Claude Code First, specification version 0.1.0. The plain-language guide

Version 0.1.0. Grade R2/S3.

F205 Git metadata does not show which commits an agent composed: the co-author trailer is on a few percent of commits, unevenly

Claim

In projects where an agent is the primary author, a Claude co-author trailer is on 0 to about 7 percent of commits, clustered in time rather than applied as a policy, and the author name Claude appears only on a small set of commits, most of them carrying a claude.ai session link; so neither field can separate the agent's commits from the person's.

Mechanism

Commits are made under the owner's git identity. The trailer is present when the commit was composed in a way that adds it and absent when the owner's commit tool writes the message, which no source says adds one; so its presence tracks how a commit was made, not who made the change. Commits made in a web session carry the author name Claude and a session link. No project at the pins states an attribution policy. This mechanism is an inference from the counts below; the sources do not state it.

Observations

All counts are this study's, made at the pins with the commands given in the case studies.

• roadkeep. 76 of 1,946 commits carry a Co-Authored-By trailer, from 2026-07-29

[roadkeep@40847570] to 2026-09-01 [roadkeep@595894af], and none of the 423 after it; before that date the share was 76 of 1,523. 69 of the 76 have a body with no hyphen-led line, against 124 of all 1,523 commits to that date, while generated bodies are bulleted lists, as in [roadkeep@13272828]. Four commits are authored as Claude, for example [roadkeep@7b653977] and [roadkeep@372c52a7]. No committed file states an attribution policy.

• freewilly. 17 trailers in 444 commits. Twelve are the commits of one pull request, merged

from the branch "alegauss/claude/bloco-h-roadmap-tmtjgk" [freewilly@7968cc5], all authored as Claude and each carrying "Co-Authored-By: Claude Opus 4.8" [freewilly@915e22e] and "Claude-Session: https://claude.ai/code/session_" [freewilly@915e22e]. The skill states that "Every word in this repository was written by a model" [freewilly@c1c2eaf:.claude/skills/ai-writing-freewilly/SKILL.md#L8].

• winwright. 33 of 708 commits carry a Claude trailer, 24 of them dated 2026-08-24; for

example "Co-Authored-By: Claude Opus 5" [winwright@15dc5db]. No commit is authored as Claude.

• Shio. 119 trailers in the 1,668 commits from the adoption commit [shio@6bf11b754] to the

pin, none before it. 55 commits are authored as Claude, from 2026-04-11 [shio@c652a2cba] to 2026-08-16 [shio@daecef641], each with a claude.ai session link. The commits that bypassed the tool to stage by path carry a trailer, as in [shio@871410748] and [shio@745823e62].

• polyweave. No trailer in 132 commits, and one author name on all of them

[polyweave@6d1c136].

Independence

Counted once. The trailer's presence depends on how a commit was made, and the ways are shared: the owner's commit tool, the harness's own commit convention, and web sessions. The provenance of the pattern in each project is therefore not independent, and where it is unclear ../grading.md says to count once. The web-session commits in freewilly and Shio are one harness behaviour (rule 3).

Counter-evidence

None found at the pins. The trailer is not absent everywhere: in Shio, 119 commits do carry it, so it identifies some agent commits; the claim is that its absence identifies nothing. The design of the roadkeep case read the stop after 2026-09-01 as a change of policy; the pinned sources hold no policy (an attribution rule exists only in uncommitted work outside the pin), so that reading is not supported and is not used here.

Grade

R2: one origin, observed repeatedly over separate periods, in roadkeep alone from 2026-07-29 to 2026-09-01 and then not at all. S3: the shares are counted at the pins with repeatable commands, git log <pin> -i --grep=Co-Authored-By --format=%h counted with wc -l for the trailers and git log <pin> --author=Claude --format=%h for the author name, and what they count is the finding's own effect: how many commits carry the attribution. The mechanism in the Mechanism section is an inference and is not raised by these counts.