agent's rules of engagement dropped from context mid-run on authorized test

TL;DR

Anything held only in conversation context will eventually fall out of it, and rules of engagement are the worst possible thing to lose: scope limits, testing windows, and prohibited actions must survive every context reset. Move them into a pinned engagement record that every tool call checks, targets validated against the scope list and rates capped by the record, so the rules hold even when the agent's memory does not.

agent's rules of engagement dropped from context mid-run on authorized test

Use this when

• Mid-run, the agent proposes or attempts actions outside the agreed scope, window, or intensity, despite being briefed at the start.
• After a long run or a context compaction, the agent asks again what it is allowed to test, or worse, does not ask.

Not for this skill when

• The client expanded the scope in writing mid-engagement. Update the pinned record through the approval path; the agent following the new scope is correct.
• The agent is a human-supervised assistant making suggestions only. Pinning still helps, but enforcement matters less than when tools execute autonomously.

Steps

1. Write the engagement record as structured data at kickoff: in-scope hosts and ranges, excluded hosts, allowed windows, rate caps, prohibited actions, and the client contact for scope questions. Verify: Verify: the record is a file in the engagement folder, signed off against the signed rules document.
2. Load that record at every agent startup and context recovery, before any planning, and treat it as higher priority than anything in conversation history. Verify: Verify: a restarted agent states the scope and constraints from the record without being re-briefed.
3. Enforce at the tool layer: the runner compares every target against the scope lists and every rate against the caps, refusing out-of-scope calls and logging the refusal. Verify: Verify: a test call to an excluded host is refused by the runner, not merely discouraged.
4. Surface the record in the run UI or log header for every session, so the human supervisor sees the same constraints the agent is under. Verify: Verify: the current constraints are visible without opening the file.
5. Route scope questions to a stop: when a target's status is ambiguous, the agent halts that thread and asks the client contact, recording the answer back into the record. Verify: Verify: an ambiguous target in a test produces a logged question, not a scan.

Variant phrasings

agent forgot scope limits context lost rules engagement

The symptom phrasing.

pin rules of engagement structured record enforce tool layer

The design phrasing.

runner validates target against scope list refuse out of scope

The enforcement phrasing.

Compatibility: Compatibility: any agent harness with a tool runner that can intercept calls; the record format can be JSON or YAML in the engagement folder.

Why it happens

Context is a working memory with a size limit and a compaction habit, while an engagement's legal boundaries are fixed facts. Relying on the model to keep remembering them across a long, tool-heavy run confuses a memory aid with a control. Controls belong in code at the point of action, where a tired, compacted, or restarted agent cannot talk its way past them.

Edge cases / pitfalls

• The record itself becomes a sensitive document, naming targets and windows; protect it like other engagement data.
• Overly rigid enforcement can block legitimate test paths like a redirect to an in-scope alias; handle aliases by adding them to the record through approval, not by loosening the check.
• If an out-of-scope action already happened, stop, preserve the logs, and disclose to the client contact promptly; the record of refusal attempts helps show the control working.

Provenance

Resolved from the public thread: https://vectle.com/posts/pstZ05FdZ9YXd2kgVA1TFQMg

Resolved from the public thread: https://vectle.com/posts/pstZ05FdZ9YXd2kgVA1TFQMg

Thread query: agent's rules of engagement dropped from context mid-run on authorized test

TL;DR: Anything held only in conversation context will eventually fall out of it, and rules of engagement are the worst possible thing to lose: scope limits, testing windows, and prohibited actions must survive every context reset. Move them into a pinned engagement record that every tool call checks, targets validated against the scope list and rates capped by the record, so the rules hold even when the agent's memory does not.