| Commit message (Collapse) | Author | Age | Files | Lines |
|---|
| |
|
|
|
|
|
|
|
|
| |
The C-; L keymap and the transient menu had drifted into two different letter schemes for the same commands, the fetch sub-map carried both issue-fetching and view-movement, and pearl-switch-account was bound nowhere.
Give views their own group under v: l/L run local/Linear, c/e/k create/edit/delete, u/U publish local/current, d save a Linear view locally. Fetch (f) is now issue sources only. A new workspace group (w) binds switch-account plus the setup commands, which previously lived only in the transient. v D is reserved, unbound, for the not-yet-built set-default-view.
The transient already had an issue-only fetch group and a dedicated views group, so its changes are additive: switch-account and copy-issue-url were both missing from the menu and are added, and save/save-all/edit-description take s/S/d to match the keymap's hot keys. Matching those three cascades a few view and edit letters because the transient is a flat keyspace where every command needs a unique key.
Suite 789 to 792, compile and lint clean.
|
| |
|
|
|
|
| |
I made the builder's State prompt a completing-read-multiple, the way the Labels prompt already is, so you can pick several workflow states for one filter. The selection maps the way the spec describes: pick one and :state stays a scalar string, pick several and it becomes a list (state.name.in), pick none and there's no state constraint. The "[ Any. ]" sentinel is filtered out of the result like it is for labels.
Phase 2 of docs/multi-state-filter-spec.org. A stubbed-builder test covers the one/many/none mapping. Suite, compile, and lint green.
|
| |
|
|
|
|
| |
I made :state polymorphic, mirroring :state-type. A string still compiles to state.name.eq. A list now compiles to state.name.in, so (:state ("Todo" "In Review")) matches issues in either state. The reverse-compile inverts state.name.in back to a scalar (one name) or a list (several), so copy-down no longer refuses a Linear view filtering on a set of named states. The validator accepts a string or a list of non-empty names, the same shape it already enforces for :labels, and a list-valued :state keeps its precedence over :state-type and :open.
Phase 1 of docs/multi-state-filter-spec.org. Suite, compile, and lint green.
|
| |
|
|
|
|
|
|
| |
I added a test that reads pearl.el and asserts the user-facing "saved query" phrase appears nowhere, so the rename can't silently regress. Internal helpers keep the hyphenated saved-query symbol fragment, which a user never sees. The banned thing is the two-word user-facing phrase in any prompt, message, transient label, or docstring.
This is the cross-cutting net over Phase 8. The per-phase tests already cover the account guard, cross-store dedup and dispatch, reverse-compile round-trips and refusals, and the metadata-preserving writer. The live create, edit, run, publish, save-locally, refuse, and delete end-to-end needs a real Linear workspace and is filed as a manual-verify checklist.
Suite at 777 green, compile and lint clean.
|
| |
|
|
|
|
|
|
|
|
| |
I implemented copy-down, the one direction that needed real engine work. A live probe showed CustomView.filterData is readable but stored as Linear's and/or view-filter tree, not the flat IssueFilter Pearl emits, so pearl--reverse-compile-issue-filter is a normalize-then-match: it flattens the top-level and, unwraps single-branch and/or, reads a one-element in as a scalar, then matches each conjunct to one authoring key. Every helper returns (ok . plist) or (refuse . reason).
The contract follows the spec's dimension table: distinct :state and :state-type, canonical numeric :priority, :open t mapped from the exact open-state nin while generic nin refuses, and a multi-value in on any singular key refuses rather than silently narrow the view. Anything outside Pearl's AND-only model refuses with a structured reason that names the construct.
pearl-save-linear-view-locally lists favorited Linear views, fetches the chosen view's filterData through a small async helper, reverse-compiles it, and saves a forked local view through pearl--save-local-view: no tracking link, :copied-from-view-id provenance only, active account stamped. On a refusal it points the user at pearl-run-linear-view; on success it states the fork's independence. Bound at f d (down) and the transient Views column.
Phase 6 of the views spec. 30 tests cover per-dimension round-trips (build->encode->read bridge), normalize-from-real-JSON trees, the full refusal set, and the command-side. Suite, compile, and lint are green.
|
| |
|
|
|
|
|
|
| |
I extended the account guard from edit and run to the two remote-mutating commands. pearl-publish-local-view now refuses a local view tagged to another account before any customViewCreate or customViewUpdate, and pearl-delete-local-view refuses before the confirm prompt or any customViewDelete. Publishing or deleting under the wrong account would resolve ids against the wrong workspace or target the wrong remote, so the guard fires at the command boundary, before Linear is touched. publish-current-view inherits the guard by delegating to publish-local-view.
This completes the round-6 finding that one guard should cover every read and mutate of a local-view entry, not only edit and copy.
Phase 5 of the views spec. Tests confirm both commands error on a cross-account entry and make no Linear call. Suite, compile, and lint are green.
|
| |
|
|
|
|
|
|
|
|
|
|
| |
dispatch
I reworked pearl--pick-source-candidates for the three Phase 4 behaviors. Labels now read [local] for a private local view, [local -> Linear:SCOPE] for a published one, [linear] for a Linear view favorite, and [KIND] for a non-view favorite, so the label itself shows where a view lives.
Dispatch follows the source-of-truth rule (decision 16): every local view, tracked or not, dispatches as a :type filter and runs its local authoring filter, never the Linear mirror. A tracked entry carries its :linear-view-url so the rendered buffer can still open the mirror. Before this, a published local view ran server-side, which contradicted the editable-local model.
Cross-store dedup: a Linear view favorite whose customView id matches a tracked local view's :linear-view-id is dropped, so the editable local entry is the one shown. A local and a Linear view that merely share a name stay distinct.
Phase 4 of the views spec. Tests cover the labels, the local-filter dispatch, and the dedup and same-name cases. Suite, compile, and lint are green.
|
| |
|
|
|
|
|
|
| |
I put the view operations under the verb a user would reach for. Create, edit, and delete local views are c v, e v, and k v. The fetch group gains run-local-view at f l and keeps run-linear-view at f v, with publish at f u and save-locally reserved at f d. Since f p already runs issues by project, publish and save-locally take u and d, which read as up and down and mirror the copy-up and copy-down lifecycle.
