Both aiming modes announced their name and then went silent, so there was
no way to know what the arrow keys were pointing at. Prompted by "it says
targetting, melee, but when I press the arrow keys it does not tell me
what I am currently selecting."
They look like a pair and are not, which is the whole shape of the fix.
Throw has a real free-roaming cursor on level.cursorX/cursorY, so its
descriptor is ingame_observe plus one clause: whether the tile is a legal
target. That clause is recomputed from level.FoV, level.interactableArea
and level.canThrowOnCell rather than read off the state's targetAllowed
upvalue, because ingame_targeting_throw.lua:228 sets targetAllowed false
for the whole 0.3s cursor glide and only the animation's onFinish
restores it -- and mid-glide is exactly when we speak, so it would have
said "blocked" on every single move. The same class of trap as the
examine-cursor popup lag, and the same answer: read the model, not the
view.
Melee has no cursor and no target state at all.
ingame_targeting_melee.lua:184-195 takes one direction key and fires the
callback in that same keypress, so nothing is ever "currently selected"
and there is no focus for a read() to return. Its descriptor speaks only
on entry, listing the eight adjacent tiles -- which are precisely the
legal targets, the game's rule being Chebyshev <= 1. Arrows still commit
immediately; the game plays as it did. A two-step confirm cursor was
considered and rejected for that reason.
That entry line is surroundings.scan()'s sentence, so scan() splits into
describeNeighbours() returning the text and scan() speaking it. onOpen
grows the ability to be a function rather than a fixed string, since the
neighbours differ every time.
menus.selfTest needed a branch for a descriptor with neither read nor
probe -- melee is the first -- or it would index a nil probe and take the
whole self test down with it, which is the one report that must never
fail quietly.
Neither mode is reachable by its key in the prologue: b is filtered out
by the room's level.allowedActions, and t falls through to
gotoInventory() when there is nothing throwable. Found that by writing
the testbed block as keypresses and watching it land in the inventory. It
now drives both states directly, borrowing levelWindow, hpbar and
levelDescriptionWindow off the ingame state's upvalues the way gotoAction
would pass them.
129 testbed checks, 0 failures. The throw block asserts the focus text
CHANGED across a cursor move rather than merely being non-nil, which is
the assertion that would have caught the targetAllowed mistake had it
been made. Still unverified against real content: only "can throw" has
been heard, not "blocked", "out of range" or "no line of sight".
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Prompted by "do sighted players have a minimap we don't have access to?"
No widget, but two map surfaces were only half exposed: the world map
screen, and the level view itself, which IS the explored-floor map --
rememberedFoV keeps everything ever seen drawn, and a sighted player pans
over it and reads the fog edge at a glance.
The world map dropped the three things the picture carries: what kind of
place each area is (the icon shape), which way it lies, and how long
travelling there takes. The last is the number the decision turns on and
was only ever in the side panel. states/worldmap.lua is now a read() plus
details() rather than probe/render -- necessarily, because render() gets
no state table and the travel time needs the screen's currentAreaId, an
upvalue like options and selectedOption. Travel time is the game's own
getDistanceMessage, the same call the screen makes. The blurb runs to
several sentences, so it goes on = and the settle timer instead of the
focus line.
Two gotchas: options[i].icon is already an image object, so the icon id
string has to come off the model area; and describe.direction only tests
the sign of each delta, which on the map's floating-point coordinates
called almost everything diagonal, hence a 25% deadzone carrying the same
4/3 x-correction the game's own distance maths uses.
New frontier.lua answers "where haven't I been", which nothing did:
surroundings is local, examine is one tile, and the tracker only finds
things it already knows about. A frontier tile is seen, not blocking
movement, with an in-bounds unseen CARDINAL neighbour -- cardinal because
the diagonal test fires on the unseen tile behind every wall corner, so a
lit corridor would report an unexplored way per corner. Clustered
8-connected and reduced to each cluster's nearest member, so a wide
doorway is one way but two openings that both lie north stay two. Reads
only FoV/rememberedFoV and reports only THAT an opening exists, never
what is beyond it -- exactly what the fog edge shows.
Shift+End names the nearest exit without disturbing the tracker
selection. Note for later: level.exits does not exist at runtime, it is
packed as nil and never assigned. The primitive is the game's own
getListByTag("exit"), which already gates on FoV or rememberedFoV, skips
hidden tiles and sorts by distance. Exits were already a tracker
category, so this is a shortcut rather than a new capability.
