Release notes | Pumpkin

Pumpkin 2026.10.574

Pumpkin's own release checks pass on a busy machine: the hook output check no longer starts helper programs, and the phone pairing checks ignore the local store's lock files.
A settings refresh that keeps losing to changes gives up exactly at its time budget instead of squeezing in one more read.

Pumpkin 2026.10.573

Add models from OpenRouter with your own OpenRouter key and use them in chat, Auto and helpers. They need this release's local proxy.
An organization owner or admin can turn on an audit log of agent actions: who ran which tools, where files changed, and model use counts.
Audit records leave out prompts, replies, commands and file contents.
An organization can set monthly spending caps. Pumpkin credit caps stop requests at the limit for the whole organization or each member.
Own-key spending caps warn at 80 percent and stop new requests at 100 percent on each computer.
Pull request review and security review skills now ship with Pumpkin, with ready loop templates.
First automatic review rounds are dry runs, and security fix pushes never go to a default branch.

Pumpkin 2026.10.572

GPT models on a Codex subscription no longer see Codex's own built-in skills, which pulled them into summarizing docs instead of doing the work.
When you paste news, docs, a post or an error with no question, GPT models now treat it as work to do in your project, not something to summarize.

Pumpkin 2026.10.571

Pumpkin Remote can notify your phone when a chat is waiting for your answer, even while the app is closed. Turn it on from the Remote app's Status page.
On iPhone, add the Remote app to your Home Screen first. The notice is encrypted on your computer and the site only passes it along.
Phones and computers connect faster: connection details are exchanged while setup is still running instead of after the full list is ready.
A phone whose clock is off no longer fails to connect; requests are timed by the server's clock.
The Remote app hides controls a computer running an older Pumpkin cannot handle, and the Status page shows that computer's remote protocol version.
Other editors and phone apps can follow one written remote protocol specification shipped with Pumpkin.

Pumpkin 2026.10.570

A remote server that misses its start deadline now keeps the timeout error. Late log output cannot mark a stopped server ready.

Pumpkin 2026.10.569

Release builds of the local proxy work again on every platform.

Pumpkin 2026.10.568

Release builds of the local proxy work again on every platform.

Pumpkin 2026.10.567

Canvas pages cannot use their own link to read or send messages through another chat's canvas.
Built-in behavior rules load from the server, with short fallbacks on disk. Signed Mac builds also hide more code details.

Pumpkin 2026.10.566

The browser can type in form fields that move or sit in frames with no screen redraws. Hidden, locked, or read-only fields still block typing.
When a chat moves after an update to start a helper, its pause note explains why and says it will resume on its own. It no longer tells you to stop.

Pumpkin 2026.10.565

The browser can type in form fields that move or sit in frames with no screen redraws. Hidden, locked, or read-only fields still block typing.

Pumpkin 2026.10.564

A model you pick stays on that model, even if its subscription runs out. Auto alone can choose a different model. Paid access to your picked model still costs money.
Chats keep the newest images the model can read. Older images become short notes with saved links the model can open again.

Pumpkin 2026.10.563

Settings save errors show more detail when a file cannot be read or its path is broken. The local log keeps the full error.
On Windows computers where Pumpkin was installed with the setup program, Pumpkin commands typed in WSL work again. Before, they stopped with a file check error.

Pumpkin 2026.10.562

Agent tool inputs are checked the same way everywhere: an unset field may be sent as null, and a bad value gets one clear correction hint instead of differing messages per tool.

Pumpkin 2026.10.561

Settings save errors now show file access and path problems, and keep the full error in the local log.

Pumpkin 2026.10.560

Codex chats now read your saved answer before going on after a question, even when other work ends while you wait.

Pumpkin 2026.10.559

Chats that run through a Codex or Claude Code subscription no longer leave a copy of the conversation setup in those tools' own history files.

Pumpkin 2026.10.558

Chats you have not used for 15 minutes leave memory and load again when needed, so the background service uses less memory. A new setting lets you change the wait.

Pumpkin 2026.10.557

Auto now compares longer-running costs before switching to a paid model after a subscription limit, and checks again when a cheaper suitable route becomes available.
Auto avoids long context when the conversation fits the normal window, and switching models clears the earlier model's context-loss count.
Auto keeps complete model-choice explanations for troubleshooting while keeping each chat's private text within that chat.

Pumpkin 2026.10.556

A chat set to Fable or another model tries your other subscriptions first when its own subscription cannot run. Paid use now shows a notice saying why.
When Auto finds no model it can run, the error now says which check removed them, such as skill floor or the subscription setting.

Pumpkin 2026.10.555

Auto no longer stops in long chats with many screenshots. Before, old images counted against every model and the chat showed a wrong helper error.
When a request really holds too many images for every model, the error now says so and what to do.

Pumpkin 2026.10.554

A loop question you answered late in Slack now leaves the Loops question list as soon as the next iteration starts, even while that iteration waits on a new question.

