CHANGELOG.md

Changelog

Release history for codeaf. Unreleased changes live in docs/changes/unreleased/.

v0.8.2 — 2026-10-23

Fixed

• the installer says why GitHub's API refused, and retries once without a token GitHub refused — #1890 · build docs

v0.8.1 — 2026-10-21

Fixed

• model usage accounting covers compatible providers and separates estimates from measurements — #1887 · engine docs

v0.8.0 — 2026-10-14

Added

• ai& is offered as a provider of its own, asked for by key and nothing else — #1737 · chat engine

• memories belong to an owner, every write walks one door, and failure is visible — #1745 · chat engine docs

• ordinary memory retains evidence and conditional context before actions — #1777 · chat engine docs

• sec, sec-af's security audit built into codeaf, audits a repository or its changes — #1781 · chat engine docs build

• review, pr-af's code review built into codeaf, reviews a GitHub pull request and posts on your yes — #1784 · chat engine docs build

• senior-dev checks CMake projects itself and takes --verify-build and --verify-test — #1793 · engine docs

Changed

• Teams previews, shared conversations and scoped deletion — #1750 · chat engine remote docs

• Choose a provider during setup and pick only available models — #1789 · chat docs

• launches offer a new codeaf in a quieter, non-blocking way, and installs run in the background — #1790 · chat

• Clear settings categories and AI team navigation without resetting preferences — #1792 · chat engine docs

Fixed

• a reply that is only a number and a full stop draws instead of vanishing — #1746 · chat

• the hosted drive passes on a Mac whose task column starts open — #1756 · build

• PlanDB run workers can search and fetch through native tools — #1783 · engine chat

• a store written before memory owners existed opens again on the next launch — #1787 · chat engine

• senior-dev reads its window from codeaf offline, refuses tiny models, keeps progress on compaction — #1791 · engine chat docs

• Keep AI team hover responsive with large conversation previews — #1794 · chat

• Built-in provider lists accept appended catalog rows — #1795 · chat docs

• senior-dev keeps its session store out of the folder and writes back a removed .senior-dev — #1800 · engine docs

• senior-dev's ending no longer says the build and tests passed when it ran none — #1801 · engine docs

• senior-dev help fits eighty columns and is checked before merging — #1810 · engine build docs

• senior-dev checks selected-service model capacity before starting — #1811 · engine docs

• task check declarations are refused before work when they cannot be rerun — #1812 · engine docs

• the belt worker page fits its budget and is checked before merging — #1813 · engine build docs

• updater acceptance follows resolved install paths and current settings — #1814 · build

• security audits distinguish incomplete analysis from a completed search — #1815 · chat engine

• hosted acceptance checks actual task endings and retains diagnostics — #1816 · build

• review agents reserve answer time and retain failure evidence — #1817 · chat engine docs

• review resolves model words and linked roots and formats final answers — #1818 · chat engine docs

• shell programs price qualified model names on their selected service — #1819 · engine docs

• review rejects malformed consistency claims and retains partial coverage — #1820 · chat engine docs

• provider connections read complete bounded model catalogs — #1822 · chat engine docs

• concurrent provider-memo saves preserve every learned fact at restart — #1824 · engine chat

v0.7.1 — 2026-10-05

Changed

• Usage receipts distinguish routing, model families and missing counts — #1755 · engine remote build docs

v0.7.0 — 2026-10-04

Changed

• ctrl+enter queues a follow-up, and queued messages can be taken back — #1714 · chat engine

• senior-dev leaves no wip commits, writes its own commit message, and gets fixes handed back — #1716 · chat engine docs

• first run answers clicks and a second enter, drops the telemetry notice, lists free models when low — #1720 · chat docs

• touched packages run concurrently and attribute flaky or inherited failures — #1734 · build docs

Fixed

• the @ list finds recent conversations on every box without taking over prose — #1725 · chat docs

• Token usage reaches telemetry while a session is still running — #1731 · docs engine

• forgotten OpenAI keys and stale folder environments no longer hide a connected key — #1752 · chat engine docs

v0.6.0 — 2026-10-01

Added

• senior-dev can be told something while it works, from its page or the chat, until it hands in — #1689 · engine chat docs

• skills kept for OpenCode and Goose are read too, including their ~/.config folders — #1691 · chat resident

Changed

• Routine coding replies lead with the outcome and stay short — #1428 · chat docs

• Home copies the current conversation name from its second menu action — #1433 · chat docs

• Sessions filters whole conversations by name, project or nested task — #1434 · chat docs

• a reply cut off mid-stream is dropped and asked again — #1497 · chat engine

• Home points to the standing place for a waiting order's due time — #1544 · chat docs

• the size budget is measured on darwin/amd64 and raised to 68,850,000 bytes — #1695 · build

• Every recognised slash command is chipped wherever it stands, in the box and the sent transcript — #1699 · chat docs

• dev and staging builds report to OpenRouter as their own apps; a release stays AgentField AI — #1723 · chat engine build

Fixed

• Provider keys no longer reach a model's shell commands — #1486 · engine

• a late task start is an uncertain receipt — #1539 · chat remote

• Startup and home conversation openings keep the interface responsive — #1662 · chat engine

• the teams cap, header, keyboard and wall show the right thing — #1668 · chat

• a job's log path is spelled the way the session spelled its folder — #1687 · chat engine

• a note or say to a program that cannot read it is refused without storing the words — #1689 · chat

• the model picker keeps the cursor's row on screen at every width — #1696 · chat

• the current model wears no band of its own and every model list owns the wheel — #1696 · chat

• a typed /senior-dev names its ceiling and its ending cards wear the badge — #1697 · chat docs

• no zero rates, no raw ids as names, and the manual matches the screen — #1698 · chat docs

• Flags after the brief, run-engine token counts, long-home ssh paths and auth failures — #1700 · engine remote docs

• chat sign-in cancel, picture names, speech spend, palette lists and team wording are accurate — #1702 · chat engine docs

• a carried-on run counts only its own files, and a published branch is pushed, not merged — #1705 · engine chat docs

• a terminal resize leaves no text from the previous frame — #1709 · chat docs

• a stopped run names the work its worker already committed — #1710 · chat engine docs

• settings says checker, no zero durations or byte counts, and help and manual match the screen — #1711 · chat docs

• skills context stays out of messages, rewind, reopened chats and titles — #1713 · chat

• A shell senior-dev run with flags after the brief runs instead of failing to parse — #1717 · engine

v0.5.1 — 2026-10-01

Fixed

• the first launch of the day no longer waits on the model list before it draws — #1724 · chat engine

v0.5.0 — 2026-09-30

Added

• custom connections are named, sit beside each other, and switch from the Providers tab — #1089 · chat engine

• Skill facts gain trust, cost card and digest; briefs gain an ordered skills list — #1327 · engine resident chat

• a ChatGPT plan connects as the Codex model service, in the browser or with codeaf connect — #1336 · chat engine docs

• /skill puts skills in front of a conversation by hand, with a chip that says which ones are on — #1346 · chat

• the chat shows which skills each message carried — #1348 · chat

• while a run is live the chat's turn opens on the run's rows, so a change of mind reaches the work — #1357 · chat engine docs

• Chat and task pages share ten animated working-logo studies — #1384 · chat docs

• Report total session token usage in anonymous telemetry — #1387 · chat build docs

• Clicking a Home project selects it for the next message — #1422 · chat docs

• the places bar has a chats word that goes back to the conversation — #1429 · chat docs

• a team can have a manager chat, with a Traffic rail of who told whom — #1429 · chat engine docs

• a handle on the Traffic opens the member at the message — #1429 · chat docs

• staging promotion checks the selected build, and stageaf installs it beside codeaf — #1438 · chat build docs

• codeaf reads the OpenRouter balance and starts a near-zero account on free models — #1440 · chat

• senior-dev, a coding agent built into codeaf, takes a whole task in the folder itself — #1488 · chat engine docs

• a Traffic row shows how long ago it happened, and a press opens that message — #1500 · chat docs

• Run non-interactive shell commands with ! and see their output live — #1654 · chat engine remote docs

Changed

• one fzf-style fuzzy matcher ranks every quick-search picker, and the settings search takes spaces — #1089 · chat

• The message-box rules — #1334 · docs

• the node belt is the default again — #1340 · engine chat

• one message carries the skills its own words choose, and a person's attachments always ride — #1345 · chat

• the prompt says what a skill is, and a missed skill name says what the shelf holds — #1347 · chat

• the belt measurement and the design for the chat as manager over the plan store — #1354 · docs

• the worker harness is the default belt again, and a run has no worker bound — #1355 · engine chat

• a note on a task reaches that task's running worker, and the chat can read the notes back — #1356 · engine chat docs

• a conversation against one machine waits for it instead of giving up on it — #1358 · engine chat

• persistedCount says in its own doc that it destroys the difference between unset and zero — #1369 · engine

• the Slack thread limit is written down once, and the description a model reads interpolates it — #1371 · engine remote

• Home returns to double-space navigation and the ask-here choice — #1388 · chat docs

• one list of twenty-two tips, said two ways — home rotates them, a conversation ranks them — #1389 · chat docs

• Claude Code plugin and Codex skills reach the chat, memory off included — #1396 · chat engine remote

• codeaf always signs what it writes, and names only the bare model in Assisted-by — #1412 · chat engine resident docs

• a run's rows on the rail and its page look like every other task — #1420 · chat

• the places bar sits where the chat strip does — #1429 · chat docs

• the tab strip reads Home, the team chip, the manager, then the tabs — #1429 · chat docs

• team Traffic is threaded, the newest thread at the top, and a manager's question keeps its answers — #1429 · chat engine docs

• engine --status, the older engine gives up the slot, held chats always move, --max-hours ends it — #1430 · engine remote chat

• a task's crew is routed per task, and /crew is the panel of what is allowed and pinned — #1436 · engine chat docs

• Every task opens the one task room, on either engine — #1485 · chat docs

• a conversation's tip covers the project on the keys row, with a cross, after 15 quiet seconds — #1487 · chat docs

