P38 forge: F0 follow-ups (placeholder departure texts, git command comparison, F1 CI job) #297

Activity

1. tonibergholm-codento commented on Oct 22, 2026

   

   tonibergholm-codento

   on Oct 22, 2026

   CollaboratorAuthor

   More actions

   More follow-ups from the Codex review of f57cbc4 (PR #294), tracked here instead of another review round; the first two belong in the F1/F2 spec update before those slices start:

   • Spec, F2: define the default-branch reconciliation authority: on PATCH write bare HEAD first, then the row; on a mismatch at startup the row is the authority and the reconciler resets HEAD from it (a first push to a branch other than the default leaves empty: true, recorded).
   • Spec, F3: specify contents writes with new_branch: the protection policy applies to the target branch (a new branch has none), creating it from a protected base is allowed (recorded: 201 with new_branch where a write to the protected branch is 403), and the base ref is read once and used as the CAS old-value.
   • Spec, F2: make the free-space reserve a capacity reservation: admission reserves min(request length, --max-pack-size) (or the pack cap when chunked) against free space plus the owner quota before spawning receive-pack, releases it on completion, and refuses with 507/Insufficient Storage when it cannot reserve.
   • crates/forge/conformance/FINDINGS.md: the "22/22 per version" summary line predates scenarios W and X; update it to the current 24/24 per version (and the cross-version 2/24).
2. tonibergholm-codento commented on Oct 23, 2026

   

   tonibergholm-codento

   on Oct 23, 2026

   CollaboratorAuthor

   More actions

   F1 (#298) follow-ups, tracked here:

   • NAT64 (64:ff9b::/96) sources are keyed as plain IPv6 /64s in the password-failure budget; only matters behind a NAT64 gateway.
   • Failed-login delay is applied before the response is written, so a client that disconnects skips the wait; the per-source budget (30/min) is what bounds guessing.
   • The source budget map clears when it exceeds 10k live entries (documented shortcut: in limits.rs); move to an LRU if a distributed flood is ever seen.
   • check replays only A#0-5 and B#0-24 (slice-pending.json); R and X run for real once F2 adds repo creation, at which point the pending file shrinks to empty. F1 covers their recorded exchanges in crates/forge/tests/e2e_a.rs.
   • Windows: the crate is unix-only (empty lib, stub main); CI's Windows job compiles it as empty, not exercised locally.
   • Local toolchain differs from CI's pinned 1.96.1 (Homebrew 1.99); run clippy on the pinned toolchain before pushing.
3. tonibergholm-codento commented on Oct 23, 2026

   

   tonibergholm-codento

   on Oct 23, 2026

   CollaboratorAuthor

   More actions

   Triage by slice (so this stays a checklist, not a pile):

   Done

   • F1 skip departures and the CI job that checks the in-tree server: feat(forge): P38 F1 server skeleton, auth, bootstrap CLI #298 (sc-forge conformance, --pending).
   • F2 spec decisions (row is the default-branch authority, capacity reservation with 507, new_branch policy) and the FINDINGS 24/24 + 2/24 line: committed on the F2 branch 9e16e87, lands with the F2 PR.

   Doing in F2 (git steps run for real from then on)

   • Compare the recorded git command in check for git steps.
   • Replace placeholder message_contains words for the repo, branch and ref-namespace refusals with sc-forge's real texts. Already real after feat(forge): P38 F1 server skeleton, auth, bootstrap CLI #298: password, auth, hasorgoruservisible, not found: no such API route.

   Later slices

   • F4: mergeable/html_url strictness, combined-status answers, the remaining placeholders (merge, state, rename).
   • Not scheduled, low value: compact fixture JSON (3.5 MB). Say if you want it as a tiny separate PR.

   F1 follow-ups (NAT64, delay on disconnect, map reset, Windows, toolchain) stay as listed above; none block F2.
4. tonibergholm-codento

   added a commit that references this issue on Oct 23, 2026

   feat(forge): P38 F2a repositories API, lifecycle, collaborators, bran…

   ...

   3571fb8
5. tonibergholm-codento commented on Oct 23, 2026

   

   tonibergholm-codento

   on Oct 23, 2026

   CollaboratorAuthor

   More actions

   F2a (#301) follow-ups:

   • F2b: compare the recorded git command in check for git steps (C and N run clones for real from then on); K, D, M can join the gate once info/refs, commits and contents exist; K#54-56 counts and the M rename departures (422) and 404-text overrides (not found placeholders in departures.json) get their reviewed departures then.
   • Open question: does Gitea also require a repository scope on /user/repos? Recorded labels only show write:user/read:user.
   • Rollout note: migration 2 has no down path; rolling back means restoring a pre-F2a forge.db, and the next startup moves every repo directory to orphans/ (recoverable).
   • Not tested: Linux execution of the waitid/killpg branch (compiles in CI, exercised on macOS only); power-loss durability beyond fsync ordering and injected crash points.
   • chore(deps): bump ctutils to 0.4.5 (RUSTSEC-2026-0385) #302 closed as superseded (the ctutils 0.4.5 bump landed in feat(forge): P38 F2a repositories API, lifecycle, collaborators, branches #301).
6. tonibergholm-codento commented on Oct 23, 2026

   

   tonibergholm-codento

   on Oct 23, 2026

   CollaboratorAuthor

   More actions

   F2b (#304) follow-ups to track: per-user transport quotas (F5; a per-user cap would throttle Engine's single service account); default-branch PATCH vs a push that deletes the new default can leave HEAD dangling (needs a receive-pack/PATCH lease, F4); --git-memory-limit (Linux RLIMIT_DATA) is untested in CI; no test that a killed push leaves no ref lock or quarantine directory (SIGTERM-first is implemented, unverified); a test that a transport killed mid-stream (delete during a clone) gives the client an error, not a clean end; deferred git exchanges pin the step name but not the command line; the principal of a deferred HTTP exchange is not pinned.
7. tonibergholm-codento

   added a commit that references this issue on Oct 25, 2026

   test(forge): conformance departures carry sc-forge's real refusal tex…

   ...

   4cda9b3
8. tonibergholm-codento

   self-assigned this

   on Oct 25, 2026
9. tonibergholm-codento

   closed this as completedin #317on Oct 25, 2026