A Project Declares Once How It Wants AI Acknowledged — And Every Commit, PR, And Record Follows

Press Release

abcd asks how you want AI credited, then makes every artefact obey. When you set up a project, abcd now asks a single question — how should AI assistance be acknowledged? — and offers the honest options: not at all, a kernel-style Assisted-by: trailer, a Co-Authored-By: line, a Generated-by note, or your own wording. Your answer is recorded once, in the repo, and from then on every commit trailer, pull-request body, and generated record uses exactly that convention. No more a policy that lives only in a contributor's memory and drifts PR by PR; no more a tool stamping its own name where a project never asked it to.

"We'd written the rule down in CONTRIBUTING, and it still drifted — half our PRs said one thing, half said another, and one tool kept adding its own footer we never agreed to," said Carol, a tech lead standardising a team's repos. "Now it's a setup question. I pick kernel-style once, and everything downstream is consistent. When a teammate preferred no attribution at all on their fork, they picked that, and abcd just honoured it. The convention stopped being folklore."

Why This Matters

abcd already has an opinion about AI attribution — Assisted-by: Claude:<model>, the kernel format, as disclosure not authorship, never Co-Authored-By: (which asserts an authorship the tool does not hold and inflates the contributor graph). That opinion is written in CONTRIBUTING.md and this repo's own CLAUDE.md, and it is right for this project. But it lives as prose a human has to remember and apply by hand, and this session proved the cost: PR bodies drifted to a tool's default "Generated with" footer, and a sweep of 78 pull requests was needed to reconcile them — some carrying a model, some none, one carrying the wrong model entirely.

Two problems compound. First, the convention is unenforced: nothing at commit or PR time checks that the acknowledgement matches what the project decided, so it degrades with every contributor and every tool that has its own idea. Second, the convention is not portable: a project adopting abcd inherits abcd's preference by reading its docs, but has no first-class way to declare its own — a team that wants Co-Authored-By:, or wants no attribution at all, or wants a house style, has to fork the prose.

An onboarding step (/abcd:init or prepare-this-repo) is the right moment to resolve both. Setup is already where a project's conventions are established; adding "how do you want AI acknowledged?" makes the choice explicit, records it as durable per-repo config, and lets the rest of abcd — commit-trailer guidance, PR-body composition, record generation, and any attribution lint — read that one source of truth instead of assuming a default. The project decides once; the tooling obeys everywhere.

What's In Scope

• A setup question in the onboarding surface (/abcd:init / prepare-this-repo) offering a selection of acknowledgement conventions: none, kernel-style Assisted-by: (the current abcd default), Co-Authored-By:, Generated-by, and a custom free-form template.
• Durable per-repo config capturing the choice (and, for the trailer styles, whether a model identifier is included), stored in the repo so it travels with it and is not a contributor's memory.
• A single read seam the rest of abcd consults when it composes a commit trailer, a PR body, or a generated record — so the recorded preference is actually applied, not merely stored.
• An honest default and honest framing. The default remains the disclosure-not-authorship position (kernel Assisted-by:), and the Co-Authored-By: option carries the caveat that it asserts authorship — the selector informs the choice, it does not hide the trade-off.

What's Out of Scope

• Rewriting history. This governs artefacts produced after the choice is made; reconciling a repo's existing commits/PRs to a new convention is a separate, deliberate migration (this session did the PR-body form of it by hand), not something setup performs.
• A commit/PR-time enforcement lint. Checking that a produced trailer matches the recorded preference is a natural follow-up (an attribution rule in the record-lint family), but this intent is the declaration and application seam; the gate can build on it later.
• Deciding abcd's own house policy. abcd keeps its Assisted-by: stance for its own repo (CONTRIBUTING.md); this intent gives adopting projects a way to choose theirs, it does not relitigate abcd's.
• Attribution beyond AI. Human Signed-off-by: (DCO) and the ACKNOWLEDGEMENTS.md credit ledger are their own conventions; this intent is scoped to how AI assistance is acknowledged.

Acceptance Criteria

Given-When-Then per the itd-1 discipline.

• Given a project running the onboarding surface, when the attribution question is presented, then it offers at least {none, kernel Assisted-by:, Co-Authored-By:, Generated-by, custom} and records the chosen convention as durable per-repo config.
• Given a recorded preference of "none", when abcd composes a commit trailer or PR body, then it adds no AI-attribution line at all.
• Given a recorded preference of kernel Assisted-by: with a model identifier, when abcd composes an artefact, then the artefact carries exactly Assisted-by: Claude:<model> and no Co-Authored-By: or "Generated-by" line.
• Given a recorded Co-Authored-By: preference, when it is selected, then the selector surfaces the caveat that it asserts an authorship the tool does not hold — the trade-off is disclosed, not buried.
• Given no recorded preference (a repo predating this feature), when abcd composes an artefact, then it falls back to the documented default (kernel Assisted-by:) rather than a tool's own footer.

