thursday-agent 0.17.2 → 0.19.0

package/guide/calls.md

@@ -5,177 +5,137 @@
 Three ways in, and any of them also answers a call she placed:
 
 - Tap her face.
-- Say the wake phrase, "hey thursday" unless they changed it. It is off to begin with, because it
-  keeps the microphone open for as long as the tab is, and the listening is the browser's own
-  speech recognition: Chrome's sends what the microphone hears to Google, which is not this app's
-  to change. It is heard loosely, in English, so it is written in English words; a phrase in
-  another language is not saved.
+- Say the wake phrase, "hey thursday" unless they changed it. It is off to begin with. It keeps
+  the microphone open while the tab is, uses the browser's own speech recognition (Chrome sends
+  what it hears to Google), and listens in English only.
 - Press the shortcut, Alt + Shift + T (⌥⇧T on a Mac) unless they changed it. It is off to begin
-  with too, and works only while the app's tab has focus and nothing is being typed.
+  with too, and works only while the app's tab is in front and nothing is being typed.
 
-Both are switched on, off or rewritten in **Settings › Thursday › Starting a call**. The first-run
-setup offers the wake phrase too, with a try.
+Both are switched on, off or changed in **Settings › Thursday › Starting a call**.
 
 She speaks first, with a short greeting, in the language the user speaks, and switches when they
-do; the browser's language plays no part. Every call is a new one: she does not open on old work
-or on how the last call ended. She has the most recent part of her earlier calls — about the last
-twenty turns — and picks a subject up when the user does. Anything older is gone unless she kept
-it in memory (`memory.md`). A call she places herself opens on why she called.
+do. Until she starts speaking — three seconds at most — she does not hear the room, so words said
+in that moment are lost. Every call starts fresh, but she remembers the last part of her recent
+calls and picks a subject up when the user does. Anything older is gone unless she kept it in
+memory (`memory.md`). A call she places herself opens on why she called.
 
 ## What she does on the line, and what goes to a bot
 
-She answers from what she keeps about them, searches the web, runs a single command on their
-computer, and holds the conversation. Anything that takes more than a few seconds — a browser, a
-file to make, several steps — goes to a bot and comes back while the call carries on. Which bot is
-hers to pick; she says who took it.
+She answers from what she knows about them, searches the web, runs a single command on their
+computer, and holds the conversation. Anything that takes longer — a browser, a file to make,
+several steps — goes to a bot while the call carries on. She picks the bot and says who took it.
 
-While she is working the line under her face says so, and the answer can take a while. The
-microphone stays open the whole time, and the bars of its meter stay at the head of that line to
-show it: anything said meanwhile is heard and answered once the work is done, so there is no need
-to repeat it.
+While she is working, the line under her face says so. The microphone stays open: anything said
+meanwhile is heard and answered once she is done, so there is no need to repeat it.
+
+**Share screen**, on that line during a spoken call, shows her a screen, a window or a tab; the
+browser asks which. A small copy of it stands at the top right while it is shared. Nothing of it
+is sent while it is only shared: she looks when what they ask needs it — "what does this error
+say?" — and then a picture of it as it is at that moment goes to the model behind her, nowhere
+else. **Stop** on the line, the browser's own bar, or the end of the call stops sharing. When the
+browser refuses — on a Mac, the browser needs Screen Recording in System Settings › Privacy &
+Security to share a window or the whole screen — its reason shows, and nothing is shared.
 
 ## Ending it
 
-Saying they want to hang up ends the call when she takes it that way, which she does not always.
-Tapping her face ends it for certain, and so does the shortcut. A call with nothing said for 40
-seconds ends by itself, counting down on screen in the last 10. The call is billed by the minute
-while it is open, silence included. Work already handed to a bot carries on.
+Saying goodbye usually ends the call, but not always. Tapping her face ends it for certain, and so
+does the shortcut. When nothing is said and she is neither talking nor working for 25 seconds, the
+call ends by itself; the last 10 seconds count down on screen. A spoken call is billed by the
+minute while it is open, silence included. Work already handed to a bot carries on.
 
