A third registry beside CommandRegistry and HookRegistry, per
docs/config-registry-framing.md. Unblocks the per-buffer auto-pair
toggle, the first of the five backlog items the missing config surface
was gating.
Substrate (src/config_registry.rs):
* ConfigRegistry keyed by name with definition order preserved, R42
mandatory descriptions, R50 typo detection, duplicate rejection,
and SourceLocation provenance -- the command/hook vocabulary.
* Closed scalar kinds: boolean, integer, number, string, enum. Owned
Rust values; Lua tables, functions and userdata are never stored.
Integer exactness is checked by value, never math.type, so the
luajit and lua54 builds agree.
* Two scopes. get(name, buf) resolves buffer-local -> global ->
default; get(name) with no buffer resolves the global chain only
and never consults an ambient buffer. Buffer-locals live in a
registry-owned side table purged at after_buffer_removed, beside
the keymap purge already there.
* An override is ALWAYS stored, even when equal to the value it
shadows; only value_epoch and listener dispatch key on effective
change. Without this a buffer pinned to the current value stores
nothing and a later global set flips it -- the pin silently never
existed. equal_valued_local_override_is_still_stored_and_shields_buffer
fails against the naive reading.
Bindings (src/lua_bindings/config.rs):
* define/get/set/set_local/reset/is_set/describe/list/on_change.
Spec tables are read raw, so neither an unknown key nor a
metatable-provided value can smuggle a field in.
* Listeners commit inside the borrow, snapshot, drop the borrow, and
only then re-enter Lua -- verified by holding the borrow and
watching the test panic with "RefCell already borrowed". A raising
listener is logged without blocking later ones or rolling back, and
a depth bound turns an accidental cycle into a pointed error.
Listeners persist until explicitly disposed; there is no Gc path,
matching the rest of the codebase.
* StartupOnly freezes off the existing InitCompleteFlag at write
time, so this arc adds no editor.rs call at all.
Adopters, each defining its own key so SourceLocation names the owning
module: editing.auto-pair (pair.lua, read per-buffer against the typed
edit's SOURCE buffer), editing.trim-on-save (editops.lua),
autosave.interval-ms (autosave.lua). No public function is removed or
deprecated, and both migration wrappers keep their legacy coercion --
trim_on_save("yes") still enables, interval_ms(1500.7) still floors to
1500 -- coercing before handing the strict registry a conforming value.
M-x describe-setting renders into *help*, modeled on describe-command.
Framing revision 3 records four defects implementation found in the
document itself: acceptance 30 and 31 contradicted each other; the
planned builtin/runtime/config.lua had nothing to hold and would have
broken the source-location contract had it held the one helper it might
have; F5 asked define to police a call it cannot see, moved to
set_local; and list() ordering was underspecified.
No protocol change; SUPPORTED stays [6..18]. No wire surface. Zero
changes to src/editor.rs.
Gates: fmt, clippy -D warnings, --lib (1683), --lib --features crdt
(1857), the new config_registry_acceptance (13) plus auto_pair (45),
editops (72), autosave (29) and m9_6 (25), m4 --skip basedpyright
(114), PMACS_REQUIRE_GPU=1 pmacs-gpu (109), the lua54 backend build,
and the full workspace sweep (2795 tests, exit 0). git diff --check
clean.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Finding-by-finding (framing revision 7; every fix bite-verified via
scripts/bite against the pre-fix tree):
1. Stored coordinates must be finite integers, and both cursor walks
are movement-bounded — they clamp at EOF, and the column walk
clamps at the target row's EOL instead of marching onto later
rows. An astronomical %d+ capture can no longer hang the editor.
2. The grep panel gains the same immediate buffer.after-edit
recovery trigger as the compile slots: M-x buffer.undo after a
COMPLETED search is marked synchronously.
3. The rustc arrow rule uses the framing's ([^:]+) spelling — paths
with spaces capture whole.
4. All pattern captures are collected (index 4+ reads the real
capture, not nil-as-column-0); capture indexes must be positive
integers; a rule naming a column its match didn't produce rejects
the match.
5. emit_text_raw is module-local — a user global could shadow the
helper the terminal-event path depends on, and its error consumed
the terminal event before pump cleanup/forget ran.