The transient now has a Views column grouping create, edit, run-local, run-linear, publish, and delete, separate from the source-fetching commands. I updated the spec's keymap to the implemented letters so the two stay in sync.
Phase 3 of the views spec. Keymap tests assert the new bindings. Suite, compile, and lint are green.
|
| |
|
|
|
|
|
|
|
|
| |
I added pearl-create-local-view and pearl-edit-local-view, plus the writer and guards the spec calls for. pearl--save-local-view replaces the old rebuild-from-scratch writer. It copies an existing entry and updates only the filter, sort, and order, so a published view keeps its :linear-view-* tracking link and a copied-down view keeps its provenance. A new entry gets the active account stamped under pearl-accounts.
pearl--require-local-view-account is the one guard every read or mutate command shares, and pearl-run-local-view now calls it instead of an inline check. pearl--resolve-local-view-name carries the Replace/Rename/Cancel collision policy for create, edit-rename, and copy-down. The ad-hoc builder's save offer routes through the preserving writer and reads "save this as a local view".
Edit re-runs the builder when you choose to rebuild and otherwise keeps the stored filter. Full per-dimension pre-seeding of the builder is deferred, since it needs the id-to-name reverse resolution that copy-down's reverse-compile introduces. I filed it as a follow-up.
Phase 2 of the views spec. Tests cover metadata preservation, the account guard, the collision policy, and the metadata-preserving rename. Suite, compile, and lint are green.
|
| |
|
|
|
|
|
|
| |
I renamed Pearl's "saved query" surface to the local/Linear view vocabulary, with no obsolete aliases, since Pearl has no users to protect. The defcustom pearl-saved-queries is now pearl-local-views. The commands pearl-run-saved-query, pearl-delete-saved-query, pearl-sync-saved-query-to-linear, pearl-run-view, and pearl-publish-current-source become pearl-run-local-view, pearl-delete-local-view, pearl-publish-local-view, pearl-run-linear-view, and pearl-publish-current-view.
User-facing prompts, messages, and docstrings drop "saved query" for "local view", and the publish command reads "publish" instead of "sync". Internal GraphQL and helper names keep their "query" terms, which a user never sees. A naming-regression test asserts the new symbols exist, the old ones are gone rather than aliased, and no public command exposes "query".
Phase 1 of docs/local-and-linear-views-spec.org. No behavior change. Suite, compile, and lint are green.
|
| |
|
|
| |
The setup check gated on pearl-api-key, which is normally unset once pearl-accounts is configured, so under multi-account it wrongly reported "API key is not set" and skipped the connection test even when the active account resolved a key fine. Under accounts mode it now resolves the active account's context, names it, and runs the test. An unusable setup (no default, missing key) reports the reason instead of signaling. Legacy single-account behavior is unchanged. The resolved key is never printed.
|
| |
|
|
|
|
|
|
|
|
|
|
| |
I added a named-account layer for working more than one Linear workspace from one Emacs.
Before this, everything that identified a workspace was a single global: pearl-api-key, pearl-graphql-url, pearl-org-file-path, and the lookup caches. Nothing stopped a work command from running under personal credentials or a work fetch from landing in the personal file, and switching accounts meant re-customizing the key, team, and file by hand and clearing the cache.
pearl-accounts maps a name to a per-workspace plist (credential source, org file, default team, optional endpoint), and pearl-switch-account makes one active. Account state flows through an explicit context rather than mutating globals. Every request snapshots its account at dispatch through pearl--graphql-request-async and re-establishes it around the callbacks. A switch mid-fetch can't bleed into a request already in flight: the result finishes into the account it was dispatched under. I centralized this in the one request primitive, so the leak surface is a single function instead of every call site.
Rendered files carry a #+LINEAR-ACCOUNT marker, and a buffer guard refuses a command run from one account's file while another is active, naming both, so a work edit can't push under personal credentials. An unmarked legacy file lets reads and refreshes through and acquires its marker on the first refresh. Mutations wait until then. Credentials resolve through auth-source, an env var, or an inline literal, and a resolved key is never persisted through Customize or logged. The active account shows in the mode line. A saved query can carry an :account so it refuses to run under the wrong workspace before any lookup.
With pearl-accounts unset, everything behaves exactly as before, off the legacy globals.
|
| |
|
|
|
|
|
|
|
|
|
|
| |
The previous commit (eb56295) shipped the buggy pre-fix version of pearl--merge-issues-into-buffer along with a too-weak test. The intended fix and the strengthened test were sitting uncommitted in the working tree when I ran git commit, because I forgot to re-stage pearl.el and tests/test-pearl-merge.el after addressing the critical correctness finding from pre-commit review. The test that landed asserts each touched marker leads to some LINEAR-ID heading, which passes against the buggy code (the markers point at the wrong heading, but it's still a valid heading).
This commit is the fix and the real test, which together honor what eb56295's message described.
Code change: pearl--merge-issues-into-buffer captures a type-nil copy of the alist marker BEFORE calling pearl--replace-issue-subtree-at-point, instead of pushing the original type-t marker after the replace. The alist marker is insertion-type t (it needs to advance past replaces of EARLIER subtrees so it stays anchored to its own heading), so reusing it after the current subtree's delete-then-insert captures the post-advance position, which is the NEXT subtree's heading. A type-nil copy made before the replace stays anchored at the deletion start and lands on the new heading after the re-insert.
Test change: the returns-touched-markers test now resolves each marker to its heading's LINEAR-ID and asserts the set matches {a, b, c} exactly (for 2 updated + 1 added), instead of just checking each marker reaches some heading. With the buggy code the IDs would resolve to {b, c, c} (a's marker advanced to b's heading, b's marker advanced to c's added heading). The strengthened assertion fails on the bug.