Pumpkin 2026.10.553

Auto in a main chat now only picks models rated 6 or higher unless you lower the minimum in Auto Agent Selection. Helpers stay at 4.

Pumpkin 2026.10.552

On WSL and SSH hosts, Pumpkin now removes old installed versions it no longer runs. Before, each update left about 240 MB behind.
An unanswered question no longer keeps an old Pumpkin version running in the background after an update. The question stays open and answerable in the new version.
Old loop runs are archived like chats after a few idle days, so their analysis files stop filling your disk. Opening one from the Loops tab restores it.
Loop runs have their own cleanup settings: scratch data goes after 3 idle days, runs are archived after 3, and the newest 3 runs of each loop always stay.
Pumpkin now cleans up old backups of code-search indexes it manages, while keeping edit records and backups whose owner is unknown.

Pumpkin 2026.10.551

Updates keep helpers working when a chat needs a new helper. Paused work resumes with results, input and open questions intact. The chat explains the move instead of saying Pumpkin closed.
Helpers now send their later job and task results back to the chat after an update, while keeping prior costs and queued input.

Pumpkin 2026.10.550

In Remote, click a window title to hide or show its chats and loops. Your phone or browser keeps each window's choice when the list updates or the page reloads.

Pumpkin 2026.10.549

File tools now find the right data folder when a chat uses a data-folder shortcut. Unknown path variables show a clear error instead of making a folder with the variable name.

Pumpkin 2026.10.548

Codex subscription models now use the full window Codex allows (872K on GPT-6), not a fixed 272K. Long chats stay on your Codex subscription instead of switching to a paid model.

Pumpkin 2026.10.547

On Windows, Codex and Claude Code now work when Node or the CLI was installed after Pumpkin started. Before, the account check failed until you restarted the background service.

Pumpkin 2026.10.546

A helper agent's result no longer reaches its chat before the helper finished saving, which could leave the chat's cost lines out of order.

Pumpkin 2026.10.545

Clicking Sign in again while a sign-in is still waiting opens the same sign-in page instead of showing a conflict error.
When you are signed out, actions that need your Pumpkin account now say to sign in instead of showing an internal error.

Pumpkin 2026.10.544

A message a helper agent sent that never reached its chat is now delivered on the next turn, so the chat's busy dot clears instead of staying on.

Pumpkin 2026.10.543

A busy Pumpkin server with many open folders no longer freezes when it checks Codex or Claude sign-ins. Before, every run on it could time out.
A Pumpkin server now shuts down when asked, even when part of it is stuck.
Many workspaces starting at once no longer leave one that can never stop.

Pumpkin 2026.10.542

On Windows, Codex and Claude Code sign-ins work again when your user folder name has a space. Before, every request to them failed and Auto quietly used another model.
On Windows, wait checks that run a command now use PowerShell, the same shell as other commands.
Canvas updates accept small edits and data even when the agent sends them as text.

Pumpkin 2026.10.541

Loops no longer open canvas tabs. Nobody watches a loop, so its canvases stay closed and you can open them later with View: Open Canvas. Loop agents also stop reaching for canvases on their own.

Pumpkin 2026.10.540

When Auto's helper replies without naming a model, Auto asks again and then falls back as it does when the helper is down, instead of failing the turn.

Pumpkin 2026.10.539

Archiving big chats uses far less CPU and finishes faster.

Pumpkin 2026.10.538

Pumpkin no longer re-checks every folder in your workspace each time it loads project rules, so it uses far less disk and CPU while chats run.
Auto turns start a little sooner: memory search now runs while Auto picks a model.

Pumpkin 2026.10.537

Auto keeps using your Codex and Claude Code subscriptions when a long chat outgrows their window. It compacts older history instead of switching to a paid model.

Pumpkin 2026.10.536

When several Claude Code chats refresh the same sign-in at once, Pumpkin now waits and retries instead of stopping the chat. On Auto it switches to another model if the retries fail.

Pumpkin 2026.10.535

Pumpkin no longer re-reads your settings and MCP config files several times a second while idle.
Semantic search no longer re-reads every project's settings once a minute.

Pumpkin 2026.10.534

The background cleanup of old chats no longer logs a memory warning.

Pumpkin 2026.10.533

A server no longer counts a run that failed to start as still working. Updates that wait for quiet time now run on time.
The server status can list what each busy item is.

Pumpkin 2026.10.532

With Settings open on Semantic search, Pumpkin no longer re-checks the search index several times a minute while you chat.
Semantic search health checks and searches on WSL no longer stall for 30 seconds waiting on a DNS lookup.
Windows: saving shared usage data no longer fails when two writers meet on a lock.

Pumpkin 2026.10.531

Pumpkin reads and parses far fewer files while you work. Chats, search, skills, memory docs and data folders now come from shared caches that refresh when files change.
Searching past chats no longer re-reads every chat file. Search snippets keep their exact text and stay in step with the match.
Semantic search no longer runs with a default memory cap. The cap made it use far more CPU on large indexes.
Very long sub-agent replies no longer slow down the chat list.