 ## Writing to her instead
 
 The **+** at the left end of the pill in the bottom right corner — or the `/` key — opens a line at
-the foot of the screen. It opens on her, always. What is sent to her there starts a call in
-writing: her answers come up beside her face as on a spoken
-call, what she is doing shows with them, and the line stays as the way to write back. She has the
-same memory, tools and bots as on a spoken call; there is no voice and no microphone, and nothing is
-billed by the minute.
-
-- **Her answers keep their shape**: a list or a table comes a line at a time, and what she names is
-  a link — a file opens over the call as a card's does, a page on the web in a new tab.
-- **Files** go with the words: the paperclip, a paste, or a drop anywhere on the window — at most 8
-  at a time, 25 MB each. A picture handed over in writing is one she can look at herself — a
-  receipt, a screenshot, a chart — so "what does this say?" is answered on the spot.
-- **Who it goes to**: the chip at the left of the line, or `@` and a name at the start. A bot
-  picked there gets that one message; after it the line is hers again, and **Esc** with a bot
-  picked goes back to her.
+the foot of the screen. Sending her a message there starts a call in writing: her answers show
+beside her face, and the line stays open for writing back. She has the same memory, tools and bots
+as on a spoken call, with no voice and no per-minute billing.
+
+- **Her answers** keep their shape (lists, tables), and what she names is a link: a file opens over
+  the call, a web page in a new tab.
+- **Files**: the paperclip, a paste, or a drop anywhere on the window — at most 8 at a time, 25 MB
+  each. She can look at a picture herself, so "what does this say?" is answered on the spot.
+- **Who it goes to**: the chip at the left of the line, or `@` and a name at the start. A bot picked
+  there gets that one message, then the line goes back to her; **Esc** with a bot picked also goes
+  back to her.
 - **What it runs on**: the GPT Subscription when one is signed in, else the OpenAI key. The small
-  **runs on** button under the line says which before anything is sent, and picks another model
-  for writing to her — any provider with a key. The pick is remembered by this browser and is only
-  for writing; the spoken call's models stay in **Settings › Thursday**. With nothing to run on it
-  says so and sends nothing; keys are in **Settings › API keys**.
-- **When a turn breaks** — a plan's limit, a refused key — her face says ERROR and the provider's
-  own words show in red under the line, over **Send it again** and the way to **API keys**. Sent
-  again on the same model, she keeps what she had already done — a job handed over before it broke
-  is not handed over twice — and goes on from there; moved onto another, she answers afresh. When
-  an OpenAI key is set and was not what broke, the button reads **Send it again on your OpenAI key**
-  and moves the line onto that key; it never switches by itself, since what a turn costs changes
-  with what it runs on.
-- **Writing again while she is still answering** does not wait: the words show at once, join what
-  she is doing and are read before her next step. Whatever she had no step left to read gets an
-  answer of its own.
-- **Work she handed over** comes back in writing too: the moment a bot asks or finishes she is
-  told and answers it, and nothing rings while the written call is on. After a turn breaks,
-  nothing comes in by itself until one goes through; after a long run of them with nothing from
-  you, the rest waits for your next message.
-- **Ending it**: **Esc** ends it — saying goodbye does not, since there is no line for her to drop —
-  and so does starting a spoken call. Opening a thread in the corner does not: the line steps out
-  while the corner is open, her answers keep arriving beside her face, and the corner says
-  *Thursday is still on the line* across its top, with **Back to her** to fold it in one press.
-
-It is kept with the other calls, marked *in writing*, and she reads it back on the next call like
-any other.
-
-During a spoken call the line writes to bots only — but a file put down then is one she is told
-about as it lands: its chip says *she knows it is here*, and what to do with it can simply be said.
-On a spoken call she sees only where the file was put, and hands it to a bot. A line left open with
-nothing in it closes when a spoken call picks up; one holding words or files stays.
-
-## Seeing and hearing the words
-
-**Settings › Thursday › Captions** picks how the words show: **Both sides** (hers on the left,
-theirs on the right, where clicking an earlier line of hers reads it again) or **Her last line**
-(one caption under her face, only what she said). Both sides is where it starts; a narrow window
-always shows her last line.
-
-What she is doing — a search, work handed to a bot, a note she reads or writes, a command — shows
-as a line that names what it is about: which note, which command, which picture. With both sides
-showing it stands under her words, the last three steps with the older ones fainter, and the steps
-stay with that answer for the rest of the call. What she is thinking about stays under her face.
-
-When she searches the web, the pages she read show under her face while she answers, and each
-opens in a new tab.
-
-## Her voice, her face, her models
-
-Everything here but her face is in **Settings › Thursday**, and changes apply from the next call.
-
-- **Face**: on a Mac, iPhone or iPad she is drawn in Apple's emoji, and on every other system in
