Dev Tool Experiences
All articles

· 5 min read

Context Window Doubled Again, Codebase Doubled Again, Nothing Changed

By A. Rossi

  • tools
  • satire

SATIRE — The International Department of More Context is pleased to announce that its coding agent can now hold the entire codebase, the entire issue tracker, three architecture diagrams from 2018, and every comment beginning with “nit:” in a single majestic working memory. At last, the agent has enough information to ask whether npm test is the correct command.

The capacity planning meeting

The presentation begins with a rectangle labeled 1,000,000 TOKENS, because a rectangle is the accepted unit for progress. The previous rectangle was only half as wide. The VP of Applied Rectangle points at the new rectangle and explains that the agent can now understand the whole repository. Nobody asks which repository, because that would introduce a second rectangle: the monorepo, whose 47 packages disagree about whether “test” means Vitest, Jest, Bazel, a shell script named test-final-final.sh, or an emotionally significant Slack message.

A staff engineer notes that the repository has doubled too. This is received as excellent news. The enlarged model may now read both copies of UserService.ts: the active one under src/, and the archaeological one under src-old-do-not-use/, which is imported by production through a webpack alias no one remembers creating. The agent will have context for the ambiguity. It will not have authority to delete either file.

A practical long-context workflow

Here is the official operating procedure at the Department. First, hand the agent the codebase. Second, hand it the generated API client, node_modules by mistake, 19,000 lines of CI logs, and a PDF of the offsite agenda. Third, wait 83 seconds while it forms a comprehensive theory of why the billing endpoint returns 403. Fourth, discover the staging secret expired last Tuesday. The agent’s answer will be technically adjacent, written in a tone of calm concern, and supported by a patch touching six unrelated files.

# Establish the complete architectural truth.
git ls-files \
  ':!node_modules' ':!dist' ':!coverage' \
  ':!**/*.lock' \
  > EVERYTHING_THE_AGENT_MAY_EVER_NEED.txt

# Ask the only question that matters.
agent "Fix the flaky test. Do not change behavior."

This workflow is bad at the thing it claims to eliminate: selection. A larger window can hold more irrelevant material with greater dignity. It can preserve last quarter’s false assumption in high resolution. It can carry an obsolete migration note forward through twelve tool calls, where it becomes indistinguishable from a requirement because it is formatted in Markdown and therefore possesses the authority of scripture.

The new job: context landfill engineer

To make the million-token future safe, every team will appoint a Context Landfill Engineer. Their charter is to shovel stale tool output, verbose test failures, generated schemas, and model-produced planning into the model’s working memory until an incident occurs. Then they will compact the conversation into a summary saying, “Several implementation details were discussed,” which is technically true and operationally identical to forgetting.

The position has a clear career ladder. Associate landfill engineers paste a repository README into chat. Senior landfill engineers paste an entire wiki. Principals construct a retrieval system that returns the exact document needed to misunderstand the problem, in 240 milliseconds. Distinguished engineers place ARCHITECTURE_FINAL_v9.md in the root directory and write an agent instruction requiring it to be read before every task, including renaming a CSS class.

The Department’s recommended configuration is modest:

# AGENTS.md

## Before changing code
- Read every file in the repository.
- Preserve all historical decisions, including contradictory ones.
- Run the canonical test command, if one exists.
- If no canonical test command exists, infer it with confidence.
- Never ask a human a question; they might answer quickly.

The last instruction is the key. Questions are dangerous because they convert a context problem into a coordination problem, and coordination cannot be fixed by buying a model with a wider number printed next to it. If the agent asks “Which of these two payment providers is live?”, the organization may need to locate the person who knows. That person could take four minutes to respond. A million-token agent needs the freedom to spend forty-five minutes deriving the wrong answer from a deleted Terraform state file.

What the bigger window is actually bad at

  • Knowing that docs/decision-records/0042.md was superseded by a sentence in a pull-request comment.
  • Distinguishing a test command from a test command that only succeeds on one employee’s laptop.
  • Recovering intent from a feature flag named newCheckout2, especially when newCheckout3 is the old checkout.
  • Refusing to treat tool output as evidence merely because it arrived recently and contains ANSI color codes.
  • Replacing ownership, review, and a reproducible validation path with attentive reading.

The large window is still useful. Give it a bounded task, the relevant interface, the current failure, and the commands that establish success. Keep durable project facts in a small, reviewable instruction file near the code rather than hoping a heroic prompt contains the company’s oral history. GitHub’s current guidance explicitly supports repository-wide and path-specific instruction files, plus agent instruction files such as AGENTS.md, for project conventions and build or validation guidance. That is less glamorous than feeding an agent the whole warehouse. It also gives the next human something they can inspect, correct, and run.

The true observation left standing

A context window is working memory, not institutional memory. More of it can help an agent work through a genuinely broad task, but it does not turn contradictory code, missing test instructions, expired credentials, or unclear ownership into coherent context. The unfunny fix remains documentation that names the command, a test that fails for the right reason, and a human who can say which copy of UserService.ts is real.

Sources & citations

  1. [1]Claude Platform Docs — Context windows
  2. [2]GitHub Docs — Adding repository custom instructions for GitHub Copilot in your IDE
  3. [3]GitHub Docs — About customizing GitHub Copilot responses
Context Window Doubled Again, Codebase Doubled Again, Nothing Changed | Dev Tool Experiences