All 674 ert tests pass. make compile and make lint are clean.
|
| |
|
|
|
|
|
|
|
|
|
|
| |
A merge refresh (the "merge keeping edits" choice on a dirty buffer, or pearl-refresh-current-view on a clean one) used to call pearl--restore-page-visibility at the tail, re-folding the whole buffer to its #+STARTUP overview. That collapsed every subtree the user had expanded, including the comment they were just editing, and disturbed point. The edit-then-merge flow felt jarring.
The fix: pearl--merge-issues-into-buffer now returns a marker for every subtree it re-rendered or appended, alongside the existing counts (:touched-markers (M1 M2 ...)). The merge call sites fold just those subtrees via a new pearl--fold-touched-subtrees helper. Kept (locally-edited) and untouched subtrees stay exactly as the user had them, and point is preserved because the localized fold doesn't touch the rest of the page.
The full-rebuild paths (Branch A no-buffer, Branch B clean-buffer replace, Branch C-discard) still call pearl--restore-page-visibility. They re-render the whole buffer so the global re-fold is the right behavior.
Pre-commit review caught a critical correctness bug in the first draft: the alist marker from pearl--issue-subtree-markers is insertion-type t (it needs to advance past replaces of EARLIER subtrees so it stays anchored to its own heading), and reusing it after pearl--replace-issue-subtree-at-point captured the post-advance position, which is the NEXT subtree's heading, so the fold collapsed the wrong subtree. The fix copies the marker as type-nil BEFORE the replace, so it stays anchored at the deletion start and lands on the new heading after the re-insert. I strengthened the marker test to verify the IDs at the touched markers match the expected set ({a, b, c} for 2 updated + 1 added), not just that they're some valid headings. The original test would have passed against the bug.
Five new tests in test-pearl-merge.el cover: merge returns markers with the correct IDs at their headings, skipped (locally-edited) issues don't add to touched-markers, fold-subtree-at-marker hides body content while leaving the heading visible, fold-touched-subtrees is a no-op when pearl-fold-after-update is nil, and fold-subtree-at-marker tolerates a nil marker silently. All 674 ert tests pass. make compile and make lint are clean.
|
| |
|
|
|
|
|
|
|
|
| |
Pearl was rendering the Comments subtree heading as `*** π¬ Comments [N/M]`. Craig flagged that it looks strange, and the renderer should work cleanly before it gets glyphed. I dropped the emoji prefix so the heading reads `*** Comments [N/M]`. Partly reverts 41a3396. The overlay-glyph idea (a content-aware display overlay on the leading stars) stays tracked as a separate task.
I touched pearl--format-comments (the new-subtree literal) and pearl--append-comment-to-issue (both the new-subtree insert and a couple of stale docstring references). The append-locator regex stays tolerant of the legacy `π¬ Comments` heading and of the older pre-2026-05 `Comments π¬ N/N` trailing-glyph layout. Buffers rendered in any of the three states still locate their Comments heading on append.
I added one regression test (test-pearl-append-comment-locates-legacy-emoji-prefixed-heading) that seeds a `*** π¬ Comments 1/1` heading and asserts append still finds it, bumps the count to 2/2, and doesn't add a second Comments heading.
I updated assertions in test-pearl-comments.el, test-pearl-list-comments.el, and test-integration-acceptance.el to expect the no-emoji form. I also updated docstrings on pearl-fetch-comments-in-list, pearl-list-comments-count-cap, pearl--format-comments, pearl--bump-comments-count-marker, and pearl--comment-count-marker. The README's sample buffer block now matches what the renderer emits. All 669 ert tests pass. make compile and make lint are clean.
|
| |
|
|
|
|
|
|
|
|
| |
pearl-sync-saved-query-to-linear used to rebuild the entire scope-candidates list a second time (full sort + 2N row allocations) just to rassoc the chosen scope plist back to its display string for the success message. The picker had thrown away the display string the user picked, so the caller had no choice but to re-derive it.
I changed pearl--sync-saved-query-pick-scope to return (DISPLAY . PLIST) instead of just PLIST. The caller now reads scope-label from (car scope-pair) and scope from (cdr scope-pair). The rassoc is gone, and so is the second pearl--sync-scope-candidates call.
To keep the re-sync path label-consistent with the picker (so a "Synced X (Team: Engineering, visible to the team)" message matches what the picker would have shown for the equivalent first-time sync), I extracted pearl--sync-scope-label as the single source of truth for the label format. The picker, the re-sync branch, and any future caller that needs to describe a scope all go through the helper. Personal scope always renders "[ Personal, only I see it ]" regardless of the shared flag (a Personal + shared combination is meaningless on Linear). Team scopes render "[ Team: NAME, visible to the team ]" or "[ Team: NAME, only I see it ]" per the flag.
I added five tests covering the new behavior: scope-label for personal (ignores shared) and team (both shared and private), pick-scope returns the team pair, pick-scope returns the personal pair, pick-scope returns nil on cancel. All 668 ert tests pass. make compile and make lint are clean.
|
| |
|
|
|
|
|
|
|
|
| |
A saved query with no filter constraints compiled to nil via pearl--build-issue-filter. The sync command then passed (cons "filterData" nil) to customViewCreate, which JSON-encoded as filterData=null. Linear's CustomViewCreateInput.filterData expects a JSON object, so the API rejected with an opaque error message after the user had already clicked through the scope picker.
I added an early refusal in pearl-sync-saved-query-to-linear: when filter-data is nil, signal a user-error naming what's wrong before the scope prompt fires. The message tells the user to add a constraint and re-sync rather than leaving them to puzzle out an opaque API rejection.
I restructured the let* nesting to put the empty-filter check between the bindings (where filter-data is computed) and the scope prompt (where the user is asked to pick a destination). The scope binding moved out of the outer let* into an inner let so it doesn't evaluate until after the filter-data guard runs.
Two tests cover the guard. The first-time-sync test (no :linear-view-id on the entry) asserts neither the scope prompt nor customViewCreate fires. The re-sync test (entry has :linear-view-id) asserts customViewUpdate doesn't fire either. The re-sync test guards against a future refactor accidentally hoisting the check into the create-only branch. All 663 ert tests pass. make compile and make lint are clean.
|
| |
|
|
|
|
|
|
| |
The prior implementation rebuilt the spec from a fixed key set (:filter, :sort, :order, plus the four :linear-view-* sync keys and the optional :linear-view-url) and silently dropped anything else. No keys outside that set live on a saved-query entry today, but the persistence layer was coupled to the current schema. A future :description, a user-added annotation, or any new pearl field would vanish on every re-sync.
