An operating discipline for autonomous agent fleets — and the launcher for one.
This repository has no application code. Its artifacts are the product: five role prompts, per-role standing goals, fleet instruments, and the Daintree recipe that launches nine agents into one workspace.
Give this line to a coding agent working in the repo you want the fleet in:
Clone https://github.com/nForma-AI/nForma-NEXT, then read
onboard.mdat the root — it contains your full instructions for onboarding me through this tool.
That is the whole instruction. onboard.md carries the rest: it detects what is
already present, copies the seven files the later steps invoke, re-scopes the goal files to your
repository, validates the recipe, gives every role its own worktree, and ends by running
scripts/fleet-preflight.sh as the acceptance test.
⇒ Every step is idempotent. Running it twice changes nothing the second time, and a step that skips says so — because a silent skip and a step that ran are otherwise the same output.
| pane | fundamental question |
|---|---|
TEAMLEAD |
is this work reaching a justified terminal state? |
ARCHITECT |
is this technically correct, and does the evidence bind? |
DEVOPS |
does the substrate do what we believe it does? |
DX |
is the way we work getting better, and how would we know? |
DEV1–DEV5 |
implementation |
| # | file | the question it answers |
|---|---|---|
| 1 | docs/FOUNDING-THESIS.md |
why does this repository exist, and what was measured to justify it? |
| 2 | prompts/README.md |
who are the roles, and how does the fleet launch? |
| 3 | tools/README.md |
what can be measured here, and what does each instrument refuse to claim? |
| 4 | goals/README.md |
what must a role's standing goal contain, and who owns it? |
CLAUDE.md is the map for an agent already working here. It is a map, not doctrine —
if it and a README disagree, the README wins and the map is the defect.
⛔
exit 2means established nothing, and must never be read as all clear.
Most of what this repository has learned is a variation on that. A check that never ran and a check
that passed produce the same silence. A missing validator and a passing one produce the same
silence. A pinned tool that is current and one frozen forty commits back both report pinned.
⇒ The instruments here are built to refuse a verdict rather than report a clean one they cannot
support — which is why several of them exit 2 more often than they exit 0.
| path | contents |
|---|---|
onboard.md |
the entry point for installing this fleet elsewhere |
docs/ |
the thesis the rest is derived from |
prompts/ |
the five role prompts |
goals/ |
per-role standing goals, and RESERVED-ACTIONS.md — the single source for what is reserved to whom |
tools/ |
fleet instruments (Python, run directly) |
scripts/ |
recipe validator, worktree manager, preflight, conformance checkers |
.daintree/recipes/ |
the ten-pane launcher |