-  letters: other systems draw emoji in their own style, and letters are lighter to draw. No setting
-  changes her face. At rest she is a face of smoke: it keeps moving inside her and drifts off her
-  edge on a wind that comes and goes. She opens her eyes as the app opens and again every twenty
-  seconds or so, waking one of three ways — a sigh of smoke, shaking off a crust of dust, or turning
-  round to face you — and before she closes them she does one thing: looks about, winks, dozes off,
-  her head goes round once, or she lets out a long breath of smoke that fills the space around her.
-  When her head turns, only her face turns; the smoke round her stays where it is. While her eyes
-  are open her smoke grows thick and far-reaching, and she closes them slowly, looking down, as if
-  falling asleep. That happens only while she is resting. She can spell a short word: CALL while she
-  rings, OK as finished work comes in, a goodbye as a call ends (BYE, SEE YA, NIGHT and a few
-  others, by the hour), and whatever she chooses to show. While she works it is a single tail
-  circling.
-- **Voice**: the voice runs on GPT-Live 1, with 22 voices to pick from; clicking a name plays it.
-  It is the one setting here a written call or a phone does not use.
-- **Style**: who she is to them — Bright, Calm, Straight or Rough, and the only difference is how she talks. The row
-  shows the one she is on; opening it lists the rest with a line each. It never changes what she
-  can do. Last in the list, **Your own** opens a box for saying it in their own words on top of
-  the one picked — how she talks, how much she says — which wins wherever the two differ. What to
-  call them is not this: said on a call, she keeps it in memory.
-- **Backend**: the model that thinks and uses tools behind the voice, and its **effort** — how
-  hard it thinks (`setup.md`, Models). Left as it came, the model moves to the newer one when an
-  update brings one; a model picked here stays until another is picked. Under **tools**, **Search the web** is on unless switched
-  off: a question about today — the weather, a price, a score — is answered on the line instead of
-  going to a bot. **Read skills herself** is off to begin with; on, she opens the same skills a bot
-  reads, and each one she opens costs the call a page of reading. **instructions** is a box about
-  how work should be handed over and what to check first.
-
-Voice, style, the backend and its settings belong to the app, so she is the same on every browser
-and from a phone. Captions and starting a call belong to this browser.
+  **runs on** button under the line shows which, and can pick another model, from any provider with
+  a key, for writing only. With nothing to run on it says so; keys are in **Settings › API keys**.
+- **When a turn fails** — a plan's limit, a refused key — her face says ERROR and the provider's
+  reason shows in red under the line, with **Send it again** and a link to **API keys**. When an
+  OpenAI key is set and was not the problem, the button reads **Send it again on your OpenAI key**.
+  It never switches by itself, since the cost changes with what it runs on.
+- **Ending it**: **Esc** ends it (saying goodbye does not), and so does starting a spoken call.
+  Opening a thread in the corner does not: the corner says *Thursday is still on the line*, with
+  **Back to her** to return.
+
+It is kept with the other calls, marked *in writing*.
+
+During a spoken call the line writes to bots only. A file put down then is one she is told about —
+its chip says *she knows it is here* — and what to do with it can simply be said; she hands it to a
+bot.
+
+## Seeing the words
+
+**Settings › Thursday › Captions** picks how the words show:
+
+- **Both sides** (the default): hers on the left, theirs on the right. Clicking a line of hers reads
+  it again.
+- **Her last line**: one caption under her face, only what she said.
+
+A narrow window always shows her last line.
+
+What she is doing — a search, a hand-off to a bot, a command — shows as a short line. Pages she
+read on the web show under her face, and each opens in a new tab.
+
+## Her voice, face, style and models
+
+These are in **Settings › Thursday › Models**, in two parts, **Voice** and **Backend**. Changes
+apply from the next call.
+
+- **voice**: 22 voices; clicking a name plays it. Only the spoken call uses it.
+- **style**: Bright, Calm, Straight or Rough. It changes only how she talks, never what she can do.
+  **Your own** adds their own words on top — how she talks, how much she says — and wins where the
+  two differ. What she calls them is not set here: tell her on a call and she remembers it.
+- Under **Backend**, **model** and **effort**: the model that thinks and uses tools behind the voice, and how
+  hard it thinks (`setup.md`, Models).
+- **tools**: **Search the web** is on unless switched off, so a question about today — the weather,
+  a price, a score — is answered on the line. **Read skills herself** is off to begin with; on, she
+  can open the same skills a bot reads, and each one she opens adds a page of reading to the call.
+- **instructions**: how work should be handed over and what to check first.
+
+Her face has no setting: on a Mac, iPhone or iPad she is drawn in Apple's emoji, elsewhere in
+letters. At rest she is a face of smoke that now and then opens its eyes, and she sometimes spells
+a short word, such as a goodbye when a call ends.
+
+Voice, style and the backend settings belong to the app, so they are the same in every browser and
+from a phone. Captions and **Starting a call** belong to this browser.
 