I switched to copy-then-plist-put: shallow-copy the existing spec, plist-put the four sync keys on top, and conditionally plist-put :linear-view-url when this call provided one. Arbitrary keys survive untouched. The URL-preservation behavior on a re-sync that omits view-url is unchanged: the prior :linear-view-url is already in the copied spec, so the when-view-url guard skips the put and leaves it in place.
I added one regression test that seeds an entry with :description and :user-tag alongside the standard keys and asserts all of them survive a mark-synced call. All 661 ert tests pass. make compile and make lint are clean.
|
| |
|
|
|
|
|
|
|
|
| |
This is Phase 5 of docs/saved-query-sync-spec.org (Β§ New: pearl-publish-current-source). When I'm reading a local saved query rendered in pearl and decide it should be a Linear view, I now have a one-chord publish instead of the M-x round-trip through pearl-sync-saved-query-to-linear with the name lookup.
The new command reads the active buffer's #+LINEAR-SOURCE header and dispatches to pearl-sync-saved-query-to-linear with the name pre-filled when the source is a :type filter that matches an entry in pearl-saved-queries. Other shapes refuse with a clear message: no source header at all, the source is already a :type view (no publish needed), the type isn't filter, the source has no usable name (transient ad-hoc filter, or whitespace-only name), or the recorded name doesn't match any local saved query.
I bound the command under C-; L f P in pearl-fetch-map. The transient gets ("U" "upload current source") in the Fetch group because P is taken in the transient's flat namespace, and the verb-prefix shape revisit task tracks reshaping that.
7 new tests cover each refusal branch and the dispatch path, including a regression for the whitespace-only :name case the pre-commit review surfaced. The prior check used string-empty-p, which let " " fall through to the no-matching-saved-query branch with a visibly-blank name in the error message. string-blank-p routes it to the cleaner no-name branch. The keymap and transient menu tests gained coverage for the new binding. 660 ert tests total green. make compile and make lint clean.
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
/ [view]
This is Phase 4 of docs/saved-query-sync-spec.org (Β§ Extended pearl-pick-source candidate label). The picker now reads where each saved query lives at a glance:
- `[saved] Name`: local-only saved query, dispatches as :type filter (unchanged)
- `[saved β Engineering] Name`: synced to a team-scoped Linear view
- `[saved β Personal] Name`: synced as a personal Linear view
- `[saved β ?] Name`: synced, but the team id no longer resolves (deleted/renamed on Linear) or the teams cache wasn't loaded
- `[view] Name`: Linear-authored view favorite (unchanged)
A synced entry's source plist gets :type view, so the existing view-dispatch branch in pearl-pick-source routes through pearl--query-view-async against the linked Linear view. Refresh and pagination honor whatever the view's filter is on Linear now, not the local authoring filter pearl built it from.
I added pearl--saved-query-scope-label to resolve the scope string from (spec, teams) and extended pearl--pick-source-candidates with an optional teams argument. pearl-pick-source fetches pearl--all-teams lazily, only when at least one saved query carries :linear-view-id, and feeds the result in. An all-local configuration pays no extra network cost.
While in the picker code I also closed a latent gap the review surfaced: pearl-open-current-view-in-linear was about to break on synced-saved-query sources because the source plist had no :url. I extended pearl--customview-create-async and -update-async to fetch `url` in the response, pearl--save-query-mark-synced to take an optional VIEW-URL arg and store it as :linear-view-url, and the picker source plist to surface :url when the entry has one. Entries synced before this commit have no stored URL. A re-sync repopulates it. Mark-synced preserves a prior URL when VIEW-URL is omitted, so a re-sync whose API response somehow lacks the URL doesn't erase the working one.
15 new tests across test-pearl-favorites.el (the picker labels for team-scope, Personal, unknown-team, no-teams-arg, URL surfaced + URL absent, mixed local/synced) and test-pearl-saved-query-sync.el (mark-synced stores URL when given, preserves prior URL on re-sync omission). The existing 4 pick-source tests stay green. They pass either no teams arg or nil, so the scope-resolution branch is dead in their cases. 653 ert tests total green. make compile and make lint clean.
Lower-severity findings from the review filed as follow-ups: the `?` label conflates "unknown team" and "no teams data" (could be split). pearl--all-teams cache has no TTL so stale team names persist across renames. The lazy teams fetch reuses the "Fetching favorites..." progress message during its own roundtrip. Synced entries silently drop the local :sort/:order because Linear's view doesn't carry them in v1.
|
| |
|
|
|
|
|
|
|
|
| |
Phase 3 of docs/saved-query-sync-spec.org. pearl-delete-saved-query on a synced entry (one carrying :linear-view-id) now asks a second question after the local-delete confirmation: also delete the linked Linear view? Yes calls customViewDelete, no unlinks only and leaves the Linear view in place. Local-only entries take the unchanged single-prompt path.
If customViewDelete fails on the API side, a fallback yes-or-no prompt asks whether to drop the local entry anyway. Accepting orphans the Linear view, and the success message names the view id explicitly so the user can clean it up by hand. The same view-id-in-message pattern fires on the timeout branch, since a timeout doesn't tell us whether the delete actually completed server-side.
I added pearl--customview-delete-async (the mutation, parallel to the create and update helpers from Phase 2) and pearl--delete-saved-query-local (factors the cl-remove + customize-save-variable that both the local-only and the synced-unlink branches need). The new pearl--delete-saved-query-do-linear-delete carries the API path so the top-level command stays readable.
I added 7 tests in test-pearl-saved-query-sync.el covering the customViewDelete success and failure parses, the synced yes/yes path (API call fires + local removed), the synced yes/no path (unlink only, no API call), the API-failure-then-delete-anyway path (orphan message names the view id), the API-failure-then-keep path (asserts the API call fired so a refactor can't silently route through the unlink branch), and the local-only entry unchanged path. The existing pearl-delete-saved-query tests in test-pearl-adhoc.el stay green. Their fixtures don't carry :linear-view-id, so they take the local-only branch. 638 tests total green. make compile and make lint clean.
