Beagle SCM

SUBS-022: a RELATIVE argv[0] breaks sub-recursion — execvp(self) after chdir into the sub mount fails (BEDOGEXIT)

When be is invoked via a RELATIVE path (e.g. ./build-debug/bin/be), its submodule recursion forks, chdirs into <wt>/<subpath>, then execvps itself — but argv[0] (and the bin dir HOMEResolveSibling derived from it) is still relative to the ORIGINAL cwd, so after the chdir it no longer resolves and the child execvp fails → BEDOGEXIT, the whole recursing verb aborts. Hit twice this session (a be post default sub-recursion, and the bare be diff: investigation). Method: Issues.

Input

Context

• Repro: cd $HOME/todo/<wt> && ./build-debug/bin/be post -m '…' on a tree with a mounted submodule → be: BERecurseInto: execvp …: No such file or directory / BEDOGEXIT. Workaround used by the worker flow: be post --nosub (top-level only), or invoke be by an ABSOLUTE path / bare name on PATH.
• Code: beagle/SUBS.c:90 be_resolve_self → HOMEResolveSibling (SUBS.c:88-89: "walks PATH for a bare argv[0] and uses dirname for an ABSOLUTE one … falling back to the bare name"). A RELATIVE argv[0] (has a /, not leading /) hits the dirname arm → a relative bin dir. BERecurseInto (SUBS.c:82) then fork + chdir(<wt>/<subpath>) + execvp(argv_c[0], …) (SUBS.c:170; same at :231 BEIndexMount, :331) — the relative path is now wrong post-chdir → ENOENT.
• Bare name (PATH-resolved by execvp) and absolute path both work; ONLY a relative invocation breaks. This bites local build-*/bin/be invocations, which is exactly how the fix-and-land worker trees run be.

Goals

• A relative argv[0] is resolved to an ABSOLUTE self path ONCE at startup (against the launch cwd), BEFORE any sub-recursion chdir, so every execvp(self) / execvp(<dog>) child resolves regardless of invocation form. Bare-name and absolute invocations unchanged.

Constraints

• Resolve against the ORIGINAL cwd before the first chdir; don't re-derive from a post-chdir cwd. beagle/SUBS.c (+ maybe HOMEResolveSibling) only; no behaviour change for absolute/bare-name invocations. Add a recursion-via-relative-path test.

WIP

Design decisions

• DONE (2026-06-16) — SITE MAP CORRECTED: the three SUBS.c sites are not the only spawns; the failing post recursion is a SEPARATE bepost_spawn_sub (beagle/BE.cli.c:2746, HOMEResolveSibling at :2758). ALL chdir'd spawns (post/get, sniff subs, woof) funnel through HOMEResolveSibling — so the fix is ONE point there, not in SUBS.c. In dog/HOME.c:1027-1053 (app code, NOT abc/): in the has-slash branch, detect a relative argv0 (u8csFind('/') present, not leading /) and realpath it against the launch cwd (via PATHu8bReal, fed the original NUL-terminated argv0) BEFORE taking the dirname — resolution runs before any chdir, so the launch cwd is still correct. Bare names keep PATH resolution; absolute argv0 unchanged; realpath failure falls back to prior behaviour. Slice endpoints only, no ptr-arith.

TODOs

• Repro-first: a hermetic submodule tree; run a recursing verb via ./rel/path/be; assert the sub child runs (no BEDOGEXIT) — RED today.
• Absolutize a relative argv[0] before any chdir; fix the three spawn sites.
• Test: relative, absolute, and bare-name invocations all recurse into the sub.

Blockers and bummers

• Surfaced by the worker flow (relative build-debug/bin/be); independent of the submodule MODEL — it's a self-path-resolution bug at the recursion spawn.

Outcome

• LANDED 453c55c2 (FF child of 3eba8e76; 306/306; no --force): one-point fix in dog/HOME.c:1027-1053 (HOMEResolveSibling realpaths a relative argv0 before dirname) — covers every chdir'd sub-spawn (post/get/sniff/woof) uniformly; beagle/SUBS.c gets a 3-line doc comment only, no abc/ change. New beagle/test/be-post-relpath.sh (rel/abs/bare invocation forms, RED→GREEN: rel hit the exact BEDOGEXIT, all 3 green after). 306/306. Awaiting gate. Msg: SUBS-022: realpath relative argv[0] before sub-recursion chdir.