UX changes: home navigation and consistent message-box controls - #1071
Draft
ZeroPoint95 wants to merge 25 commits into
Draft
ZeroPoint95 wants to merge 25 commits into
ZeroPoint95 wants to merge 25 commits into
Conversation
… a control on the seam The gate had an install scope (the settings rows) and a launch scope (--yolo) and nothing a person standing in a conversation could move. The conversation has its own posture now: `alt+y` walks asks → guardian → YOLO → asks, a press on the ◇ cell after the thinking rung is the same step, and /approvals (alias /yolo) prints the stops or sets one by name — ask, guardian, yolo, deny, auto. The posture is live and sticky in meta.json, so it survives /resume; `codeaf resume --yolo` outranks the saved word for that launch. internal/session holds the word ([Agent.SetApprovalPosture]) and asks cmd/codeaf's v3ApprovalGate for the finished policy, so every stop is the same v3PolicyMode the flag builds through: exceptions, shell rules and both floors are identical at every posture. `guardian` folds the guardian row into the ladder; `deny` is never a stop on the wheel. The row's YOLO badge is off the status row while the seam carries the chip and back on it — only when the gate is open — on the frames whose seam has no chip (the welcome, a task's page). The phone sheet says the posture at every posture. The door crosses the wire: MethodResolvedApproval, MethodSetApproval, Welcome.Approval and Facts.Approval, with an old engine leaving the cell a reading and every door saying whose rules decide. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
…, once #1063 gave /status and the phone sheet an `approvals` row read off the profile's word, and the approvals chip gave them one read off the conversation's dial; rebased together the page said the gate twice. One row now, in the ladder's words — ask, guardian, allow, deny, what /approvals takes — from the same resolver the chip draws from, so the cell, the badge and the page cannot disagree. The manual says so, and the task page's heading names the `guardian allowed` line so that question still lands there. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
ZeroPoint95
force-pushed
the
claude/codeaf-autonomy-control-2f07c4
branch
from
September 16, 2026 21:17
731dbd1 to
6372adb
Compare
…engine has a door A session assembled without an approval gate has the rebuild method and refuses it, and that refusal silenced the plain push — the person was asked twice for one banked command (TestABankedBashRuleReachesTheRunningEngineBeforeTheAnswer). The door is asked first now. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
ZeroPoint95
marked this pull request as ready for review
September 16, 2026 21:31
|
|
…model, the rung and the gate The rule above the box was three different lines. Home's said the folder and the model; the other places said a note or a bare rule, with a `here ~/codeaf` chip on the box row; a conversation's seam said the model, the thinking rung and the approvals chip, each a door. `alt+y` and `ctrl+v` did nothing outside a conversation, and the gate's word was a chip on one screen and a badge on another. Every place with a draft now draws the conversation's own seam over its box: `◎ new conversation in ~/x · <model> · ⠿ <rung> · ◇ <gate>`, spelled by the same functions the seam uses, with the same doors — `alt+o`/press on the model, `ctrl+v`/press on the rung, `alt+y`/press on the gate, `alt+w`/press on the folder. The pins are the draft's and ride onto the conversation `enter` opens from any place, through the session's own doors, off the loop. The model and the rung last as long as the window; the folder and the gate are spent, because an open gate is a safety claim about one conversation. The unpinned cells say what a fresh conversation on this install would run at — the install's row folded with the model's level, and the rows as they stand or `--yolo` — read through session.Agent.StandingApprovalPosture and carried over the wire once at the welcome. A place's own note takes the rule from the chords, whole, and the draft's cells give way before it, so the foot keeps one height. The scope chip is gone from every box row: the rule says the folder, and the composer layer opens its task in the same folder the rule names. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
ZeroPoint95
changed the title
…d gate over its box The room's seam said `room · esc/←← main` and nothing else; the node's model and machine sat on the status row's identity cluster as a door, its rung was on the panel, and the conversation's YOLO badge was drawn on a row about work the gate does not govern. The one box a person opens to correct running work was the one that looked least like the others. The room's seam is the box seam now — `room · esc/←← main · task glm-5.2 · ⠿ high · ◇ on its own · via friendli` — laid out by the seam's own ladder. The model and the rung are doors on the node while it can still be moved (a press, or ctrl+v), through the doors the panel already used. `◇ on its own` is a reading: a node runs every tool without asking and has no wheel, so alt+y inside a room says so instead of moving the conversation's gate behind a page that does not draw it. The status row names the task alone and never the badge; the rider rides the seam. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
…er come off the other places Every place had a message box at its foot, saying `say what you want done` under a hint that said `enter talk about it · alt+enter send it off as a task`. On six of the seven it was a box in a slot: `enter` opened the row under the cursor, typing on tasks went to the filter at the top of the list, on spend the words went nowhere, and a task could be sent from a page that could not show its answer. The owner's ruling: one box, on home. Tasks, standing, memory, spend and search now end in their note on the rule and their hint, with no box, no `new conversation in`, no alt+enter and no task clause on the map. Typing still filters tasks on its control row, searches on search (the query is the first row of the body while there is one) and filters memory (the head row echoes the words); spend and standing take no text. Two bare spaces still open home from a place with nothing to type into. The shared composer, the resting-sentence and box-on-body hooks and the tasks type-word are deleted; the foot-height law is two shapes, home's and everyone else's. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