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
pearl-sync-saved-query-to-linear promotes a local pearl-saved-queries entry to a Linear Custom View via customViewCreate, or updates an already-synced entry in place via customViewUpdate. This is Phase 2 of docs/saved-query-sync-spec.org. Re-syncing keeps the stored scope. First-time sync prompts for the destination (team + visibility), with each candidate spelling out the complete end-state. Same-name collisions in the chosen scope prompt Replace / Rename / Cancel.
A successful sync extends the entry with :linear-view-id, :linear-view-team-id, :linear-view-shared, :linear-view-synced-at. If the customize-save-variable persist fails after a successful API call, the message names the orphan view's id explicitly so the user can re-sync with Replace to reconcile.
I bound the command under C-; L f S in pearl-fetch-map and added ("S" "sync saved query") to the transient's Fetch group. filterData is the existing pearl--build-issue-filter output verbatim. The 2026-05-28 live probe confirmed the shape passes through customView.filterData with no translation needed.
Four related fixes ride along:
- Extended pearl-get-teams-async to fetch `key' alongside `id name'. It shares pearl--cache-teams with pearl--all-teams, so an async-first populate would otherwise leave the cache without `key' and silently break filter-team lookups.
- Normalized the stored :linear-view-shared flag to a strict boolean at the re-sync read site. `:json-false' is truthy in Elisp, so an unnormalized stored false would flip the view to team-shared on re-sync without asking.
- pearl--sync-saved-query-await now returns `(:success nil :timeout t)' on a wait-for elapse, so the caller can distinguish a timeout from a Linear-side rejection and tell the user the request may still complete server-side.
- The success path now emits a single message via pearl--sync-record-or-orphan-error. The prior code clobbered the orphan-recoverable view-id with a second message.
I hoisted the prompt-sentinel block (pearl--filter-cancel etc.) above its first user. The prior order produced a forward-reference compile error the moment any other change touched the file.
I added 31 tests in test-pearl-saved-query-sync.el covering the scope candidates, the customViewCreate/Update parsing, the entry persistence (including the :json-false overwrite case), the orphan-id error path, the timeout sentinel, and the get-teams-async cache parity. The keymap and menu tests gained coverage for the new binding. 631 ert tests total green. make compile and make lint clean.
Lower-severity findings filed as follow-ups: stale pearl--cache-views across a schema upgrade, empty filter encoding to null filterData, the pearl--save-query-mark-synced rebuild dropping unknown plist keys, and the scope-label rassoc rebuilding the candidate list to look up a string the picker already had.
|
| |
|
|
|
|
|
|
|
|
| |
Two UX refinements on the just-shipped prompt surface, both following the principle "the prompt should describe what it's doing and the default should be the most-common choice."
First, the sentinel that was uniformly "[ None. ]" really meant two different things at two kinds of prompt, and the label only matched one of them. Filter-dimension prompts (team, state, project, labels, assignee) treat picking the sentinel as "no constraint on this dimension." Every value matches, which is *any*, not *none*. The saved-query prompts (delete, run) treat picking the sentinel as "don't act." That one is *cancel*, not *none* or *any*. Renamed accordingly: `pearl--filter-any` ("[ Any. ]") for the five filter dimensions and `pearl--filter-cancel` ("[ Cancel. ]") for the two saved-query prompts. The generic helper became `pearl--with-sentinel SENTINEL CANDIDATES` so each call site picks the label that fits its case. The predicate became `pearl--filter-sentinel-value-p` (recognizes either sentinel or empty/nil) so the cancellation logic is unchanged.
Second, three non-destructive yes/no prompts ("Open issues only?", "Save this filter locally...", "Save N fields across M issues?") moved from `y-or-n-p` to a new `pearl--read-yes-no` helper. The helper renders a completing-read over ("yes" "no") with the most-common choice as the default and topmost candidate, so RET takes it without typing. Default is "yes" for all three (each is a do-the-thing-I-asked confirmation), but the helper takes a DEFAULT arg so a future prompt where "no" is more common can opt in. The destructive prompts (delete issue, delete saved query, delete comment) stay as `yes-or-no-p`. Typing "yes" there is a deliberate safety affordance, not friction worth removing.
Tests cover the sentinel-value predicate across both sentinels + empty/nil + real values, the `pearl--with-sentinel` helper, the `pearl--read-yes-no` t/nil return and default-ordering behavior, and the three save-test stubs swapped to mock the new helper.
|
| |
|
|
|
|
| |
`pearl-saved-queries' was write-only from a user's perspective. Save was bound to the builder, run was bound to a key, but removing a stale entry meant `M-x customize-variable RET pearl-saved-queries RET` and a manual delete from the Customize buffer. `pearl-delete-saved-query` under `C-; L k q` (and `q` in the transient's Delete group) closes that gap: completes over the saved query names with `[ None. ]` at the top as a cancel, confirms, removes the entry, and persists via `customize-save-variable`. Linear is untouched. Saved queries are local. Pushing them up to Linear as custom views is a separate spec.
While in the same path, `pearl-run-saved-query` got the same sentinel cancel. Both prompts use the keep-order completion table so `[ None. ]` stays at the top regardless of the user's completion sorter. Both error cleanly when `pearl-saved-queries` is empty rather than offering an empty picker. Six tests cover the new command (delete-on-yes, no-at-confirm, sentinel-cancels, unknown-name-errors, empty-list-errors) plus run's sentinel cancel.
|
| |
|
|
|
|
| |
Wrapping each builder prompt with `pearl--with-none` put the sentinel first in the candidate list, but the completion framework's default sort (vertico-sort-history-length-alpha, prescient, ivy's sorter, etc.) re-sorted alphabetically and any real candidate starting with a letter before "[" pushed the sentinel down. Linear's "Software Engineering" outsorted "[ None. ]", so the new pattern failed at the very first prompt.