• one side column holds Tasks and Traffic in every chat — #1494 · chat docs

• a standing card names its kind, and each button says what it will do — #1494 · chat engine

• A hovered tab's full title sits on the blank row and ends under its tab — #1496 · chat docs

• tui3 head grows an air row under the top navigation line — #1509 · chat

• unify provider vocabulary, background-list connected provider models, and ease adding providers — #1513 · chat docs

• clarify when accepting a standing card repeats work — #1520 · docs

• chat keeps messages readable and folds operational work across every conversation view — #1607 · chat engine docs

• the installer walks a person through the first minute and leaves codeaf ready in the same terminal — #1635 · build docs

• compaction can end with a summary, refused requests recover, and a summary leaves a visible line — #1658 · chat engine remote

• senior-dev works in a private worktree copy that starts from where you are — #1674 · chat docs

Renamed

• Home names its standing-order panel standing — #1415 · chat docs

Fixed

• a compaction pass that found nothing to do no longer takes the conversation away — #1341 · chat engine

• a replay draws the record once, whatever was already on the screen — #1342 · chat

• a quiet reply from the only machine there is gets asked again, with a wait — #1343 · engine

• a run's row on the tasks place is the store's task, wears its state and opens its page — #1359 · chat engine

• the picture hand refuses a prompt that is not there — #1372 · engine chat

• skills that are switched off say so instead of looking like a feature that was never built — #1380 · chat engine

• /cost and every multi-row note keep their columns — the gutter moves every row, not just the first — #1390 · chat

• one machine is waited on, a pool is never one machine, and the suite lock is seen and freed — #1392 · engine chat build

• the chat's run digest leads with the live run, in the side list's words, on every message — #1393 · chat engine docs

• the task page's keys, its note and its tab mark say one thing, and three catalog warms are owned — #1395 · chat engine

• a Codex model carries its 272k window, and a Codex call row names Codex as its endpoint — #1402 · chat engine docs

• run workers honour machine limits, keep private records, and finish at their step cap — #1410 · chat engine

• run workers wait for their step to be recorded and their notes delivered before acting again — #1410 · engine docs

• a new task never picks up an earlier run, and every run ending is written on its record — #1411 · chat engine

• codeaf do on the run engine edits in place, commits nothing, and stops at a price — #1416 · engine chat

• tasks approved together are one run, and a run that could not start says so — #1417 · chat engine

• a large paste no longer freezes the box for twenty seconds — #1419 · chat

• a deliverable path on another machine no longer refuses the task — #1421 · engine

• The home tab keeps its spelling and position across views — #1424 · chat docs

• Closing a task removes it from the current Sessions results immediately — #1425 · chat docs

• Home conversations keep their working mark while tasks run — #1426 · chat docs

• with the manager in front the right column is the traffic — #1429 · chat docs

• Conversation drafts and attachments stay with their conversation — #1431 · chat docs

• a team's wrap-up survives a restart, and a cap packet is raised once — #1494 · engine

• worker commits carry attribution when a run ends — #1495 · chat engine

• the /crew panel's weak-checker warning counts a pinned checker — #1501 · chat docs

• a connect offer shows its sign-in reason together with the waiting state — #1502 · engine

• a reminder's time reads with a middle dot, and Traffic ages stay short past a month — #1503 · chat

• a message that carries skills shows your words, not the skills' description sheet — #1504 · chat

• enter takes the answer the pointer stands on, on a standing card — #1506 · chat

• one wake per member reply, team_stop reaches background members, a demoted manager loses its verbs — #1514 · chat engine remote

• a task's worker can stop its own job, and a run's checker checks its own copy — #1515 · chat engine

• wall tiles say what each conversation spent, and the teams surface stops saying untrue things — #1516 · chat engine

• runs land every file by its real name, --one-model reaches every run seat, caches stay out — #1517 · chat engine

• crew follow-ups — checker ceiling by seat, route pins beside the tier row, landing line at the end — #1518 · chat engine docs

• a devaf first run reaches its own binary, shows the usage notice on screen, prints no raw lines — #1519 · build chat engine

• senior-dev works in plain folders and resolves shell model names — #1530 · chat

• telemetry --help exits 0, the manual names six --help groups, a wall tile shows its first age — #1533 · chat docs

• a side row hides its clock while its page is open; senior-dev manual matches its row and card — #1534 · chat docs

• Home search's card follows the pointer and returns to the cursor's match — #1541 · chat docs

• a spend under a hundredth of a cent reads <$0.0001 on every receipt, never $0.0000 — #1542 · chat docs

• task proposal countdown properly handles keypress and avoids starting declined task — #1547 · chat

• standing execution enforces git branch isolation and worktree validation — #1560 · engine

• plandb done refuses terminal tasks and manager tasks derive project workspace — #1562 · engine resident

• Slash command completion executes bare /task, /redo, and /workspace immediately — #1563 · chat

• interactive chat turns and task crew delegations enforce the daily spend rail — #1595 · chat engine

• standing order project scoping normalization, live rail allowance words, and task activity labels — #1596 · chat engine

• manager conversation initialization, handle fallback, and approval posture inheritance — #1597 · chat engine

• plan-born task worker briefs carry the project's standing orders — #1598 · chat engine

• bound foreground bash spill files and preserve their initial output — #1603 · chat engine

• bound background logs and keep recursive searches out of runtime output — #1603 · chat engine

• blank terminal briefs stop before planning or model work — #1604 · engine docs

• Single-model chat defers independent Model Pool judging — #1610 · chat

• one-time standing approvals preserve the action and report approval truthfully — #1612 · chat engine

• Session close joins its routing-cache refresh — #1616 · engine

• Held tasks stop cleanly and recover their exact working folder — #1619 · engine chat

• First standing approval preserves the shared timer's owner — #1622 · engine

• paid reasoning-ceiling retries remain in session and budget accounting — #1625 · chat

• CLI help names task model choices and unattended chat budgets — #1626 · docs

• first-run judge fixture joins its usage writer before home cleanup — #1627 · chat

• hedge fixtures hold the primary before its first visible word — #1627 · engine

• Opening a task room gives its note box the keyboard — #1628 · chat docs

• Home tasks keep the project displayed before submission — #1629 · chat

• A reopened hosted conversation draws a completed reply once — #1632 · chat engine

• Home tasks, standing work, and traffic keep their context and navigation — #1632 · chat engine

• a check naming issue_.go is sent back, and a file nothing makes starts no paid fix — #1634 · engine docs

• the manual's standing, pool, task-card, team-start and first-run pages say what the code does — #1639 · chat docs

• a task that lands mid-answer is said after it, and Done draws every task its heading counts — #1640 · chat docs

• a stopped background job reads as ended at once, and its log-folder cleanup runs just after it — #1641 · chat engine

• Tab selects the highlighted slash command on Home and in conversations — #1642 · chat docs

• Completed task forks retire their Furrow timelines before their directories — #1651 · engine

• Conversation names skip shell commands and task-only chats remain resumable — #1667 · chat engine docs

• Budget refusals name the limit that stopped the conversation — #1671 · chat engine docs

• a closed conversation has no memory pass still reading or writing its brain — #1673 · engine

• a stopped answer can still be opened after reopening its conversation — #1681 · chat docs

• do --json answers an early refusal, no unasked merge after senior-dev, manual truth — #1682 · chat engine docs build

Removed

• the installer says nothing about telemetry — #1488 · build docs

• Search lives on Home instead of a separate page and slash command — #1650 · chat docs

v0.4.1 — 2026-09-22

Fixed

• a waiting message stays waiting when you leave the chat, and still goes when its answer ends — #1385 · chat

v0.4.0 — 2026-09-22

Added

• the worker harness — the bash belt, the plan store on SQLite, the shell doors — #1109 · engine chat build

• the worker harness — pause, person notes, finish gate, archive, spend, crew arms — #1109 · engine build

• a run nothing is driving can be carried on, in the copy its work is already in — #1305 · engine

• a dev build installs beside codeaf as devaf and follows the dev channel — #1333 · build chat docs

Changed

• UX changes — consistent message boxes, Home sessions and task trees — #1071 · chat engine remote

• The model picker is a table with headed columns, not a ragged tail of facts — #1107 · chat

• the worker harness replaces the task engine under /task and do — #1109 · docs engine chat

• A task is one worker and a plan store; the older node engine is deleted — #1109 · engine chat build

• the worker harness — the work seat, the belt's pages, parent wakes, the run's store — #1109 · engine chat remote

• Decision dialogs support updated requests and concurrent clarification — #1146 · chat engine remote docs

• the curated lexer set loads on first use, not all 279 at init — #1193 · chat

• the task rail and the task page draw a belt run's plan as a tree, and the work tab is designed — #1204 · chat engine

• a belt landing no longer wakes the chat, the conversation gains a work tab, three hand-off rules — #1205 · chat engine

• the installer prints a receipt, a three-line telemetry notice, and the bare export PATH line last — #1208 · build docs

• chat answers number their steps, say where a job stands, and end on one next action — #1209 · chat

• a proposal learns every problem in one refusal, and a task stands where its brief plainly says — #1214 · chat engine

• The telemetry off switch quiets the Model Pool too, and codeaf telemetry show prints both streams — #1214 · chat engine docs

• a run's step rows show the work and leave out the run's own bookkeeping — #1257 · chat engine

• quick-search rows draw the matched letters in bold, and internal/fuzzy says where each term hit — #1321 · chat

• one fzf-style fuzzy matcher ranks every quick-search picker, and the settings search takes spaces — #1321 · chat

• the commit co-author links to the CodeAF account, and assisted-by names the model — #1325 · engine docs

Fixed

• a belt check runs on the crew checking seat, not the thinking seat — #1207 · chat

• a run's result is read from the store when the root's worker returns after the run ended — #1210 · engine

• a landing's report counts as evidence, and a check refused for its cd says the form that passes — #1212 · chat engine

• a reopened conversation keeps its tasks, and a task file that cannot be read is never overwritten — #1213 · chat engine

