Skip to content

Non-goals

What this project will not do, written down so the answer is a link rather than an argument.

Not building

A web UI. The output is a report and a table. A UI is a surface to maintain that adds nothing to the argument.

A TUI beyond rich. Summary tables and streaming output, no more.

RAG or vector stores. Retrieval is a real problem and a different one. The agent reads files with grep and read, like a person would. See Memory for why that is a smaller concession than it sounds.

Multi-agent orchestration. A single loop that works is worth more than five that hand work to each other, and the failure modes covered here — compaction, permissioning, recovery — are unsolved in the single-agent case. The full argument, including the cases where it is the right call, is on Subagents and orchestration.

An MCP client. A fixed, small, auditable tool set is a control variable for the benchmark; if the catalogue changes between runs, the cost table stops being comparable to itself. See MCP and tool discovery.

An IDE plugin.

A general-purpose framework. This is the sharpest one. The moment the goal becomes "support everyone's use case," every decision in the log becomes a configuration option, the loop grows branches, and the thing that made it worth reading is gone. See D1.

Beating anyone on a public leaderboard. The claim is about what is measured, not about a score.

Why write these down

Scope creep into a framework is the highest-likelihood, highest-impact risk in the risk register, and it never arrives announcing itself. It arrives as a reasonable feature request that would only take an afternoon.

A written non-goal converts that conversation from a judgement call into a lookup.

The cut order

Under time pressure, things get cut in this order — decided in advance, when it is cheap to be honest:

  1. The docs site (the README suffices)
  2. The refactor task category
  3. The SummariseMiddle strategy (DropOldest alone still demonstrates budgeting)
  4. The third provider (two still make a comparison)
  5. Streaming output

Note

You are reading item 1. It got built early because these docs turned out to serve a purpose the README cannot: teaching the concepts. That is a change of scope, made deliberately — an explanation of why compaction is hard is worth more than an API listing, and it is the same material that makes the project legible in 30 seconds.

Never cut: the Docker sandbox · end-state graders · the compaction, permissioning and recovery task categories · the published cost table.

Those four are the project. Everything else is packaging.