The rule under the tasks list read `9 finished today · 191 earlier`, and the section headings a few rows up already say what the page holds — one figure on one frame twice. The rule is bare now and carries only `nothing matches` while a filter has emptied the page. The reading's tally stays as the count the headings are held against in the tests; nothing draws it. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
The welcome box has no seam over it, and the status row's badge was the one thing on that otherwise empty frame saying the gate was open: a red word in the bottom-left corner. The owner asked for it to go. The row says nothing while the greeting is up, at every posture; the seam and its gate cell come up the moment the greeting goes (the first keystroke, or on the very first conversation the first message). A frame with no greeting and no cell — an engine with no approvals door — still carries the badge when the gate is open. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
…on home and in a conversation alike Home's foot was the rule (the draft's facts, its chords on the right), the box and a hint; a conversation's was the rule (name, model, rung, gate, the keys on the right), the box and the status row of numbers. The owner ruled the two boxes should end the same way: the last row is the keys, and nothing else. So the seam carries the facts on the left and the numbers on the right — the bill, the meter, the state word, every law they had — and the row under the box is the keys: the state's own, the doors at rest, and on home the draft's chords beside its sentence. The bill and the meter are pressed on the rule now, the home door on the keys row. The numbers never wrap: they climb the old ladder, then the link, the badge and the question chip, and the state word is the last thing standing; the name and the model give way only when not even that fits beside them, and a via machine outlasts the cheap figures but not the dear ones. A room's status-row chip is gone (the breadcrumb bar names the task); the greeting and a frame with no rule put the state word on the keys row's right; the phone tier keeps its deck. The ladder is climbed lazily from one set of pieces per frame and only the rung drawn is painted, which takes the one-screen scroll from 229 allocations to 206. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
… the machine beside the model Home's box row is `› type to search or start something new` and its foot is keys only (`↑↓ pick · enter open · …`): the promise about typing belongs where the typing lands, and the lowest line is for keys. The phone tier's box row says the same sentence. The conversation's seam no longer carries the conversation's name — it took the room the numbers need, and the tab strip and the breadcrumb bar say it — and nothing stands in for it. The `via <machine>` rider now rides the model (`glm-5.3-flash · via deepinfra · ⠿ high · ◇ asks · main*`), in a room too, so the two halves of "who is answering" read as one cell; the rung's door is measured past the rider. The ladder's name-cutting rungs are skipped where there is no name. Owner rulings of 2026-09-17. Manual, change entry and the e2e word table follow. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
…lash (baidu)` — and stays The seam's `· via baidu` segment becomes brackets on the model itself, whole or nothing, in a room too: `task glm-5.2 (friendli)`. It no longer ages out ten minutes after the last answer — who answered last is a fact until somebody else answers — while the sheet's `served` row keeps its window, since the figures beside its name are about one answer. The direct-service check is asked of the conversation's own model (the node's in a room) rather than of the empty id the seam passes as the row's spelled name. Two regression tests pin the two rulings. Manual and change entry follow. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
…swered last A window opened onto a conversation the host was already holding drew the model and no machine until the next answer: the lane desk lives in the host's process and only live news crossed. The host now keeps each conversation's last landed sighting and puts it first on every new news outbox, with its age on the frame (`LaneWire.AgeMS`, omitted when zero, ignored by older peers), so the seam names the machine at once and the sheet's `served` row files it at its true age. Rescues in flight and withdrawals are not kept. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
… for it The engine host replays a conversation's last sighting during the welcome, which is before the surface has registered its lane reader, so the replay was dropped and a window attaching to a held conversation still named no machine. The desk holds the last landed sighting told while nobody was listening and hands it to the next reader, once; a rescue in flight is not held. Verified with the stub model over a real engine host: the second window's seam reads `deepseek-v4-flash-latest (stub)` before any typing. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
…nd not stood on, and no heading opens the search `where you were` is headed `threads` and `running` is headed `tasks`, on the grid and on the phone lane. The `threads` heading no longer opens the search place — the box under home already searches, and /search and alt+7 still reach the place. Nothing under `spend` is a cursor stop or answers a press: its three lines are a new [homeReadout] row kind, so the arrows step over them and a click on one does nothing; the heading still opens the spend place. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
