Maintaining Engagement State

An engagement lasts weeks. A session does not. The gap between those two facts is where work gets repeated, findings lose their evidence, credentials become unattributable, and artifacts get left on production because nobody wrote down that they were placed.

This is not administrative overhead bolted onto the real work. Provenance is what makes a finding reportable, and the artifact register is what makes cleanup possible. Neither can be reconstructed afterwards from terminal scrollback.

Assume everything not written to the record is lost: not merely forgotten, but silently lost, so the next session confidently redoes it or contradicts it.

When to Use

• Any engagement or investigation continuing past a single sitting
• Resuming after a break, or picking up someone else's work
• Approaching a context boundary with unrecorded state
• Handing off between analysts, shifts, or teams
• Before running anything that writes to a target
• When you cannot answer "which host did this credential come from?"

When NOT to Use

• Writing the client-facing deliverable — use reporting-security-findings; this skill produces the raw material that skill consumes
• Deciding whether a finding is real — that is the relevant testing or analysis skill; record the outcome here either way
• Chain of custody for evidence that may reach court — legal custody has formal requirements beyond this; use responding-to-incidents and follow the organisation's counsel
• A single self-contained task finished in one sitting with nothing left on target

The Six Records

Keep these in one file, in the repository or case folder, updated as you go — not reconstructed at the end.

1. Scope and authorization

In-scope targets and explicit exclusions, the testing window, the named contact, and where the authorization letter lives. Copy the exclusions verbatim; paraphrasing scope is how people test the wrong estate.

Record one more field, because "in scope" and "reachable" are different questions. Egress mode says what this engagement is allowed to talk to:

Name the mode before the first command, not after something reaches further than you meant. It is the field that catches the two classic mistakes: a "static-only" malware triage that resolves a C2 domain, and a scoped web test whose scanner follows a redirect off-estate. When a skill in this repo says an action needs an isolated lab, it means lab_only or stricter.

Authorization is a gate, not a note. If authorization is pending or absent, reading documentation and planning are fine; scanning, hooking, and exploitation are not.

2. Credential provenance

For every credential, hash, token, or key: what it is, where it came from, how it was obtained, and what it has been used against.

cred-04  | svc_backup NTLM hash
         | source: SAM dump, WS-014 (10.2.3.14), local admin via CVE-2021-XXXX
         | obtained: 2026-07-27 14:02
         | used on: FS-02 (success), DC-01 (failed, no local admin)
         | provenance: CONFIRMED

Provenance is not bookkeeping. It answers three questions you will be asked:

• Is it real? Credentials of unknown origin may be planted — see recognizing-deception. If you cannot say where it came from, you cannot rule that out.
• Was it in scope? A credential harvested from an out-of-scope host taints everything reached with it.
• What is the blast radius? The client's remediation is "rotate these specific secrets", which requires the list.

Mark anything you cannot trace as provenance: UNKNOWN and treat it as suspect rather than quietly using it.

3. Access inventory

Each access obtained: host, account, privilege level, the method, and whether it still works. Access decays — sessions die, passwords rotate, hosts reboot. A list of what you had is not a list of what you have.

4. Artifact register — the cleanup list

Every write to a target, recorded at the moment you make it. Files dropped, services or tasks created, accounts added, registry keys, shares, firewall rules, cloud resources, config edits.

art-07 | WS-014 | C:\Windows\Temp\c2.exe        | dropped 14:02 | REMOVED 18:40
art-08 | DC-01  | scheduled task "WinUpdateSvc" | created 15:10 | PENDING
art-09 | AWS    | IAM user "svc-audit-tmp"      | created 16:22 | PENDING

Written afterwards from memory, this list is always incomplete, and what it omits is left on the client's production estate. Two independent parties need it: you, to clean up; and the client's blue team, to distinguish your artifacts from a real intrusion — during and after the engagement.

For an incident investigation the same register covers what you touched on evidence systems, so your activity is separable from the adversary's.