• a check is one command as the shell reads its quotes, and its refusal says what passes — #1215 · chat engine

• a review conclusion is judged by the commands the worker declared and the checker ran — #1220 · chat

• the conversation's tasks tool reads a run's tasks, by the names the rail shows — #1222 · chat engine

• unread profile config keys are reported once at load — #1223 · engine

• a resident dispatch pass does not run again the leaf it just released — #1225 · engine

• a place sweep note can no longer write after the process it belongs to closes — #1227 · chat

• a hosted conversation reads and steers its run — #1230 · chat engine remote

• a task's page opens with the task, and a run's rows on the rail open their pages — #1231 · chat

• a run works in a copy of its own, and its work comes home when it ends — #1232 · chat engine

• a self-finished codeaf do run seats its review before it ends — #1233 · engine

• every run a conversation has made stays readable — #1234 · chat engine

• a wait whose stream has stopped arriving paints at the spinner cadence, not thirty frames a second — #1235 · chat

• a reconnect reuses its registered identity instead of introducing itself again — #1236 · chat

• chat names unread config keys once before using default models — #1239 · chat engine

• a task's page says only what is true of it — #1240 · chat

• what is typed after a press on a run's row goes to its page, never the conversation — #1244 · chat

• unread-config notice skips keys the product itself retired — #1245 · engine chat

• stop it ends a run, from the card, from the run's own page, and from the chat — #1250 · chat engine

• a press on a run's row opens its page at once, and an open page reads on a beat — #1251 · chat

• drawing the side list never waits on the engine for a run's rows — #1255 · chat

• a run or a part you stopped reads stopped on its page, not incomplete — #1258 · chat engine

• a call that did not run is no longer drawn as a step that ran — #1260 · chat engine

• a task's step numbers continue across a wake — #1261 · chat engine

• a time limit a person sets now ends a run, by the same road as the cost limit — #1262 · chat engine

• the dim line under a step row is only drawn when it must be the row's own — #1265 · chat engine

• a worker that has stopped making progress is told once, and then ends, well short of the step cap — #1266 · engine chat

• closing chat now stops and joins the pool judge sweep — #1267 · chat

• a run's dollar limit now holds while a worker is still working — #1268 · chat engine

• a parked task is finished only by a worker woken for it — #1271 · engine

• a headless launch and task-owner open reach a restarting host, not a crashed one — #1272 · chat engine

• a backslash pair inside double quotes is text to the one-command reader — #1273 · engine chat

• closing a process cancels and joins lazy catalog warming before a late cache write — #1274 · engine

• no run answers done while a finished piece of work has had no review — #1275 · engine chat

• closing chat stops the place sweep before it can change the home again — #1276 · chat engine

• an ending names the limit that caused it — #1279 · engine chat

• a run's dollars reach the conversation's own total — #1280 · engine chat

• a run is handed what is left of every dollar limit — #1281 · engine chat

• a cancelled command is ended by the call that ran it — #1282 · engine chat

• a long declared check says why it was refused — #1283 · chat engine

• the heavy-suite lock is held for the suite, not for what the suite leaves behind — #1284 · build

• a usage flush says whether it drained or ran out of time — #1285 · engine

• a row that needs you names what it needed and opens the item — #1288 · chat

• a task's page draws its brief through the reader the transcript already uses — #1289 · chat

• a standing firing runs behind the gate a task node runs behind — #1290 · engine

• a live run counts as work, so its conversation is not retired out from under it — #1291 · chat engine

• a session stopped on a sub-harness question says so instead of working — #1293 · chat engine

• work nothing is driving comes back interrupted, not failed and stopped — #1296 · chat engine

• a run's working copy is written down, so carrying it on adopts the work instead of deleting it — #1298 · chat engine

• the interrupted page no longer offers two answers nobody can press — #1304 · chat

• a watch stopped for permission says the line, and no longer says a person asked — #1314 · chat

• a sign-in and a harness ask say what they are waiting for, instead of only that somebody is needed — #1315 · engine chat

• the front tab and the tab beside it agree about whether one conversation needs a person — #1316 · chat

• the caret parks in the box the person is typing into, and hides on surfaces with no box — #1321 · chat

• a landing whose accept is already in flight stops asking the person to accept it — #1322 · engine chat

• a heavy suite takes both the file lock and the older directory lock, so a checkout behind — #1324 · build

• the drafted-with footer links to the codeaf repository — #1329 · engine docs

• a landing that turned the work back offers its answer without standing in needs you — #1331 · engine chat

• a changed directory in a nested module is not a package of the root module — #1332 · build

• hosted chat stops the live run it can see — #1376 · chat engine remote

v0.3.0 — 2026-09-18

Added

• A model family toggle under the crew, so a preset can answer from open weights or from everything — #1073 · chat engine

• custom connections are named, sit beside each other, and switch from the Providers tab — #1089 · chat engine

• The crew's seats are picked from a row — table, catalog or learn — #1093 · chat engine

• Three pure packages — a crew picker, an additive tally and a measurement-index reader — #1093 · engine

• The pool client gains the relay's real addresses, an install nonce, and the index signer's key — #1093 · engine

• Three more internal/pool packages — an outbox, a settings resolver and a signed-document fetcher — #1093 · engine

• a landed task is scored by a model outside the crew into the own sheet, outbox and judge seat — #1093 · chat engine

• A pure package that scores a finished run seat by seat, by a judge outside the crew — #1093 · engine

• A pure package that records a run's judged scores into an own sheet and outbox rows — #1093 · engine

• A Cloudflare Worker relay that folds installs' rows into a signed, versioned pool index — #1093 · engine

• The Model Pool's index is seated at start-up from a seed in the binary — #1093 · chat engine

• A tier row that says auto computes its model from the catalog — #1093 · chat engine

• The Model Pool reaches its first surface — a model_pool setting and a codeaf pool verb — #1102 · chat engine

• a headless exec run leaves a pending judge record at its tail for the pool's sweep — #1137 · engine

• the headless task door leaves a pending judge record at its tail — #1138 · engine

• a codeaf do errand leaves a pending landing for the Model Pool's judge — #1139 · engine

• pool status says what is waiting for a judge and what the last sweep did — #1142 · engine

• The seed index is regenerated from the relay by one in-repo command — #1148 · chat engine

• pool show --cells lists the held index's cells with their installs and dims — #1149 · engine

• a chat surface now runs under a soft memory limit drawn from the machine — #1158 · chat

• the relay can purge fixture vendors and refuse them at the door — #1167 · engine

• pool status says the install identity, the judged total, and that nothing has landed — #1169 · chat

• A non-verified codeaf do run names its kept branch and its verdict — #1182 · engine

• The tasks text and the record line name a failed node's kept branch and verdict — #1184 · engine

• A non-verified codeaf exec or codeaf run run names its kept branch and its verdict — #1186 · engine

• a job subtree is cut when it passes its CPU or process bound — #1190 · chat engine

Changed

• the coverage round buys once, and the envelope says when the requested work was done — #1093 · resident docs

• The crew's default family is all, and both families' worker, careful and mastermind seats moved — #1093 · chat engine

• Six cells of the fixed crew tables moved to the measured seats — #1093 · chat engine

• The fuel meter asks an installed tariff first, and its table holds the ids this build ships — #1093 · engine

• A re-dispatch after running out is granted more room, and one that banks nothing new is refused — #1093 · resident docs

• Every usage row names the seat the call ran under — #1093 · engine

• The crew picker blends a measured pool rating into a seat's quality by its observation count — #1101 · engine

• The README telemetry section is one collapsed block at the end of Docs — #1115 · docs

• a launch compiles 78 fewer patterns, and a surface holds 8 Ps instead of every core — #1125 · chat engine build

• the relay carries the acceptable metric beside the judge's opinion — #1129 · build

• the relay writes a submit batch once per distinct cell and day — #1136 · engine

• pool show and verify say both metrics the index carries — #1141 · engine

• the seed index follows the relay's first clean publish — #1180 · chat engine

• a settle turn runs under the checker's own bound, not the run's wall — #1183 · chat

• the status word says working while a task subtree still turns at rest — #1191 · chat

Renamed

• the /connect row for a custom service reads Custom OpenAI-compatible API, not Something else — #1107 · chat docs

Fixed

• a finishing task can read the evidence its context has archived — #808 · chat engine

• a question's own answer box answers to the whole caret vocabulary, not a hand-rolled copy of it — #1032 · chat docs

• an answered landing carries its answer's fate, and an in-flight settle holds the question down — #1086 · chat engine

• the headless doors wait for the catalog rows a pick reads — #1093 · engine

• No git command runs without a directory, and the checkout guard reads a linked worktree's own share — #1093 · engine

• A node retry asks the router to avoid the upstream machine that just failed it — #1094 · engine

• Usage counts from a hosted chat carry real numbers, and nothing without a version ever leaves — #1111 · chat docs

• a service connected mid-conversation reaches the conversation that is open — #1114 · chat engine

• a decision card under --yolo takes its own default instead of waiting for somebody — #1117 · chat

• a design's thread stops its parts before it closes, and a node machinery cut says what cut it — #1118 · chat engine

• a bare auto row names the pick word's rung, so a seat computed under learn says learned — #1119 · chat engine

• a bare auto row under picked from = catalog reads the catalog's figures alone — #1120 · chat engine

• the do door seats its worker from the settings the run will use — #1122 · engine

• a launch's pool writers stop when the process closes — #1124 · chat

• The pool index's min_installs floor counts a cell's installs, not its rows — #1127 · engine

• The pool relay folds a retried batch once — #1128 · engine

• A row the relay refuses by line leaves the outbox and the rows around it are sent — #1130 · engine

• The published pool index holds only the cells that meet its min_installs floor — #1131 · engine

• A run a process death left unjudged is judged on the next chat start — #1132 · engine

• the outbox file is compacted back to the rows it still holds — #1133 · engine

• The finished-tree check finds a project cloned into a subdirectory — #1135 · engine

