Blast Radius

Who Should Read

A backend developer whose coding agent's long sessions cause high token costs and context overflow

Core Mechanics

• It predicts a prompt's reach first
  • Blast Radius couples context and code channels to estimate how far an incoming prompt reaches.
  • That estimate decides which records to evict and how far into the codebase to reach.
• NECROPHORESIS makes eviction reversible
  • Dead context is archived verbatim before being removed from the window.
  • It can be restored later, so nothing is lost the way it is with summaries.
• Recurring dead records get buried
  • Recurring Dead Matter (RDM) finds transcripts that keep showing up.
  • These keep occupying context, so they are buried first.
• It sits beneath HCRC
  • Blast Radius decides which records to bury underneath HCRC.
  • It is a layer under existing context management rather than a standalone tool.

Evidence

• Token use dropped
  • Token consumption fell 17-26% across seven OpenAI models.
• Lowest overflow
  • It had the lowest overflow rate among the tested policies.
• Reversal was byte exact
  • Archived records could be restored byte for byte.
• Buried records were never recalled
  • Of 450 buried bodies, 378 were recurring dead matter.
  • Zero of them were recalled.

How to Apply

• Count repeated records in your agent logs today
  • Pull recent coding agent session logs and count how often the same tool output or transcript repeats.
  • Mark high-repeat items as candidates for cleanup.
• Archive verbatim instead of deleting
  • Save records to evict into a file or DB as-is instead of overwriting them with summaries.
  • Leave only an archive ID in context so you can restore on demand.
• Scope reach when a prompt arrives
  • Pick the relevant code files and past records first, and build context only from them.
  • Move everything outside that scope to the archive.

Try It

# Paste into Claude or ChatGPT (simplified version of the paper's method: find recurring dead context)

Below is a coding agent session transcript.
{{transcript}}

# Instructions
1. Find blocks that repeat or are no longer used for later work, and list their numbers and reasons in a table.
2. Mark each block as 'archive verbatim and remove from context'. Do not summarize or rewrite them.
3. For the next request {{next prompt}}, list only the blocks that should remain.

# Try next
Ask whether any removed block might be needed later, and if so write the condition for restoring it.

Terminology

Related Papers

• Mercury 3: Diffusion at scale

  Scaling diffusion LLM 40x gives Mercury 3: 4–8x faster than speed-tier models, similar quality

• Timestral: Mistral's time-series foundation models

  A 209M time-series model ranks first on three benchmarks and runs on CPU.

• DeepSeek V4.1 Flash 1017 (Max)

  DeepSeek V4.1 Flash 1017 Max scores 45 on the Intelligence Index at 211 tok/s and $0.30 input.

• If you don't run your own LLM proxy, how do you know whats going on?

  Running your own LLM proxy shows which keys, models, and cache misses drive your usage.

• DeepSeek: DeepSeek v4.1 Flash 1017

  DeepSeek V4.1 Flash 1017: $0.05/M input, 1M context, built for agents

• DeepSeek-v4.1-Flash-1017

  DeepSeek V4.1 Flash minor update lifts coding and agent benchmarks while cutting run time and output length

• Show HN: Llmspeed – feel how fast a local LLM runs before you buy the hardware