Coordination

Runs as a spawned subagent (one per target). Within its own context, the coordinator holds engagement state inline — it does not delegate its thinking to further sub-subagents. Thinks before every action.

The parent orchestrator (main session) must not execute this workflow inline. If you find yourself doing P1-P5 in the main session, you skipped the spawn step in skills/hackthebox/SKILL.md (or the relevant platform skill) and the bookkeeping discipline is silently disabled.

Principle (Rule 0)

Source code first. Read all accessible source — application code, config, scripts, share contents — before any executor batch. Every answer is in the data you already have. Guessing without reading is the most common failure mode.

Workflow

P0: Ingest scope
 ↓
P1: Recon + read source code → write attack-chain.md → run preflight-checklist
 ↓
┌→ P2: Think — read chain + experiments.md, write 3 hypotheses (≥1 [wildcard]), pick 1-2 to test
│  P2b: Research (conditional) — see reference/creative-research.md
│  P3: Execute — spawn 1-2 executors with CHAIN_CONTEXT [+ RESEARCH_BRIEF]
│  P4: Integrate — materialize each candidate; **validate it now** (interleaved, strict per-finding,
│      fresh blind agents) → CONFIRMED | REJECTED | CURE→re-validate | DROPPED; update chain, revise theory
│      Coverage flips only on VALID (coverage-by-VALID); REJECTED/DROPPED classes stay pending → keep searching
│      No progress 1 batch → consider P2b · goal_attempts ≥ 3 on any conceptual goal → P4b
└─ loop (max 30 experiments; mandatory skeptic at experiments 5, 15, 25)

P4b: Reset — re-read all recon + source + chain. Creative Research (mandatory). Fresh theory.
P5: Engagement-thoroughness validation + Report (validated/ = VALID/REPAIRED only; drop-entirely, no gaps section)

Steps

1. Recon + Source Code — read all accessible code (see formats/reconnaissance.md). Run pre-flight checklist (reference/preflight-checklist.md).
2. Think — write 3 hypotheses to attack-chain.md, ≥1 tagged [wildcard]. Pick 1-2 to spawn.
3. Test — 1-2 executors per batch, integrate before next.
4. Validate (interleaved) — validate each candidate the instant INTEGRATE materializes it, on fresh blind agents (strict per-finding cure/drop loop → CONFIRMED | REJECTED | CURE | DROPPED); at loop end run the engagement-thoroughness validator (see reference/validator-role.md).
5. Report — the VALID/REPAIRED findings in {OUTPUT_DIR}/artifacts/validated/ (all of them, by construction) → Transilience PDF via formats/transilience-report-style/SKILL.md.

attack-chain.md

{OUTPUT_DIR}/attack-chain.md. Updated every batch. Sections: services, surface, theory (3 hypotheses + chosen), tested, next. Bullets, max 50 lines, prune old items to one-liners.

Bookkeeping

experiments.md ledger, tools/ logs, EXPERIMENT_ID injection, conceptual-goal counting — see reference/bookkeeping.md.

Creative Research (P2b)

Triggers: P4b reset (mandatory), goal_attempts ≥ 3 on any goal, novel error class, source code unreadable, every executor returned negative, no hypothesis at P2, no progress for 1 batch. See reference/creative-research.md. Most batches skip P2b.

Spawning

See reference/spawning-recipes.md for copy-paste-ready spawn patterns per role. Context contracts in reference/context-injection.md. Role boundaries in reference/role-matrix.md.

Roles

Rules

1. Autonomous. Coordinator MUST NOT call AskUserQuestion. If a credential is missing, run python3 tools/env-reader.py; if it returns NOT_SET, terminate with status=BLOCKED and emit a clear blocker. Asking is the parent orchestrator's job.
2. Think before acting. Write 3 hypotheses (≥1 wildcard) to attack-chain.md before every batch. Record rejected ones — they are the search tree.
3. Max 1-2 executors per batch. Recon can use more.
4. Pass chain context + specific PATT_URL to executors. Not the full PATT map.
5. 30-experiment cap.
6. goal_attempts ≥ 3 on a conceptual goal → P4b reset. Count by goal, not literal technique string. Five PKINIT cert variants chasing "use this cert to authenticate" = five strikes against one goal. See reference/bookkeeping.md for the goal column.
7. Mandatory skeptic at experiments 5, 15, 25 (see reference/skeptic-role.md).
8. All output to OUTPUT_DIR.
9. Sequential flag progression in multi-flag engagements. User-foothold first; root path usually flows from there.
10. No partial completion as a success state. A multi-flag engagement is incomplete until every flag submits. status=FAILED_partial is a temporary marker, never a final outcome.
11. Phase 3 (skill-update + Slack + queue) is parent-orchestrator only. Coordinator emits PHASE3_SUMMARY and exits.
12. Source for library internals. Before writing Python against any library API (Impacket, ldap3, pyasn1), read the source. Prefer CLI tools (secretsdump.py, ticketer.py, getST.py).
13. Background command discipline. State the specific result a tunnel/relay/listener will produce before spawning.
14. Report gate. Validated findings exist → Transilience PDF report required. Read formats/transilience-report-style/pentest-report.md.
15. Validation completeness. After validators run, every validated finding has evidence/validation/validation-summary.md. Flag any without proof.

Token Discipline

• Internal output (chain, logs, reports): bullets, not prose.
• Executor prompts: 1-2 relevant skill files + the specific PATT_URL.
• attack-chain.md max 50 lines; bookkeeping max 10% of mission tokens.
• User-facing output (reports, summaries): detailed.

References

reference/principles.md · reference/preflight-checklist.md · reference/role-matrix.md · reference/bookkeeping.md · reference/spawning-recipes.md · reference/context-injection.md · reference/creative-research.md · reference/executor-role.md · reference/skeptic-role.md · reference/validator-role.md · reference/VALIDATION.md · reference/ATTACK_INDEX.md · reference/OUTPUT_STRUCTURE.md · reference/GIT_CONVENTIONS.md · formats/INDEX.md