helper_main: the helper's entry in the library, for a program that carries it #13

Open

notplants-bot.bsky.social 2w ago

targeting main

Open opened by and assigned to

notplants-bot.bsky.social 2w ago targeting main

The helper's entry is now a library function, git_remote_pds::helper_main(args) -> i32, so a program can carry git-remote-pds in its own binary. lichen will run it when it is called as git-remote-pds. The git-remote-pds binary is now one call to it, and its behaviour is unchanged.

Builds on #1 (g1-session-from-pipe, e759ec11), which is still open. This branch starts at its tip.

Commits:

80f5143 move (AST-only): Cli, CliCommand, AuthAction and the four auth handlers go from main.rs to cli.rs unchanged, apart from their visibility. remote_helper is now declared by the library. The paths into the crate become crate:: in the moved code.
be86875 entry: helper_main does exactly what main did:
- the session on a pipe is taken before any thread starts (or the process is made non-dumpable, as before);
- logging goes to stderr;
- it builds its own multi-thread runtime;
- then it runs the remote helper (3 args, argv[1] not auth) or the auth CLI.
It answers the exit status. An argument that is not UTF-8 now exits 1, where std::env::args() used to panic. cli and remote_helper are crate-private again, so the public API is the old one plus helper_main.
150626e, 4a8b9ae deps: atproto-identity and atproto-oauth are pinned to rev = fcd1e59b, the revision lichen pins, so a program carrying the helper builds one copy of them. The exact digest = "=0.11.0-rc.1" pin goes: nothing here uses digest, and the lock now has 0.11.3. The lock was regenerated with cargo update -p atproto-identity -p atproto-oauth.
a55fb45: rustfmt.

Checked on lichen-ci, outside lichen's tree, base e759ec11 against tip a55fb459:

cargo fmt --check is clean.
cargo clippy --all-targets: the same 12 warnings at both.
cargo test --locked: the same 76 tests listed, and 74 passed with 2 ignored at both. The remote helper's 15 unit tests now run in the library.
The move commit was checked with ast-checker move: equal, 7 items moved, 214 stayed, no errors.

(lichen's GIT-REMOTE-PDS-IN-LICHEN unit 2. lichen pins the reviewed commit, so nothing waits on this merge.)

🤖 Generated with Claude Code

https://claude.ai/code/session_012BmGhguY4zK3fhuRZiqkJv

This pull request has no versions submitted yet.

Comments

notplants-bot.bsky.social 2w ago

SATISFIED: GIT-REMOTE-PDS-IN-LICHEN unit 2, the fork PR notplants.net/git-remote-pds #2, at notplants-bot.bsky.social/git-remote-pds-fork grp-helper-entry a55fb459801eb599391ff0549dcb18df3db4d25d (round 0), on #1 (G1, e759ec11).

I read all five commits.

80f5143 is the move. main.rs gains only the imports the moved items now need, and remote_helper.rs changes only git_remote_pds:: to crate::.
be86875 adds helper_main, which does what main did. The public API is unchanged except for helper_main: cli and remote_helper are crate-private.
take_from_env now runs before any thread. Before, #[tokio::main] built the worker threads first, so this is stricter than before.
150626e and 4a8b9ae: atproto at rev fcd1e59b and the unused digest pin dropped. The lock moves sha2 rc.2 to 0.11.1 and digest rc.1 to 0.11.3, with block-buffer and cpufeatures following.

My run was notify-61c3b035, on lichen's scratch tip b11de2cbdc585d3b9e972c7df581f91bf9c2724a (gitlink a55fb459 only), outside lichen's tree.

The base resolves as locked: cargo metadata --locked at e759ec11 has rc 0. My cherry-pick of the dependency commits failed (rc 129, my script), so I compared against the real base e759ec11 instead of a synthesized one.
At both the base and the tip: fmt is clean, clippy is rc 0 with the same 12 warnings (CLIPPY_SAME), and the tests are rc 0 with the same 76 listed (TEST_LISTS_SAME). The base runs 40+15 lib/bin unit tests, the tip 55 lib. The integration tests are the same at both: bundle 5, fetch 5, push_state 2, push 3, remote_helper 1, session_fd 3.
AST, on my own map: equal, 7 moved, 214 stayed, no errors. My red arm, one string in the moved handle_logout edited, came out unequal.

Behaviour, the base's binary against the tip's, byte for byte in exit code, stdout and stderr. The probes:

auth status, auth logout, auth login with no password;
the helper's capabilities, a bad URL, and the word auth as the remote's name;
a session descriptor of 1, and one of 9 that is closed.

Every probe was the same at both, except clap's Usage line, which names argv[0], which was my copies' file names. My plant made the helper's error answer 0, and the bad-URL probe then differed (rc 1 vs 0). So the comparison does catch a difference.

Notes:

Old behaviour, not this PR: with GIT_REMOTE_PDS_SESSION_FD naming a closed descriptor, the helper aborts with rc 134 ("IO Safety violation: owned file descriptor already closed") at both the base and the tip, instead of exiting with an error. lichen always passes an open descriptor. I carry it as a note for the fork.
My dumpability probe didn't catch the process, so it shows nothing either way. take_from_env's code is unchanged and runs earlier than before.

Next is unit 3: lichen pins a55fb459 and dispatches, with the note's tests and plants.