Pumpkin 2026.10.530

When a server updates, running turns and jobs finish first again, instead of being stopped early.

Pumpkin 2026.10.529

Subscription quota is worked out in one place. Routing, Auto, Settings and the chat header badge now agree on what is used up.
Old or undated quota readings show as unknown instead of looking fresh, and a known usage number stays visible while an account is on hold.

Pumpkin 2026.10.528

A server owner can now pick one safety wrapper for all commands the agent runs. If that wrapper is missing or broken, commands stop instead of running without it.
When the server's sandbox rules change, the old runtime and its jobs are stopped instead of being reused under the old rules.

Pumpkin 2026.10.527

A loop can now use its own writing style instead of your chat style. Loops that write emails or reports no longer copy how you like chat replies.

Pumpkin 2026.10.526

Long chats show the model in use in the model menu and the chat header again.
The header badge drops the words Subscription and Paid. The color shows which, and the hover still says it.

Pumpkin 2026.10.525

When a subscription runs out mid-turn, Auto picks again with that route removed and keeps going, on a paid model if needed. Only Subscription Only blocks paid picks.
Auto no longer stops a turn with "paid fallback was stopped".

Pumpkin 2026.10.524

The Semantic search health check now runs a real search. It warns when searches are slow and reports a problem when they fail, instead of saying healthy.

Pumpkin 2026.10.523

Semantic search no longer stalls about 30 seconds per search on machines whose DNS does not answer TXT lookups (seen on WSL). Memory lookup works again there.

Pumpkin 2026.10.522

Auto keeps working when its task helper is slow or its provider is down: it retries, then tries a helper on each other provider in your order.
If no helper answers, Auto reuses the chat's last approved model, or the model it has approved most often lately. The error names the helper, attempts and cause.

Pumpkin 2026.10.521

Clicking a Pumpkin desktop notification works again. It brings forward the window that has the chat and opens that chat. Before, the click did nothing.
Chat links open in the right window again.

Pumpkin 2026.10.520

Semantic search limited to a folder now finds results in that folder. Before, it often came back empty on large indexes, and memory lookup fell back to keyword matching.
Folder paths relative to the open folder now search the right place.
The Semantic search health card checks again within a minute after a problem, so it clears soon after things recover.

Pumpkin 2026.10.519

After WSL restarts, Pumpkin drops the old WSL connection instead of retrying it forever. The red WSL warning clears once the new connection works.

Pumpkin 2026.10.518

Unsupported session folders no longer block syncing known chat files. Unknown refs stay saved, and bad paths are rejected.

Pumpkin 2026.10.517

Usage logs keep all chats and turn counts when more than one device writes at once. New builds avoid counting the same saved batch twice.

Pumpkin 2026.10.516

Auto uses one helper and route flow for chats, side threads, child agents and model switches. Ready subscriptions win before paid routes in Priority mode.
Empty API credits skip wasted retries and go straight to an approved fallback.
Codex starts with the right Node path. Older Codex builds keep working when they lack the model-list command.

Pumpkin 2026.10.515

Large chat views and model lists do less repeated account and settings work. Settings edited outside the editor may show up on the next message.
Old chat pages get cleaned up even when those chats stay idle.
Missing notice pages leave a saved record of what was lost.
Code search updates changed files without rewriting its whole symbol file each time.

Pumpkin 2026.10.514

Each chat can run up to 32 child agents at once, up from four.

Pumpkin 2026.10.513

A background job stopped by a Pumpkin restart can be dismissed from Background jobs without running it again. The agent can clear it too.

Pumpkin 2026.10.512

Background commands keep running through a Pumpkin update. The new version picks them up, so the agent can still wait on them and read their output.

Pumpkin 2026.10.511

Long turns get a softer final check. Done work can stay done, review steps are respected, and old quotes are not fresh consent.

Pumpkin 2026.10.510

Codex and Claude Code subscription accounts load again on Windows without WSL. Sign-in no longer ends with "Saved account state is unreadable".

Pumpkin 2026.10.509

Shared servers keep saved Codex accounts and new chats working after old workspaces close.

Pumpkin 2026.10.508

Saved Codex accounts keep working on shared servers after old workspaces close. One stale workspace no longer blocks sign-in or new chats.

Pumpkin 2026.10.507

Chats synced from another computer no longer get stuck on "missing notification segment". Sending, forking and syncing work again, even when the other computer never sent its old notice pages.

Pumpkin 2026.10.506

Agents can save SQL queries to the Data page. Ask the agent to save a query and it appears in your saved queries, shared like its connection.

Pumpkin 2026.10.505

Agents can refresh one MCP server's tool list mid-chat. After you add or change a tool on a server, the agent can load it and call it without a restart.

Pumpkin 2026.10.504

A long turn no longer hides its answer. When Pumpkin keeps the agent working after it replied, that earlier reply stays visible above the folded work.