• the manual said the status row's ledger stays the conversation's; it is the tree's — #1143 · docs chat

• The wall refuses a first replan round it cannot hold — #1144 · resident

• an unstamped binary says it was built outside a git checkout, not without make build — #1151 · build docs

• the terminal audit names the gitignored build products a run left in the deliverable tree — #1152 · chat engine

• status says when the relay dropped a row and why — #1154 · engine

• a job teardown never signals a process group whose pid was reused — #1155 · chat engine resident

• an exhausted leaf continues the plan it has instead of a second full planning pass — #1156 · resident chat

• pool status's index line says what this run's own fetch did to the cache — #1161 · chat

• the soft memory limit now takes the cgroup bound, not just physical memory — #1162 · chat

• the pool spells a seat by its bare model id — #1163 · engine

• The relay bounds judge severity and publishes mean and sd on one scale — #1165 · engine

• a job's shell cannot reach the tmux server hosting codeaf — #1166 · engine

• An unattended run cannot push or move a branch it did not create — #1168 · engine

• an exec run's worker spend reaches the usage ledger — #1170 · engine

• the task door honors a stored crew word, and keeps its pins under the learned pick — #1171 · chat engine

• a landing card under --yolo takes its default instead of parking the run — #1174 · chat

• the model warm stops when the process closes — #1179 · chat engine

• the auditor's calls say they are the auditor's and which node they check — #1181 · engine

• a parked worker is handed back with a record at a third of its allowance — #1187 · engine

• the unconnected custom row on /connect reads Custom OpenAI-compatible API again, not custom — #1195 · chat

v0.2.1 — 2026-09-17

Added

• Every release carries THIRD-PARTY-NOTICES.md beside its binaries — #1051 · build docs

• a launch says when a newer codeaf is out, and /update installs it and restarts — #1060 · chat build

• Anonymous usage counts, on by default with a notice printed once — #1095 · docs

Changed

• The switcher's foot is one instruction and its name column takes the spare width — #1040 · chat docs

• The switcher card lists conversations in the tab row's order, not by recency — #1040 · chat docs

• A closed tab leaves the switcher's list at the same moment it leaves the row — #1040 · chat docs

• ctrl+k kills to the end of the line, and the conversation switcher moves to alt+k — #1040 · chat

• the spend place is three tables in four columns each — no bars, one right edge, money in cents — #1043 · chat

• A home panel stands where its content puts it, and the quiet ones gather in a rail on the right — #1046 · chat docs

• The README is the launch story and the reference it used to be is docs/GUIDE.md — #1049 · docs

• The install story, the compiled-in manual and the branch rules say what is true on release day — #1050 · docs build

• The thinking wheel comes back to auto, so the chip can reach the state a fresh install is in — #1057 · chat

• ctrl+c leaves on the press that lands — #1061 · chat docs

• Workspace documentation uses accessible references — #1064 · docs

• a wrong completion check is answered with [no change], and the answer before it stays in view — #1067 · chat engine docs

• the drafted-with footer names CodeAF and still links to agentfield.ai — #1075 · resident docs

• The manual and the guide say the install proxy is live, because it is — #1076 · docs chat

• task commits are authored by the agentfield-bot account; the old local address is legacy — #1088 · chat engine

• The wordmark is drawn CodeAF, four rows tall, with the capitals on the d's cap line — #1106 · chat

Fixed

• The switcher's close key says close tab, because put away is home's word for archiving — #1040 · chat docs

• The switcher's fold says where the tabs stop, and every refusing row says why — #1040 · chat docs

• Taking a conversation from the switcher leaves the place you took it from — #1040 · chat docs

• An empty target is nothing to open, and never a window on the working directory — #1042 · chat

• the installer takes the newest channel build, not the first one GitHub lists — #1056 · build

• The folder walk's budget bounds every entry, and the suite never walks your home — #1058 · chat

• "ask before running" reaches the gate you are already behind, and /status names it — #1063 · chat

• a job root working in the person's directory is offered its overflow file outside that tree — #1096 · engine resident

• The README shows how to pin a release version again — #1103 · docs

• a release download finishes over a real network, on its own clock — #1110 · chat build

v0.2.0 — 2026-09-15

Added

• The sizing reading and the handover's brief say what they are while you wait — #101 · engine chat

• ctrl+k is the switcher, and the task page folds a family under its root — #105 · chat

• a conversation can be about places beyond the one it opened in, and work there lands as a diff — #109 · chat engine build

• sent pictures draw in the transcript, and a drop converts however the terminal spells it — #115 · chat

• a foreground command backgrounds itself at a threshold, and the running turn teaches one hint line — #134 · chat engine

• settled tasks grade the models, and the divider reads that record before it assigns a tier — #155 · chat engine

• missing OpenRouter profiles connect in the browser — #206 · chat

• a defect issue now owes a replication a stranger can run and an acceptance that starts end to end — #209 · docs

• an ffmpeg library and edit_video — a join that cannot come out silent — #254 · chat engine docs

• a real-model lane that proves a family's parts work apart and land, on a repository and a folder — #280 · engine

• a person can open aforge's own manual — aforge manual and /manual, with no key and no spend — #297 · chat docs

• a real-model end-to-end lane proving the manual is reachable through the live chat surface — #309 · chat engine

• aforge logs shows the whole row, both token counts, five filters, JSON and a body — #359 · build chat

• one switch keeps the full record of a run, in a folder of its own that nobody else can read — #361 · chat engine docs

• a task brief names the journal path and line of the person's original words — #475 · engine chat

• /status --json prints the status note as one JSON object — #484 · chat

• continue re-arms the same failed task instead of proposing a new one — #622 · chat engine docs

• ctrl+w closes the tab in front, the way every browser closes one — #653 · chat

• long-running compact steps show quiet elapsed time — #653 · chat docs

• Open a New chat start page from plus without starting work — #653 · chat

• ctrl+shift+t reopens the tab you just closed — #653 · chat

• each compact step carries one still action icon in a fixed gutter — #653 · chat engine docs

• every chat tab holds its own engine connection, so opening one stops none of the others — #657 · chat engine remote

• Organize existing records through independent logical collections — #661 · engine chat docs

• a running turn now says what went up and what came back, in tokens, at the right edge of its work — #666 · chat

• The message box selects text with the pointer, and ctrl+z takes back what you typed — #669 · chat docs

• one object for every decision the engine hands a person, and one door every answer goes through — #712 · engine

• The model can ask through one shared question door — #728 · chat engine

• A question opens into a page you can read, compare, annotate and ask back at — #732 · chat engine

• questions reach home, another page, a sheet and an unattended window as one decision — #733 · chat engine

• one block above the box for every question, with esc meaning later rather than no — #734 · chat

• A pinned Back to main row returns directly from nested tasks — #771 · chat

• ctrl+r in /model fetches today's model list, and aforge models --refresh does the same from a script — #787 · chat

• a second place models come from, connected from /connect — #800 · chat engine docs

• The terminal's own tab says where you are in aforge, in Terminal.app and iTerm2 too — #809 · chat docs

• A quick task does the small job where you are — no copy of the folder, no branch, no check — #811 · chat engine docs

• A small model gets a smaller prefix, derived from its own window and the crew's worker seat — #812 · chat engine

• the chat signs the git work it does for you, in the resident's own words and bytes — #812 · chat engine resident

• a service's subscription door is tried before its metered one — #861 · chat engine docs

• Every call can be watched — one seam for a call's whole life, fed from the readings it already has — #868 · engine

• The tasks place draws the cursor row's record beside the list, and answers it there — #902 · chat

• home headings carry a dim explainer beside the word — #907 · chat

• a chooser change is judged by regret on ten days of log, not by a screenshot — #922 · engine docs

• an answer is a message to the conversation, so the work carries on while you decide — #954 · engine chat remote

• a launch sweep names live sessions on an older build — #1015 · chat engine

• /drafts, the page over the kill ring — #1018 · chat

• a fourth routing mode, simple, sends exactly what you asked for and nothing else — #1019 · engine chat docs

• a picture's step row names the image model, and opening it shows the prompt — #1026 · chat

Changed

• dev is the trunk, staging is a pointer, and nothing publishes by itself — #82 · build docs

• money has one editor — the Spending tab — and every door lands on it — #110 · chat engine docs

• web search is Firecrawl by default, keyless, on both surfaces — #113 · chat engine

• the forming block shows the brief being written, and several tasks forming share one block — #125 · chat engine

• the reflex tier ships on mistralai/mistral-nemo instead of nex-agi/nex-n2-mini — #140 · chat engine

• a quiet conversation left alone is let go of, so a window's memory can come back — #151 · chat

• the write-your-notes rule is held rather than advised, and a turn that will not comply stops — #154 · chat engine

• a landing's claims are its checklist, and a divider is checked on its parts' work too — #156 · engine

• the leash judges what a task produced, not what it was called, and it has to open the tree — #159 · engine

• a task starts from its parent's folder as it stands, and its brief is checked against it first — #162 · engine chat

• both doors that write a brief carry the handoff manifest, and the prefix came out lighter for it — #179 · engine chat

• a worker stopped by the write-your-notes rule lands under its own words — #182 · chat engine

• a repository task inherits the whole world, its .env included, and still comes home as a merge — #183 · engine chat

• home's detail card is four groups with one gap law, and it is 56 cells wide — #219 · chat

• no two parts of one division may own the same path, and it is refused at admission — #236 · engine

• every part of a division opens on the same world, whoever wrote the parts — #240 · engine

• a fold marker names the journal path the model can grep, not a store id — #247 · engine chat

• the deliverable-file path law is said once, in system.md — #248 · engine

• nobody waits long — one hazard controller on every call, and a wait that says why — #253 · chat engine docs

• a parent commits its work to the family branch before it hands out the parts — #262 · engine

• web search settings land on the next search, and every search names its back end — #263 · chat engine

• the crew has a worker seat, and its presets are open-weight models picked off the catalog's scores — #278 · chat engine docs

• a proposed task waits while you type, a bare no means no, and the default window is fifteen seconds — #279 · chat engine remote

