Claude Code First, specification version 0.1.0. The plain-language guide
Version 0.1.0. Grade R1/S3.
F410 Measured on one authoring task, editing projected files cost an agent less than writing through the API, and the saving was in what the agent writes, not in what it reads
| Field | Value |
|---|---|
| Status | active |
| Topic | Environment, permissions, and agent-facing products |
| Grade | R1/S3 |
| Cases | Shio |
| Contrary | none |
| Harness facts | none |
| Use | background |
Claim
Shio benchmarked one canonical authoring task, renaming a product in the body of ten of thirty posts, through its projected content files and through its HTTP API on one fixture, and the file path cost less in all and several times less in what the agent itself wrote, while it received almost as much as the API path did.
Mechanism
On the file path the CLI makes the HTTP calls, so they never reach the agent's context; what does is the command output, a Grep of the tree, and for each file a Read and an Edit. The Read is not optional: the harness refuses an Edit on a file the session has not read, so every changed file is read whole, which spends most of what the projection saves on the way in. The Edit then carries only the changed line. Through the API the agent receives the posts once but must send each changed post's whole body back, so the difference lands on output tokens, the dearer kind.
Observations
- Shio. SH1141 counts both paths on one fixture, only what reaches the agent's context
and in two directions, and holds the result as ceilings and ratio floors [shio@0dd3b9119:shio-app/src/test/resources/token-budgets.properties#L1213-L1216]. The file records the figures, "Measured 2026-09-25: files 4072 tokens (760 written by the agent), API 5245 (2390 written)." [shio@0dd3b9119:shio-app/src/test/resources/token-budgets.properties#L1205], and the reading, "1.29x in all, 3.14x of what the agent writes." [shio@0dd3b9119:shio-app/src/test/resources/token-budgets.properties#L1206] The test is [shio@0dd3b9119:cli/test/authoring-files-cost.test.mjs], and the commit is past Shio's pin, which corpus.md lists with its reason.
Independence
One origin, one occasion. The benchmark was written for this study's open question, by a session working for it, which the Counter-evidence weighs.
Counter-evidence
- The benchmark is one task shape. A task whose edit is most of the file, rewriting a post
rather than a line of it, would close the gap, and the benchmark does not measure one.
- It is generous to the API on purpose, with the scripted instance's lean response shapes
and a write that sends the body field alone; the real endpoints carry more, so the ratio is a floor on this shape.
- It models the agent's tool calls rather than recording a session: the Read and Grep
outputs are reproduced in the harness's format, not captured from one.
- It was written to answer the question it answers, by an agent working on this
specification, which is a threat to its construct that a benchmark written for Shio's own needs would not carry.
Grade
R1: one project, one occasion. S3: counted figures, held by a test whose floors would fail if the file path lost its edge, but not yet seen to fail.