6. stdin/group spec fields reject wrong Lua types as hard errors;
group is matched as a raw Value because mlua's bool conversion
applies Lua truthiness ("true" would silently coerce).
7. resync also nils the public line_start_byte — total pre-marker
anchor invalidation includes the byte anchor.
8. The inherited cwd resolves through
pmacs.instance.identity().working_directory; the header always
names a real path and relative error files get an explicit base.
9. New AnsiParser::finish() + parser:finish() (additions #5): a
truncated multibyte sequence at process EOF surfaces as U+FFFD
before the exit marker instead of vanishing.
10. The built-in default rules are a private deep copy — in-place
mutations of the public table no longer survive the "using
built-in defaults" degradation.
Eleven new tests (r1f1a/b–r1f10); bites: 9 fail against pre-fix
compile.lua, r1f2 against pre-fix default.lua, r1f6 against pre-fix
lua_bindings/mod.rs — all clean assertion failures. Gates: fmt,
clippy workspace all-targets, lib 1522, crdt lib 1696, compile
acceptance 45, crdt acceptance 1, m4 101, GPU 59, workspace sweep
2493/0, git diff --check.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01VoiEyuPjoBhvwACf8HAnLB
builtin/runtime/compile.lua (Q#CM1-CM6, CM8-CM11): streaming
intercept-read-only *compilation* / *shell-command* slots fed by a
Lua-side ANSI parser (SGR to overlay spans, CR/BS/erase progress
collapse); once-per-newline error parsing over a validated,
fail-closed rule table (rustc arrows, gcc/clang, Python, generic;
severity override + keyword sniff; sub-1 captures discarded);
buffer-revision external-edit guard with desync marker and anchor
epochs, checked before every producer write, before byte-anchor
use, and immediately via buffer.after-edit; unified error.next /
error.previous dispatcher with last-claim-wins sources and a
diagnostics fallback (M-g n/p unbind-then-rebind — hence the
loader's ordering contract after lsp.lua; C-x ` bound; M-! bound);
buffer-local RET/n/p/g/q/C-c C-k plus all seven undo/redo chords as
status no-ops; tombstoned pump teardown honoring forget's
terminated-only contract; q-target never captures a generated
buffer; overlay retained per incarnation, cleared per run,
re-attached from buffer.after-switch.
builtin/commands/default.lua (Q#CM7): project.search's
*search-results* becomes a first-class locations buffer — read-only
with bypass writes, RET/n/p/q + undo no-ops + round-trip input,
structured-match locations (line-1, match_start as col, paths
resolved against the search root), per-write revision checks so a
batch cannot mask an external edit, on_removed stream cancel +
guards for kill-mid-search, root retention across interactive
supersedes from inside the pathless panel, and an error-source
claim per search.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01VoiEyuPjoBhvwACf8HAnLB
Arc 2 (docs/kill-ring-framing.md, rev 3 — three review rounds). Kills
accumulate in a ring; consecutive kills append; C-y yanks the head; M-y
right after a yank cycles older entries; C-k (kill-line) exists at last.
The ring is daemon-global (Emacs-daemon model); chains and yank sessions
are per-frontend.
The substrate (Q#KR2): EditorCore.command_history maps FrontendId ->
{this, last} command. Every input path updates it --- the rev-1 design
treated dispatch_key as the only input path and review falsified that
twice:
keybound command dispatch_key Run arm rotate
typed char (round-trip) self-insert fallback rotate
unbound key dispatch_key unbound arm break
GPU optimistic edit handle_remote_crdt_op break
pointer gesture dispatch_mouse + dispatch_pointer break
inbound OS paste unified paste route break
menu item menu_invoke_active rotate
M-x accept pmacs.command.invoke_interactive rotate
invoke_interactive gives Emacs's execute-extended-command semantics
(M-x kill-line then C-k appends; C-k then M-x kill-line does not); the
public pmacs.command.invoke stamps nothing. Wheel scroll deliberately
does NOT break (mwheel-scroll vs mouse-set-point, as in Emacs).
Three shipped bugs fixed en route (Q#KR10):
- Semantic-path Paste was dropped ("no grid-less effect yet"), and the
GPU always negotiates semantic render --- GPU Ctrl-V was a no-op. Paste
is now a dispatcher-level arm serving both attachment kinds.
