Skip to content

A Dumpster You Throw Everything Into

Evidence Level: F (directly proven by source code)
Analysis Baseline: 4f843556

TL;DR

build_agent_context packs soul.md, memory.md, skills/, session_context, focus, triggers, and enterprise_info into a single system message—no layering, no caching, no prioritization.

1. Evidence

# agent_context.py
system_prompt = static_content + dynamic_content

# caller.py:557-559
api_messages = [LLMMessage(
    role="system",
    content=static_prompt,
    dynamic_content=dynamic_prompt,
)]

What Goes In

Type Content Processing
Static soul.md Truncated to 30K chars
Static memory.md Truncated to 2K chars
Static skills/ index Progressive disclosure
Static enterprise_info Full load
Dynamic session_context Compact summary
Dynamic focus Currently active items
Dynamic triggers Trigger conditions
Dynamic timestamp Current time

All merged into a single LLMMessage(role="system").

2. Consequences

  1. No layering: system / task / memory / tool result all mixed together—LLM can't distinguish priority
  2. No caching strategy: static content (soul.md, skills/, enterprise_info) recalculated every round, never cached
  3. Truncation loss: memory truncated to 2000 chars, excess discarded
  4. No independent updates: any static part changes → entire system message must be rebuilt

3. Cross-Article Interactions

With "One Pipe" (018): tool outputs (pip progress bars, npm dependency trees) flow directly into messages[] → mix with soul.md, memory, skills in the same context window → critical information drowned in noise.

With "Fake Caching" (020): static content recalculated every round → but mixed with session dynamic content → impossible to cache even if you wanted to.

4. Relationship to AI Coding

  1. "Concatenate everything" is the simplest implementation: packing all content into one big string is far faster than designing a layered architecture
  2. No caching awareness: AI won't proactively say "this content hasn't changed for several rounds, it should be cached"—that requires an architect's holistic judgment about performance
  3. Locally correct, globally catastrophic: each _read_file_safe call is correct in isolation—but combined, they create an unmaintainable dumpster

Key Code

File Lines Content
agent_context.py throughout build_agent_context — everything concatenated
agent_context.py 467 memory truncated to 2000 chars
agent_context.py 455 soul truncated to 30000 chars
caller.py 557-559 static + dynamic merged into one system message