Beagle SCM

POST-021: git-scheme push (be post/be put ssh://) fails with BNOROOM

be post ssh://localhost/src/dogs and be put --nosub ssh://localhost/src/dogs?/main from a tree AHEAD of the git upstream fail the push outright: after recursing the abc submodule the parent push dies fatal: the remote end hung up unexpectedly + Error: BNOROOM + Error: BEDOGEXIT, and the upstream is NOT advanced. be head against the same remote fetches fine, so the read side is healthy — only the send/receive (push) path is broken. Locally-committed work is stranded: it cannot be published, which blocks the fix-and-land loop (isolated agents can't clone a base that includes it). Distinct from POST-020 (cosmetic stderr on this same path, already landed).

Input

Context

A tree N commits ahead of its git remote cannot push; the wire dies mid-send.

• Repro: in /home/gritzko/beagle.new (cur 25fb62bd DOG-003, a clean ff 8 commits

  ahead of ssh://localhost/src/dogs?/main), run be post ssh://localhost/src/dogs or be put --nosub ssh://localhost/src/dogs?/main.

• Output: the abc submodule push runs (synthetic ?.dogs/ → git default branch,

  lists ABC-004/PTR-009 + add abc/BIT.h …), then post: 0 change(s), then fatal: the remote end hung up unexpectedly / Error: BNOROOM / BEDOGEXIT.

• Contrast: be head ssh://localhost/src/dogs?/main fetches OK (Total 0 (delta 0)…)

  and correctly lists the 9 ahead commits — the READ path is healthy.

Goals

be post / be put to an ssh/git remote must complete the ff-push without BNOROOM.

• The remote ref advances to the local tip; a re-run be head shows 0 ahead.
• Repro test: push a tree a few commits ahead of a git remote, assert it advances

  (mirror the git-receive-pack drop-in harness, keeper/KEEP.cli.c).

Constraints

• ABC idioms; ssh/git remotes use --nosub (submodules manual). ff-only, never forced.

WIP

Design decisions

(pending root-cause)

TODOs

• Find the BNOROOM raiser on the SEND path. Leads: keeper/KEEP.c:742

  (if (u8bIdleLen(out) < outsz) return BNOROOM — a fixed-cap out buffer), the git push send in keeper/WIRE.c / keeper/KEEP.exe.c (receive-pack peer) / keeper/KEEP.cli.c. BNOROOM is an abc buffer overflow → a send/pkt buffer is likely undersized for the pack stream (cf. the dog/WEAVE.c rms NOROOM fixed in DOG-003 — same undersized-buffer class).

• Check whether the MULTI-LINE commit message on 25fb62bd trips the wire encoder

  (work.mkd: wire messages "may be multi-line once the parser allows"); if so fix the encoder, or enforce the 1-line/≤64-char commit rule and re-test.

• Check the submodule-recursion interaction (the abc sub-push precedes the failing

  parent push; does the parent inherit a spent buffer / bad state?).

• Repro-first (CLAUDE.md §17): a failing push test, then fix.

Blockers and bummers

• Complete breakage of git-scheme publishing. The DOG line is stranded local-only on

  beagle.new; DIS-043 and onward can't land via the shared remote until fixed.

Outcome

• (open)