- That arm keys off the dispatcher's AUTHENTICATED source; the old grid
arm trusted the client-supplied payload frontend_id, letting a forged
id paste into another frontend's active window (unit-tested).
- Paste, M-x-invoked commands, and menu-invoked commands never fired
buffer.after-edit (each runs outside dispatch_key's revision check),
so LSP/syntax/autosave missed those edits. A shared
with_after_edit_check helper now wraps all three sites; scope is
honest --- active-buffer compare, sound for these paths, not a general
any-buffer guarantee (buffer-aware edit epoch deferred).
The ring (killring.lua, Q#KR4-7): entries carry stable monotonic ids.
Append requires last_command in the kill family AND this frontend's
last_kill_id == the head's id --- A-kill/B-kill/A-kill pushes fresh
instead of corrupting B's entry. Yank sessions store {buffer, start,
stop, entry_id, text}: M-y validates last_command + live session + same
buffer + slice(start,stop) == text (out-of-bounds reads as changed ---
pcall'd; an early test caught the guard throwing on an upstream
deletion instead of refusing), rotates by locating the entry_id's
CURRENT position (positions shift under other frontends' pushes; ids
don't), verifies the applied replace (intercepts may alter it; accepted
post-hoc semantics), then goto_byte. Failed kills clear last_kill_id;
failed/refused yanks create no session, so a second invalid M-y cannot
ride the first's name-stamp.
OS clipboard: ring head mirrors to the ACTING frontend's OS clipboard
only (pending_clipboard's existing shape; frontends may be different
machines). External content joins the ring at yank time via the
clipboard_get slot check (an OS copy reaches the daemon only when
pasted). New core seams: clipboard_set(bytes) / clipboard_get.
Lifecycle (Q#KR11): SessionDetached prunes command_history and fires the
new frontend.detached hook (raw id); killring.lua drops that frontend's
tables.
pmacs.killring.max([n]) validated (non-finite rejected --- math.huge
would defeat the cap; shrink trims immediately), default 60.
Deferred, named: word kills (M-d/M-BS/C-BS/C-h/C-DEL discard bytes ---
needs bytes-returning deleters), C-SPC/set-mark, clipboard watching,
ring browser/persistence, C-u C-y / C-M-w, buffer-aware edit epoch,
Lua-visible intercept probe.
Tests: tests/kill_ring_acceptance.rs (24) --- chain mechanics incl. all
break rows, the M-x three-direction matrix, per-frontend interleaving
(A-kill/B-kill/A-kill; stable-id rotation under B's pushes; eviction
mid-session; upstream-edit invalidation), menu Cut via real right-click
+ menu pointer (feeds ring, fires after-edit once, chains with C-k),
external-paste integration, cap validation + shrink-trim, detach
cleanup. Plus daemon unit tests: forged-id paste lands in the
authenticated source's window and leaves the claimed frontend's chain
untouched; optimistic CRDT op breaks only the source's chain.
Gates: fmt + workspace clippy clean; lib 1500; crdt 1672; killring 24;
cua 5; query-replace 16; completion 9; autosave 29; desktop 11;
persistence 5; clobber 6; m4 90; m8 10+15; m10/m11 crdt; GPU 58;
git diff --check clean.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
EditorCore::save() wrote unconditionally via save_atomic and only THEN
recorded the new FileMeta. It never compared the on-disk identity against
the one the buffer read. So another editor's writes, or a `git checkout`,
were destroyed without a word --- the single worst data-loss path in the
editor, on its most-used command.
The comparison seam already existed and no caller used it: FileMeta is
PartialEq (mtime + size, sized so same-second edits still differ) and
file_io::current_meta reads it. The file_io module docstring even said
callers "should" compare before saving. Nobody did.
save() now refuses when writing would destroy content the buffer has
never seen:
* the buffer recorded a meta and the on-disk meta differs --- someone
else wrote the file;
* the buffer recorded NO meta (a `[new file]`, or a path set without
reading) yet a file now exists --- it was created underneath us.
A *missing* file is not a clobber: there is nothing to destroy, so
recreating a deleted file saves normally. An unstattable path falls
through and save_atomic reports the real error.
On refusal the status line says what happened and how to override, the
buffer keeps its unsaved edits, and `buffer.after-save` does not fire.
