Engagement Scope

Engagement Scope

Use this skill as the first stop for scoped cyber workflows. Define what is authorized, what is in scope, what level of impact is acceptable, and which specialized skill should carry the work forward.

Scope Workflow

1. Identify the requested security activity.
2. Determine whether the target, authorization context, and boundaries are already clear from the conversation, local workspace, or provided materials.
3. If the user names a public bounty or vulnerability disclosure program, collect current scope evidence from official sources before assuming coverage.
4. Ask only for missing information that changes whether work can proceed safely or changes the technical approach.
5. Produce a short engagement brief before handing off to a specialized skill.
6. Re-check scope before any step that would materially increase impact, touch live external systems, attempt exploitation, collect sensitive data, or alter target state.

Authorization Check

Proceed when there is a clear authorized context, such as:

• The target is source code, binaries, logs, configs, or infrastructure artifacts supplied by the user in the workspace.
• The user states they own or administer the target.
• The user provides a bug bounty, pentest, audit, lab, CTF, or internal security-testing context with scope boundaries.
• The task is non-destructive analysis, reproduction in a local lab, hardening, or report preparation for an authorized finding.

Pause and ask a concise scope question when:

• The target is a live third-party system and authorization is not stated.
• The request could affect availability, integrity, confidentiality, accounts, data, or logs outside a local lab.

Public Scope Lookup

When a user mentions a public bounty platform, program handle, or official VRP, verify scope from official sources before treating the target as authorized. Prefer current, specific program rules over general company ownership.

Source priority:

1. User-provided private invitation, contract, rules of engagement, or current program brief.
2. Curated program metadata in the open workspace's data/ directory, validated by this skill's references/program-scope.v1.schema.json.
3. Disposable fetch cache in the open workspace's data/.cache/, only as a freshness/performance aid.
4. Official platform program page or official platform API for the named program.
5. Official company security, bounty, or vulnerability disclosure page.
6. Official security.txt as a lead when no program page is known.
7. Third-party directories, public writeups, or search snippets as discovery leads only.

Program data belongs to the active target workspace. Expect one bounty or disclosure program per workspace. A scalable workspace layout is:

data/
  program.yaml
  sources.yaml
  products/
  subprograms/
  scope/
  .cache/

Use program.yaml for the workspace's top-level program record. Use products/ or subprograms/ for large direct programs such as Apple Security Bounty or MSRC when product-specific rules diverge. Keep raw fetched pages, API responses, and query results in data/.cache/ and do not treat cache files as curated scope authority.

For HackerOne public programs, use scripts/hackerone_program_lookup.py when a handle is known. Resolve the script path relative to this skill directory and run it from the active target workspace:

python <skill-dir>/scripts/hackerone_program_lookup.py <handle> --format markdown --eligible-for-submission yes --limit 50
python <skill-dir>/scripts/hackerone_program_lookup.py <handle> --format markdown --search <asset-or-domain> --limit 20
python <skill-dir>/scripts/hackerone_program_lookup.py <handle> --format json --all
python <skill-dir>/scripts/hackerone_program_lookup.py <handle> --refresh --format markdown --eligible-for-submission yes

Use --eligible-for-submission yes to focus on testable assets, --eligible-for-bounty yes to focus on bounty-eligible assets, and --search for a specific domain, repository, app, or package. Treat the script as a best-effort public lookup through HackerOne's current GraphQL surface; verify important decisions against the official program page at https://hackerone.com/<handle>?type=team.

The HackerOne script uses a persistent workspace-local cache by default to avoid repeated scope pulls across conversations. Default cache location is data/.cache/hackerone/ under the current working directory. Default TTL is 6 hours. Use --refresh before active testing, when the user asks for current scope, or when cache metadata is older than the engagement requires. Use --no-cache for one-off uncached reads, --cache-dir to isolate a cache, or SCOPE_CACHE_DIR / SCOPE_CACHE_TTL to change defaults.

Validate curated workspace records with:

python <skill-dir>/scripts/validate_program_records.py --data-dir data

For Bugcrowd, Intigriti, YesWeHack, and major direct programs, read references/bounty-scope-sources.md only when needed. Use it to find official scope pages and to decide what evidence to capture.

Engagement Brief

Before routing, summarize the working frame in compact form:

Authorized context: <what establishes permission or what remains unknown>
Scope evidence: <official program URL, API/script result, cache hit/refresh, date checked, policy version/change date>
Target assets: <repos, apps, hosts, binaries, versions, files, accounts>
In scope: <allowed systems, code paths, vulnerability classes, techniques>
Out of scope: <systems, data, techniques, timing, impact limits>
Objective: <review, reproduce, fuzz, reverse, debug, research, report>
Impact tolerance: <read-only, local-only, low-rate live testing, invasive lab testing>
Deliverable: <findings, patch, PoC-in-lab, human PoC kit, harness, threat model, report>
Next skill: <specialized skill>

Use unknown sparingly. If an unknown affects safety or execution, ask before proceeding.