• the task engine has a ceiling on how many endings one function may have, and it only turns one way — #284 · engine

• No call sets a temperature — every request leaves sampling to the provider — #308 · engine chat

• the manual's wire lane prints what each pass spent, summed off the ledger as it asks — #331 · engine

• The reply guard reads a kilobyte, watches the thinking, and a reply you stopped mid-soup is not kept — #332 · engine chat

• a task's page runs the conversation's reducer, and learns durations, retries and notices — #337 · chat

• a task page reads a correction as a correction — steers are elbows and the engine reads the record — #351 · chat engine

• a task's page folds its settled work into phase chips, and ctrl+e, a click or a scroll up opens them — #355 · chat

• a model that has been watched thinking is waited against what it really does — #365 · engine docs

• the proposal hold of — #390 · engine chat remote

• the split gate is off by default, and the count and the plan's sizing stay as pins — #435 · engine

• the chat manual can answer how to run aforge with nobody watching — #436 · chat docs

• background jobs are their own section, named, paged, and stoppable — #439 · chat

• a person's manual answer heads each section with a line no page can contain — #446 · chat docs

• the chat manual stops saying a nested question is hidden, and names the folder that would not freeze — #447 · chat docs

• the money pages count the rows /cost draws, and an order is stopped in the words people ask in — #451 · chat docs

• the lanes page answers the floating alias, the headless run, and the third routing word — #453 · chat docs

• the manual's stemmer meets itself by Porter's measure, and "much" is a stop word — #458 · chat engine

• the surface paces the wire's lumps, never its own fold — #473 · chat

• the touched-packages job no longer runs on pull requests — #499 · build

• Check-refused tasks now say incomplete and keep their reason — #512 · chat docs

• the finished-work heading carries the asker's words, so the plain floor keeps its slack — #516 · chat docs

• what every surface says, and what it does when it does not fit — #518 · chat build docs

• the touched packages run on every pull request again, and check is green only when they are — #557 · build docs

• caption sentences now head work, and ctrl+e opens their outline before machinery — #563 · chat engine

• a task never writes a protected branch, and repository work is always on a branch of its own — #589 · engine chat

• a body-on call log keeps a long debug session instead of rotating at 32 MB — #622 · engine docs

• Named chat breadcrumbs and readable expanded task answers — #653 · chat

• live conversation work rolls through three compact step descriptions — #653 · chat docs

• Keep streaming prose compact until the response confirms an answer — #653 · chat engine

• Conversation work keeps attributable direction, independent views and retrievable results — #653 · chat engine docs

• A dedicated tab strip over the chat, with the task trail on its own row beneath — #653 · chat

• Name conversations promptly with distinct full and tab titles — #653 · chat engine remote docs

• Give conversation navigation a padded header and a Home button — #653 · chat

• Separate chat navigation from task metadata and make tab dismissal explicit — #653 · chat

• keep useful step descriptions visible between actions — #653 · chat docs

• Ctrl+t starts a new chat and the task roster moved to alt+t — #653 · chat

• Leave default reasoning to the model — #653 · engine chat

• conversations retain their work, context and delivery across navigation — #653 · chat engine remote docs

• Keep running task work compact and expandable — #653 · chat

• Keep many conversation tabs readable with a scrollable strip — #653 · chat

• Separate the conversation switcher from the page beneath it — #653 · chat

• one browser chooses folders and files, with a real preview beside them — #657 · chat engine remote

• Work starts from the original request and checking keeps recoverable evidence — #660 · chat engine

• Small edits across several files can finish in the conversation — #664 · chat engine

• Model generation uses provider defaults until you choose an override — #665 · chat engine resident docs

• a task row answers one question in one word, and no decision stays with the model past a turn — #668 · engine chat

• a task tries the check on another model, one merge round and one rerun before it asks you — #673 · engine

• the chat manual learns the three tiers, and stops teaching five vocabularies for one state — #679 · chat docs

• first-run setup groups models and spending before the first conversation — #680 · chat

• the landing card asks one question, in one word, with two answers — #689 · chat

• every row of work but the landing card draws one tier, one glyph and one word — #694 · chat

• A your-call landing opens its report with the question its row is asking — #696 · engine chat docs

• the chat pages follow the rail and the roster, and the ! leaves every row of work — #709 · chat docs

• Task pages group scoped model and Stop controls below the task tree — #710 · chat engine docs

• every mark a person sees comes from one vocabulary, in the tier their terminal can draw — #711 · chat docs

• a task whose files moved under it asks you to resolve it, instead of saying nobody could check it — #714 · engine chat

• images start collapsed and share explicit preview and original-file controls — #744 · chat remote docs

• The box at home is a draft, and the rule above it says which conversation — #747 · chat

• The thinking rung is on the seam beside the model, pressable, and it crosses --host — #747 · chat

• The conversation's name and model are on the seam, and the status row is a ledger — #747 · chat

• the icon law reaches every surface package, not just the chat — #753 · chat resident docs

• the approval question is the question block now, answered by digits, with esc meaning later — #758 · chat

• the stop card and the close-a-tab card are the question block now, answered by a cursor and enter — #758 · chat

• moving a conversation out of another window asks with a card, and home can ask at all — #758 · chat

• the task proposal is the question block now, answered by digits, and its card is the assignment — #758 · chat

• tasks can search and read earlier conversations without filesystem discovery — #769 · chat engine docs

• the connect offer asks on the question block and its key is typed in the message box — #780 · chat engine

• the harness offer and a finished design ask on the question block, and five old keys are gone — #780 · chat engine

• the standing card asks on the question block, and esc on it now means later rather than no — #780 · chat engine

• Home is seven panels, each answering one question, under a four-place bar — #783 · chat docs

• → walks into a model's lanes, a fold always opens, and a pinned lane rides the model's name — #785 · chat docs

• A wide read is quick tasks, a wide change is one task, and independent pieces start in one breath — #811 · chat engine docs

• A harness message says how to read itself, so the page stopped explaining messages — #812 · chat

• The page keeps existence, the load reply keeps the prose — long mechanics are pulled, not pushed — #812 · chat engine

• A tool description is a contract, and a schema law fails the build when one turns back into prose — #812 · chat engine

• A tool result is cut to a tenth of the model's window, and each description quotes its own cap — #812 · chat engine

• Pushes publish dev, staging and rc builds, and a curl line installs any of them — #813 · build docs

• A settings row for the prompt profile, a smaller ask schema, a job footer said once per change — #844 · chat engine

• The token column shows what moves — ↑ is the request size, the moving side glows, tasks draw it — #849 · chat engine docs

• the call log says who answered and what a failure cost, and a committed census reads it — #852 · engine build docs

• every wait is spoken within a second, and a ceiling that fires always acts — #853 · chat engine docs

• One dispatcher, one plan — a failed call is bounded by one deadline — #858 · engine chat

• The rung is a move, the race reads the plan, and the person's patience is time — #864 · engine chat

• A warm wire and a socket that never waits — the reader reads, and a keystroke crosses — #872 · engine remote chat

• A task may hand out twenty pieces, tasks nest three deep, and the goal is the whole job's wall time — #874 · engine chat

• propose_task is a staged tool, so its card can go up before the reply that carries it is whole — #880 · chat engine

• a task's worker starts as soon as its folder exists, and memory and sizing are read beside it — #883 · engine chat

• home's needs you is two groups — the questions that stopped something, then what to check — #900 · chat

• The tasks list becomes a table — two columns, a sort key you can see, and every fold shut — #905 · chat

• A leaf's grant is written in one place, with the measurement saying it is too small beside it — #918 · engine chat docs

• The turn decides whether to become a task, and a task it starts inherits what it already read — #923 · chat engine

• A leaf's grant is 220,000, measured, with the band it sits in written down — #931 · engine chat docs

• A question is drawn one way, in one frame, with the amber on its marks alone — #933 · chat

• one call asks for one set of machines, whichever door its request left by — #934 · engine chat docs

• the chooser demands the set it admitted rather than suggesting it — #937 · engine chat docs

• With exactly one stoppable row on screen, x raises its card without alt+t first — #943 · chat

• questions from one step are one panel with tabs, and permissions from one batch one frame — #946 · chat docs

• /task starts at once, and its brief and its width are read beside the worker — #950 · chat engine remote docs

• the gate grades a grave call irreversible, and a permission's pointer opens by the stakes — #953 · engine chat

• An answer's evidence stands beside the answers, and the page a question opens is two panes — #972 · chat

• the controller-acted list moves onto the phase type, beside Waiting() — #982 · engine

• Home's box is three rows tall, typed in or not, and one command clears every engine — #1000 · chat engine docs

• auto lends the routing to OpenRouter and takes it back when its answers turn bad — #1003 · chat engine

• The turn fold keeps the newest slice of a repeatedly read file and points the older ones at it — #1015 · engine docs

• simple is the routing row this build ships with, and it is the whole algorithm — #1019 · engine chat docs

• the README sells the chat flow and says which install road is actually open — #1021 · chat docs

• the thing behind a model that answers you is called a provider, everywhere a person reads — #1027 · chat engine docs

• an unconnected account is reached through one question and the same turn carries on — #1037 · chat engine docs

Renamed

• the four pipeline verbs live under aforge plan, and aforge run <name> is the saved program — #518 · build docs

• a flag about money is not spelled like a flag about tokens, and a duration flag takes a duration — #518 · build docs

• The product, its binary, its module path, its state root and its variables are codeaf — #1041 · chat engine build docs

Fixed

• a declared door is something that can actually run — #92 · engine

• a hint on a failed tool result is a command with evidence behind it, or it is silence — #93 · engine docs

• a reply waiting on a job it started is waiting, and carrying on stops after three — #95 · chat engine docs

• work only a person can do goes to the person, not to a worker — #97 · engine

• a checkpoint that arrives after the close writes nothing — #99 · engine

• a watch that fires wakes the conversation, and only the firing does — #102 · chat engine

• a background job's page tails its log while it runs, and stops saying finished over work that is not — #106 · chat