 ## When work has something to say
 
-While a call is open, a question from a bot, and a result or a progress line that comes in during
-the call, reach the conversation by themselves — on a spoken call in a quiet moment, in writing as
-soon as they come — and she says them. Work that had already finished before the call opened is
-not read out: it is on the screen — the card, the pill, the thread — and she looks it up when asked.
-What is done on screen while a call is open — a bot answered, a job stopped — she is told of, spoken
-or in writing, so she does not ask it again.
+During a call, a bot's question or result reaches her by itself — on a spoken call in a quiet
+moment — and she tells the user. Work that finished before the call opened is not read out; it is
+on screen, and she looks it up when asked.
 
-With no call open, a finished job lands as a card in the bottom left corner, and if the app is not
-the window in front, as a browser notification that brings it forward when pressed (the browser
-asks for permission when a call is placed). With no app tab open at all, the computer's own
-notification says it instead; on a Mac pressing that one opens Script Editor rather than the app.
+With no call open, a finished job lands as a card in the bottom left corner, plus a browser
+notification when the app is not the window in front. With no app tab open, the computer's own
+notification says it instead (on a Mac, clicking it opens Script Editor, not the app).
 
 **Settings › Thursday › Starting a call › She calls you** makes the screen ring instead:
 
-- **When a job needs me** — where it starts — rings when a bot asks something.
-- **Whenever a job ends** rings for every ending as well.
+- **When a job needs me** (the default) rings when a bot asks something.
+- **Whenever a job ends** rings for every ending too.
 - **Never** leaves it to the notification.
 
-It needs this tab open. It is one call however much is waiting: while it rings, the screen says
-under her face whose work it is about and what it asks — with the answers the bot offered and the
-other work waiting by name — and one round button, **Answer**. Pressing one of those answers
-answers the bot there and then, with no call, and the ring moves on to whatever else waits.
-Tapping her face answers too, and so
-does the wake phrase when it is on; **Esc** is "not now". She opens by saying why she called, then
-tells the rest one by one. A ring nobody takes becomes a short missed list in the same place, with
-**Call back**, until it is called back or cleared with Esc; the work itself stays in the room
-either way. She cannot switch this on herself: when they ask to be called back, say where it is.
-The ring sounds once, as it starts, and the screen keeps ringing without it; a browser plays no
-sound on a page nobody has clicked yet, so right after a reload it rings on the screen only.
+It needs this tab open. While it rings, the screen shows whose work it is and what it asks, with
+the bot's suggested answers — picking one replies without a call. **Answer**, tapping her face or
+the wake phrase picks up; **Esc** is "not now". A missed ring stays on screen with **Call back**
+until called back or cleared with Esc. She cannot switch this on herself: when asked to call back,
+say where it is.
 
 ## History
 
-**Settings › Thursday › History** has two tiles. **Call history** is every call that was kept, one
-after another, oldest at the top, each with the jobs it opened and where they stand; a single call
-can be deleted, or all of them with **Delete all**. A call is kept for three months and then goes
-by itself. **Reset history** deletes every call, every job and everything she remembers, for good;
-keys, bots and connectors stay, and so does what each bot keeps for itself.
+**Settings › Thursday › History** has two tiles. **Call history** lists every kept call, each with
+the jobs it started and where they stand; one call can be deleted, or all with **Delete all**.
+Calls are kept for three months, then removed by themselves. **Reset history** deletes every call,
+every job and everything she remembers, for good; keys, bots and connectors stay, and so does what
+each bot keeps for itself.