The standard Emacs hook for "I sorted these on purpose" is a completion table that returns `(metadata (display-sort-function . identity))` for the `metadata` action. The new `pearl--completion-table-keep-order` helper does exactly that, and the five builder prompts (team, state, project, labels, assignee) now wrap their candidate lists through it. The framework keeps its default sort everywhere else. Only these five prompts opt out, in the place where author-order is meaningful.
|
| |
|
|
|
|
|
|
| |
The builder's state / project / team / labels / assignee prompts each relied on the "empty input means no constraint" convention, with a parenthetical hint in the prompt label ("Team (empty for any):"). The convention is invisible, has to be remembered, and reads as a non-choice rather than a choice. Now every prompt carries `pearl--filter-none` ("[ None. ]") as the first candidate and uses `require-match` so the user picks from the list. Selecting the sentinel is the same logical opt-out as empty input was, but the choice is on screen rather than in the user's head.
Two small helpers (`pearl--with-none`, `pearl--filter-none-value-p`) and one defconst keep the sentinel string in one place. The prompt labels lost their parenthetical hints since the affordance is now in the candidate list. Five unit tests cover the sentinel predicate and the list helper across the normal / boundary / "real value" cases.
The pattern generalizes well past the builder: any time pearl prompts the user to pick between a value and "no value," the no-value option belongs in the list, not hidden behind empty input.
|
| |
|
|
| |
The four-line preamble enumerated `M-x pearl-save-issue` / `M-x pearl-create-comment` / etc., even though pearl-mode binds the whole verb-prefix keymap under `C-; L` and that's what users actually type after the first session. Replaced with two lines: one for inline edits + save, one for the verb-prefix groups (edit, refresh, menu, pick-source). Reads as a key reference rather than a command list. The build-org-content test now matches the affordance line on "save" instead of the literal `pearl-save-issue` so the next rephrase doesn't trip it.
|
| |
|
|
|
|
|
|
|
|
| |
Pearl now treats Linear sources (favorites, custom views, ad-hoc filters, saved queries) as one surface. `pearl-pick-source` is the unified picker: it lists every Linear favorite alongside every locally saved query, and dispatches by kind. View / project / cycle / label / user favorites resolve to runnable filter or view sources; issue / document / dashboard favorites open in the browser since they aren't lists.
The filter compiler grows id-based forms (`:label-id` and `:assignee-id`), so favorites that name a Linear entity (a user, a label) compile to rename-proof queries rather than fragile name matching. The ad-hoc builder gets a matching member-assignee option ("me / member / any"), so a one-off "what's a teammate working on?" doesn't require building a saved query first.
Two pre-existing builder bugs surfaced during verify and got fixed in the same pass. The builder previously passed display names straight to the compiler, but `:team` compiles to `team.key.eq` and `:project` to `project.id.eq`, so Linear rejected every builder run that pinned a team or project. Now the builder resolves the picked name to its key / id before assembling. And `pearl-list-issues-by-project` previously injected `:assignee :me` into the filter, hiding every teammate's work in the project. It now returns the whole project.
The save prompt on the builder was rewritten so the local-only nature is unambiguous, and the README grew a Sources section plus a keymap-table entry for `f s`.
|
| |
|
|
|
|
| |
The keybinding prefixes now each name an action verb: f fetch, e edit, c create, k delete, o open, y copy. The url group is gone, its actions split into open and copy, and the direct lowercase c (new comment) is dropped so c can head the create group. Two commands carried non-verb names, so I renamed them: pearl-new-issue is now pearl-create-issue and pearl-add-comment is now pearl-create-comment. The old sync helper that already held the pearl-create-issue name moved to the private pearl--create-issue, where it belongs as internal plumbing.
I also swept "ticket" to "issue" across the labels, transient, docstrings, and README. Linear's own term is "issue" and the buffer mirrors Linear, so the command names already said "issue" while the labels said "ticket". Now they agree. The transient's per-entry keys still trail the keymap verbs, which stays its own task.
|
| |
|
|
| |
The compose buffers (edit-description, add-comment) and the smerge conflict buffer popped wherever display-buffer happened to put them. Now they open in a side window, defaulting to the bottom at 30% of the frame, with two defcustoms to control it: pearl-compose-window-side (bottom / top / left / right) and pearl-compose-window-size (a fraction of the frame, or an absolute line/column count). pearl--compose-display-action builds the display-buffer-in-side-window action and both pop sites pass it through pop-to-buffer, so Pearl always applies the side window over any display-buffer-alist entry for these buffers.
|
| |
|
|
|
|
| |
A comment whose markdown doesn't survive md->org->md (a heading, single-asterisk italics) was flagged changed on every save, because the dirty scan compared org-to-md of the rendered body against the stored markdown hash. When the comment was authored by someone else, that surfaced as "1 read-only comment skipped" on every save even though nothing had been edited.
Descriptions already dodge this: 9424b84 gave them LINEAR-DESC-ORG-SHA256 and compared Org-to-Org. I did the same for comments. pearl--format-comment now writes LINEAR-COMMENT-ORG-SHA256 (the rendered body hash), and a new pearl--comment-dirty-p compares the current body against it, falling back to the old markdown hash for comments rendered before the baseline existed. Both detection sites use it: the save scan and the edit-current-comment no-op check. The markdown hash stays for the remote-conflict gate.
|
| |
|
|
|
|
|
|
| |
An issue's Linear labels now appear as Org tags on its heading (** TODO [#B] Title :bug:backend:), so the org-native gestures work: filter by tag, build a tag agenda, sparse-tree on :bug:. Before this, labels lived only in the :LINEAR-LABELS: drawer, which Org's tag machinery can't see.
pearl--label-name-to-tag slugifies a label name to a tag: downcase, non-[[:alnum:]_] runs become a single underscore, Unicode letters preserved. pearl--label-tags builds the deduplicated set, first occurrence winning on a collision. The renderer appends them, and pearl-edit-labels rewrites them through pearl--set-heading-label-tags after updating the drawer.
