cxcap intent audit finds 3 currency-rounding touchpoints, while separate focus on money.py hotspot exposes 42 files in demo repo

Run `cxcap audit . --focus core/money.py` and the tool reads what is on disk, then answers a narrow question about one file: who imports it, and how far those imports carry. On the demo repository the answer was 23 outside files importing a 42-line helper, plus 19 more files reached through those ^[7]. The run needs no index and no model, and the whole thing installs as one binary with `cargo install cxcap` ^[9].

Two numbers describe that same file. It holds 2.6 percent of the repository's measured complexity, and the files it can reach come to 42 of 83, a little over half the project ^[7][1].

For 42 to mean anything in a different codebase, the shape has to match. The demo is synthetic, built for the article, with auth, billing, notifications, API and workers across 83 files of Python and TypeScript ^[4]. Currency formatting is a favourable pick for a demonstration, since one money helper tends to get imported by everything that prints a number. The author's version of that point: "Rounding is never just rounding when one helper formats every price, receipt, refund, and digest in the system" ^[20]. The count sees imports. Coupling that arrives through a plugin registry, or through a key resolved at runtime, does not show up in it ^[7].

The intent command is the cheaper half of the workflow. `cxcap audit . --intent "change currency rounding"` returned 3 likely touchpoints and a reasoning surface of 15 files across 5 components ^[5], five files of surface for every touchpoint ^[2]. All the figures are verbatim output from v1.0.2 runs with nothing estimated, according to the post ^[8]. The post does not report a before-and-after: no run where an agent did the same task with the audit and without it.

The token argument explains how the cost arises, and it arrives without a number attached. A more complicated repository can require an agent to inspect more files, ingest more context, make more tool calls, reason across more relationships and retry failed approaches, and agent platforms meter usage by tokens, model, context and workload ^[17]. "I won't claim any percentage savings on anyone's AI bill. That hasn't been proven," the author wrote ^[16]. The shorter line the author does commit to: "Technical debt can now have a token bill." ^[18]

Supporting evidence comes from elsewhere. GitClear's dataset of 623 million code changes over the AI-adoption period shows block duplication up 81 percent, within-commit copy/paste up 41 percent and error-masking constructs up 47 percent, against refactoring line moves down 70 percent and long-term legacy maintenance down 74 percent ^[13]. Whether that correlates with AI adoption is GitClear's to interpret and not his to overclaim, the author writes ^[14]. He reads DORA in the same direction, with some of the time saved on creation reallocated into auditing and verification ^[15].

What you install is small. What you adopt is a step in the loop, which the author describes as task, inspect local context, implement, test, fix, ship, next task, with every step passing while coupling and duplicated behavior accumulate ^[2]. The audit goes in front of implement. Its output is a verdict of LOW, MODERATE, HIGH or SEVERE, summarizing how strongly measured complexity should constrain the next change ^[11].