• a background job cannot draw on the chat — #107 · engine

• the forming task card moves for the whole arrival — #114 · chat

• another window's running task says which life it is in, and a task may not read your login — #121 · chat engine docs

• A live stage keeps drawing, and the mark's reading is one of them — #122 · chat engine

• tool-grammar leaks are cut and rerouted, and /model says what it leaves running — #131 · chat engine

• enter on home's typed sentence opens a new conversation, and never sends it into the old one — #132 · chat

• a resumed conversation opens knowing what it has already spent — #135 · chat

• the money on the status line is the whole tree's, live, not the conversation's half of it — #153 · chat engine

• refusals name a door, the composer owns the letters, and the check reads the ground — #157 · chat engine

• resource policy is derived from what the process measured, not from constants — #158 · chat engine

• a picture dropped on home becomes a picture, on every box that can take one — #160 · chat

• a typed drop lands on whichever box has the keyboard, and a start in a typed folder carries the tray — #165 · chat

• a fork's hands and an adaptive run's nodes are on the machine's day total once, not twice — #180 · engine chat

• a context fill somebody set decides where a conversation folds; one nobody set still does not — #181 · chat engine docs

• the crew reaches the headless doors, and every run says which voice chose its models — #186 · engine docs

• two task nodes can no longer be handed one journal, which is what made five session tests flake — #187 · engine

• only work demotes an answer, and a turn's end settles the whole turn — #190 · chat

• the tmux TUI suite runs again — 9 of 9, with one table of words guarding every needle — #191 · chat build

• the sweep tells furrow to forget a swept session's forks — #197 · engine

• a task's card names where the work happened in plain words, from the rung that made it — #198 · chat engine

• four flaky tests now measure the claim rather than the machine they ran on or the hour — #199 · engine build

• a divided parent reads its two facts as one, and stops buying a turn on an empty request — #200 · engine

• a standing end that stands before the item's own first firing is refused, not stood up — #203 · chat engine docs

• a standing card's hint is built from the chips it drew, so it cannot name an absent answer — #204 · chat docs

• an answer asked from home renders as markdown, through the transcript's own renderer — #216 · chat

• the task tree's total goes with the conversation it is about — #217 · chat docs

• a tool call in the transcript is one row, and the row is measured in the bytes it draws — #218 · chat

• the move-it-here story is on the row and the card, and an idle window lets go in a quarter second — #220 · chat

• dragging over text selects characters, not rows — and a double-click takes the word — #224 · chat

• one owned decision about what the final answer is, and a turn's end settles it once — #234 · chat engine

• A task that hands parts out lands what the whole family wrote, not the parent's slice of it — #237 · engine chat

• a task family on a plain folder gets a family tree of its own, and its parts get worktrees off it — #239 · engine

• parts work in a copy of their own, not a copy of the repository — #251 · engine chat

• a file dragged from Windows into a WSL session becomes a picture, not a path in the message — #257 · chat

• a task working on a plain folder no longer writes over an edit you made while it ran — #270 · engine chat

• steering a task with nobody inside is refused with the reason instead of swallowed — #275 · engine chat

• a landing that cannot save the work keeps it and asks for your look, rather than deleting it — #277 · engine chat

• a growth governor may refuse a round, never a review's finding — #287 · engine resident

• a landing carries your uncommitted work or names the files, and a writing turn becomes a task — #288 · engine chat

• spending is written down per call, so an interrupted turn's money is on every surface — #290 · engine chat

• a leaf that ran out of its tokens is never settled on a judge's say-so — #291 · engine resident

• a landing that needs your look asks on every surface at any depth, and a timeout says who could not — #292 · engine chat

• a manual page read is bounded like a file read, and the rest of the page is asked for by heading — #295 · chat engine

• a page's own title is a search term, so a plain question reaches the page it names — #298 · chat docs

• generated manual packs are ignored again, and a build refuses to track them — #299 · build

• the shipped default keeps one ledger key instead of three, so it routes with priors — #300 · engine

• the manual quotes the code's own figures, and a moved figure names the page — #301 · chat docs

• a ground step owns every cell it covers — a sweep lights code spans and chips — #303 · chat

• a turn spent watching its own pieces is not a turn to hand over — #304 · engine

• two mints of one subharness at the same instant can no longer both win — #306 · engine

• the mint gate is taken under a deadline, so a mint refuses rather than waits — #310 · engine

• a crew older than a seat still fills it, and the run says which row it took — #311 · build engine docs

• looking something up in the manual searches your own words, not only the ones the model composed — #313 · chat engine

• a conversation seats a crew older than the work seat too, and says so once — #314 · chat engine docs

• the belief write leaves the send path, and no lock in internal/lane waits — #317 · chat engine

• an empty profile directory is the ordinary launch, so the crew draws and the notices remember — #320 · chat

• the ordinary launch is shown its front door, its rung and its own gate — #326 · chat

• a node that ran out with work recorded is offered again, and the sentence names the bound — #327 · engine resident docs

• a task on an attached repository seals, and work aforge starts is named before it is announced — #333 · engine chat

• a part of a division owns what its done-condition names, not what its brief mentions — #336 · engine

• AFORGE_DEBUG records the calls, the tools and the choices, not just a header — #338 · chat engine docs

• one model has one ledger identity, at every door and in the file already on disk — #353 · engine

• the headless doors fetch a lane sheet of their own — #354 · engine resident

• the tagged e2e lane runs on any machine, and the folder landing test is green on a Mac — #356 · build

• the model-call log never loses a row over a number JSON cannot spell — #357 · engine

• stopping is bounded at ten seconds, it says so on the way, and there is a door behind it — #358 · chat engine

• a failed headless run keeps its record, and every headless call names its tag and node — #363 · engine

• the background timer is a (home, program) pair, and a launch speaks only for its own — #366 · chat

• a machine that refuses a model is struck at once, dropped from the pin, and drawn as refused — #368 · chat engine

• a compile that comes back with no goal keeps the person's own words instead of ending the run — #369 · chat resident engine

• the standing package builds for windows again, through the lock and pid packages the tree has — #378 · build chat

• a headless run's total includes the plan model's reading of the request — #387 · engine resident

• a background reading over a connection lets go of its latch when it dies, and says so once — #396 · chat remote

• the checkpoint fixtures answer the namer by shape, so the early name spends no scripted step — #400 · engine

• a usage frame is folded into the call's count, never swapped in for it — #405 · engine

• the chat journal and the demo seeder post through the one thread door, and its sweep is green — #406 · chat build

• the standalone harness-design tool renders a brief again, and both doors read one machinery map — #410 · chat build

• the settings sheet's completeness is measured on the laid-out sheet, its calm on the window — #413 · build

• The goroutine sweep reads the defers, and every bare spawn in the guarded tree runs under guard.Go — #414 · chat engine build

• an observation is retired when it stops being needed, and a re-read of retired bytes gets the bytes — #415 · engine

• a file the request names is never scratch, so a leaf changing it is never at a standstill — #416 · resident

• a base is asked whether it publishes a lane sheet, never read for its hostname — #419 · engine chat resident

• the machine writes a device into its list before it says the pairing held — #420 · remote docs

• the background check walks its items on a machine that has no API key yet — #421 · chat build

• a brief that fails to write is retried, then composed, and the plan it belongs to is kept — #422 · chat engine

• an oversized node that cannot run in parallel is divided into ordered stages — #424 · engine

• the planner is told what a goal names against one worker's reach, and no pass hands over more — #425 · engine

• once nothing is changing the job stops spending — no gate, no repair, no resume — #430 · resident engine

• concurrent hedge arms no longer freeze while recording their call rows — #438 · engine

• quitting stops the task graph, so a node admitted a moment before never runs on unreachable — #448 · engine

• --one-model now reaches the mark reader and the brief writer, which had no model at all — #449 · chat engine

• every door that opens the chat surface parks the standard logger in chat.log — #450 · chat remote

• the headless ↻ line says why a claim went back on the queue, not only how much it picked up — #454 · engine resident docs

• a reading covers what the request names, and is retaken only when the tree changed — #460 · engine

• a constraint the person states is a law of the run, held at the gate and never advice — #461 · engine resident

• The system prompt names exactly the tools the agent reading it can call — #462 · chat engine

• the tui3 harness stops waiting on ticks that cannot reach it and names every command that parks — #463 · chat

• a run ends when the request is satisfied, not when the plan runs out — #464 · engine chat

• a settled task with green checks is not carried on as unfinished — #468 · engine chat

• isOpenRouter reads the alias marker off the model before the prefix check — #470 · engine

• a lane that attaches mid-job is handed the live job, not the checkpoint stub — #472 · chat

• A hand is told which of the tools on its caller's page are actually its own — #474 · chat engine

• a gated task inherits every prerequisite report, bounded and counted, and never cuts its own brief — #476 · engine chat docs

• two spaces go home from every place, and the memory card editor keeps its keyboard — #486 · chat

• an open /model list no longer rewrites via and speed on every paint — #491 · chat

• a shaper that spends its whole window thinking starts the task unshaped with a note — #493 · engine chat

• aforge do takes its wall as a duration with a unit, and a bare number is still seconds — #495 · engine docs

• the landing accounts for its own change, on every belt — #496 · engine

• internal/session compiles on windows — the session side of the standing fix — #501 · engine

• under --one-model the status line names the flag, and no crew receipt is posted — #503 · chat docs

• An unattended run decides on what landed and what ran, and stops when nothing moves — #507 · chat

• asking a command for help is not a failure any more, and every command has a usage — #523 · build

• exec --json says why a run failed, not only that it did — #524 · engine

• the spend lab opens on the fixture's clock, not the wall's — #526 · chat

• the day's money on the top line is the day's money on the spend place, wherever you are standing — #527 · chat

• a reopened conversation redraws the jobs it ran, and its job page is reachable again — #529 · chat

• a pinned lane the router refuses for a model yields to another lane, and says so — #533 · chat engine

• An unattended run reads every ending, and a landing that cannot come home is not offered again — #534 · chat