A panel heading that names a place — needs you, tasks, since you left, spend, scheduled — underlines its word while the mouse is over it, so a person can tell it is a door before pressing. The underline is on the word alone, never the count, the explainer or the clause; projects and threads, which open nothing, never wear it; and the pointer on a heading moves neither the cursor nor the marked heading, which still follows the keyboard alone. The press and the hover share one lookup of the headings the frame drew. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
… needs-you rows drop the enter word, and a heading that opens nothing is dim The tasks panel is the tasks place in miniature: every task and adaptive run that started or landed inside the last twenty-four hours, from every conversation, plus anything still running, newest first — a title and a time like a row of threads, up to ten rows and then N more. Enter or a click on a row opens the task itself through the door the tasks place's own list takes: the live page for work this window runs, the record inside the tasks place otherwise; since you left's task lines take the same door. Background jobs and firing watches are not tasks and left the panel; a firing watch keeps its row on scheduled, and its card, ctrl+e, ctrl+x and ctrl+v follow the item there rather than the row's kind. No row of needs you says enter (or enter on standing) at the right of its description any more, and a question's answers are cut at three with nothing after them. A heading that opens no place — projects, threads — is painted dim, word and explainer alike, so it does not read as a door. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
…der the pointer, and the thread named in them The tasks heading carries no count and no explainer, and threads no explainer: both are the bare word. A needs-you row is its amber mark, its title and its wait at rest; what it is asking, with its answers, shows only while the row is being read — under the pointer, or under the cursor when nothing is pointed at — like every other description on the field. A digit still answers the top question from anywhere. A landing's line and a watch's line lead with the thread they belong to, spelled as threads spells that conversation; a question's title already is its thread. A growing row draws its second line under the pointer as well as the cursor now, on every panel. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Every needs-you description is headed by the thread the row belongs to — thread: <name>, spelled as threads spells it — then a blank line, then the question or the landing's files and sentence with the answers at the right, in the description column and under the row alike. The cdot prefix is gone. A row that names its thread grows three lines under itself; a column too short for that gives up the two extra reserved lines before it gives up a row, and lets the rows under a read row move down for the moment instead. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
…projects is read and not pressed ← and → no longer cross home's columns: → opens the row's own strip and ← closes it, on every column as at one, and the foot is the design's sentence alone (it named ctrl+o open folder while the strip was out of the arrow's reach; the chord still works). ↑ off the top row of home stays there rather than climbing onto the tab bar, which is reached by a press, tab or a chord; every other place keeps the walk. The projects heading carries no explainer, and the panel has no interactivity left: its rows are not cursor stops, a press on one does nothing, and the enter that started a conversation in the folder and the strip that offered its chats and its folder are gone — ctrl+t on a row is the way to a new chat in a folder. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
ZeroPoint95
changed the title
ZeroPoint95
marked this pull request as draft
September 18, 2026 13:44
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Home and conversation screens used different message-box layouts, and home mixed navigation targets with information-only panels. This change gives them consistent controls and makes home clearer to scan and navigate.
UX changes
threadslists conversations;taskscombines live work with root tasks begun or finished in the last 24 hours, newest first. Task rows open their task page. Needs-you descriptions identify their thread and retain visible answer controls even when the field has few rows.project: <path>. Model names stay bold and bright, provider names sit beside them, and effort usesmodel:effortwithout a badge. Project paths preserve their absolute root or home-directory tilde and truncate on the right.space space homelast. Home omits pick, Enter, Ctrl+O and Tab hints, conversations omit Tab, and home model selection uses/modelor a click rather than Alt+O. The task panel's hide control sits above+ /task./approvalsand/yoloalso set named postures. Conversation choices persist and work over remote connections. Home draft choices carry into the conversation they open. Reattaching to an engine replays the last serving provider instead of leaving the seam blank until another answer.Modifiers display as Option on macOS. Manuals and the change entry describe the resulting behavior.
Validation
make test-quick: build, vet, formatting, packed manual, change-entry validation, manual gates and repository laws.TMPDIR=/private/tmp go test ./cmd/codeaf/;go test ./internal/remote/.go test -tags=e2e -run '^$' ./internal/e2e/; live-model terminal tests were not run.make build; merged binary reports67639fba2.The launch suite's two existing
/varversus/private/varpath assertions also fail on the incoming branch with the default macOS temp path. A short canonical temp directory avoids that fixture mismatch and the Unix socket path-length limit.