The expensive part is often the cargo.
A Spotify engineer describes a familiar coding-agent bill: a frontier model reads thousands of lines to answer one narrow question, or emits a predictable test file by copying a pattern already sitting next door. In the published Java-monorepo scenarios, the Shunt plugin’s bulk-reader path reduced the main model’s consumed tokens by 82–94%, with a reported mean near 90%.
Keep the dyno sheet attached. Those are project-published measurements from four scenarios, not a universal savings rate. The worker model, repository shape, summary format, cache behavior, and question all matter. Delegation also adds a network round-trip. Below the plugin’s default 350-line threshold, the overhead can be worse than simply reading the file.
Do not route by prestige. Route by how much judgment the cargo contains.
The hook is the product.
The first version used advice in CLAUDE.md. Advice is context, and context can be ignored. The current plugin registers PreToolUse hooks. A full read over the configured line threshold is blocked and redirected to a bulk-reader skill; targeted reads with an offset or limit still pass.
That distinction is excellent interface design. The hook is a hard guardrail. The skill explains the detour. The script handles transport and output cleanup. Anthropic’s own documentation makes the same boundary explicit: project instructions are context, while a PreToolUse hook is the mechanism for blocking an action regardless of the model’s choice.
Do not shunt the steering wheel.
Spotify’s write-up is refreshingly clear about failure modes. The worker missed a subtle thread-safety bug that the main model found. Summaries did not provide reliable enough line numbers for edits. Very large requests hit command-line payload ceilings, and remote calls add latency and timeout risk.
So the useful split is not “cheap work” versus “smart work.” It is compressible work versus decision-bearing work. A bulk inventory can be summarized. A reference-shaped config can be drafted. But exact edits require exact spans; debugging requires causal reasoning; architecture and safety decisions deserve the full driver.
Build the return lane before the shortcut.
- Set a visible threshold. Use observed file size and latency, not an invisible vendor default.
- Keep targeted reads legal. A summary should narrow the search, not become an irreversible loss of detail.
- Pin the worker contract. Name the model or mode, instructions, temperature, tools, payload ceiling, and timeout.
- Separate read from write. A cheap worker may summarize broadly; repository mutation still crosses a validation gate.
- Print the receipt. Record what left the main context, what came back, and which model retained final responsibility.
Shunt is currently Claude Code-specific for enforcement, though the Portal marketplace also describes workflows for Codex and Cursor. The portable idea is bigger than one plugin: context is a budgeted supply chain. Put the forklift on bulk cargo. Keep a human-readable manifest at the dock.