You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
Blob-size and object-count clone filters (P27) #217
sc clone --filter is prefix-only, matching the sparse path-boundary machinery. A size-based filter (git's --filter=blob:limit) and an object-count limit are a separate axis, deferred in ROADMAP.md (P27).
Needs a decision. A size filter leaves blobs absent by size rather than by path, so the "filter records what a clone asked for, not what is present" rule and the presence-based partial-clone model (ADR-0048) need to cover a second kind of gap: lazy fill by id already exists (ADR-0047), but sc backfill <prefix> does not name a size gap.
What is the user-facing repair for a size gap (sc backfill --all, sc cat, a --min-size form)?
Does the server evaluate the filter, or the client? (A server-side filter changes the GetPack request, so check the wire version.)
Interaction with protected content, which is opaque by size.
Acceptance criteria (once decided)
sc clone --filter blob:limit=<size> produces a working partial clone; checkout materializes small files and reports or lazily fills large ones.
Tests for push, verify and gc on such a clone, and for the gap error text.
PR #262 ships the blob-size filter (sc clone --filter blob:limit=<size>, ADR-0055, wire v6). Still open on this issue:
A limit-aware sc fetch. Fetch stays unfiltered, so the limit bounds the initial transfer, not later growth. It needs per-path exemptions for backfilled files in the request.
An object-count clone limit. Deferred: the transfer set is id-ordered, so "the first N objects" has no stable meaning, and git has no such filter.
Also open (ADR-0055 open questions): a server-side size index so a limited request need not stream-measure blobs.
feat(repo,cli): limit-aware sc fetch; close the object-count limit (#217) #269: limit-aware sc fetch. A fetch from the promisor origin sends the recorded limit, so later history growth stays bounded and every fetched tip stays complete (except a blob the clone already gapped, which a reverting tip names without it). sc fetch --no-blob-limit opts out for one command. The sender skips measuring blobs the client already holds, so a no-op fetch measures nothing. Per-path exemptions for backfilled files turned out to be unnecessary (ADR-0055 decision 9).
Not shipping: the object-count limit (ADR-0055 decision 10). Git has no such filter. The transfer set is ordered by object id, so "the first N objects" is an arbitrary, server-dependent set and a client cannot tell a cut-off object from a missing one. The deterministic readings are other features: the newest N commits is a shallow clone, and the N smallest blobs is the size filter again. blob:limit and path prefixes cover the need of bounding a clone.
Still open, tracked in ADR-0055 and not part of this issue: a server-side size index so a limited request need not stream-measure blobs.
Activity
feat(repo,cli): blob-size clone filter, wire v6 (ref #217)
docs: ADR-0055 accepted, partial-clone docs and blob-limit demo (ref #…
fix(repo): let a size-only clone restore its full view; review follow…
tonibergholm-codento commentedon Oct 20, 2026
PR #262 ships the blob-size filter (
sc clone --filter blob:limit=<size>, ADR-0055, wire v6). Still open on this issue:sc fetch. Fetch stays unfiltered, so the limit bounds the initial transfer, not later growth. It needs per-path exemptions for backfilled files in the request.Also open (ADR-0055 open questions): a server-side size index so a limited request need not stream-measure blobs.
test,docs: assert fixture cleanup, split the CLAUDE.md refusal clause…
fix(repo): harden blob filter parsing and gap messages (ref #217)
test,docs: named sessions on a size-limited clone (ref #217)
feat(repo,cli): blob-size clone filter (sc clone --filter blob:limit=) (
docs(adr): note the cost of a limited fetch against a pre-fetch v6 se…
tonibergholm-codento commentedon Oct 20, 2026
Closing out #217 with PR #269 (merging it closes the issue).
Shipped:
sc clone --filter blob:limit=<size>(ADR-0055, wire v6).sc fetch. A fetch from the promisor origin sends the recorded limit, so later history growth stays bounded and every fetched tip stays complete (except a blob the clone already gapped, which a reverting tip names without it).sc fetch --no-blob-limitopts out for one command. The sender skips measuring blobs the client already holds, so a no-op fetch measures nothing. Per-path exemptions for backfilled files turned out to be unnecessary (ADR-0055 decision 9).Not shipping: the object-count limit (ADR-0055 decision 10). Git has no such filter. The transfer set is ordered by object id, so "the first N objects" is an arbitrary, server-dependent set and a client cannot tell a cut-off object from a missing one. The deterministic readings are other features: the newest N commits is a shallow clone, and the N smallest blobs is the size filter again.
blob:limitand path prefixes cover the need of bounding a clone.Still open, tracked in ADR-0055 and not part of this issue: a server-side size index so a limited request need not stream-measure blobs.
feat(repo,cli): limit-aware sc fetch and --no-blob-limit (#269)