Pumpkin 2026.10.503

Busy chats save and process changed rows.
Large-chat overflow trims avoid repeated counts.
Runtime startup and settings builds are safer.

Pumpkin 2026.10.502

Chats from another computer open even when saved notice files have not synced yet.
Chat sync brings all saved chat files. Update all paired computers; older versions cannot download these chats.
Missing files must sync from the source computer before you can send in the chat.
Follow-up messages work while a child chat closes. Usage logs keep each model's key source.

Pumpkin 2026.10.501

Headless runs can use their own Anthropic API key while keeping it out of saved chats and tool environments.
Chats retry when a model says it is at capacity or faces high demand.

Pumpkin 2026.10.500

Agents now accept Autopilot answers and approvals as yours. Autopilot acts with your full authority.

Pumpkin 2026.10.499

The chat header badge names the subscription account in use, such as Claude Code · Ezoic · Subscription, so you can tell your accounts apart.

Pumpkin 2026.10.498

The chat header shows the provider now in use. Green means Subscription; amber means Paid. The badge follows route changes as they happen.
When there is room, the badge shows the plan and quota used for that account. Old readings stay marked, and full quota details stay in the hover.
Closing the host waits for chat saves already in flight.

Pumpkin 2026.10.497

Repeated chat saves reuse known item sizes and skip extra notice checks while keeping edits and saved content exact.
WSL and SSH starts keep a live startup lock safe when another window checks an old owner.

Pumpkin 2026.10.496

Reuse prepared session context on unchanged saves while keeping nested edits and disk repair safe.
Shutdown waits for session saves already running.
Memory lookup runs alongside chat setup while keeping results fresh.

Pumpkin 2026.10.495

Saves reuse known folder names instead of listing unchanged folders.
Search saves write new text in small files while keeping full results.
Auto model choice and memory loading run together when every allowed model uses memory.

Pumpkin 2026.10.494

Memory title reads run in bounded groups during chat setup, while lists with limits still read only what they need.
Slow workspace setup gets time to finish and skips old-runtime cleanup. Timeouts name the last setup step.
Auto shows its live model pick in the selector and clears the old pick when a new turn starts.
Windows CLI failures name the check that failed while keeping private data out of the error.

Pumpkin 2026.10.493

Pumpkin links the code changes and issues it mentions, and finishes work instead of ending the turn on a plan or a promise.

Pumpkin 2026.10.492

Chats send again while a subscription account has a usage-limit hold; turn setup no longer fails with "Converting circular structure to JSON".

Pumpkin 2026.10.491

Auto settings now offer Subscription Only, Subscription Priority, and Normal. Priority is the default. These choices guide Auto picks; models you pick and helper calls stay allowed.
The task helper now picks all model fits. Code only sorts those fits by cost, quota and access. Your model notes stay in force for each switch.
A Fable limit leaves other models on that account free to run when the helper approves them. A full account limit still blocks that account.
Failed helper picks stop new model switches. Image counts stay within each model's limit for chats and child tasks.
A pinned chat shows its quota switch notice once while the same model route and hold stay in place. The notice names the task helper's approval.

Pumpkin 2026.10.490

Chat stays quiet when its prompt cache resets or saved user text grows large.

Pumpkin 2026.10.489

Turn setup reuses unchanged rules, skills, agents and memory files instead of reading them again.
Independent setup work runs together, including source loading while Auto picks a model.

Pumpkin 2026.10.488

Chats reuse shared settings and key checks when several turns start at once. Unrelated settings changes no longer reload every chat.
Usage logs finish flushing without an extra minute of waiting, while failed writes still retry safely.
The Windows-to-WSL command launcher reuses checked installed files instead of copying the runtime on each command.

Pumpkin 2026.10.487

Chats share fresh settings and key reads, reducing repeated work when several chats run at once.
Chat lists reuse saved cost totals. Large cost details load in pages and can be opened from the chat menu without losing any values.
Usage logs keep lines when a read fails and retry them before shutdown.

Pumpkin 2026.10.486

The chat Media panel loads again when Pumpkin keeps its chat files in a folder outside the project and the chat uses its own work copy.

Pumpkin 2026.10.485

Claude Code models no longer block actions you approved, such as deploys. Pumpkin's own approvals now decide, not Claude's auto mode check.

Pumpkin 2026.10.484

When a workspace admin enables Slack thread events, replies in an existing Pumpkin thread can continue the same chat without another mention.
When Slack thread details lack a chat link, mentioning Pumpkin again resumes the saved chat for that thread.

Pumpkin 2026.10.483

Auto now lists a model that runs on your Codex or Claude seat as its own provider, beside the cheapest API model of the same maker. The helper judges capability; Pumpkin ranks cost.

Pumpkin 2026.10.482

An Auto turn now saves every model the helper was offered, with its skill, price and route, beside the pick, so a surprising choice can be checked.

Pumpkin 2026.10.481

Auto's model list now prices each model by the route it would use. A model served by your Codex or Claude seat, or by the local model, shows no per-token price instead of its API price.