5. Findings with evidence

Each finding: what, where, the reproduction steps, the evidence captured, and the impact demonstrated. Capture evidence when you see it — the state that proved it will not survive remediation, a reboot, or a rotated credential.

6. Dead ends

The record most often skipped and most valuable across sessions: what was tried, against what, and why it failed.

dead-03 | Kerberoast svc_sql | hash cracked? NO after 6h, 14M candidates
dead-04 | SMB signing DC-02  | enforced -- relay not viable
dead-05 | /admin on portal   | 403 from all tested source ranges

Without this, the next session re-runs the six-hour crack. Recording negative results is also the difference between "we found nothing there" and "we did not look" in the final report.

Session Boundary Discipline

Before a session ends — planned or not — the record must answer:

1. What access do I currently hold, and does it still work?
2. What have I left on target that is not yet cleaned up?
3. What was I in the middle of?
4. What did I rule out, so nobody repeats it?

Write these down as you work, not at the end. The session that ends unexpectedly is precisely the one whose state was never captured, and an engagement can lose its context without warning.

Handoff

A handoff is the same record plus three additions: current position, the immediate next action, and anything time-sensitive (a running crack, an expiring token, a scheduled task that fires at 02:00). If the receiving analyst has to ask "what were you doing?", the handoff failed.

Across Engagements

The six records reset each engagement. Two things are worth carrying between them, and both fail the same way — reconstructed from memory, they are wrong.

What the machine can actually do. Before a skill runs a tool, the tool has to exist, at a known path, at a version whose flags match what the skill assumes. Guessing /usr/bin/ghidra or jadx and debugging the failure as if it were the target's fault is a standard way to lose an hour. Keep a local, machine-specific inventory — tool, real path, version, and for anything that speaks over a socket (an MCP server, a helper daemon) whether it is not just installed but registered and answering. Present and ready are different states: an MCP binary on disk with no running server is not a capability. This inventory is per-machine and does not belong in a shared repo; treat it the way you treat ~/.ssh/config — yours, gitignored, assumed by nothing you publish.

What past cases taught. A short, de-identified journal entry per finished case — the scenario, the technique that worked, the dead ends and the hours they cost, the tooling surprise — turns the next similar case from a cold start into a lookup. Keep an index over the entries by scenario, by technique, and by target trait, and read that before starting new work, not the entries themselves; the index is the cheap part. Seed it with a handful of worked reference cases so it is useful on day one rather than after the tenth engagement.

Two disciplines make the journal safe to keep. De-identify at write time, not before sharing — real domains, IPs, tokens, and PII become placeholders the moment the entry is written, so an entry that is never meant to leave still cannot leak if it does. And keep it local by default: contributing lessons to a shared or public location ships engagement detail to a third party, and "the model anonymised it" is not a control you would defend to a client. If a journal is ever shared, that is a deliberate, reviewed decision per entry — not a default step in the workflow.

Rationalizations to Reject

• "I'll remember where that came from." You will not, and neither will the next session. Unattributable credentials are unusable in a report and indistinguishable from planted ones.
• "It's all in my terminal history." History has no reasoning, no timestamps you can trust, no failures, and no record of what you concluded. It is also routinely lost with the session.
• "I'll write the cleanup list at the end." The end is exactly when memory is worst and the list is longest. Artifacts written from memory are always incomplete, and the omissions stay on production.
• "Nobody needs my dead ends." They are the difference between a reported negative result and an unexamined gap, and they stop the next session spending six hours re-cracking the same hash.
• "I'll clean up as I go, so no register is needed." Then the register costs you one line per artifact and proves it. Without it you cannot demonstrate the estate was returned to its prior state.
• "The client only wants the findings." Until an alert fires next month and they need to know whether the account you created was yours. Unclaimed artifacts get investigated as intrusions.
• "Recording failures makes the engagement look weak." An engagement with no recorded negatives looks unmethodical, not successful. Coverage is a finding.

References

• reporting-security-findings — consumes this record to produce the deliverable
• recognizing-deception — why unattributable credentials are dangerous, not merely untidy
• responding-to-incidents — the investigation-side equivalent, including formal evidence handling