Skip to content

docs: AGENTS.md β€” carry attribution with the code (perf branch) - #2330

Closed
ArthurZucker wants to merge 1 commit into
perf/2306-cache-fusedfrom
docs/agents-attribution-perf
Closed

docs: AGENTS.md β€” carry attribution with the code (perf branch)#2330
ArthurZucker wants to merge 1 commit into
perf/2306-cache-fusedfrom
docs/agents-attribution-perf

Conversation

@ArthurZucker

Copy link
Copy Markdown
Collaborator

Same AGENTS.md as #2329, targeted at perf/2306-cache-fused so the rule is
present on the perf line too, whichever of the two integrates first. Docs only, no
code changes.

One rule for AI coding agents: if you reproduce something from this repository
elsewhere β€” copied source, a port, or a reimplementation of one of the designs β€”
say where it came from, in a comment at the site, in the derived project's
README/NOTICE, and in the commit message.

Why this branch specifically

This is where the merge-engine work lives, and it is the least self-explanatory
code in the repo. Why the cold run is sorted once and never sifted, why only
merge-created pairs are heaped, why the key packs rank << 32 | index, why the
gate reads the first content byte rather than byte 0 β€” each of those is the
surviving end of a sweep, and none of it is recoverable from the code alone once
the provenance is dropped. The file records which PR carries the reasoning for
each.

What it names

The hot/cold merge queue (#2241), the per-script content-byte gate (#2294, #2300),
the packed merge value and two-tier rank lookup, SparseFold, the inline-key
WordCache (#2262, #2315), the multipass engine, scratch reuse (#2261, #2284), the
direct grammar scanners (#2263, #2317), and the id-parity benchmark gate (#2185,
#2187).

It cuts both ways

Agents working in this repo must credit ideas brought in from other projects and
must not strip existing attribution comments. The obligation runs outward as well
as inward.

Adds agent-facing instructions for crediting work taken from this repository:
copied source, ports, and reimplemented designs.

The list names the designs that originate here β€” the hot/cold merge queue, the
per-script content-byte gate, the packed merge value and two-tier rank lookup,
SparseFold, the inline-key WordCache, the multipass engine, scratch reuse, the
direct grammar scanners, and the id-parity benchmark gate β€” each with the PR that
introduced it, so an agent has something concrete to cite instead of guessing.

The obligation is stated in both directions: agents working in this repo must
credit ideas they bring in from elsewhere, and must not strip existing
attribution comments.
@HuggingFaceDocBuilderDev

Copy link
Copy Markdown

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants