Test Claude Code Token-Cut Workflows on TypeScript

The demo is not your repo

Your team is staring at a Claude Code bill and a cheaper worker model that may cut token spend. Do not make it the default until your TypeScript repo has been through a bounded audit. AI coding costs are projected to exceed the average developer's salary by 2028. A quarter of engineering leaders already pay $200 to $500 per developer per month for tokens, and some are past $2,000.

The examples used Gemini 2.5 Flash as the worker model, although the Portal model field can accept other configured models. The shunt plugin stops Claude from reading a file when its length is above a user-adjustable cutoff, set to 350 lines unless changed. Run those mechanics on your TypeScript repo before making it the default.

The failure mode is the other half of the bill. The worker model failed to detect a thread-safety defect, while Claude found it within seconds after receiving the relevant context. If the cheap path saves tokens and then ships that defect, the savings become a problem you own.

Run the audit in a repeatable order

• Baseline Claude Code token usage, latency, and cost for a normal work cycle on your TypeScript repo. Track the baseline in a spreadsheet with date, task type, model, tokens, dollars, and wall-clock time. Include interactive sessions and batch runs, because a token-saving workflow can look great in some modes and bad in others. Do not optimize a number you cannot see.
• Reproduce the worker-model routing on a branch or a staging environment. Use the same model field, the same shunt cutoff, and the same bulk-read path as the reference setup. The branch should let a developer trigger the cheap path without changing the main agent's behavior. Record the cutoff and model choice in the branch notes so the next person can repeat the setup. Do not mix another change with the routing change; change only the routing variable, or the delta table becomes a result you cannot defend.
• Measure the same tasks through the main and cheap paths, then compare tokens, cost, and latency. Produce a delta table: tokens saved, dollars saved, median and worst-case latency, and timeout rate. Delegated responses typically took 10 to 30 seconds, and Portal limits a single invocation to 30 seconds. If your tasks hit that ceiling, split the work before you celebrate the savings.
• Run code-quality and failure-mode tests on your own TypeScript patterns. Include concurrency, async state, and long files. Log defects in a short list with severity, reproduction steps, and whether the cheap path caught it. Ask the cheap path to explain its reasoning on a known-bad example. If it cannot say why a pattern is unsafe, treat that as a signal, not a curiosity. If the cheap path misses the class of bug your team ships, the token cut is not free.
• Set a default or no-default rule in writing. Write a policy note: which task types may use the worker model, which must stay on the main model, and what metric triggers a rollback.

The evaluation used a Java codebase with four test cases and reported average token reductions of roughly 90% for the bulk-read path. Treat the cut as a routing result, not a quality guarantee.

Make the default decision explicit

The default should match the measured risk. If the cheap path saves meaningful money on low-risk tasks and does not raise the failure rate on your TypeScript code, make it the default for those tasks only. If it saves money but misses subtle defects, keep it as an opt-in for bulk reads and leave correctness-sensitive work on the main model.