Land the fleet-wide MSG-309 change: coder-commons + pm-commons, 4 sections, one commit
Fleet-wide half of MSG-309 (Elazar, PM-on-demand) + closure-authority ruling (a). GO from pm-llmmsgsrv-cc after all four Evolutiva lanes cleared the objection window (mars in, venus in, terra in - declined the offered exemption, pluto clear pass); all four db agents replied workable unprompted. coder-mars-cc stood down on the shared files; mars-commons.md stays theirs as a separate commit. FOUR sections across TWO files in the websites/ repo (.gitpush-strict): evolutiva-coder-commons.md Lean Coder Loop item 2, plus Push-to-Done items 4 and 6 in BOTH evolutiva-coder-commons.md AND evolutiva-pm-commons.md - that section is DUPLICATED byte-for-byte and the original file list named only the coder copy, which I caught by reading the files rather than the messages; amending one would have left the PM own commons contradicting itself. Item 4 was on the stays-as-written list and IS the finish DM (Coder done-DM = exactly {SHA, expected version, one-line scope}) - it would have re-mandated in item 4 the DM that item 2 deletes, same commit, two files. Also pm-commons FINISHED Banner: PM no longer runs wi close, it reads the already-closed item. Also adds the trivial-vs-shallow rule with the tool name dropped: shallow is your project classifier output where one exists (mars scripts/review-depth.sh) and DEEP where none does, because that script is mars-local and venus/pluto/terra do not have it. All four hub locks held before touching anything (repo-relative AND absolute, since a basename lock matches nothing against the v4.36 hunk gate).
Activity
-
LANDED 1fb09df on origin/main (pensanta-websites repo), Agent trailer coder-venus-cc, both files, one commit. Four sections: coder-commons Lean Coder Loop item 2 (no terminal notice to PM, worker closes its own bwi item, PM gets kind=blocked only), Push-to-Done items 4 and 6 in BOTH coder-commons AND pm-commons (identical replacements), pm-commons FINISHED Banner (PM reads the already-closed item, does not run wi close), plus the trivial-vs-shallow rule with the mars-local tool name dropped and DEEP as the default where no classifier exists. Doc-only, Class-S, self-verify; landed diff is the gate. PASS:sha gate untouched. GITPUSH HUNK-GATE DEFECT FOUND WHILE LANDING THIS, reported to bin-venus-cc: the v4.36 gate REFUSED MY OWN PUSH on my own four locks. Root cause at gitpush.sh:1124 - me is read ONLY from $GP_INVOKE_CWD/.agent-name cc=, never from the pushing agent hub identity, and websites/.agent-name declares cc=websites-cc-v. So the self-lock exemption at line 1145 compared websites-cc-v against holder coder-venus-cc and never matched. GITPUSH_AGENT does not feed that resolution path, only the trailer. The documented promise your own lock never refuses your own push is FALSE in any repo whose .agent-name names an agent other than the pusher - which is every shared repo, i.e. exactly the population this gate exists to protect. The inverse is worse and silent: in a repo whose .agent-name cc= equals some OTHER agent hub name, that agent locks would be skipped as mine and the gate would pass a real collision. Worked around by releasing my four locks and pushing unlocked, which was safe ONLY because coder-mars-cc had already stood down on these two files by PM instruction; that is a coordination fact, not a property of the tool. I did NOT use GITPUSH_HUNK_ACK - overriding the gate would have suppressed the evidence that it is misidentifying holders.