Pumpkin 2026.10.480

The phone and web views refresh only the chats that changed: the runtime no longer re-lists every chat of every workspace on each 30-second heartbeat.
The per-chat prompt cache ledger is stored in pages: a round appends a few kilobytes instead of rewriting a half-megabyte file.
Token counting for the context budget reuses frozen history instead of copying it on every call; counting now costs about a third of what it did.
Background job records are stored one file per job; a streaming job no longer rewrites a multi-megabyte registry on every output chunk.

Pumpkin 2026.10.479

After a Claude quota reset, choose **Retry after reset** on that account in Settings › Models. The next turn can try Claude again; usage counts keep their true age.

Pumpkin 2026.10.478

Settings › MCP has a Refresh Tools button on each server. It reloads that server's tools without a restart, and running chats see them on their next step.
If a refresh fails, the old tools stay in use and the server row shows the error.

Pumpkin 2026.10.477

Auto's model list now says a signed-in Codex or Claude seat and the local model have no per-token price, instead of "unpriced", so the helper weighs a seat as the cheap capable choice it is.

Pumpkin 2026.10.476

Clicking a chat notification with several Pumpkin windows open now brings up the window that holds that chat and opens it, even on a Mac.

Pumpkin 2026.10.475

Old loops return to their saved workspace when their moved chat history proves the link. History and replies stay intact. Loops without enough proof still need Take over.
Loop pre-checks in saved workspaces run in the first code folder again, so relative script paths work.

Pumpkin 2026.10.473

The pumpkin agent run command works in an open saved workspace with several folders: pass its state folder, and the chat lands in that workspace.
Passing a saved workspace file to the command now says which folder to pass instead.

Pumpkin 2026.10.472

When a background job ends just before the agent waits for it, the agent now gets its result once instead of twice.

Pumpkin 2026.10.471

A Fable limit now lets a chat use Opus on the same Claude Code account. A limit on the whole account still blocks all its models.
Chats can switch again when the next subscription runs out. Each switch keeps the original skill floor and skips routes already tried.
A held model can use another ready subscription even when no paid API key is set.

Pumpkin 2026.10.470

When Fable's Claude Code quota runs out, a chat now uses another subscription such as Codex instead of the paid API.
The paid API runs only when no subscription model at most one skill lower is ready. The chat shows a notice when it switches.

Pumpkin 2026.10.469

On a Mac, Check for Updates with an update already downloaded no longer says pumpkin.club cannot be reached.
When an update check fails, Pumpkin says why, such as signed out or server error, instead of blaming the network.

Pumpkin 2026.10.468

A chat can no longer get stuck with the error "session metadata or archive catalog exceeds the live history budget": a save now always lands, however large the chat's bookkeeping grew.
A single message or activity entry larger than a history page is stored as a page of its own instead of failing the save.
Settled notifications (finished jobs, watchers, agents) move out of the chat's main file into pages, so long chats stop growing that file with every save.
During a runtime upgrade, a chat still served by the previous runtime keeps its Media tab and activity column working; requests are translated for the older runtime.
The Media tab checks only the files of the page it shows: opening the Media tab of a chat with 35,000 rows no longer looks at 25,000 files first.
An opened, pinned or tagged chat no longer jumps to the top of the sessions list for a moment, and paging the list no longer skips it.
Searching chats after a restart reads a small per-chat search file written with each save instead of parsing every transcript; the first search no longer takes seconds.
The web storage sync pass no longer checks every chat's files every 15 minutes; only chats that changed are looked at.

Pumpkin 2026.10.467

Agents start again after an update without a computer restart. If the background service from before never reconnects, Pumpkin now replaces it after two tries.

Pumpkin 2026.10.466

Loop Questions lists questions that still need your answer first. Answered questions waiting for the loop's next run move to their own group below, with their saved answer.

Pumpkin 2026.10.464

Find in Chat: a search button in the chat header (or Ctrl+F / Cmd+F) lists every message that contains your words; click one, or press Enter, to jump to it with the words highlighted.

Pumpkin 2026.10.463

Desktop chats reconnect again. Since 2026.10.420 some windows kept failing to reattach their agents. Your chats and worktrees were never changed.

Pumpkin 2026.10.462

Headless CLI runs release their workspace when they finish; the server no longer stops accepting runs after 128 of them.

Pumpkin 2026.10.461

Windows: the installer for the 2026.10.459 changes below; releases 459 and 460 could not build it.

Pumpkin 2026.10.460

Windows: the installer for the 2026.10.459 changes below, which that release could not build.

Pumpkin 2026.10.459

Opening a chat with a very long activity history is fast again: the runtime sends the folded state (files, tasks, agents, canvases, watchers) and a short tail instead of every row.
The Media tab shows the first 100 items with a Load more control, and refreshes only what changed instead of rebuilding the whole list after every turn.
Media lists every saved data file; the 500 row cap added in 2026.10.430 is gone.
The end of turn data scan reads only folders that changed and no longer writes one activity row per file it finds.
Session saves during a turn are coalesced; most bookkeeping writes join the turn's save instead of writing the file at once.
Interrupted and startup session listings come from the session index, not a directory scan.
Fewer stray temporary files: every atomic file write in the runtime goes through one shared writer.

