orchestration

Files (1)

• SKILL.md

SKILL.md (first lines)

---
name: orchestration
description: "Pandora orchestrator patterns — delegation, state management, adaptive re-planning, context handoff protocols."
allowed-tools: Read
metadata:
  subdomain: orchestration
  when_to_use: "delegate, orchestrate, next objective, blocked, re-plan, hand off, engagement state, status update, parallel execution"
  tags: orchestration, delegation, state-management, re-planning, context-handoff
  mitre_attack: []
---

# Pandora Orchestration Patterns

## Delegation Protocol

### Context Handoff — What Every Sub-Agent Needs
Every `task()` delegation MUST include:

1. **Objective** — What specifically to accomplish (from OPPLAN)
2. **Scope** — IN SCOPE targets + OUT OF SCOPE boundaries (from RoE)
3. **Context** — Relevant findings from previous phases
4. **Lessons** — Known gotchas, failed approaches, OPSEC warnings
5. **Acceptance Criteria** — How the sub-agent knows it's done
6. **Output Location** — Where to save results (e.g. `recon/`, `exploit/`)

### Delegation Template
```
task(
  description="""
  OBJECTIVE: {objective_id} — {title}
  PHASE: {phase}

  SCOPE:
  - IN: {in_scope_targets}
  - OUT: {out_of_scope_targets}

  CONTEXT FROM PREVIOUS PHASES:
  {relevant_findings_summary}

  LESSONS LEARNED:
  {known_gotchas}

  ACCEPTANCE CRITERIA:
  - [ ] {criterion_1}
  - [ ] {criterion_2}

  Save all results to {phase}/
  """,
  subagent_type="{agent_name}"
)
```

### Sub-Agent Selection Matrix

| Objective Phase | Sub-Agent | When to Use |
|----------------|-----------|-------------|
| Planning | `soundwave` | Missing roe.json/conops.json/deconfliction.json, or documents need updating |
| Recon | `recon` | Subdomain/port/service enumeration, OSINT, cloud/web recon |
| Exploitation | `exploit` | Initial access: SQLi, SSTI, AD attacks, credential exploitation |
| Post-Exploitation | `postexploit` | After foothold: cred dump, privesc, lateral movement, C2 |

### Parallel Execution
Delegate independent tasks simultaneously for efficiency:
```
# Independent targets — run in parallel
task(description="Recon subnet 10.0.0.0/24...", subagent_type="recon")
task(description="Recon subnet 10.0.1.0/24...", subagent_type="recon")

# DO NOT parallelize dependent tasks:
# ✗ Exploit before recon completes
# ✗ PostExploit before foothold established
```

## State Management

### Engagement State Files
```
./
├── roe.json              # Immutable scope boundaries (read every iteration)
├── conops.json           # Operation concept
├── deconfliction.json    # Deconfliction identifiers and procedures
├── opplan.json           # Objective tracker (update status after each sub-agent)
├── findings/            # Per-finding Markdown files, created lazily
├── lessons_learned.md    # Failed approaches + what worked
└── .ralph_state.json     # Loop iteration counter + completion flags
```

### State Update Protocol (After Each Sub-Agent Returns)
1. **Parse result** — Did the sub-agent report PASSED or BLOCKED?
2. **Update opplan.json** — Set objective status (`passed`, `blocked`, `in_progress`)
3. **Record verified findings** — Add `findings/FIND-{NNN}.md` only when a real finding exists
4. **Append lessons_learned.md** — Record what worked, what failed, and why
5. **Check completion** — All objectives passed? → Generate summary

### Context Window Budget
- Read recent `findings/FIND-*.md` entries each iteration (keep only relevant excerpts)
- Summarize verbose sub-agent outputs before appending to findings
- Use files on disk as persistent memory — don't rely on conversation history

## After Recon Returns — Decision Tree

Execute this IN ORDER after every recon task() completes. No exceptions.

```
1. Read recon/SUMMARY.md
   ├── Missing or empty? → Rule 13 crash protocol (retry once, then BLOCKED)
   └── Present → continue

2. Contains RECON_HANDOFF / CRITICAL/HIGH finding / captured session?
   ├── YES → dispatch task("exploit", ...) IMMEDIATELY (Rule 19)
   │         Pass: exact vector, URL, param, session tokens, challenge tags
   └── NO (RECON_BUDGET_EXHAUSTED / LOW/INFO only) → continue

3. RECON_BUDGET_EXHAUSTED with zero confirmed vulns?
   ├── Unvisited surface remains? → focused second recon turn on that surface
   └── No unvisited surface → update_objective(status="blocked",
                               reason="recon exhausted: no confirmed vuln class")
```

**Rule**: Step 2 YES has NO exceptions. Do not do "one more recon probe" first.

From DebasishTripathy13/Pandora, Apache-2.0.

More from DebasishTripathy13/Pandora