package/guide/index.md

@@ -14,9 +14,9 @@ When a wish needs a setting, say which one and where, rather than agreeing to it
 | `calls.md` | Calling her or writing to her, what she does herself, how a call ends, her voice, face, style and models, call history |
 | `bots.md` | Bots, the work they take, threads, stopping and questions, what they made and where it is kept, routines |
 | `memory.md` | What she keeps about the user, and how they change it |
-| `setup.md` | The first run, the buttons and Settings list, keys, models, skills, connected services, sign-ins, its own window, where the files live, what it costs |
+| `setup.md` | The first run, the buttons and Settings list, keys, models, skills, connected services, sign-ins, its own window and how it is started, where the files live, what it costs |
 | `phone.md` | Writing to her from a phone through a chat app: setting it up, who is let in, what reaches the phone |
-| `trouble.md` | A call that will not open, a job that stopped, a bot that cannot do something, where to ask when nothing here answers |
+| `trouble.md` | A call that will not open or ended, the wake phrase, a job that stopped, a bot that cannot do something, starting over, where to ask when nothing here answers |
 
 Two things hold everywhere:
 

package/guide/memory.md

@@ -2,36 +2,32 @@
 
 Memory is notes, and a note is lines of plain text — facts. `profile` is who they are,
 `preferences` is how they want things done and said, including what to do in a particular
-situation, and the rest are one note per person, project or topic. Each note is listed by one line
-saying what it is about, with the names they use for it. Profile and preferences are read whole on
-every call; every other note is opened when its subject comes up.
+situation, and the rest are one note per person, project or topic. Profile and preferences are read
+on every call; every other note is opened when its subject comes up.
 
-She writes memory herself as things come up in a call, what tells her most about them first:
-what they loved or could not stand and why, the way she talks included; what they are going
-through; good news; their stories and the people in their life; then plain facts and tastes. Bots
-read it and never write it; a bot keeps what a job taught it in its own memory instead
-(`bots.md`).
+She writes memory herself as things come up in a call: what they loved or could not stand, what
+they are going through, the people in their life, facts and tastes. Bots read it and never write
+it; a bot keeps what a job taught it in its own memory instead (`bots.md`).
 
 **Settings › Memory** is the whole of it. The notes are grouped — **You**, **People**, **Projects**,
 **Topics**, **Other** — with a count of notes and facts at the top, and **Filter memory** narrows
 them by any word. There the user can:
 
-- open any note and see each fact, when it was saved and who wrote it (you, on a call, a bot);
+- open any note and see each fact, when it was saved and who wrote it (you, or on a call);
 - change a fact or delete it, and **Add a fact** to a note;
-- change the line a note is listed by (not for profile and preferences, whose line is fixed);
+- change the line a note is listed by (not for profile and preferences);
 - make a note of their own with **New note**: its kind (Person, Project or Topic), **Name**,
   **Summary** (the line it is listed by) and first **Facts**;
 - delete a whole note with **Forget this note**;
 - type what changed instead of editing by hand: **Edit with a model** opens a box, *Tell memory
-  what changed*, with its own pick of model, and each edit shows as it lands.
+  what changed*, with its own pick of model.
 
 Each note says when it was last read back, and notes that are rarely read are drawn fainter.
 
-Nothing is hidden: what she knows about them is what is on that screen, plus the most recent part
-of her earlier calls (`calls.md`). Asking her to forget something removes the line.
+Nothing is hidden: what she knows about them is what is on that screen, plus the last part of her
+recent calls (`calls.md`). Asking her to forget something removes it.
 
-Memory holds up to 400 facts, and 50 in one note. When it grows past that she says so once on a
-call and goes through what looks out of date with the user, the heaviest notes first, deleting
-only what they name. Forgetting a fact takes it out of everything she reads; the earlier wordings it
-replaced stay in the database until its note is deleted. **Reset history** in Settings › Thursday
-deletes all of her memory along with every call and job; what each bot keeps for itself stays.
+Memory holds up to 400 facts, and 50 in one note. Past that she says so once on a call and goes
+through what looks out of date with the user, deleting only what they name. **Reset history** in
+**Settings › Thursday › History** deletes all of her memory along with every call and job; what
+each bot keeps for itself stays.