This is render-only. The :LINEAR-LABELS: drawer stays the source of truth, hand-edited heading tags are ignored by save and the dirty scan and get rewritten from Linear on the next change or fetch, and the way to change labels is still pearl-edit-labels. Bidirectional tag editing (heading tags back to Linear) is out of scope.
|
| |
|
|
| |
I added pearl-copy-issue-url. It copies the Linear URL of the issue at point to the kill ring and the system clipboard, reading the stored LINEAR-URL drawer property so there's no network call, and works from anywhere inside the issue subtree. kill-new carries it to the clipboard through interprogram-cut-function, so nothing shells out to a clipboard tool. Bound at u c in the url group, beside u o (open in browser).
|
| |
|
|
|
|
|
|
| |
Pearl is now drivable entirely from the keyboard. pearl-mode turns on automatically in any buffer Pearl renders, detected by the #+LINEAR-SOURCE header on org-mode-hook, and binds the command map under pearl-keymap-prefix (default "C-; L"), so the keys are live without any global setup.
I filled out pearl-prefix-map so every day-to-day command is reachable. The hot-path commands bind directly under the prefix (list, refresh view, refresh issue, save, edit description, new comment). The rest sit in category sub-maps (fetch, edit, new, delete, url). The common ones appear in both places, so edit-description is the direct d and also e d. Delete moved from d to k so d is the direct edit key, and the old save sub-map is gone now that save is the direct s and S.
The org-mode-hook is registered at load rather than through an autoload cookie. An autoloaded hook would fire before pearl loaded, calling a void function, and would pull the whole package into every org buffer the user opens. Freshly fetched buffers also enable the mode directly in the writer.
|
| |
|
|
|
|
| |
The buffer should mirror Linear: a value on the page appears exactly as Linear stores it, so the heading title matches what you'd see opening the issue in Linear itself. I flipped pearl-title-case-headings to default nil so titles render verbatim. I also stopped running the view name in the #+title through the title-caser, which was coupled to the same flag. The smart-title-case path stays as opt-in tidying for anyone who wants it.
Two transformations remain, both forced by Org's syntax rather than chosen: state names become TODO keywords (Org keywords can't contain spaces, so the real name stays in the LINEAR-STATE-NAME drawer), and descriptions and comments convert between Markdown and Org. Documented under "Fidelity to Linear" in the README.
|
| |
|
|
|
|
| |
pearl-add-comment and pearl-delete-current-issue guarded position with the nearest-heading check, so from inside a comment subtree, or any heading without a LINEAR-ID, they errored "Not on a Linear issue heading". Both docstrings already claimed they work from anywhere in the subtree, so the behavior was just wrong.
The edit and save commands already climb to the issue via pearl--goto-issue-heading-or-error. I switched these two to the same guard, so every ticket-scoped command resolves the enclosing issue from anywhere in its subtree, including from inside a comment. The redundant LINEAR-ID re-checks are gone since the guard guarantees it. Walk-up tests added for both.
|
| |
|
|
|
|
| |
After a refresh, merge, or comment add, pearl folds property drawers so the page stays scannable. When org-tidy-mode is on, org-tidy owns drawer display with its own compact inline symbol, so a plain native fold doesn't match. Worse, org-tidy's overlays go stale when pearl rewrites the buffer, leaving the new drawers untidied.
I added pearl--hide-or-tidy-drawers, which re-tidies via org-tidy-buffer when org-tidy-mode is active and falls back to the native fold otherwise. I routed the three drawer-hide paths (full render, single-issue refresh, comment add) through it. org-tidy stays an optional soft dependency, guarded by bound-and-true-p and fboundp, so a setup without it behaves exactly as before.
|
| |
|
|
| |
Adding a comment left three rough edges. Its property drawer rendered expanded while every other drawer in the page was folded, so I fold the new subtree's drawers right after inserting it. The Comments heading's count went stale on add: a freshly created subtree now shows π¬ Comments 1/1, and an existing π¬ Comments N/M bumps to N+1/M+1. And the fallback that creates the Comments subtree when an issue had none was still inserting a glyph-less "Comments". It now leads with the π¬ like the rest.
|
| |
|
|
| |
Running pearl-list-issues on a buffer with unsaved edits used to print a passive "refresh deferred" message and do nothing visible, so it read as a no-op. Now it prompts: (d)iscard and rebuild, (m)erge by LINEAR-ID keeping the edits, or (q) cancel. Branch C of pearl--update-org-from-issues dispatches on the choice: discard rebuilds in place, merge hands off to the same by-id merge that pearl-refresh-current-view uses, and cancel leaves the buffer untouched.
|
| |
|
|
| |
The merge refresh protected a subtree from being overwritten only when its description body had changed, since pearl--subtree-dirty-p is description-only. A local edit to the title, state, priority, assignee, labels, or a comment looked untouched and got re-rendered away. I added pearl--subtree-has-local-edits-p, which wraps the comprehensive per-field detector the save path already uses, and pointed both merge protection sites at it. A merge now keeps any ticket you've touched in any field. The same gain flows to pearl-refresh-current-view.
|
| |
|
|
|
|
| |
The comment overlay started at the line beginning, so it sat on top of the leading stars. With a package like org-superstar that dims those stars (and composes a bullet from them) via a face, our overlay's face overrode it and the dimmed stars reappeared in our color, a stray "***" before the bullet. The darker, body-spanning read-only overlay made it obvious, but the overlay had covered the stars since comment highlighting first landed.
I start the overlay just past the leading stars now (the point right after the `^\*+ ` match), so the stars and the bullet keep org's own fontification and only the heading text and body get our face. The highlight tests check the face on the heading text and assert the stars are left uncovered.
|
| |
|
|
| |
A comment you can't edit only had a grey heading, and the grey (`shadow`) was light enough to blend with the body text of your own comments. Now the read-only overlay covers the whole comment subtree, heading and body, so the comment recedes as a unit. And `pearl-readonly-comment` uses a darker background-aware grey (gray40 on a dark theme, gray60 on light, `shadow` as the fallback). Editable comments keep the heading-only overlay so your own comment text stays normal-colored, with the green heading as the "this one's yours" cue.
|
| |
|
|
|
|
| |
`pearl-state-to-todo-mapping` and its derived regex stopped doing real work once keyword rendering moved to slugifying the Linear state name (`pearl--state-name-to-keyword`) and the keyword-to-state direction moved to resolving through the team's workflow states. The defcustom fed only `pearl--get-todo-states-pattern`, which had no caller left, and the two cache vars existed only to serve that dead function. State extraction reads `org-get-todo-state` off the buffer's `#+TODO` line, so nothing live touched the mapping.