"remembered" moves from verbose to normal verbosity. tileFacts reads
actors only when a tile is currently VISIBLE, so a remembered tile is
silent about whatever may be standing on it now, and at normal verbosity
that silence was indistinguishable from "nothing there". A safety flag,
not a nicety.
Also adds descriptor.probeNames: a read() descriptor has no probe table,
so menus.selfTest could only report "custom read()" and the version-bump
canary was lost on exactly the screens reaching for the most upvalues.
worldmap and research now declare their game locals and get the same
PROBE BROKEN check.
120 testbed checks, 0 failures. It caught two things review had not. The
world map said "You are currently here." twice, because
getDistanceMessage and getLastVisitedMessage return the same sentence for
the area you are standing in -- fine as two panels on screen, a stutter
in one utterance. And the first frontier check walked a step east and
demanded the numbers change; it failed on a run where nothing was wrong,
the step landing but both openings keeping the same Chebyshev distance.
Replaced with the deterministic property: every reported distance must
equal Chebyshev from the player's current position, nearest first.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The research tree spoke only the node's name and whether you could afford
it. Everything that tells you whether it is worth affording -- the
knowledge cost, the recipes it unlocks, the description -- is drawn in
the hint windows beside the tree and was never spoken, so "cannot afford"
gave no way to find out what was missing or by how much.
states/research.lua now implements details(), the same = / Shift+= hook
the item grids use: cost with the shortfall spelled out, what it unlocks,
then the description, one block each so [ and ] walk them back. Shift+=
adds the name line and the whole knowledge purse across all four types.
The focus line also takes its name from the node's title instead of
humanising the id, so it is "Sewing techniques", not "sewing techniques".
Everything is read from the state's currentNode upvalue -- the display
table updateCurrentNode() assembles for the hint window -- rather than
models/research.lua, whose nodes table is file-local and whose unlocks
are recipe ids, which would have named the recipe instead of the item it
produces. The shortfall is computed against journalModel.knowledge,
required lazily because models.journal drags in much of the game.
The tree has no hotkey and the prologue may hold no writing desk, so the
testbed enters it the way the desk does: ingame.interface("research"),
the exact call in the desk's tryInteractFunction. Require it as
"states/ingame" and not "states.ingame" -- states.lua uses the slash
form, and the dot form loads a second copy of the module whose
levelWindow is nil, which research.draw indexes on its first frame.
98 testbed checks, 0 failures.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Learning a recipe spoke the word "Popup" and then nothing.
There are two separate popup mechanisms and only one was covered.
patches/popupstack.lua wraps popupStack.add, but views/fadingPopup.lua
registers just four types -- journal, loot, bestiary and tutorial -- so
the recipe/lore/map keys in that patch could never fire. Those three are
modal instead: states/{ingame,inventory,research} build one from
modules/{recipe,lore,map}Popup.lua and switch to states/ingame_popup.
ingame_popup had a name in stateNames.lua and no descriptor, so
announceScreen said "Popup" and the focus poll had nothing to add.
The popup modules are factories exposing only draw(), so every string is
a closure local. The new descriptor takes popupObject off the state and
reads them as upvalues of popupObject.draw, detecting the type by which
locals exist since the object carries no type tag. Reading order follows
draw order, newlines are flattened (the lore footer is one sentence with
hard breaks in it), and the blocks go into the console history so [ and ]
walk a long fragment back; = re-reads it.
The prologue fires no popup, so the testbed raises them itself via
ingame.recipePopup and ingame.lorePopup -- both display-only, so nothing
is learned and the world is untouched. That check is this descriptor's
version-bump canary: the self-test can only report "custom read()" here,
because the probe is on the popup object rather than the state.
88 testbed checks, 0 failures.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Pressing a hotbar key ran hotkey.action(state), which for an equipment
binding calls an itemActions entry's onActivation directly -- Wield,
Drink, Switch on, Pack away. None of those reach console.print or a
popup; most only play a sound. So the whole hotbar was silent: no item
name, no verb, no confirmation the key had registered at all.
patches/hotbar.lua wraps the hotbar VIEW factory, not models/hotkey:
models/hotbar.lua captures newHotkey as a file-local at require time,
and models/world.lua captures newHotbar the same way, both long before
our entry point. views/hotbar.lua is required only by the three states,
which load lazily -- the same ordering argument as patches/slider.lua.
The verb is computed BEFORE the action runs. The itemAction conditions
key off item.isEquipped, which the action itself flips, so asked
afterwards they return the opposite verb and a press that wielded would
announce "Unwield".
