bc-dgx7.44: B7e-signoff — End a delivered run — the pull request

[mordam]

Sequences the four steps every worker brief in this repo already ends a delivered run
with — deliver the branch, leave a debrief, rename the session, print the closing
marker — behind one command instead of five hand-typed invocations. A session audit
(bc-1kwl.32) found the same four steps done by hand at the end of five sessions
(bc-1kwl.32, bc-ibt8g.1, bc-xl7n.120, bc-xl7n.118, bc-36xx.27), no two the same way, and
filed this bead as "the command that would have replaced it".

New: lib/signoff.js (the sequencing — chooseWorktree, queuedTitle/fallbackTitle,
markerLine, parseDeliverOutcome, ownedMarkerFor, and runSteps, a tiny sequencer
that stops without touching later steps the moment one throws) and bin/b7e-signoff
(the CLI that wires real subprocesses — node bin/deliver.js, ~/.claude/rename-session.sh
— and the real lib/memory.js#debrief into it).

Two things it fixes structurally rather than by telling the next session to be careful:

1. bin/deliver.js decides the pull request's head from the branch checked out where
   it runs, so delivering from the main checkout dies with "refusing to open a PR from
   main into main" even when the actual work is finished one worktree over — hit by
   three of the five sessions in the audit. b7e-signoff resolves the bead's own live
   worktree itself first (its branch is worktree-<slug>-<tag>, where <tag> is the
   bead id with its workspace prefix and punctuation stripped — the same rule
   lib/notinmain.js's tagOf/ownsBranch already use), and hands deliver.js
--dir <that worktree> explicitly. Zero live worktrees claiming that tag, or more
   than one, is a refusal naming the pattern it looked for — deliver.js's own refusal
   never gets a chance to fire, because it is never reached with the wrong directory.
2. beadcause-memory debrief "…" and a hand-typed rename have both, at different
   times, taken prose as a literal argv token inside a Bash tool call's double-quoted
   command string — and a backtick or $(...) in that prose is resolved by the shell
   before the command ever runs (bc-ibt8g.1's own stored debrief still carries
   (eval):1: no such file or directory from exactly this). --summary/--debrief/
   --tests/--risk/--left all take a file path (or - for stdin) here — never
   inline text — so nothing this reads is ever re-parsed by a shell.

What I chose against: shelling out to ~/.claude/rename-session.sh in "set" mode
directly with a hand-built name, the way every prior session did — instead this reads
the session's current name (--show) and only ever prepends QUEUED- to it
(lib/signoff.js#queuedTitle), which is what fixes the other historical bug (five
sessions each hand-truncated the new name to a different length, twice cutting it
mid-word). With no current name to read at all it falls back to the bead's own title in
full, never hand-shortened. I also chose to have this command override --marker to
REVIEWED whenever the delivery ended in a question card rather than a merge (auto-merge
off, --review, an in-app edit) — deliver.js's own header is explicit that nothing
after a merge can be owed by work that has not merged, and I'd rather this enforce that
than trust every future caller to remember it.

What I'm not fully sure of: the rename step shells out to ~/.claude/rename-session.sh
(overridable with --rename-script, which is how the suite tests it) rather than using
lib/retitle.js's in-process retitle() — that function needs a session object
resolved from liveSessions(cfg), and finding "which live session is me" needs a
process-tree walk rename-session.sh already does correctly in bash; I judged
duplicating that in JS wasn't worth it for this bead, but it does mean this command's
own rename step depends on a script outside the repo rather than on code this repo
tests directly (test/signoff.mjs proves the logic — queuedTitle/fallbackTitle — but
the real script's own process-tree walk is untested by this repo, same as it always
was).

b7e-signoff is deliberately not on DEFAULT_TOOL_LIST — it delivers, comments and
renames, which is write-shaped exactly like b7e-apply/b7e-gate, and dispatch (the
one agent that list governs) never delivers anything.

Tests: node test/signoff.mjs — 19/19 passing; node bin/b7e-gate — full suite 423/423 green in 8m47s (run b7e-signoff-dgx744-20260824T133420326Z-5v09)

Files changed — 7 files · +1219 −0 · against `main`

test/signoff.mjs   +627 −0
bin/b7e-signoff    +304 −0
lib/signoff.js     +197 −0
README.md          +81 −0
lib/toolbelt.js    +8 −0
package-lock.json  +1 −0
package.json       +1 −0

---

Opened by a beadcause worker session on bc-dgx7.44 — b7e-signoff — End a delivered run — the pull request, the debrief and the rename, in one call. It merges itself once the checks report; merging is what closes the bead. If this is still open, something stopped that, and the reason is on bc-dgx7.44 and in Adam's inbox.

bead: bc-dgx7.44

[mordam]

A beadcause worker opened this and does not merge its own work. It is on the merge queue as bc-m6l22.

[mordam]

The beadcause merge queue tried to merge this and could not: it has been tried 3 times and stopped at the same place each time. 1 check failing (test). A merge queue will not merge over a check the branch broke — if it is a flake, that is your call. Tried 3 times — that was the last. It is Adam's call now — see bc-m6l22.

[NeanderthalMan]

beadcause-resolver: stood down — the worktree for worktree-b7e-signoff-dgx744 is locked by another live resolver (pid 37220, still running per ps), reason "resolver pid 37220 #694". Left the tree and the branch untouched; the branch still conflicts with main.

[NeanderthalMan]

beadcause-resolver: stood down at step 2 — worktree-b7e-signoff-dgx744 is locked by another live resolver, reason "resolver pid 65543 #694" (pid 65543 is a running claude, started 14:00:21 ADT today). Did not touch the tree, did not unlock, branch left as it is.