agent-ops
OPS-177
· child of OPS-173 gitpush.sh split by stack: retire the 2408-line monolith, per-stack scripts, bump-after-gate, deploy health poll Backlog
gitpush.sh becomes a retirement stub naming the chooser doc
Done normal
bvbin-venus-cc
Elazar's spec: gitpush.sh replies that it is retired and points at a document that says which script to run. Constraints: exit NON-ZERO so a scripted caller fails loudly rather than continuing as if it pushed; print the chooser table inline as well as the doc path, because an agent that must open a second file to learn its next command has paid the token cost the split exists to remove; and keep the stub to a screen. Write the chooser doc as part of this item, not as a follow-up - a stub pointing at a path that does not exist yet is worse than the monolith.
Activity
-
parent=#3117
-
Stale reference to fix while writing the stub, found reading gitpush.sh:1140 for the OPS-183 correction: the error text names 'audit-mba-l-ca', and the roster has audit-mba-l-cc. A -ca/-cc slip in user-facing help text sends a reader to an agent that does not exist. Trivial; noting it so it dies with the monolith rather than being copied into the successors. (Elazar's own recipient list had the same slip in the other direction - coder02-terra-ca for coder02-terra-cc - so the harness suffix is a general trap, not one typo.)
-
Stale reference to fix while writing the stub, found reading gitpush.sh:1140 for the OPS-183 correction: the error text names 'audit-mba-l-ca', and the roster has audit-mba-l-cc. A -ca/-cc slip in user-facing help text sends a reader to an agent that does not exist. Trivial; noting it so it dies with the monolith rather than being copied into the successors. (Elazar's own recipient list had the same slip in the other direction - coder02-terra-ca for coder02-terra-cc - so the harness suffix is a general trap, not one typo.)
-
DELIVERED at bfde099: gitpush.sh v5.5 stub + gitpush-chooser.md + gitpush-retire.test.sh (13 checks, 0 failed) + index.md. Every constraint in the spec is asserted, not just implemented: [2] non-zero AND no commit created, because "refused" and "refused after committing" print identically; [3] all four lane commands inline, 17 lines; [4] placed BEFORE argument parsing, which is what --help needs - it exits 0 on its own, so a stub after the arg loop lets it straight through; [5] CONTROL, flag unset still pushes, without which a simply-broken gitpush.sh satisfies everything above. Chooser doc written in this item, not deferred, per the spec's own last sentence. Forced-red matrix, 4 mutants, each reddens ONLY its target: exitzero -> [2a][4a][4b]; docponly -> [1c][3a][3b]; afterparse -> [4b]; alwayson -> [5a]. One iteration was needed: [6] originally resolved gitpush-chooser.md from the SUT's own path, so it reddened under all four mutants including the two that never touch the doc. A false-red destroys attribution exactly as thoroughly as a false-green. It now resolves from the INSTALLED script. Six monolith suites re-run green after the edit (bumporder 12, onlystage 7, scoped-anyrepo 29 +3 controls unavailable, prestage-abort 17, gate 20, core 15; 0 failed). WHY THIS DOES NOT CLOSE AS A RETIREMENT. The flag is not the default and flipping it is not a code change in this checkout: 1. ~/.claude/git-workflow.md on venus, whey and lezama still names `gitpush.sh "what changed"` as THE command. Per-host, outside sh.git, nw lane. Flipping first breaks every agent's push mid-session, including the agents who would push the fix. 2. Five executing callers: ship.sh:114, deploy.sh:191, gitpush-locked.sh:66, channel-rule-install.sh:59, fixcodexgitpush.sh:7. deploy.sh is the one that is NOT mechanical - it runs `gitpush.sh --patch`, bumping the file named by .bump, where gitpushhtml.sh bumps through sitebump.sh. Which is correct depends on the consuming project and is not knowable from here. 3. Six suites take the monolith as their SUBJECT, ~100 checks. They retire with it or their cases move; either way deliberate work. Asking: close this row on the provable stub and open a separate cutover row for 1-3, or hold it open? The stub is rehearsable on any host today. Also closes the 2026-08-31 note: the audit-mba-l-ca / coder02-terra-ca suffix slips are gone from the whole checkout, confirmed by search, and were not copied into any successor.
-
Stub delivered and proved at bfde099; closed on the stub per pmmaster-evolutiva-cc 2026-10-05. The cutover (flip the default, git-workflow.md on three hosts, five callers, six suites) is OPS-278, a separate row under the same parent - distinct cross-host scope, not part of 'did the stub work'. gitpush.sh still pushes normally with the flag unset, which case [5] asserts; nothing changed for any existing caller today.
task
6w ago by wi-cli-venus
1w ago
2026-10-05 22:01