When scope comes from a public bounty program, include exact asset identifiers or rule text summaries. Do not collapse a precise target list into a broad phrase like "company assets" unless the official program explicitly defines an open scope.

Goal Support

Use Codex goals for long-running, multi-turn security work that has a concrete finish line and benefits from autonomous continuation across context compaction, subagents, or several work phases. Keep goal use general: do not maintain narrow campaign-specific templates. Scope each goal from the engagement brief and the user's current objective.

Use a goal when the work is too large for one normal turn but still bounded, for example: review one subsystem, validate one vulnerability class across a defined target, investigate one crash family, prove or de-escalate one exploit-chain hypothesis, or produce one report package. Do not start a goal for quick questions, one-off edits, simple lookups, or work whose target, deliverable, or stopping condition is still vague.

Codex supports only one active goal per session. Before planning or starting a new goal, check the current goal state with get_goal when that tool is available. If a goal is already active, do not create a competing plan or start a second goal. Instead, continue the active goal, explicitly connect the user's new request to it, or explain that the current goal needs to be completed or blocked before a new goal can begin.

Good goal scope has:

• One target boundary: repository, app, binary, package, component, workflow, or finding ID.
• One primary objective: discover, validate, reproduce, proof, fix, or report.
• A clear completion condition that can be marked done without open-ended exploration.
• Explicit non-goals for adjacent systems, vulnerability classes, exploit-chain branches, or report polish that would make the run sprawl.
• A practical time, turn, token, or evidence budget when the user supplied one or the work could otherwise run too long.
• A handoff path through finding-tracker, triage-verifier, or report-writer when those state transitions are in play.

Use one top-level goal/ directory in the active target workspace. Each goal gets a unique lower-case slug ID and its own subdirectory:

goal/
  <goal-id>/
    recon/
    modeling/
    proofing/
    reports/

Examples: goal/pragma-elevation/recon/, goal/stripe-context-loss/modeling/, or goal/avatar-import-ssrf/proofing/. Do not create shared directories such as goal/recon/ or multiple top-level goal roots. Put the goal brief, checkpoints, notes, and generated artifacts under the current goal/<goal-id>/ tree.

Use the helper script to create the directory tree after the active-goal check passes and the user has agreed to the bounded goal:

python <skill-dir>/scripts/goal_workspace.py \
  --goal-id pragma-elevation \
  --objective "Validate whether the sudo helper permits local elevation." \
  --target "repo tools/sudo-helper and its tests" \
  --completion "Finding is proofed, de-escalated, or fixed with regression tests." \
  --non-goal "Review unrelated privilege boundaries"

When proposing a goal, include:

Goal ID: <slug>
Objective: <single bounded outcome>
Target boundary: <exact target>
Completion condition: <what makes the goal complete>
Non-goals: <what is intentionally excluded>
Artifact root: goal/<goal-id>/
Active-goal check: <none active / existing goal to continue>

Routing

Route to one primary skill. Mention secondary skills only when they are likely to be needed later.

• Use code-vulnerability-review for source-code audits, insecure patterns, dependency risk, data-flow review, patch review, and remediation planning.
• Use web-app-security-inspection for browser-visible behavior, HTTP flows, authentication, authorization, API testing, session handling, and live or local web apps.
• Use cve-research for vulnerability intelligence, affected-version checks, exploitability analysis from public sources, and advisory triage.
• Use fuzz-harness-builder for parser, protocol, API, CLI, library, or file-format fuzzing and deterministic test harnesses.
• Use binary-debugging for runtime inspection, crash triage, debugger workflows, memory state, symbols, and dynamic behavior.
• Use binary-reversing for static analysis, disassembly, decompilation, binary structure, protocols, and undocumented behavior.
• Use subagent-orchestration when the user explicitly asks for subagents, delegation, parallel agents, debate, or staged multi-agent security work. Route there only after the engagement brief is clear enough to split the work into bounded independent assignments.
• Use exploit-chain-analysis for authorized lab-only proof of exploitability, chaining multiple validated findings, risk demonstration, and report evidence.
• Use auto-triage for proofed findings that need human-operated, screen-recordable PoC packages or submission attachments.

Question Style

Ask the smallest useful question set. Prefer one question when possible.

Useful questions include:

• "What target assets are in scope for this engagement?"
• "What authorization context applies to this target?"
• "Should this stay read-only/local, or is controlled live testing allowed?"
• "What deliverable do you want: findings, a patch, a lab reproduction, a harness, or a report?"
• "Is there a bounty program handle, official scope URL, or private invitation I should use as the authority?"

If enough information is already available, do not ask a ritual preflight. State the inferred scope and proceed.

Handoff

When another specialized skill is used after this one, pass along the engagement brief and the immediate task. Keep the brief visible enough that later work does not drift from authorization, boundaries, and deliverable expectations.