Pumpkin 2026.10.458

Headless runs keep private temp files until all child runs finish, then clean them up safely, including when startup fails.
Headless runs wait briefly for MCP tools before starting. Slow connections do not block the run, and sign-in prompts still appear when a connection finishes later.

Pumpkin 2026.10.457

Headless server: a run with explicit read roots gets its own private temp folder; the server's shared temp folder is no longer readable or writable by that run.

Pumpkin 2026.10.456

A loop question card that fails to settle twice in a row logs its error once, not once per attempt.

Pumpkin 2026.10.455

Headless runs: each turn's JSONL route record names the model, provider and account that actually served it, and a new record follows every mid-turn model switch.
Semantic search: a path equal to the workspace root searches the whole workspace instead of failing with "does not belong to any project".

Pumpkin 2026.10.454

A loop deleted and recreated with the same name between two reloads no longer receives a stale question from the old loop.
Loop question logs always carry a reason, even when the error has no message.

Pumpkin 2026.10.453

A loop question asked while the loop is being reloaded or taken over by another window is no longer lost; it is recorded and reaches the next run.
When a loop question card cannot be settled and is released, the log line now says why.

Pumpkin 2026.10.452

Headless server: a run started right after another in the same folder keeps its MCP tools; a new folder joining no longer drops other runs' MCP access for a moment.
Auto no longer sends a request to a model whose input limit is below the request size; when a provider refuses the size, Auto picks a fitting model once and continues the turn.

Pumpkin 2026.10.451

Closing a loop question card whose run is still winding down now logs once and settles once, instead of twice.

Pumpkin 2026.10.450

Replies already read for a loop question card are no longer lost when the card's save fails once; the save is retried and the card keeps the replies.

Pumpkin 2026.10.449

Headless server: semantic search results inside a run's allowed folders are kept again (every hit had been dropped), and call-graph entries are judged by the repository that holds the file.

Pumpkin 2026.10.448

A loop whose folder was removed, deleted, or taken over by another window now releases its open question cards so they get their usual review.
Replies already read for a loop question card are kept when the loop is taken over by another window.
Closing a loop question card that failed to settle is retried for up to ten minutes once its run is idle, then released, instead of staying open until a restart.

Pumpkin 2026.10.447

Saving a chat writes only what changed: the prompt-cache ledger and the cost records of closed turns are stored once beside the chat instead of being rewritten on every save.
Token counting over an unchanged conversation reuses its memo instead of re-walking every message, so busy chats cost less CPU between rounds.
Checking a CLI account's quota or sign-in status no longer rebuilds the whole settings inventory, which made Settings and the model picker slow to answer.
Web storage sync works again for workspaces with thousands of chats; its saved state was being refused as too large and every sync started over.

Pumpkin 2026.10.446

Headless server: semantic search results that cannot be checked against a run's allowed folders are now withheld instead of shown.
Headless server: subagents and side threads keep the run's folder and account limits, and the account limit also holds when a turn falls back to another model.

Pumpkin 2026.10.445

A loop question card you closed while its run was still winding down now closes on its own once the run is idle, instead of staying open until a restart.
Looking up a lost Slack post for a loop question no longer restarts from the beginning on every try, and gives up after a bounded number of pages instead of running forever.
Deleting a loop releases its open question cards, so they get their usual review instead of staying claimed.
The note on a closed loop question says whether anything was passed to the next run.

Pumpkin 2026.10.444

An agent browser shows its handle (b1, b2, …) at the left of its address bar, so a separate browser window with one tab shows it too.

Pumpkin 2026.10.443

The Data page has a New query icon beside Run. It clears SQL and rows, keeps your connection and chat, and lets Undo bring the SQL back. Save starts a new file.

Pumpkin 2026.10.442

Loop questions now keep one record per question across the chat card, the Loops tab and the Slack thread. An answer in any of them reaches the next run, and a question is never asked twice.
A loop question survives a Pumpkin restart or crash: the Slack thread is found again, a question that never posted is posted once, and a reply after the wait ended is still read next run.
Mark resolved on a loop question reads every Slack thread for it first, so no reply is lost.
Loops built on an older Pumpkin can still read their state files, and older Pumpkins can read files written by this one.

Pumpkin 2026.10.441

Updating Pumpkin no longer cuts off running chats when a loop could not run because its model is turned off.
After an update, the old background service now exits instead of staying alive with its memory and helper programs.

Pumpkin 2026.10.440

Fireworks (Kimi, GPT OSS) works again from AWS networks: those reach Fireworks edges whose certificates chain to Amazon roots, which Pumpkin now pins alongside Let's Encrypt.