Also narrates the empty slot, the silent no-applicable-action branch,
binding and unbinding (sound-only in game), and action-type slots. Says
nothing when the item is missing from the inventory -- the game prints
that itself and patches/console.lua speaks it.
78 testbed checks, 0 failures. The prologue gives the player no items,
so the live path is exercised by planting a torch, binding it to slot 1,
pressing 1 twice and putting the world back.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The script ended with `git rm -r accessibility/publish` plus a commit, which
was created fresh with the current time on every run. That changed the tip SHA
even when nothing else had, making the old tip a sibling rather than an
ancestor of the new one -- so every publish was a force-push. It also was not
achieving its stated purpose: `git rm` only removes a file from the tip, so
publish.sh was still present in every older commit of the published repo.
Fold the exclusion into the filter instead, as one filename callback covering
keep, rename and drop. The machinery now leaves the whole history, and the
rewrite is deterministic -- verified by publishing twice into separate
directories and diffing git rev-list --all: all 24 SHAs identical.
Publishing is now `git push`, and a rejected push means something real: that
the dev history itself was rewritten.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The tutorial teaches "press s to see the contents of containers", and it was
silent. For a sighted player s sets levelWindow.itemsHighlightEnabled, which
draws the top item's icon on every loot-bearing tile in FoV.
We watch that flag rather than binding s ourselves -- it is game-claimed, and
keybindings.getCommand() ignores modifiers -- and speak on its rising edge, so
the player's own binding, hold vs toggle mode and the toolbar icon all work.
The flag is checked both at end of frame and inside the keypressed wrapper,
because in hold mode a quick tap goes up and down within one frame.
Reads the whole tile.loot list, not just the icon's top entry, nearest first,
with each container pushed to the console history for [ ] review.
71 testbed checks, 0 failures.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Tuned by ear: nearVolume 1.0 -> 0.8 and farVolume 0.25 -> 0.05. At the old
values a wall out at the full radius was still loud enough to compete with
footsteps and speech, so the tones read as a wash rather than as a cue. The
wider gap between near and far also makes distance easier to judge, which is
the point of the volume ramp.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The item grids said which item you were on ("a torch, slot 3") but nothing
about what it was -- no description, no condition, and no "what does this
refine into", which is the whole of alchemy.
The game already composes all of it in one place: item.getPopupMessage()
(modules/item.lua:199) returns the cursor popup's line list, so read that
rather than re-deriving anything from templates.
It also already gates it behind a dwell -- the popup only fades in after
examinationPopupFadeInDuration (0.4s), so a sighted player who arrows past a
slot never sees the details either. Mirror that instead of inventing a
modality: stop on an item for 0.6s and the details speak themselves, queued.
"=" reads them on demand, Shift+= keeps the name line too. ("r" was the
obvious mnemonic but is a game-claimed scancode, and getCommand() dispatches
on scancode ignoring modifiers.)
New generic hook descriptor.details() returning a LIST of blocks, so each is
also pushed into the console history and [ ] \ walk a long description back
afterwards -- the journal page trick. Only _itemgrid.lua implements it so far.
The other half is the hint panel craft/refine/trade draw permanently beside
the grid, which the popup never contains: recipe components, and price, which
appears nowhere else in the game. The three panels share a shape, so one
reader covers them. Affordability survives only as a red colour in the hint
table and decides whether enter does anything, so it is decoded and spoken.
59 testbed checks, 0 failures: a synthetic ivy_leaves item as the version-bump
canary for getPopupMessage, hasUpvalue over every hint-panel local, four
assertions on the pure settleStep, and an end-to-end "=" press. Not yet
exercised against a live item -- the prologue inventory is empty.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The documented way to run the testbed began with
rm -rf ~/AppData/Roaming/LOVE/Alchemist
which is the player's actual playthrough. It was written down as the
official procedure in four places, so every run of the test suite ate a
real save -- including twice in the session that produced the previous
commit. Reported by Talon.
The testbed does genuinely need an empty save directory: it drives a new
game from the main menu through the prologue, which has already happened
in any real save. So tests/run-testbed.sh MOVES the save aside (never
deletes it), runs, discards the test run's directory and moves the real
one back. Restore is on an EXIT trap plus INT/TERM, so it also happens on
a failed run, a crash and Ctrl-C. If the restore itself fails it says so
loudly and leaves the backup on disk rather than pressing on.
