Skip to main content

Documentation Index

Fetch the complete documentation index at: https://mintlify.com/XxYouDeaDPunKxX/SensecraftXStudio/llms.txt

Use this file to discover all available pages before exploring further.

Contract activation is not the same as reading a file. An assistant can read AGENTS.md once, store its content in working memory, and then proceed from first-visibility reasoning as if the file were background text. Activation means the file is the operative frame — it governs how every next response is read, bounded, and approached, not only which individual actions are permitted. Until activation is established, the contract exists but does not govern.

​
What Activation Means

The contract defines five numbered rules for activation:
1

The file is an active operational contract

AGENTS.md is not passive text or documentation. It is a live governing frame for the session.
2

Read the file in full before proceeding

Partial reading is not activation. The file must be read end-to-end before any task work begins.
3

The file governs framing, not only action

Activation means this file governs how the next response is read, framed, and approached — not only what actions are permitted.
4

The contract remains active for the session

This contract remains active for the session unless the operator explicitly states it does not apply to the current task. Drift is present when this file is not the active frame for the next response — treating it as a one-time read, memory-only guidance, or background text is drift.
5

If drift is present, re-enter before continuing

If drift is present, do not proceed. Re-entry means: reread this file in full and activate it as the operative frame before continuing.

​
How to Instruct an Assistant to Activate

Place AGENTS.md in the target project or workspace, open a fresh assistant session in an environment that can read workspace files, and issue the minimal operator instruction: 
Read AGENTS.md in full and use it as the operative frame before acting on this workspace.
Then give the real task normally. The file should remain active as a working frame throughout the session — it is not text that was read once and then set aside.

​
Drift

Drift is present when the contract is no longer the active frame. Treating AGENTS.md as a one-time read, relying on it as memory-only guidance, or letting it recede into background text are all forms of drift. Drift does not require explicit abandonment — it can arrive silently as session momentum accumulates. An assistant that continues working after drift is present is operating outside the contract, even if the output looks plausible.
Drift looks like:
  • Responding without checking whether the current move is consistent with the contract’s scope, posture, or stop conditions
  • Using conclusions from earlier in the session without re-grounding them in the contract frame
  • Treating the contract as one input among many rather than the governing frame
  • Expanding scope, skipping authority closure, or skipping the horizontal plane checks

​
Re-entry

Re-entry is the correct response to detected drift. Re-entry means rereading AGENTS.md in full — not skimming a remembered summary — and reactivating it as the operative frame before continuing. A re-entry signal is not a failure event; it is the contract working as designed. Once reactivated, the assistant resumes from the contract frame, not from where session momentum had taken it.

​
Conflict Handling

If another local instruction in the workspace appears to conflict with AGENTS.md, the assistant must not switch silently. The correct behavior is to:
  1. Surface the conflict explicitly
  2. Name which instructions are in tension
  3. Wait for operator clarification before proceeding
Silently deferring to the local instruction — without surfacing that a conflict exists — is a contract violation. The contract does not resolve conflicts by itself; it requires the operator to resolve them.

Build docs developers (and LLMs) love

Get started for freeTalk to us