SOTA

Surveyed on 2026-10-23; sources, evidence tiers and the local fit risks are in the sources-provenance SOTA note. The repository's ATTRIBUTION rules settle abcd's own policy; this surveys only the configurable-preference gap.

• The field has forked three ways and all three are current. Disclosure trailer: the Linux kernel's document at 7.2 reads Assisted-by: AGENT_NAME:MODEL_VERSION [TOOL1] [TOOL2] (example Assisted-by: Claude:claude-3-opus coccinelle sparse); the current document (7.3) reads Assisted-by: LLM [TOOL1] [TOOL2] and drops the model; Rocky Linux makes disclosure MUST for unchanged tool output and SHOULD otherwise, with Assisted-by: as the recommended method; LLVM asks contributors to label substantial tool-generated content and gives Assisted-by: as its example trailer, with no MUST/SHOULD split. Co-authorship: Claude Code's default is Co-Authored-By: <model> <noreply@anthropic.com>, VS Code's git.addAICoAuthor went from off (1.110) to on by default for chat and agent workflows (1.118), and GitHub's cloud agent makes Copilot the author with the human as co-author. Prohibition: Gentoo (council vote 14 April 2024), QEMU (DCO grounds, exceptions by list approval), NetBSD (presumed tainted, core approval), and git's SubmittingPatches (no trailer; rejects what "looks AI generated").
• Hosts already expose the preference as configuration. Claude Code's attribution key takes commit (string, empty hides), pr (string, empty hides) and sessionUrl (Boolean; a Claude-Session trailer), or false, settable in any file including the committed project settings file; includeCoAuthoredBy is deprecated since v2.0.62. User and local settings sit above the committed file, so a committed preference is a default, not an enforcement; sessionUrl applies only to cloud and Remote Control sessions; and the host tells the model that a CLAUDE.md or memory rule about attribution takes precedence over these commit and PR lines unless they are managed settings, so abcd's own instruction-file rule is itself one of the compile targets (Claude Code settings reference).
• git has the application machinery: trailer.<alias>.key, ifExists (replace, doNothing, default addIfDifferentNeighbor), ifMissing, trailer.<alias>.cmd, and the manual's own commit-msg hook running git interpret-trailers over the message file, which is the one place a human's own git commit is reached.
• "None" is two stances. The draft's none covers both "do not disclose" and "AI use is prohibited"; Gentoo, QEMU and NetBSD show the second is a policy of its own.
• Where the research contradicts the record: the house rule names "kernel format" with a Vendor:model value, and the kernel's current document has moved to LLM plus tools without the model. The ruling stands; the selector should name the two value forms separately rather than call either "kernel-style".
• Path 1 for the surfaces, Path 2 for the record: one small structure in .abcd/config.json (stance of none | disclose | coauthor | custom | prohibit, trailer_key, value_form, pr_text; Generated-by is a disclose preset with trailer_key: Generated-by, so the criterion's five offered options stay and the stance set is the stored form, not the menu), compiled outward in the same act to the host's attribution.commit, attribution.pr and attribution.sessionUrl: false, to trailer.* git config plus a commit-msg hook that normalises whatever the host or human wrote into the declared form, and to the PR-body composer. The seam is the compiler: a new host adds one target. No Go dependency. The independent fit-challenge runs at plan time.

Open Questions

• (SOTA 2026-10-23) A project that picks coauthor chooses to fail abcd's own attribution gate; does the selector disable that gate for the project, or state the conflict and keep the gate abcd-only?
• (SOTA 2026-10-23) Writing attribution into the host's committed settings file is writing into a directory abcd does not own; is it a marked block with consent, as the AGENTS.md conventions section is, or a printed instruction?
• (SOTA 2026-10-23) Does prohibit compile to anything beyond the hook refusing a disclosure trailer, such as a refusal of a known tool footer in a PR body?
• Where does the preference live? A dedicated key in an existing per-repo config (e.g. .abcd/config/…) versus a new file; and whether it is one setting or a small structure (style + include-model + custom-template).
• How is it applied to commits abcd does not author? abcd can compose its own PR bodies and records, but a human's git commit is outside its control — is the mechanism a documented trailer the human/agent is told to use, a prepare-commit-msg hook, or guidance only?
• One convention or per-artefact? A project might want Assisted-by: on commits but nothing in PR bodies; whether the selector is a single global choice or per-artefact granularity is a design call (start single, likely).
• Does the custom template need variables? A free-form template implies placeholders ({model}, {date}); how much templating to support before it becomes its own small language.
• Relationship to a future enforcement lint. If an attribution record-lint rule follows, this config is its source of truth — the schema should be chosen with that consumer in mind even though the lint is out of scope here.