I removed the defcustom, both cache vars, and the dead function, and dropped the three pattern tests that exercised it. Removing the defcustom orphaned the mapping bindings that ~12 test setup macros still carried. They were passing through slugification rather than the map, so I stripped them too and kept the `org-todo-keywords` bindings the temp buffers actually need. Also dropped the stale README config row.
|
| |
|
|
|
|
| |
The Comments heading rendered the count marker as `Comments π¬ 5/18`, with the glyph wedged between the word and the count. I moved the π¬ to the front of the heading, rendering `π¬ Comments 5/18`, so it reads as an icon for the section rather than part of the number. `pearl--comment-count-marker` now returns just ` 5/18`, and `pearl--format-comments` leads with the glyph.
The append locator that finds an issue's Comments subtree now tolerates the leading π¬, and still matches the old `Comments π¬ N/N` layout so a buffer rendered before this change keeps working on the next add-comment.
|
| |
|
|
|
|
| |
Comment headings only got their green (editable) / grey (read-only) coloring after `pearl-refresh-current-issue`, an add-comment, or a view merge (the three paths that call `pearl-highlight-comments`). A plain `pearl-list-issues` rendered through `pearl--update-org-from-issues`, which never called it, so a fresh fetch showed no coloring until I refreshed.
I added the highlight call to branches A and B of `pearl--update-org-from-issues`, right after the content lands in the buffer. Branch C defers a dirty buffer without writing, so it stays uncolored. The regression test renders the comment fixture and asserts both headings carry the editability overlay (own to editable, other's to read-only) straight from the bulk render.
|
| |
|
|
|
|
| |
`make coverage` only reported the JSON's file size, so reading the real number meant opening the report by hand. I added a self-contained `scripts/coverage-summary.el` that parses the SimpleCov JSON and prints per-file covered/total lines with a percent, a line-weighted project figure, and a source-weighted figure. A tracked source missing from the report counts as 0% rather than dropping out silently, so a source that never got instrumented still shows up.
A new `coverage-summary` target runs it against the last report, and `make coverage` chains it onto the tail of a local run (CI emits coveralls.json, not simplecov.json, so it's skipped there). Covered by `tests/test-pearl-coverage-summary.el`: parser hit/executable/merge-key semantics, the per-file records, and the error paths for a missing or malformed report.
|
| |
|
|
|
|
| |
Cycling a ticket's TODO keyword (C-c C-t) now counts as a state edit. The dirty scan gains a keyword-cycle arm: state is dirty when the explicit id moved off its baseline (the picker arm, as before) or when the keyword diverged from the synced state name's slug. At save, the picker arm pushes the explicit id; the keyword-cycle arm resolves the keyword to a team state id by slug match, first by position, and pushes that. A keyword no team state matches is reported skipped rather than guessed.
set-heading-state now slugifies the keyword instead of the static-map lookup, so the picker writes the same keyword the renderer derives. Without that, a picker edit to a non-standard state would read as permanently cycled. That retires pearl--map-linear-state-to-org, whose only caller this was, and its tests.
|
| |
|
|
|
|
| |
A same-source refresh updates issue subtrees in place and keeps retained dirty subtrees the merge skips, so the #+TODO line has to stay honest across it -- a refresh can surface a new state, and a kept heading's keyword must still be declared. pearl--update-derived-todo-header scans the final displayed buffer (every LINEAR-ID heading's TODO keyword + :LINEAR-STATE-TYPE: drawer), unions that with the fetched teams' states, and rewrites the line via pearl--derive-todo-line. Scanning the final buffer rather than the fetched issue list is what covers the retained subtrees. A legacy heading with no state-type drawer is classified by org-done-keywords. pearl--merge-query-result calls it after the merge, beside the source-header update.
The merge and integration test setups stub pearl--team-states so the gather stays network-free.
|
| |
|
|
|
|
|
|
| |
The renderer wrote a fixed #+TODO line and mapped state names through a static defcustom, so a workspace's real states (Dev Review, PM Acceptance, ...) never showed up. Now the displayed issues' teams drive both.
pearl--gather-header-states collects each team's full state set (cached via pearl--team-states, teams in first-seen order, states by position) plus every displayed issue's own state, so a failed team fetch still leaves its issues' states declared. pearl--build-org-content takes that list and derives the #+TODO via pearl--derive-todo-line. pearl--format-issue-as-org-entry renders each keyword with pearl--state-name-to-keyword instead of the static map -- slugify reproduces the old defaults, so standard states are unchanged -- and writes a :LINEAR-STATE-TYPE: drawer so a later merge scan can classify a retained heading. The team-states query gains type and position.
The integration test stubs pearl--team-states so the render stays network-free. The merge-refresh header rebuild is the next commit. Until then a same-source refresh keeps the previous header.
|
| |
|
|
|
|
| |
Two pure functions behind the coming workspace-faithful #+TODO line. pearl--state-name-to-keyword slugifies a Linear state name to an Org keyword: upcased, non-alphanumeric runs collapsed to a single hyphen, edges trimmed, Unicode letters kept (so "Γ
ngstrΓΆm" survives), and an all-symbol name falls back to TODO. pearl--derive-todo-line turns an ordered state list into the "ACTIVE... | DONE..." string: completed/canceled/duplicate types go after the bar, the rest before, order preserved within each side, de-duplicated by slug, with the hardcoded default returned for an empty list.
There's no caller yet -- the render integration and the gather pipeline land next. This is the derivation half of the keyword spec. The sync-back half it described is superseded by save-model-v2 (state reconciles at save) and the separate keyword-cycle task.
|
