CipherSwarm Phase 5 - Agent & Server Collaboration Vision#

Table of Contents#

• CipherSwarm Phase 5 - Agent & Server Collaboration Vision
  • Table of Contents
  • Overview
  • Agent as Observant Executor
    • Structured Status Streaming
  • Agent as Local Governor
    • Self-Tuning Execution
  • Agent Fault Recovery
    • Offline Slice Recovery
  • Server as Strategic Coordinator
    • Live Plan Adjustment
    • Crack Result Feedback Loop
  • Performance Learning & Forecasting
    • Historical Agent Performance
    • Agent Capability Signaling
  • Optional Concepts
    • Slice Replay for Debug
    • Agent Karma Score
    • Adaptive Task Graphs
  • Summary

Overview#

This document outlines a future-forward vision for Phase 5 of CipherSwarm: transforming agents into adaptive, observant, and partially autonomous nodes that work with the server, not for it. These ideas extend the Sync Extensions model by elevating both agent responsibility and server strategy.

The goal is a smarter, more cooperative distributed cracking system — one that adapts to real-world conditions, anticipates failure, and evolves over time.

Agent as Observant Executor#

Structured Status Streaming#

Agents will parse --status-json internally and emit structured telemetry on an interval (e.g. every 5s).

Fields to Stream#

• Slice ID
• Total guesses completed
• Percent progress (relative to --skip/--limit)
• Cracked hashes (count + list)
• Device temperatures
• Guessrate per device
• Rejected guesses
• Host metadata (load average, memory, thermal status)

Protocols#

• Stream to /api/v2/client/status via chunked POST or SSE
• Use JWT auth to tag source agent and project

Benefits#

• Real-time dashboards
• Predictive failure detection
• Live slice reassignment if an agent lags
• Audit-grade traceability per slice

Agent as Local Governor#

Self-Tuning Execution#

Agent adjusts hashcat launch parameters based on:

• Thermal envelope
• Load average
• Historical guessrate vs benchmark

Adjustments#

• Workload profile (-w) from 1 to 4
• Manual tuning of -n, -u, -T if available
• Preemptive cooldown mode if device crosses critical threshold

Benefits#

• Avoids wasteful slicing
• Helps survive degraded environments (e.g. closet rigs)
• Reduces silent throttling artifacts

Agent Fault Recovery#

Offline Slice Recovery#

Agent writes a checkpoint file mid-task (e.g. every 15s) to disk:

• Last known hashcat status
• Command used
• Slice metadata

On restart:

• Agent uploads checkpoint to /client/recover

• Server may:

  • Mark slice complete if near 100%
  • Reassign remaining % as new slice
  • Flag slice as recovered

Benefits#

• Prevents full slice loss during crash
• Enables graceful resumption
• Server doesn't need to re-calculate or guess slice state

Server as Strategic Coordinator#

Live Plan Adjustment#

Server tracks slice completion durations and guessrate over time. It dynamically adjusts:

• Slice size
• Assigned hash types
• Task priority

Server maintains slice history:

{
  "slice_id": 42,
  "agent_id": 3,
  "duration": 68.2,
  "average_speed": "2.1 GH/s"
}

Benefits#

• Better utilization
• Easier load spreading
• Can accelerate campaign completion by tuning on the fly

Crack Result Feedback Loop#

Each crack submission:

• Triggers UI toast
• Is written to dynamic wordlist for this project
• May influence future rule generation (loopback-inspired)

Server roles#

• Deduplicate cracks
• Update stats
• Integrate with Attack templates for reuse

Performance Learning & Forecasting#

Historical Agent Performance#

Each agent stores hash-type-specific baseline speeds:

{
  "hash_type": "sha512crypt",
  "benchmark": 80000,
  "recent_avg": 60000,
  "stdev": 5000
}

Server uses this to:

• Predict duration of unstarted attacks
• Estimate completion time for campaigns
• Alert on underperformance

Agent Capability Signaling#

Agents self-report:

• Supported hash types (via --backend-info)
• Disabled devices (opt-out, failed diagnostics)
• RAM, swap, CPU, GPU memory

Server uses this to prefilter /pickup results.

Optional Concepts#

Slice Replay for Debug#

• Server can export a slice config and replay it offline for debugging
• Useful for triaging slice errors or verifying crack speed issues

Agent Karma Score#

• Long-running performance metric
• Affects priority in slice selection (stable agents go first)
• UI badge: "🌟 Veteran" or "⚠️ Unstable"

Adaptive Task Graphs#

• Campaigns can be pre-planned as DAGs (mask length → fallback)
• Slices report status back to adjust execution order
• More useful in exploratory or speculative cracking

Summary#

By promoting the agent to an active peer — one that communicates, adapts, and recovers — CipherSwarm unlocks a new generation of distributed cracking. This vision puts control where it belongs: in the hands of the orchestrator, informed by smart, situationally aware agents.