* feat(#3218): the prompt layer asks the CLI for plan counts
Seven sites across four workflows counted plans with ls and wc -l instead of
asking the CLI. A shell glob is not scanPhasePlans, so every fix that landed on
the owner missed all seven: they counted superseded plans as live, reported zero
for the nested plans layout, and missed loosely-named files. The 1762 figure of
30 plans and 24 summaries came from here.
phase find is extended rather than a verb added - 3218 is an enhancement whose
own checklist says it adds no new command, and CONTRIBUTING makes a new verb a
feature needing approved-feature. It gains plan_count and summary_count for the
live set and plan_count_all for the physical one, additively; the existing arrays
are untouched.
Both sets are exposed because the sites need different ones. Amendment 1 names
two cases; three of these sites ask a third - did the planner write files to disk
- and take the physical set, because a superseded plan is still a file the
planner wrote.
The progress.md dead route is fixed and was worse than the issue said. It read
.plans and .summaries arrays that roadmap.analyze has never emitted, so the
fallback always fired, both counts were always zero, and Route 0's
resume-incomplete-phase check had never fired at all.
The ratchet baseline is empty. Its own stale-entry check makes that
self-enforcing.
Verified on the remote runner.
* test(#3218): acknowledge the workflow growth and update the stale guard
The emitted-attribution gate named its own remedy, so it was followed rather
than pre-guessed: four workflow files grew between 200 and 770 bytes because each
replaced a shell glob with a find-phase call plus its jq extraction. plan-phase
grew most - two sites, and it takes the physical count for its did-the-planner-
write-files question. progress also carries the Route 0 dead-path fix.
plan-phase-drift-guard asserted the literal old ls shape. Updated rather than
deleted: what it protects is that a filesystem fallback exists and is reachable,
and that is intact. It is not a regression - gsd_run is already load-bearing
throughout plan-phase.md long before step 9, so the 9a and 11a fallback never
existed to survive gsd_run being unavailable; it guards against the planner
subagent's return hanging.
Three ack sources collided with the new fragment, which the gate treats as a hard
error rather than last-wins. Only the three colliding keys were removed, not the
421 spent entries, and two fragments left entryless were deleted per the
convention that an empty fragment signals nothing.
Verified on the remote runner.
* docs(#3218): document the live and physical plan counts
docs/CLI-TOOLS.md gains a find-phase counts section covering plan_count and
summary_count for the live set against plan_count_all for the physical one, plus
the null-not-zero not-found behavior. The live-versus-physical distinction is
spelled out because a caller picking the wrong one gets a plausible number, which
is the trap Amendment 1 records.
Changeset leads with what a user sees: progress and execute-plan stop counting
superseded plans as outstanding, a nested plans layout stops reporting zero, and
Route 0 resume routing starts working after never having worked.
No how-to. Nothing is enabled and nothing is sequenced - the user runs the same
command and the number is simply correct. The one new distinction is field
semantics, which is what a reference entry is for.
* chore(#3218): backfill changeset PR number
pr:0 placeholder replaced with the real number now that #3327 exists.
---------
Co-authored-by: sim <sim@local>
8b4545f3c0