This cannot live inside the mod: love.load calls config.initialize,
keybindings.initialize and user.initialize long before main.lua reaches
require("accessibility").initialize(), so by the time any of our Lua runs
the save directory has already been read. The swap has to happen around
the process.
The rm -rf is removed from all four docs that carried it, each replaced
with the script and the reason not to clear that directory by hand --
being written down as procedure is what made it recur.
accessibility/.gitattributes forces LF on *.sh, or a Windows checkout
hands bash a \r and the script dies. It sits inside accessibility/ on
purpose: a top-level one would show up in the game-file footprint check,
which still reads `main.lua | 1 +`.
Verified with a marker file in a populated save directory: 38 checks, 0
failures, marker intact, no leftover backup directories.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Pressing enter/z on an inventory item plays a click and, until now, said
nothing. The click is sound.select: the menu WAS opening every time.
inventory.keypressed -> inventoryWindow.enableItemMenu() builds a second,
nested selection over the grid (modules/item/itemMenu.lua) listing the
actions that item currently allows, filtered from itemActions.lua by
per-action conditions: Activate, Wield/Unwield, Wear/Take off, Switch
on/off, Drink, Eat, Apply, Read, Open, Prepare for throwing/Pack away.
Up/down move in it, enter activates, any other key cancels. None of it
goes through console.print or the cursor popup, so it was silent -- and
since the game has no separate equipment screen, this menu is the only
way to equip anything. Reported by Talon: the torch quest was unplayable.
_itemgrid.lua now checks window.itemMenu first and keeps the grid cell
quiet underneath, since the menu is modal over it. The menu is a closure
with no getters, so lines and currentSelectedAction are read as upvalues
of its own draw(). The item name leads only when the menu opens, and the
"just opened" memory is the menu table itself rather than a flag, which
keeps read() idempotent -- a flag would be consumed by the focus poll and
the real announcement would come out headless.
Items also now read "equipped", which is part of the focus key on
purpose: after Switch on the menu closes with the cursor unmoved, so
otherwise the re-read is suppressed as unchanged and the action still
feels silent. Now: "Switch on" -> "a torch, equipped, slot 3".
New patch (the 6th) wraps the inventoryWindow factory for the dead case:
when getActionsList returns nil -- an item with no applicable actions --
the game does literally nothing, not even the click, which by ear is
indistinguishable from a dropped keypress. Says "<item>, no actions".
The prologue inventory is empty, so the live path is unverified; the
testbed instead builds an itemMenu from a synthetic item and asserts the
two upvalue names still resolve. That is the version-bump canary for this
descriptor, not proof of the wording. 38 checks, 0 failures.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The setup is unusual enough to need its own page, and the facts were
scattered across scratch.md, CLAUDE.md and publish.sh: the mod is developed
inside a copy of a commercial game, so the dev tree deliberately has no
remote and publishing means rebuilding a filtered repo and force-pushing
that instead.
docs/00-dev-setup.md covers the two-tree arrangement and which one you are
in, prerequisites and why the game must be an extracted folder rather than
a .love (ffi.load goes through the OS loader), building a dev tree from
scratch including the vanilla-0.4.2 baseline tag and the one main.lua line,
the everyday commands, the save location, reading the [speech] console and
the startup self-test, and the publish flow with its leak check.
Linked from CLAUDE.md and the docs index. No code change.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Walking blind, the hard part is not "what is next to me" -- that is `/` --
but "what changed when I moved". Rounding a corner reveals a room, stepping
back hides it, a rat walks into sight behind you, and none of it is spoken.
Shift+/ cycles three modes:
settled (default) changes accumulate while moving, spoken once you have
stood still for 0.4s
always every step reports its own change
off
`settled` diffs against the state at the last ANNOUNCEMENT, not the last
step, so a long walk gives one combined summary that is a genuine net
change: something that came into view mid-walk and left again before you
stopped is not mentioned, and something visible throughout is mentioned
once. Summing per-step diffs would report both. A single step is the
degenerate case -- you are already standing still.
Arrivals and departures are deliberately asymmetric. Membership is current
FoV only (never rememberedFoV), but furniture drops out of FoV constantly
indoors as walls come between you and it, and "the bed is out of view"
every other step is noise. Arrivals are reported for every watched
category; departures only for creatures.
Terrain is unwatched by default -- water arrives dozens of tiles at a time.
Step detection polls playerActor.x/y as walltones does, so pathing, portals
and pushes are covered with no game function wrapped; the scan is a 14-tile
box and runs only on a step or a due settle, never per frame.
snapshot() and diffText() are pure and exposed so the testbed can assert
the wording and the FoV invariant directly instead of capturing speech.