Pumpkin 2026.10.439

An update no longer cuts off your running chats when a loop had just failed to start (for example on a disabled model). The old service hands over cleanly; the loop's failure stays on its run.
The pumpkin terminal command works again when a folder window and a workspace window both have the same folder open. It talks to the folder window.

Pumpkin 2026.10.438

A window opened from a saved workspace file shows your older chats again. They move in once from the workspace's first folder (F-243).
If a folder window on that folder is open, its chats stay there. Close it and reopen the workspace window to bring them over.

Pumpkin 2026.10.437

Interrupted chat recovery keeps your choices when you switch windows. A damaged saved chat no longer blocks the others, and chats that finish before shutdown no longer leave stale recovery offers.

Pumpkin 2026.10.436

Headless server: a run can be limited to its caller's own subscription accounts, so one person's plan never serves another; Auto then picks only from those accounts and API keys.
Headless server: semantic search over the code folders you configure; results outside a run's allowed folders are left out of its answers.

Pumpkin 2026.10.435

Chats already running no longer show up as interrupted. Recovery prompts stay in the main editor window instead of covering browser windows.

Pumpkin 2026.10.434

The chat beside the Data page SQL editor can read and write the query again. Since 2026.10.405 every turn said "The SQL editor beside this chat is unavailable this turn".

Pumpkin 2026.10.433

Headless server: a run starting no longer breaks the tools of runs already going; their tool calls had been failing with a grant error.

Pumpkin 2026.10.432

Claude Code and Codex chats show how long the model thought again. Their thinking rows read "Thought for 0ms" every time.

Pumpkin 2026.10.431

On a busy machine, open chats no longer reload their model list, agents and context over and over. Settings reads that ran late used to trigger that reload every time.

Pumpkin 2026.10.430

The Media tab no longer slows a chat whose data folder holds thousands of files. It shows the newest 500 plus every file the chat itself read or saved.

Pumpkin 2026.10.429

Long chats with many tool calls run much lighter: Pumpkin no longer re-reads every earlier tool call on every step to find repeats.

Pumpkin 2026.10.428

Data answers now ask for clear results in chat, with a short preview for large sets and a link to the full data.

Pumpkin 2026.10.427

Fireworks models such as Kimi K3 work again. Fireworks moved to a new root certificate this month and Pumpkin refused every request to it; the new roots are now trusted for every provider.
Headless runs can hand their tools environment variables: a new run option reads NAME=value lines from a file, and every shell command, hook and MCP child of that run sees them.

Pumpkin 2026.10.426

Old temp folders that Pumpkin could not delete before, because parts of them were locked against changes, are now cleaned up too.

Pumpkin 2026.10.425

A window opened from a saved workspace file now keeps its own chat list, stored next to that file (F-243).
A folder window and a workspace window that start with the same folder now work at the same time instead of taking turns.
Reordering the folders of a workspace no longer moves its chats. Chats saved before this update stay in the first folder and show in that folder's window.
The idle worktree cleanup reads every chat list on the repo first, so a chat open in another window keeps its worktree.

Pumpkin 2026.10.424

Opening a chat is faster on a busy machine. Pumpkin remembers which providers have a key instead of asking the desktop again on every open.

Pumpkin 2026.10.423

Auto no longer keeps a chat's model just because it once had a prompt cache. Once that cache has expired, the model competes at full price, so an idle chat can move to a cheaper one.

Pumpkin 2026.10.422

A loop never restarts a finished run: a late answer, a timed-out question or a helper can no longer start new work in an old run; the next run picks the answer up.
Slack stays connected: a brief relay refusal after a dropped connection is retried, and Settings says Disconnected while the relay refuses this machine.
A loop question keeps waiting through a short Slack error, and closing a question by hand keeps the replies already in its thread.
Answers read from Slack are kept when the run cannot start, and answering in the Loop questions pane no longer hides other replies in that thread.

Pumpkin 2026.10.421

A folder you exclude from semantic search stays excluded, even when a file inside it asks to include everything. One wildcard rule now covers every new worktree.

Pumpkin 2026.10.420

Busy chats save faster. The rules and setup text a chat started with are kept in one small file beside the chat, not copied into every save. Nothing is deleted.

Pumpkin 2026.10.419

Long chats load earlier messages even when the page is too short to scroll. Loading waits while you type an answer.
Unsent question choices survive chat redraws. Sending an answer twice is blocked while the first send is still running.

Pumpkin 2026.10.418

Loop question answers given after a run ended now reach the next run.

Pumpkin 2026.10.417

A workspace whose folder order changed no longer leaves an old record behind that stops its window from starting chats after an update (F-243).
When two records claim one window, the error now names both chat folders instead of saying the workspace is not ready.

Pumpkin 2026.10.416

Chats that received many subagent results save faster: delivered results are no longer written again on every save.

Pumpkin 2026.10.415

The editor side now carries the loop answer fix from 2026.10.414, so Loops tab questions show and update correctly.

