alkgit/docs/research/pocs.md

alkgit POC Plan

POC modes

POCs are not production code — fewer constraints (comments, unwrap, error taxonomy are all relaxed), and their code never merges to main. Two modes, chosen by whether the POC touches code in this repo:

• Branch mode (POC depends on alkgit code in the repo): create a git branch off main, do the work there, then merge the findings into the research docs on main and drop the branch — the POC code itself does not merge.
• Standalone mode (POC is independent of alkgit code): a scratch crate or project at /workspace/<poc-name> (e.g. /workspace/alkgit-poc1), outside the repo entirely. Findings are written directly into docs/research/ on main.

In both modes the deliverable on main is the findings write-up, not the POC source. POC source is disposable once findings are recorded.

Each POC records: hypothesis, method, result (proceed/pivot/block), and what it changes in the research docs.

POC-1: pkt-line over alkcall BiStream

Status: complete (2026-09-20) — proceed. See poc-1-findings.md. Real git 2.43 clones/fetches/fsck-clean over the alkcall Connection → BiStream → tokio-util compat → gix-packetline path; the flagged futures-io risk resolved with a two-line compat adapter.

Hypothesis: a full V2 fetch handshake (ls-refs + one fetch round) between real git CLI (as client, git clone/git fetch) and a Rust listener that bridges alkcall BiStream to gix-packetline's async codec can be completed.

Method sketch: alkcall Connection accepted over a local stream; spawn a handler that receives a BiStream; wrap read_half/write_half in the pkt-line async reader/writer; advertise V2 capabilities with honest feature list; parse command=ls-refs, emit refs from a fixture repo (created with gix), flush. Client side: git -c protocol.version=2 clone against a local bridge (POC can start with plain tcp and only simulate the alkcall side if alkcall dialing adds friction — the alkcall-fit part can be tested with Connection::from_stream).

Success: git prints its ref advertisement fetch result without error; all framing validated against real git's parser (the strictest pkt-line validator available).

Risks: gix-packetline async-io feature uses futures-io traits — confirm alkcall stream halves implement AsyncRead/AsyncWrite compatibility (they should, being tokio io objects; may need a thin adapter).

POC-2: server-side pack generation

Status: complete (2026-09-20) — proceed. See poc2-findings.md. Real git 2.43 clones/fetches (fsck-strict clean) packs generated by gitoxide's own pipeline (count::objects + iter_from_counts + FromEntriesIter) streamed over sideband; memory is O(counts), not O(pack) — 60k-object pack written through a 64 KB sink at ~11 MB extra RSS. Course correction recorded: TreeContents does not follow commit parents — the closure needs a commit-ancestry walk feeding the count stage.

Hypothesis: given a fixture repo and a want/have set, we can produce a valid pack stream in memory that git verify-pack/git unpack-objects accepts, using one of: a. gix-pack::bundle::write into a temp dir then read the file back (correctness baseline), or b. gix-pack::data::output::bytes fed by an odb object walk (streaming).

Method: build fixture repo via gix API or a scripted git (fixture creation may shell out to git — only serving-path must be git-binary-free). Try (b) first; fall back to (a) to establish the correctness baseline.

Success: git clone from a bridge that serves the generated pack completes and git fsck passes in the clone.

Decision to make: streaming composition that doesn't materialize the whole pack in memory for large repos — record memory behavior for a 10k-object repo at minimum.

POC-3: smart-http shape through alkhttp

Status: complete (2026-09-21) — proceed. See poc3-findings.md. Real git 2.43 clones/fetches/tag-checkouts over http:// through the full alkcall → alkhttp path (Connection::from_bidi(http/1.1) → HttpAdapter → axum custom routes); request bodies stream (measured via dribble probe + git's forced-chunked path), pack responses stream under back pressure with O(counts) memory on a 60k-object clone. Key protocol correction: http V2 responses end at the flush — the 0002 response-end pkt is synthesized by the client's transport helper, never sent on the wire.

Hypothesis: alkhttp can serve GET /info/refs and stream a POST body (ingest without full buffering) for git-receive-pack.

Method: minimal alkhttp service exposing a fake /repo.git/info/refs?service=git-upload-pack and /repo.git/git-upload-pack wired to the POC-1 bridge; run real git clone http://... against it; measure whether alkhttp's request-body API streams or buffers (inspect/measure, not guess).

Success: real git clones over http; documented answer on body streaming for receive-pack sizing.

Sequencing

POC-1 first (the BiStream/packetline fit is the load-bearing assumption). POC-2 next (the crux of fetch). POC-3 last (http shape). Each POC updates the corresponding research doc with results and a proceed/pivot/block note.

Status: POC-1, POC-2, and POC-3 complete and proceed — phase 0 gates all pass.

Modes applied to these POCs

• POC-1, POC-2: standalone mode — nothing in the alkgit repo exists to build on yet; run as /workspace/alkgit-poc1 and /workspace/alkgit-poc2. (If POC-1's bridge turns out to want the workspace deps, it can flip to branch mode on the repo instead — decide before writing, not after.)
• POC-3: branch mode — it exercises alkhttp wiring against this repo's skeleton; findings merge into alk-stack.md, branch is dropped after.