34/34 testbed checks, 0 failures. Game footprint unchanged: main.lua | 1 +.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The mod is 63 files; the tree it lives in is 1838 files of a commercial game
and 296 MB of history. Publishing therefore means stripping, and stripping by
hand is exactly the sort of thing that quietly leaks a build of the game.
accessibility/publish/publish.sh rebuilds the standalone repository from this
tree with git-filter-repo, then asserts that no path outside accessibility/ and
the docs survived, and aborts rather than pushing if one did.
The publish-only files live in accessibility/publish/ rather than at the root
so that the game-file footprint check stays a clean "main.lua | 1 +", and the
filter renames them to the root where a visitor to the mod repo expects them.
The script then drops the publish machinery itself, which means nothing to
someone installing the mod.
README.md is the install guide: copy accessibility/ into an extracted game
folder and add the one line to main.lua, which install.patch also applies.
CLAUDE.md now states which of the two repos the reader is in, since half of it
(running the game, the footprint check) is meaningless without the game present.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
New sessions land in what looks like a game repository and start treating the
task as game development, so the accessibility framing had to be re-explained
every time. CLAUDE.md states it up front: everything outside accessibility/ is
the vanilla game and is read-only reference, and a request phrased as a game
bug ("the journal doesn't work") means a screen reader user cannot use it.
Deliberately a pointer rather than a copy, so it cannot drift out of sync with
accessibility/docs/. It sends readers to docs/scratch.md first and carries only
what a fresh agent would otherwise get wrong: that the user is blind so no
verification may depend on seeing anything, the small-footprint rule, that
speech must never throw, the FoV rule, how to run the testbed, and how to read
the startup self-test when something is silent.
CLAUDE.md sits in the game tree, so it would show up in the footprint check
that must stay "main.lua | 1 +". Excluded there and in the three places the
docs quote that command.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Opening the journal and selecting an entry announced its title and stopped,
which is an index into content the player could never reach. Page flipping was
silent too.
The screen is three widgets at once: the tab row (q/e), the list of entry
titles down the left (up/down), and the page on the right holding the actual
text (left/right). The descriptor read only the first two.
It now reads options[selectedOption].pages[selectedPageId] and speaks its text
blocks. A page is bounded by what fits on screen, so it is safe as a single
utterance. Layout newlines are flattened first -- splitTextByPages joins the
game's own wrapped lines with \n, so left in they make the reader pause
mid-sentence at every wrap point. Page position is announced only when there is
more than one page, and the blocks are pushed into the console history so [ and
] walk back through a long lore entry after the utterance has gone by.
The tab leads only when the tab changed and the title only when the entry
changed, so flipping pages does not bury the new page under a repeated heading.
That suppression made read() stateful, which is a trap: it is called for the
same focus from four places -- the per-frame poll, the post-keypress announce,
repeatFocus, and the testbed's before/after probes -- so a poll would consume
the "changed" flag and the real announcement would come out headless. The
descriptor now caches its last (key, text) and replays it whenever the key is
unchanged, making read() idempotent; the suppression state advances exactly
once per genuine focus change.
Adds a generic descriptor.reset hook, called from menus.announceScreen. The
memory lives in file-locals that outlive the screen, so without it reopening
the journal on the entry you left it on would read the page with no title.
Testbed asserts the spoken text contains the page body rather than merely being
non-nil. 29 checks, 0 failures.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Tracker selection now survives a rescan. The scan is rebuilt every turn and
both lists are distance-sorted, so holding a bare index meant one step by the
player silently re-pointed the selection at a different rat. The tracker now
remembers identity instead -- the actor table for creatures, coordinates for
tiles -- and re-derives the indices in reanchor() after each rescan. Selection
is dropped only when the thing genuinely stops being reported (out of FoV,
killed, picked up), and the existing clamp then falls back to the nearest.
Ctrl+Home walks to the selection, reusing the exact shape of the game's own
click-to-path in ingame.mousepressed: calculatePathFromPlayer, then either
bufferInput or playerInput for a one-step path. Interruptions and threatened-
player truncation therefore behave the way a mouse click does. Most things
worth tracking stand on impassable tiles, which calculatePathFromPlayer refuses
outright, so we fall back to the cheapest of the 8 neighbours and say "next to"
rather than "to". Guarded on states.getCurrent() == "ingame": elsewhere the
input buffer belongs to another state and feeding it moves would be a bug.
