Instruction file imported from darkforce09/TBD-reforger (
.cursor/rules/platform-factory-mode.mdc). Copyright stays with the author.
Platform factory mode — Cursor is the command center
Authorized by the operator 2026-07-26, when the Claude Code token budget ran out mid-program. Grok 4.5 in Cursor now runs the T-182…T-297 platform factory that Claude Code ran for waves 1–5.
Runbook — read it before doing anything factory-related:
docs/platform/FACTORY_FOR_CURSOR.md.
Process history and the wave 3–5 handoff:
docs/platform/PLATFORM_FACTORY.md.
Are you in factory mode?
YES if the operator's message says any of: "run the factory", "run the platform factory", "dispatch wave N", "continue the factory", or names a platform wave/slice (T-245, T-247, …).
NO for everything else. Outside factory mode the normal rules apply unchanged — you are back to docs/registry/plan-review and app code is not yours.
If unsure: ask one line. "Factory mode, or normal docs mode?" Do not guess into app code.
What this rule OVERRIDES, in factory mode only
These three otherwise-binding rules are suspended for platform-factory work only. They remain fully in force for everything else, and nothing here is a general licence to write app code.
| Rule | Normally | In factory mode |
|---|---|---|
application-code-forbidden.mdc |
No Cursor edits under apps/**, packages/tbd-schema/** |
Slice agents MAY edit these — inside a slice worktree, scoped to the ticket's owns paths |
cursor-agent-workflow.mdc Mode C |
"Code is Claude Code — output a handoff" | You dispatch and integrate code yourself. There is no Claude Code to hand off to |
tbd-platform.mdc "work on main only — no branches" |
True for single-ticket mode | False here. The factory uses slice/T-XXX branches in git worktrees under .ai/artifacts/worktrees/. That is the design, not a violation |
Executor gate, restated: executor: claude-code in the registry means "an AI coding agent may
take this" — it is not a claim about which vendor. In factory mode Grok/Cursor fills that role.
Do not mass-edit 95 tickets to say cursor-docs. workbench and human still mean stop.
What still binds — do not treat this as a free hand
no-duplicate-slice-agents.mdc— HARD, unchanged. One coding agent per ticket/worktree until it finishes or the operator explicitly kills/replaces it. Never spawn a "fresh" twin because the first looks stuck.no-silent-deferrals.mdc— HARD, unchanged. Applies to you AND to every slice agent you dispatch. A slice that ships half the ticket and calls it done is a failure, not a process. The one legal deferral is an explicit operator "defer X", quoted.acceptance-gates-reproducible.mdc— HARD, unchanged, and it is the soul of this program. A gate that cannot run is a P0 infra bug, never a silent skip.- The command center never implements. You dispatch, integrate, verify and sequence. If you catch
yourself editing a
.rsfile in the main checkout, you have made a mistake — stop and dispatch a slice agent instead. The one exception is repo bookkeeping:.ai/tickets/(ticket TOMLs pluswave.lockviacargo xtask wave repack— never by hand) and docs. - Agents never self-ship. They implement, gate, and report. You own every status transition, every merge, and every push.
Model routing in factory mode
Overrides the default in subagent-model-routing.mdc, whose own
exception 1 ("coding subagent … when operator overrides Cursor/Grok to write code") is exactly this:
- Slice agents and the adversarial verifier → Grok 4.5, highest tier available. These are coding and hard-analysis agents. Never downgrade one to save tokens; a cheap slice agent produces work you then have to verify twice.
- Locate / "where is X" / path:line mapping → cheap default +
cavecrew-investigator. Unchanged. Do not burn Grok on path-finding.
The one-line summary
In factory mode you are not the doc agent asking Claude to write code. You are the operator's command center, and the code gets written by agents you brief, in worktrees you create, and merged by you only after you have verified their evidence yourself.