`M-x buffer.save-anyway` (ed.save_ignoring_disk_changes) overwrites
deliberately and re-syncs the meta, so an ordinary save works again.
Named `buffer.save-anyway`, not `save-buffer-anyway`, for two reasons: it
belongs in the `buffer.` namespace next to `buffer.save`, and the latter
outranked `buffer.save` as an M-x completion for "save" (which a lib test
caught).
This is the bug the autosave arc kept circling: Q#AS5's Fresh/Stale guard
refuses to auto-offer a recovery for an externally-changed file, but
nothing stopped save() from overwriting that same file.
Tests: tests/save_clobber_guard_acceptance.rs (6) --- refuses and leaves
their content intact, after-save does not fire on refusal, save-anyway
overwrites and re-syncs meta, unchanged files save repeatedly (the guard
must not trip on our own writes), a deleted file is recreated not
refused, and a `[new file]` buffer refuses once someone else creates the
file. Verified the tests bite: 4 of 6 fail with the guard disabled.
Gates: fmt + workspace clippy clean; lib 1499; crdt 1670; clobber 6; m1/m3
/m4 90/m5.8/m7.8/m8 10+15; autosave 29; desktop 11; persistence 5;
query-replace/completion/listview/overlay/cua green; GPU 58;
git diff --check clean.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Emacs query-replace, built on isearch with zero protocol change
(framing: docs/query-replace-framing.md).
- search.rs: find_first_from (literal) + find_first_regex_from (cached
engine, zero-width-skip) + compile_search_regex (shared smart-case
compile). Q#QR2's forward-scan-past-replacement primitive.
- QueryReplaceSession + core methods (editor_core.rs): begin (invalid
regex refuses, Q#QR2), replace/skip/all/replace-and-quit/finish;
matches run forward from next_from on the LIVE buffer, so offset
shifts and never-re-matching-replacements (a->aa) fall out for free;
current match highlighted via a single-element search_store set
(SearchMatchActive, both frontends free) + cursor reveal; quit keeps
replacements, only nothing-matched restores origin (Q#QR10).
- Dispatcher shadow (editor.rs): QueryReplaceKey (y/SPC, n/DEL, !, .,
q/RET/Esc/C-g) + dispatch_query_replace_key, the 5th modal shadow;
added to dispatch_idle disjunction (GPU round-trips keys) and fires
buffer.after-edit itself (Q#QR1 — a shadow returns before the normal
post-command check; once per !-batch).
- Lua: ed.query_replace_start/query_replace_active; query-replace /
query-replace-regexp commands (chained minibuffer.read, separate
from/to history buckets, empty-from reject / empty-to deletion);
M-% / C-M-% bindings.
- Per-match prompt via core.status → v15 StatusFacts.message band.
Tests: 7 core unit + 11 dispatch_key acceptance (replace/skip/!/./quit,
nothing-matched restore, empty-to deletion, a->aa non-loop, regex incl
invalid, after-edit fires, dispatch_idle gate, explicit M-% AND C-M-%
binding tests) + 5 search unit.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Extend the gutter with the last two of the framed modes (Q#UX4):
- Relative: each line shows its distance from the cursor line (cursor = 0).
- Hybrid: cursor line shows its absolute number, others relative
(Vim number + relativenumber).
`LineNumberMode` gains `Relative`/`Hybrid` + `number_for(line, cursor_line)`
(the per-line displayed value) and `is_on()`. `paint_line_number_gutter`
now derives each number from the mode and the cursor's buffer line
(`text_view.line_at_offset(cursor)`); the TUI re-renders the whole frame on
cursor motion, so relative numbers track the cursor for free. Gutter width
is sized by `digits(line_count)` for every on-mode, so the text never
jitters as the cursor moves.
Mode selection (chosen over a 4-way cycle): `window.toggle-line-numbers`
stays a binary off/absolute toggle; a new `window.set-line-numbers` opens
the minibuffer with an arrow-navigable completion dropdown
(off|absolute|relative|hybrid) to pick a mode directly. `set_line_numbers`
accepts all four; the getter returns them.
No protocol change here — the GPU half (which needs the mode over the wire,
protocol v14) follows. Test: number_for across all modes. fmt + clippy
clean both flavors; 1446 lib tests pass. Needs a TUI eyeball.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_014TXbAwk27agwhrNNrhLi2U
Introduce a reserved left gutter column with absolute line numbers in the
TUI/grid frontend — the foundational piece of the UX arc (docs/ux-arc-
framing.md). Default OFF (Emacs tradition), so zero layout/coordinate
change until a window opts in.