Wall tones now follow the usual proximity convention -- a near wall is loud,
one out at radius is quiet. It was inverted, which read backwards in play since
a nearby wall is the more urgent fact. config.louderWhenFurther still flips it.
Testbed covers both: selection identity across a forced rescan, and that
Ctrl+Home moves the player without breaking out of the map.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Adds accessibility/README.md, the player-facing setup and keymap reference --
the thing to read before playtesting.
Brings the developer docs in line with what was actually built:
- docs/README.md rewritten as an index with current status and a one-paragraph
architecture summary.
- docs/scratch.md gains a HANDOFF section: state of play, the commands to run
after any change, the five traps that each cost real debugging time, what is
NOT verified and why, deliberate omissions, and how to rebase onto a future
game version.
- docs/03-implementation-plan.md gains a status banner listing where the
implementation deliberately diverged from the plan (menu narration pulled
forward as a blocker, levelCursorPopup not patched, tryInteract insufficient
as a tracker filter, testbed added, no settings UI built).
- docs/04-open-questions.md gains a resolution banner: Q6 and Q8 remain open.
Also records that the testbed check count varies between runs (23-26) because
some blocks are conditional, so what matters is zero failures rather than the
total.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Game-file footprint unchanged: still one line in main.lua.
All 23 states now have descriptors; 5 patches applied.
Final two descriptors:
- keybindings_menu has two panels (commands and keys) with independent
selection, so it reads currentlyActiveWindow to decide which one has focus.
Announcing the wrong panel's row would be actively misleading, since the two
lists are unrelated.
- research is a 2D node graph rather than a list, and its nodes are drawn as
icons with no text label, so the id is the only identifier available. What
actually matters is researched / locked / cannot afford / available: without
the affordability part the screen is unusable by ear, because pressing enter
on an unaffordable node silently does nothing.
status.lua adds two pull keys -- ' for health and active effects, , for the
current action -- plus unprompted damage alerts.
Damage is polled from actor.hp rather than hooked. ingame.checkInterruptions
exists but only fires for interruptions during queued movement, so damage taken
on your own turn never reaches it. Polling catches every source (combat, poison,
fire, falling) with no game edit. Damage speaks immediately and interrupts;
healing queues, since it is worth knowing but never urgent.
patches/popupstack.lua wraps popupStack.add, the single entry point for loot,
recipe, lore, map, quest, tutorial and bestiary popups. None of these go through
console.print, so they were entirely silent before.
speech.immediate now logs to the console like queued speech does. It bypasses
the frame queue and so also bypassed the flush-time logger, which meant damage
alerts fired correctly but were invisible in the testbed transcript. Prefixes
are now: ! interrupt, * immediate, space queued.
26/26 testbed checks, history unit tests passing.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Game-file footprint unchanged: still one line in main.lua.
The slider turned out to be genuinely generic: modules/slider.lua is a factory,
so one wrapper on it narrates all four users -- craft, refine, trade and
ingame_wait (the wait-N-turns screen). Same leverage as the console.print and
inventoryCursorPopup patches.
The options screens do NOT share it, though: config_menu changes values through
config.increase/config.decrease, so it needed its own descriptor. It reports tab,
option name and value, and its focus key includes the value so left/right
re-announce even though the focused option has not moved.
states/_itemgrid.lua is now shared by inventory, craft, refine and trade, all of
which build a views/inventoryWindow. It stays silent while
confirmationWindowEnabled is set so the slider owns speech during quantity
selection instead of the two fighting.
testbed gains walkTo(), which navigates using the game's own pathfinder rather
than hardcoded direction sequences -- those drift as soon as anything moves the
player. The config test only adjusts a value when the focused option is
demonstrably safe, and puts it straight back.
The slider patch is applied and confirmed by the self-test but not exercised
live: ingame_wait is unreachable in the prologue because tutorial_bed overrides
bed's tryInteractFunction with an empty function by design, and the crafting
screens need an alchemy table and real items. Recorded as a NOTE rather than a
failing check, since it is a property of the game and not of the mod.
25/25 testbed checks, history unit tests passing.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Game-file footprint unchanged: still one line in main.lua.
Adds descriptors for actions_menu, worldmap, alchemy_table, journal, character,
inventory, and a shared one for the full-screen prose states (menu_text,
ingame_message, menu_message). The prose states also push their text into the
console history, so the existing review keys walk the manual paragraph by
paragraph rather than delivering it as one unusable utterance.