• a remainder is the original assignment plus a finding, and it divides on its size — #544 · engine

• whether a base carries a routing preference is learned from the base, never from its hostname — #547 · engine chat

• a price ceiling that empties the endpoint set no longer fails the turn — #553 · chat engine docs

• a harness saved note no longer races the done channel that tests wait on — #555 · engine

• a node is measured by the material it will read, not the file its words mention — #564 · engine

• a task waits for the command it started instead of inventing a way to poll it — #574 · engine

• Active tool context stays whole until the work uses it — #575 · chat engine docs

• the end-to-end lanes ask for a key the way the product does, so a machine with one runs them — #576 · build

• a task's contract, and a division's checks, are written where the work actually happens — #579 · engine chat

• A checkpoint hands over work it can hand over, and keeps the pieces it is still holding — #580 · engine chat

• A task's ground never climbs out of the scratch directory it lives in — #585 · engine

• an ignore list never empties the set the request is sent to — #586 · engine

• A delivery the gate could not judge is never ok — it is its own ending, said in words — #593 · engine docs

• under a steward with a wall, inline head work is bounded by a share of the wall — #595 · chat

• tool captions return, with its narrator off the scripted queue and its prompt paid for — #598 · chat engine

• A step out of a pasted reproduction is never run as a check — #599 · chat engine

• A file counts as made only when its content changed, and a stash is said out loud — #601 · chat

• docs/HEADLESS.md names the unchecked ending, and three probes hold the page that explains it — #604 · docs

• an interrupt from outside reaches Close, so nothing the session owned is orphaned — #605 · chat engine

• a finished leaf's ending is journaled, so a restart restores it as done — #606 · engine

• a leaf's done carries the commands it ran, or says no check was run — #607 · engine

• a task's landing time is a fact of the record now, and not a subtraction anybody performs — #608 · chat engine

• a leaf knows its own clock, and reading is never the reason one is stopped — #610 · engine docs

• A task's checker is told what was already red, and old red no longer refuses a landing — #611 · chat engine

• a leaf's room reads the wall it runs under, and a tree that does not build is never settled — #612 · engine docs

• continue task N reaches the continue verb, and a missing graph says so — #622 · chat engine docs

• a timeout is not a person's no, and a look does not ask in default mode — #622 · chat engine docs

• an unanswered approval waits; silence is never a no — #622 · chat engine docs

• a stalled first prompt names /model instead of hanging silent — #622 · chat engine docs

• a commit, an undo, or a one-line fix is never converted into a task — #622 · chat engine docs

• a leaf's landing reserve is bounded in tokens as well as turns, so a spent grant cannot buy another — #624 · engine docs

• A run left on its own answers to the same git list a task does — #625 · engine chat

• a landing reads the commit its branch was cut from, and not only its name — #626 · engine chat

• the wall is read while a session is idle, so hours that run out with work still going end the run — #627 · chat engine

• a leaf is judged on the job's tree, not on what it happened to write itself — #628 · engine docs

• three identical timeouts of one command end a leaf's round, with a reason that names the command — #629 · engine docs

• A command quoted in the person's ask names no check, in either spelling — #630 · chat engine

• a cut stream's bill is asked for from the provider's own receipt, and a price nobody has is said — #631 · engine chat

• a lane a person pinned is honoured at every door, not only in the conversation — #632 · engine chat docs

• a check run over an unchanged tree is not run again, and none starts that the wall cannot hold — #633 · chat engine

• a stream that trickles is hedged or cut on a named reason, not left to the wall — #636 · chat engine

• a run left going checkpoints on every door, not only where somebody is watching — #637 · chat engine

• A file the run created counts as made only while there is something in it — #638 · chat engine

• a task's report keeps the fenced tail it promised, and says when it was cut — #639 · engine chat

• a run that was handed a list answers the list, item by item, at every ending — #640 · engine chat docs

• a job's ending reaches the model whole, and the general trim is gone — #641 · chat engine

• A mark reader nobody could reach is not a witness, and the checks stand in for it — #642 · chat

• a thinking level on a class value reaches every request its seat sends — #643 · engine chat docs

• a run that has not landed carries no landing time, and the index answers with the file's last word — #644 · chat engine

• a slash written as prose is not a folder, and a task is no longer refused over one — #645 · engine chat

• an errand's ending names the gate that judged it, and one reader decides a task's audit posture — #646 · engine chat docs

• A faulted gate's row and handover name the fault that happened, not one fixed sentence — #647 · engine docs

• the status line names the gate --yolo opened — #648 · chat

• a covering refusal teaches which machines this account cannot reach — #649 · engine chat docs

• a cost that was really spent never reads $0.00, and money.go is the only place the head spells one — #650 · resident docs

• a refused headless run names the directory it worked in and says nothing reached disk — #651 · chat docs

• A file the run only read is not the change it made, so a refusal names only what it was shown — #652 · engine chat

• Learn generation work and keep successful prompt caches stable — #653 · engine chat

• Wait for delegated workers when their session closes — #653 · chat engine

• Make chat switching respond to hover and give conversation navigation room to read — #653 · chat

• Browse chat names before switching and return to the same task reading position — #653 · chat

• Select the last open chat when closing the current tab — #653 · chat

• Wait for endpoint reachability after a lost connection — #653 · chat

• A turn delivering a result this conversation owns finishes it here, not in a second task — #653 · engine chat

• Name chats from the first message without waiting for the answer — #653 · chat engine remote

• Finished task scrolling preserves collapsed work and disclosure arrows match their state — #653 · chat

• Keep hosted chat identity attached to the selected session — #653 · chat

• Keep explicit verification checks for unattended inline work — #653 · chat engine

• Bound memory preparation and detect stale engines after rebuilding — #653 · engine chat

• Keep elapsed feedback beside a completed step while the turn continues — #653 · chat

• Later acceptance restores a retained task checkout before delivery — #653 · engine chat

• reopening long conversations keeps completed work folded — #653 · chat engine docs

• Keep retained task work distinct from delivery to the requested workspace — #653 · chat

• Keep guest recovery with its owner and preserve unanswered corrections — #653 · chat

• Group task header facts and preserve saved design answers — #653 · chat

• Make task header cost easier to read — #653 · chat

• Keep intermediate task narration distinct from the answer — #653 · chat

• Preserve code blocks in task reports without repeating the answer — #653 · chat

• Read finished tasks as request and answer — #653 · chat

• Align compact task navigation with the reading gutter — #653 · chat

• Reject explicit failed completions before learning provider speed or cache affinity — #653 · engine chat

• Replace failed network attempts on the live conversation page — #653 · chat

• Footer throughput identifies the whole-turn average — #653 · chat

• same-named tags cannot disguise protected landing branches — #654 · engine chat

• Folder and attachment browsing share one modal with live columns — #659 · chat

• the status row stops losing its last cell when the turn clock ticks mid-frame — #667 · chat

• Unusable naming replies recover and compact tabs reveal their full names on hover — #671 · chat engine docs

• endpoint relaxation keeps the price ceiling until a second refusal — #678 · chat engine docs

• a turn split by a failed first step stops drawing a second work door above the split — #681 · chat

• A kept branch is delivered only when the check could find the files it compared — #682 · chat engine

• a check that answers while the send keeps failing now waits, and a media wait names its model — #683 · chat

• The New chat start page keeps its keyboard, and no question is drawn on it — #684 · chat

• Collections survive a second aforge, and a read never takes the writer lock — #685 · engine chat docs

• enter on a running conversation opens it here, and the window it left steps back — #691 · chat engine remote docs

• a first turn nobody has measured an answer for still respects the price of writing one — #693 · engine chat

• A conversation's identity is written under a lock of its own, so two windows lose nothing — #695 · chat engine

• Keep task details inside their column and short task windows usable — #697 · chat docs

• an unattended one-message run stays for the work it hands over, and says every ending — #701 · chat docs

• A state path too long for a socket answers at once, in words, and starts nothing — #704 · chat engine docs

• one sentence for every road onto a folder, and a chooser that opens where the person is — #713 · chat remote docs

• a rescue that takes over an answer withdraws the words the dead machine wrote — #715 · chat engine

• The tui3 harness only gives up on a command that may never answer — #716 · chat

• tool rounds and worker handoffs retain their own evidence — #717 · engine chat docs

• a tab and its switcher row say the same thing, and an unnamed chat has one name — #720 · chat

• Home-card terminal checks follow the current layout and asynchronous Git reading — #721 · chat docs

• A list marker in front of a naming label is markup, and a "Sure!" before one is still an answer — #723 · chat engine

• /help and seven manual pages say what this build actually does — #725 · chat docs

• A step caption keeps a filename, a version number and a path whole — #731 · chat engine

• a rebuild of the same commit is not an older aforge — #732 · build engine docs

• a plain launch sits down in the conversation the engine already holds — #736 · chat engine

• a tasks section heading says what its fold hides, and stops stating a total the foot disagrees with — #739 · chat

• the add context sheet leaves no cell of itself behind, and always draws its way out — #742 · chat docs

• a landing that is your call can be answered from a window on an engine host — #743 · chat remote docs

• One thinking ladder, walked the same way at every door a person can turn it — #745 · chat docs

• streamed reasoning is replayed as completed blocks — #746 · engine chat docs

• The up arrow recalls in every conversation, not only the first one — #747 · chat

• The live rate, the via rider and the phase words cross to a conversation on an engine host — #747 · chat

• the auto-settle floor survives a restart — a landing aforge was deciding comes back yours — #751 · chat engine docs

• an adaptive run prices a bare model id as unpriced, never at a vendor row whose name merely matches — #752 · engine

• Accepted rescue cuts reach the bounded call retry — #755 · chat engine

• a plain launch is greeted by home again, asks from it, and shows what it has spent — #756 · chat engine

• internal/config builds every provider client, and a thinking level leaves the model id — #757 · chat engine

• the model catalog cache belongs to the base it was fetched from — #759 · chat engine

• a thinking model is not a silent one, and no turn ends with nothing said — #760 · chat engine