- window.rs: LineNumberMode { Off, Absolute } + per-window `line_numbers`
field + `gutter_width()` (digits(line_count) + PAD) + `decimal_digits`.
- editor.rs: the gutter is one viewport shift at the paint site
(cell_origin.col += gutter_w, cell_size.cols -= gutter_w) — every
viewport-relative painter (text, syntax, diag underline, search) stays
gutter-agnostic. The sites that read rect.origin.col directly get a
manual +gutter_w: cursor placement, local selection, mouse hit-test
(a gutter click maps to line start, Q#UX6). paint_line_number_gutter
writes right-aligned dim digits alloc-free.
- overlay_paint.rs: remote-presence cursor/selection shift by gutter_w.
- Lua: pmacs.window.set_line_numbers/line_numbers +
window.toggle-line-numbers command.
No protocol/daemon change (frontend-local, Q#UX1). Tests: gutter render
(right-aligned digits + past-EOF blanks) + decimal_digits.
Validated: fmt clean; clippy --lib clean both flavors; 1439 lib tests pass
both flavors. Needs a human eyeball (coordinate-math change) before the PR.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_014TXbAwk27agwhrNNrhLi2U
The Lua glue that gives the menu real items to surface.
- `edit.copy/cut/paste/select-all` commands (Q#CM6) over the core
clipboard, with the Emacs kill/yank bindings `M-w`/`C-w`/`C-y` and
`C-x h` (the CUA trio's keys are already bound: `C-a` line-start,
`C-v` page-down).
- `pmacs.lsp.active_attachment()` (Q#CM5): a pure, side-effect-free
attachment lookup for the menu's `symbol`/`diagnostic` visibility
checks. Unlike `attached_for_active`, it never triggers an attach just
because the menu opened.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_014TXbAwk27agwhrNNrhLi2U
Wires regex matching into the search session and the terminal
frontend.
SearchSession gains `regex` and `invalid` flags. search_begin takes a
`regex` argument; recompute dispatches find_all_regex (regex) vs
find_all (literal), recording `invalid` when the pattern won't
compile (an invalid pattern clears the matches and shows [invalid]
rather than a stale count). search_toggle_regex flips the mode and
re-runs the current query.
Input: C-M-s / C-M-r start a regex search (search.forward-regex /
search.backward-regex commands → ed.search_start(forward, regex)).
M-r toggles literal <-> regex mid-search — a new SearchKey decoded in
dispatch_search_key, so it works the same in both frontends (the GUI
already round-trips every key while searching). The TUI prompt reads
"Regex I-search:" in regex mode and "[invalid]" when the pattern
won't compile.
Multi-line: SearchView now washes each row a match spans, mirroring
paint_local_selection's per-row clip (newline excluded so a spanning
match doesn't paint a phantom trailing cell). Single-line matches —
every literal match — touch exactly one row, unchanged. The GPU
already fans multi-line ranges per-line, so it needs no change here.
Tests: regex match / smart-case / invalid-flags-and-recovers /
toggle-reinterprets-query (core); C-M-s starts regex + M-r toggles
mid-search (dispatch); multi-line per-row wash (SearchView render).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Wires the live-typing half of in-buffer search for the terminal
frontend, on a frontend-agnostic core so the GPU (next commit) can
share it.
EditorCore gains a `search: Option<SearchSession>` (query + origin
cursor + direction) and the `search_*` methods that drive it:
begin records the origin, input_char/backspace re-run `find_all`
against the origin buffer and refocus the match nearest the origin
(failing searches anchor the cursor back at the origin), step walks
the store's active match (wrapping, also usable post-accept), and
finish either keeps the cursor + matches (accept) or restores the
origin and clears them (cancel). The matches live in the shared
`search_store`, so the decorations producer and the TUI SearchView
light up live as you type.
