pql finds the vault by walking up for a .git *directory*; a linked
worktree's .git is a *file*, so a bare `pql plan rebuild` from a fresh
worktree resolved to the MAIN checkout and left the worktree's own pql.db
empty — breaking ticket/decision reads inside worktrees (hit during the
Layer-5 batch; agents had to rebuild by hand). Pass --vault
"$(git rev-parse --show-toplevel)": the worktree root in a worktree, the
repo root on main (no-op there). Upstream pql bug filed as FR-4.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Refines 34ffcda86. The export --stage approach preserved the mutation but did more
than the bug needs: it eagerly stages changelog rows mid-checkout (surprising in
git status), and those staged rows can leak onto the wrong branch when a later
switch rebuilds by replaying the working-tree changelog.
Minimal fix: on a branch *creation* / no-op checkout (prev-HEAD == new-HEAD) the
working tree is unchanged, so pql.db already holds the committed state plus any
uncommitted mutation — just skip the whole block (do nothing). The mutation stays
in pql.db untouched and the next commit's pre-commit hook flushes it normally. Only
a real branch *switch* (prev != new) rebuilds. No git-index side effects, no leak.
Verified: same-HEAD invocation preserves an uncommitted status mutation and stages
nothing. (Credit: pql-clide's review of the upstream feature request.)
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
The post-checkout hook ran 'pql plan rebuild' (drop + replay changelog) on every
branch checkout, including 'git checkout -b'. That silently dropped any pql.db
ticket mutation not yet in the committed changelog — e.g. an in_progress status set
right before cutting a feature branch (hit this session: T-1039/T-1043 reverted to
backlog at branch creation).
Git has no pre-checkout/pre-worktree hook, so capture the changelog at post-checkout
instead: when prev-HEAD == new-HEAD (a branch *creation* — content unchanged, pql.db
already correct), run 'pql plan export --stage' (the same flush the pre-commit hook
does) instead of the destructive rebuild. Real branch *switches* (prev != new) still
rebuild to match the branch's changelog.
Verified: simulating a same-HEAD checkout with an uncommitted in_progress mutation now
preserves it (previously reverted).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Folds pql's planning logic into the version-controlled .config/hooks/* (pql's own
.pql/hooks installer is dead under core.hooksPath=.config/hooks):
- pre-commit: + `pql decisions validate` (decision-ID/format gate, supersedes the
never-built check-decision-ids TODO) and `pql plan export --stage` (flush ticket
mutations to the git-tracked changelog and stage them into the commit). Both
guarded by `command -v pql`; export is a clean no-op when nothing changed.
- post-merge: `pql plan import` + `pql decisions sync` (replay incoming changelog,
re-sync markdown decisions).
- post-checkout (branch only): `pql plan rebuild` + `decisions sync`.
- post-rewrite (rebase/amend): `pql plan rebuild`.
- install-hooks chmods the three new hooks.
Makefile decision targets repointed to pql: decisions-sync -> `pql decisions sync`,
decisions-active -> `pql decisions list --type confirmed`, new decisions-validate ->
`pql decisions validate`. Dropped the SQLite-query conveniences (coverage/orphan/
orphan-tickets); per-decision coverage is `pql decisions show <id> --with-tickets`.
db-backup/db-install and SR_DB_PATH are intentionally kept until Phase 6 so the
legacy SQLite store stays intact as the migration rollback path.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
jpmschweitzerandClaude Fable 5