• AFORGE_QUESTION_DEMO is registered as operator plumbing, so internal/config is green — #762 · build

• a window whose link was repaired takes its task rail back — #764 · chat remote engine docs

• Question handling supports six additional interaction cases — #765 · chat engine docs

• a conversation that moves mid-thought asks its question again where it lands — #766 · chat engine

• A question crosses the engine link, so the road an ordinary aforge takes has one — #768 · chat remote

• Question handling completes six additional interaction cases — #772 · chat engine docs

• home says the gate's own sentence, and names a reminder that fired while you were away — #773 · chat

• The model's task doors now reach the same work yours do — stop, settle, and a bare id — #774 · chat engine docs

• an accept whose merge was refused says so, and the landing it left behind can be answered again — #776 · chat engine

• the column and the composer say the same thing about a landing as its card does — #778 · chat

• a held row that is asking lands on its own answers, and the move card waits for enter — #779 · chat docs

• the manual stops quoting two landing receipts that no code spells any more — #781 · chat docs

• a turn that loads a tool and stops without using it is sent back once — #782 · engine docs

• answers are rows, keys are digits, a checklist is ticked on the card, a switch forgets questions — #784 · chat engine

• a long reply still writing at its lane's pace is no longer cut at the stream wall — #786 · chat engine

• every question has a pointer the arrows walk and enter takes, and the task page shows its question — #789 · chat

• the question room's answers walk, open and click, and an open one reads in three tiers — #791 · chat

• a 429 inside an open stream now paces the pool that served it, like one that arrives as a status — #793 · engine

• a model that keeps refusing is given up on, and the reply finishes on the next one in the chain — #794 · chat engine

• a failed request leaves a row where you are reading, and a finished step stops saying it runs — #795 · chat

• an errand is bounded as an errand, and a rung that is answering is not cut by a wall clock — #796 · chat engine

• a question with evidence under each answer reaches the person on the first ask — #797 · engine chat

• A model picked while a task is being checked is the next run's, not a rewrite of what ran — #802 · chat engine

• a handover no longer lists finished work as work that is left — #804 · engine chat

• a task room draws its own machine and rate, because news now says which work it is about — #806 · chat engine remote

• A reply waiting on its own running task is left alone rather than pushed on — #811 · chat engine docs

• a checking call the window refused is no longer written down as a call that was made — #812 · engine

• the job footer stops defeating de-duplication, and every footer is budgeted inside its own cap — #812 · chat engine

• the routed memory block rides at the tail, so message[0] holds still for a session — #812 · chat engine docs

• An answer the door never acknowledged is still your answer, not another window's — #812 · chat

• The lean prompt profile keys off the context window and an explicit pin, never off the worker seat — #812 · chat engine

• a cancelled arm is waited off the temp home, and a recovered lane is a candidate again — #814 · engine

• roomfeed waits for the room's way back; reopened rooms keep each call's duration — #820 · chat engine

• blanks keep [tab] next blank at a hundred columns, and a stood order refreshes the column count — #821 · chat

• a stalled check under load keeps its stall account, person manual labels match — #822 · chat engine docs

• Nested landing accept banks the question and writes you took this as done — #823 · chat engine

• A dev build publishes its release instead of skipping publish after the build — #834 · build

• a refused race recovers instead of ending the turn, and account exclusions are learned once — #835 · chat engine docs

• a waiting question rides behind the row's door instead of over it — #840 · chat docs

• A bare launch draws the spend panel and the head's money on its first frame, not three seconds later — #842 · chat

• the provider and the live tok/s no longer vanish at random — #845 · chat remote

• a keystroke is not refused at ten seconds behind an unrelated getter — #846 · chat remote

• a later task is not pointed at an earlier task's copy — #847 · engine chat

• a retry goes to a different machine, and the ledger is keyed on who answered — #850 · chat engine docs

• a window that just acted is still watching, and a stop names the door it really came through — #851 · chat engine remote

• One classifier for every failed call — the transport states facts, one function picks the move — #854 · engine chat

• a correction that fell through gets its turn drawn — the steer lane is stamped by conversation — #855 · chat

• A hosted conversation keeps drawing the turns the engine starts after this window's own first turn — #856 · chat

• the router can reach the fast lanes again, and a refusal now counts against a lane — #859 · engine chat

• a decayed belief draws no tail — #862 · chat

• Opening a task's page or switching tabs no longer counts the token figures up from zero — #867 · chat docs

• A conversation that ran a quick task reopens with its task graph — #869 · engine chat

• a question you talk past comes down, and your own words go out at once — #870 · chat engine docs

• Nothing auxiliary blocks the person — one reading beside the work, never in front of it — #871 · chat engine

• a machine is asked only when the answer is expected inside the wait its work will sit through — #873 · engine chat docs

• the frame reads cached data without disk access — #875 · chat build docs

• nothing hidden happens on a turn — one file memo, one off-path reading, one deferred write — #876 · chat engine docs

• A quick task comes through one door, its checkpoint rebuilds it, and it holds the files it wrote — #879 · chat engine docs

• A task no longer pays for a fork it will not use, and its runner reads git once a batch — #881 · engine docs

• A task being sized shows the model thinking, the seconds and the tokens coming back — #882 · engine chat docs

• a fan of tasks starts what this machine can carry, not all of it at once — #885 · engine chat

• a refused call keeps the schema's sentence for the model, and the person reads their own words — #890 · chat engine docs

• a row of the tasks place says its state once, and a root names its folder only where that is news — #895 · chat engine

• a working copy that holds work is kept, whoever did the work — #899 · chat engine docs

• A conversation's anchor writes a whole identity, so an older snapshot cannot undo it — #904 · chat engine

• plain engine-road chats keep this machine's connection, harness, and profile doors — #906 · chat engine

• every question is raised and settled through one door, and answering leaves the prompt alone — #910 · engine

• a task's check runs in its own copy, and a piece that comes home late folds into its parent — #911 · chat engine docs

• A conversation that has said nothing is called new conversation on the tasks place, not by its id — #915 · chat

• an answer is drawn the moment it is pressed, and a question takes keys only where it is drawn — #919 · chat

• a wire below its promised pace is rescued out of the call's own budget, and every attempt is watched — #924 · chat engine docs

• a named refusal is a move, an account's ceiling waits once, and a person's pick outranks the chain — #925 · chat engine docs

• a task card quotes the report's first line that says something, not its literal first line — #930 · chat

• a word you say while work is running lands at the next request, not the next turn — #935 · chat engine remote

• a planning model that thinks past its wall is asked for its answer instead of losing it all — #940 · engine resident

• The tasks tool counts another window's families, not its parts — #944 · engine

• the machine reading is divided by the lanes the whole process is running, not one conversation's — #947 · engine

• an observation the reply already answered is not raised again — #948 · chat engine docs

• a check that runs out of time is told its window, keeps what it read, and says it was the clock — #949 · chat engine

• A task's page says a call is still open, with that call's own clock and rate — #951 · chat engine

• A reading beside the work is waited on, not guessed at, and a held memory line is not lost — #952 · chat engine

• Asking back gets an answer with the question still open, and the chip counts only what waits on you — #960 · chat

• a split drawing reaches the boundary its own cut opened, and never says it carried on — #962 · engine

• a countdown on a question is let go when the conversation closes, like everything else it armed — #963 · engine

• an answer left for a closed conversation is applied as it opens — #965 · engine chat

• an idle conversation on an older engine build is named on the way in — #966 · chat engine

• a cold question about how it writes, and about what to do after installing, reaches its page again — #973 · chat docs

• A namer that answers with its plan is refused, so the task keeps the person's own words — #975 · chat engine

• key presses avoid unnecessary redraws — #977 · chat build

• Two cold questions reach their page first instead of fourth, so the retrieval floor has margin — #978 · chat docs

• a task the sizing judge called wide keeps the divide verb when the reviewer refuses its first parts — #987 · engine docs

• the permission frame names its way out wherever the pointer stands — #992 · chat docs

• the reading that tells a model what other windows did no longer writes into a closed conversation — #997 · engine

• The stale-engine refusal names the workspace in the command it tells you to run — #998 · chat engine docs

• the OpenRouter-takeover gate counts a refusal only when the router's own pick was refused — #1015 · engine docs

• a read of lines the conversation already holds answers with a pointer — #1015 · engine docs

• the cache pin is admitted into the lane chooser's enforced set rather than dropped by it — #1015 · chat

• The race can no longer pull the first taking-stock mark below half the handoff price — #1015 · chat engine

• billed arms journal their call lines, and a hopped turn seals one usage line per model — #1015 · chat engine

• the model picker's auto row stops promising a takeover the simple routing never does — #1019 · chat docs

• the chip names a machine only while the wire would demand it, and a routing change lands now — #1025 · chat engine docs

• a task-start question you answered is not asked again when you come back to the tab — #1036 · chat engine

• a conversation you put away with ctrl+e leaves home on the keystroke, not a few seconds later — #1039 · chat

• the stable release notes are rendered to fit a GitHub release page instead of copied whole — #1044 · build docs

Removed

• the v2 chat surface is gone — tui2 is now only the library v3 draws with — #111 · chat resident build

• a window holds as many conversations as you open — #137 · chat

• There is one worker — swe, the bare worker and every way of choosing between workers are gone — #235 · engine chat build docs

• the proposal hold of — #367 · engine chat remote

• the v2 chat entry point, and everything only it reached — #469 · chat

• lanestub's router-shaped second address is gone; the plain base is the only base — #554 · engine

• The caption outline of — #591 · chat engine

• five stale artifacts are gone — no binary at the root, and harnesses/ is only ever generated — #788 · build docs

• the v2 remnants — the root compositor, the model picker, the blocks engine — #790 · chat

• the v1 window and internal/command — 111 lines moved, 36,037 deleted — #792 · chat resident build

• fork is off the belt: a reply no longer copies itself into hands — #811 · chat engine docs

• The mid-stream ledger teaching of — #990 · engine

v0.1.0 — 2026-08-17