Two driver improvements, both prompted by real failures the self-test caught:
- menus.getAnyUpvalue finds a file-local across all of a state's functions and
one level into inner closures. Safe because a Lua file-local is a single
shared upvalue. Needed because actions_menu and worldmap build their lists in
open() and never reference them in keypressed, so the named probe function
found nothing. The generic probe path now falls back to this.
- menus.hasUpvalue checks that a local exists rather than that it holds a value.
getUpvalue cannot distinguish 'no such local' from 'declared but currently
nil', which made the self-test report PROBE BROKEN for worldmap, whose
'local options' is nil until the screen is first opened.
patches/popups.lua now stands down for any state that has a descriptor.
inventoryCursorPopup.set only fires when an item is under the cursor, so empty
slots and empty tabs were silent and you could not tell the cursor had moved.
The inventory descriptor reports item-or-empty plus slot and tab instead.
21/21 testbed checks, history unit tests passing.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Fixes three claims that would mislead a cold session: the levelCursorPopup
patching note (reversed in Phase 2), the tryInteract classification claim
(corrected in Phase 3), and the answered-questions list.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Game-file footprint unchanged: still one line in main.lua.
Three-level navigation as specified in D2: Ctrl+PageUp/Down cycles categories,
PageUp/Down cycles kinds, Shift+PageUp/Down cycles instances of that kind, Home
announces where it is.
- categories.lua: flag-first classification into Creatures, Items, Alchemy,
Interactables, Exits and Terrain features, reusing the game's own 'exit' tag
and excluding the six nameless mechanical templates.
- tracker.lua: scan, cache, navigation, announcements.
Caching keys on level name, player position and turn: position because
everything sorts by distance to the player, turn because creatures move and
containers empty. Scans run on keypress, never per frame.
FoV is respected throughout -- creatures only when currently visible, tiles when
visible or remembered -- so the tracker never becomes a cheat.
Home reports path length alongside the offset, via the game's own
calculatePathFromPlayer. Straight-line distance is misleading indoors, where a
chest three tiles north through a wall can be twenty steps away.
Corrects D4 in docs/05-decisions.md. I had claimed tile.tryInteract cleanly
separates 'the player can do something here' from scenery. It does not: 'wall'
carries a flavour-only tryInteractFunction ('You touch the wall. It's cold and
rough.'), so every wall segment was promoted into Interactables -- 13 in the
starting house, burying the chest and cauldron. Structural tiles
(blocksMovement and blocksVision) are now excluded before the tryInteract test;
containers and furniture block movement but not sight, so they survive.
12 kinds/19 instances instead of 13/32.
13/13 testbed checks, history unit tests passing.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Game-file footprint unchanged: still one line in main.lua.
- describe.lua: shared spatial formatting. 8-point compass and Chebyshev
distance, since the game moves 8-directionally and Euclidean would overstate
diagonal travel. Verbosity setting, because sweeping eight neighbours and
inspecting a single tile want different amounts of detail.
- states/ingame_observe.lua: narrates the game's existing examine cursor.
- surroundings.lua: probe one direction, scan all eight, or report location,
without moving anything or touching game state.
- patches/popups.lua: wraps inventoryCursorPopup.set, covering inventory, craft,
refine and trade in a single patch.
Q3 resolved as augment rather than relay: the game's descriptions carry no
position, which is fine on screen and useless by ear, so distance and direction
are appended.
Examine mode is polled rather than driven by levelCursorPopup. That popup is set
from an animation-completion callback, so unless the player has enabled the
game's quickExaminationPopup option the text arrives a full keypress late -- in
testing you heard the previous tile while standing on the next. Polling
level.cursorX/Y is same-frame and setting-independent.
hotkeys.isCtrlDown/isShiftDown are now overridable fields rather than direct
love.keyboard calls: real hardware state cannot be faked by love.event.push, so
without that seam every modifier-gated hotkey would be untestable.
11/11 testbed checks, history unit tests passing.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Game-file footprint unchanged: still one line in main.lua.
Wall tones: stop the previous tone per direction before playing the new one.
The tones are 2.33s mono; steps land far faster than that and up to four
directions sound at once, so overlapping them stacked into a continuous drone
instead of reading as a per-step cue. Cutting the old tone also means the
distance/volume you hear always reflects where you are standing now.
ingame_dialogue descriptor: speaks speaker and message on a node change, but
only the response on a selection change -- re-reading the whole speech every
time you move between responses would be unusable.