Input routing is intercepted in `EditorState::dispatch_key`: while
a search runs, every key flows through `dispatch_search_key`
(SearchKey::from_chord) instead of the global keymap — printable
chars extend the query, C-s/C-r (and Down/Up) step, RET accepts,
C-g/Esc cancel, BS shortens. This is the same dispatch path the
daemon's `FrontendEvent::Key` round-trip uses, so the daemon-side
search already works; the GPU just needs to route keys + show the
prompt (commit 4). The TUI paints an `I-search: <query> (n/m)`
prompt on the bottom row while keeping the terminal cursor in the
buffer at the active match.
C-s / C-r start the search (search.forward / search.backward Lua
commands → ed.search_start). Both keys were free in the default
map (save is C-x C-s, redo is C-x r), so isearch lands without
disturbing the CUA / Emacs editing keys — no cursor.right rebind
needed (the framing doc had flagged C-f for veto; C-s is cleaner
and Emacs-faithful).
Any edit now marks the buffer's matches stale in apply_active_edit
(M11.8), closing the headline "stale-after-edit linger" bet:
accepted highlights vanish the moment the text they described
changes, rather than painting at wrong offsets.
Tests: EditorCore-level (begin/type/step/wrap/focus-from-origin/
cancel/accept/backspace/smart-case/stale-on-edit) and dispatch-
level acceptance (C-s drives the whole loop; Esc restores; query
keys never self-insert).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Typing over a selection composed two edits in Lua —
delete_region() + insert_char() — so it recorded two undo steps:
one undo left the half-replaced text, two restored the original.
Undo granularity is per apply_edit in both modes (v0.1 pushes one
UndoEntry per edit; CRDT commits per edit via export and groups by
commit, with record_checkpoint unused), so the fix is to make
type-over one edit.
New core EditorCore::insert_char_over_region emits a single
EditOp::Replace when a region is active (cursor past the inserted
bytes, selection cleared) and delegates to insert_char otherwise.
The three type-over commands (buffer.newline / tab / self-insert)
call it via a new Lua binding instead of the delete+insert pair.
delete_region and insert_char are unchanged for their other
callers; region-aware backspace/delete already emit one op.
Verified one undo unit in BOTH modes (dual_mode
replace_is_a_single_undo_step covers v01 + crdt — a CRDT Replace is
delete-then-insert internally but one commit) plus an end-to-end
acceptance test through the key-dispatch path.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
- S-<arrows>/S-<home>/S-<end> (+C-S word/paragraph variants) extend a
selection; the TUI grid paints it reverse-video; double-click
selects the word at point.
- Backspace / Delete consume the active region (delete_region first,
falling back to single-codepoint semantics).
- Typing replaces the region: buffer.self-insert / newline / tab
delete_region before inserting. pmacs-gpu cooperates by
round-tripping keys while an own-window selection is active, so
the region-aware commands run instead of a raw optimistic op.
- tests/cua_region_acceptance.rs drives the real dispatch path:
select -> BS/DEL/char/Enter, plus the no-region fallbacks.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Land the Model Context Protocol (MCP) integration as a transport binding,
not a built-in feature. Six Lua functions plus userdata methods expose
the substance of three MCP feature areas (resources, tools, prompts), a
notification dispatcher, and a non-trivial AI-assistance example package
that meets the architectural ship gate (spec/pmacs-spec.tex:1572): zero
direct calls into the Rust core, zero special-cased MCP handling outside
the public API, source under 2000 lines of Lua.
The M9.5 -> M9.6 -> M9.7 -> M9.8 layered composition validates the claim
"AI is a transport binding, not a feature" -- pmacs-mcp-ai composes with
pmacs-mcp-prompts.render and inherits notification handling transitively
through M9.7's package, demonstrating that the AI domain is a layer
above MCP, not a thread woven through the core.
Subtask shape:
M9.1 stdio transport + initialize handshake + restart policy
M9.2 resources with in-flight + settled cache and per-uri invalidation
M9.3 tools with isError-vs-JSON-RPC-error semantics + cancellation
M9.4 prompts with required-argument validation
M9.5 notification dispatcher (on_notification, off_notification)
M9.6 tools-as-commands fixture package + 12 audit findings disposed
M9.7 prompts-as-result-buffers fixture package + tree-sitter-md grammar
M9.8 AI-assistance fixture package (363+ LoC; 17/17 acceptance tests)
M9.9 formal package audit -- PASS on all three criteria
M9.10 release: TRANSITION-M9.md + MCP-for-package-authors guide
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Levi Neuwirth