo-consistency-audit.md. Stop after each prompt, review the diff, run
validation, and commit.
If the model says planning is complete but ni status says blocked, planning is blocked.
If the model says implementation can start but .ni/plan.lock.json is missing or stale, implementation is blocked.