Two more bugs found by the testbed:
4. Nested states.switch announced in reverse. ingame.open() sees
world.pendingDialogue and switches straight to ingame_dialogue; because we
announce after realSwitch returns, the inner announcement fired first and the
outer one interrupted and clobbered it, so entering a conversation said
'Game' rather than 'Conversation'. Now we stay quiet if we are no longer the
current state.
5. The game already numbers response text ('1. (Continue)'), matching the number
key that picks it, so our own numbering produced '1. 1. (Continue)'.
9/9 testbed checks, history unit tests passing.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Pulls menu focus narration forward from Phase 4: it was a blocker, since the
screen name was spoken but not the selected option, so the game could not be
started at all.
Game-file footprint unchanged: still one line in main.lua.
- menus.lua: descriptor driver supporting both D3 backends (upvalue probe and
patched-announce), sharing one render per screen. Probe self-test included.
- Descriptors for main_menu, settings_menu, load_game, new_game, narration.
- testbed.lua: drives the game by code via love.event.push, so input takes the
same path as a human keypress. Coroutine-based, and navigates by READING focus
rather than counting keypresses, which drifts when options are disabled.
Three bugs found by the testbed and fixed:
1. love.keypressed was still bound to the pre-wrapper states.keypressed.
states.setup() captures love[callback] by value and runs before us, and
main.lua does not define love.keypressed itself, so wrapping the states table
field alone had no effect. Now reassigns love.keypressed/textinput.
2. Focus can move with no keypress: main_menu.mousemoved assigns selectedOption
from cursor position, silently overriding selection including at startup.
Focus is now polled in the update hook instead of announced only on keypress.
3. Probes sometimes need chaining. narration.keypressed only calls
continueNarration(), so the text locals are upvalues of that inner function.
Verified: main menu -> new game -> name entry -> intro narration -> gameplay,
8/8 checks passing, including console.print capture and its (x2) fold.
Known gap: ingame_dialogue announces 'Game' and reads no dialogue text.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Game-file footprint unchanged: still one line in main.lua.
Console (Q4: speak everything queued, cancelled by explicit action):
- history.lua keeps our own unbounded history, since the game's console buffer
expires entries after 3s. Mirrors the game's (xN) coalescing so what is heard
matches what is on screen. The review cursor holds position while new lines
arrive rather than being yanked to the tail mid-read.
- patches/console.lua wraps console.print, calling the real one first so a bug
on our side cannot swallow a message the player would otherwise see.
- Any keypress reaching the game silences the pending speech backlog.
Hotkeys: [ and ] review, Shift+[/] jump to ends, \ repeat, Shift+\ copy.
Chosen from scancodes the game does not bind. Note keybindings.getCommand()
ignores modifiers, so Ctrl+Up would read to the game as a movement -- our keys
are consumed and never overlay a game-claimed scancode.
Wall tones (Q5: on step, 3D-positioned, radius 5):
- Direction and distance are deliberately decoupled. The listener sits at the
camera 4 units above the plane under an inverse distance model, so placing a
tone at the wall's real position would make distant walls quieter (the
opposite of the request) and collapse the stereo pan. Instead the source sits
at a fixed radius in the wall's direction at listener height, and volume is
computed explicitly from distance.
- Step detection polls player coordinates rather than hooking movement, so it
covers walking, pathing, portals and pushes for free.
- Respects FoV/rememberedFoV; never leaks unseen geometry.
tests/test_history.lua: 16 assertions, all passing, pure Lua.
Not yet verified in actual play -- needs a human at the keyboard.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Establishes the mod skeleton and gets speech working end to end.
Game-file footprint: one line in main.lua.
- accessibility/init.lua: sole entry point. Patch registry, hooks states.switch
(announces screen changes) and states.update (per-frame speech flush, avoiding
an edit to love.update). Startup self-test prints what resolved -- the canary
for when a future game version moves something we hook.
- accessibility/speech.lua: TTS wrapper. pcall-guarded so a missing DLL degrades
to silence instead of crashing. Queues utterances within a frame and joins them,
since one game turn emits several console.print calls and speaking each with
interrupt would leave only the last audible.
- accessibility/stateNames.lua: spoken names for the 28 states.
- accessibility/vendor/love2talk: vendored TTS library, see VENDORED.md.
Loads Tolk.dll by absolute path via love.filesystem.getSource() and adds that
directory via SetDllDirectory, so the DLLs live in the project instead of
needing an elevated copy into Program Files.
Verified with lovec against NVDA: TTS available, speech channel yes,
'Main menu' spoken on startup.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