Pumpkin 2026.10.414

Answering a loop question after its run ended now reaches the next run, which uses the answer instead of asking again.

Pumpkin 2026.10.413

Send Feedback no longer stops with "Registered canonical source cannot be read" when the form checks your account. You can send screenshots and logs even when folder settings cannot be read.

Pumpkin 2026.10.412

Media now lists only images, video, audio, PDFs, data files and canvases. Code and config files the chat read or edited no longer show there.

Pumpkin 2026.10.411

On macOS the chat list now sees a chat another Pumpkin process changed, like Linux and Windows do.

Pumpkin 2026.10.410

Idle worktree cleanup no longer deletes git worktrees you made yourself, and never removes a folder open in any Pumpkin window (F-242).
To also clean up worktrees made by hand or by agents, turn on Reap Unowned in Settings › Pumpkin › Cleanup.

Pumpkin 2026.10.409

When Claude Code or Codex cannot start on your computer, the chat now says so plainly. It no longer says the model "rejected the request as invalid (HTTP 400)" (F-241).

Pumpkin 2026.10.408

A loop no longer asks the same question again: every later run sees the questions still waiting for you and is told not to repeat them, so one question is asked once, not once per run.
Time a loop spends waiting for your answer no longer counts against its run limit, so a question with no deadline really waits instead of ending the run as "timed out" after 6 hours.

Pumpkin 2026.10.407

A Claude Code or Codex turn no longer fails with "could not start claude: no such file or directory" when the program is updating itself at that moment. Pumpkin retries and finds it again (F-241).

Pumpkin 2026.10.406

New command View: Open Canvas lists every canvas in this workspace, newest first, and opens the one you pick in front.

Pumpkin 2026.10.405

The chat list, new chats and chat summaries load much faster in workspaces with thousands of saved chats. Pumpkin now re-reads only chats that changed instead of every chat each time.

Pumpkin 2026.10.404

The chat beside the Data page SQL editor connects again. Since 2026.10.386 it showed "Could not connect this chat to the SQL editor" on every Data tab.

Pumpkin 2026.10.403

Agents now get a canvas guide, like Cursor's. They reach for a canvas when the data is the answer instead of a long table, label every chart, and add filters such as a dropdown by account.
A new canvas opens in front where you can see it, and your cursor stays in the chat. Asking the agent to open a canvas that is already open brings it to the front (F-240).

Pumpkin 2026.10.402

Browser tool search finds the open-tab tool again. Proxy options now show their help text.

Pumpkin 2026.10.401

On Windows, opening a WSL folder works again. Since 2026.10.392 it failed at once with "Error resolving authority", because the setup script had grown too long for Windows to start.

Pumpkin 2026.10.400

When a subscription limit holds Auto's first choice, Auto now switches to an equal-skill model on another ready subscription before using a paid API key.
Each turn now records which route it used, any holds it skipped, and why a fallback model won.

Pumpkin 2026.10.399

Background browser tabs now get a real page size before the first click, including popups, without raising their window.
Page scripts and function sources work in browser tools. First-load requests appear in network logs, DevTools waits for its real state, and missing request times no longer show 1969.
Still pages keep the last distinct recording frame at the chosen interval.
Browser tools can use an HTTP proxy for each isolated chat session. Other chats keep their own routes, HTTPS checks stay on, and older builds refuse proxy opens safely.

Pumpkin 2026.10.398

When Auto cannot reach its helper, it now prefers a capable model on a subscription you already pay for over a cheaper pay-per-token model.
It also keeps your current model even after a login was re-added.

Pumpkin 2026.10.397

Models now has a full JSON edit view with defaults, saved overrides and effective values, including models that are off. Saving refreshes the shown values without pinning unchanged defaults.
Agents can list all models and their defaults, even when a model is off or not ready to run.
Clicking Got it on an old flag clears it without starting old work after you moved on. Timed-out question checks also yield to newer input, even during a slow send.

Pumpkin 2026.10.396

Scripts can now call an MCP tool through Pumpkin itself: pumpkin agent mcp call runs one tool with your saved sign-in, no chat and no tokens, and pumpkin agent mcp tools lists what each server offers.
Loop pre-checks run with your login shell environment and are told where the pumpkin command is, so a pre-check can ask an MCP server whether there is any work before a paid run starts.
On a WSL or SSH host, the pumpkin command bundled with the runtime reaches the background service directly.

Pumpkin 2026.10.395

When Codex hits a rate limit after its own retries, Pumpkin now treats it as a rate limit and switches to the next ready account or provider instead of stopping with an invalid-request error.

Pumpkin 2026.10.394

New Cleanup page in Settings. It gathers every way Pumpkin frees disk space, each with a days dropdown: archiving idle chats, old worktrees and old temp files.
You can now have Pumpkin delete the saved data files of idle chats, or delete archived chats after a while. Both are off until you turn them on.
A chat whose saved files include a link can be archived again. Before, it stayed in the chat list forever.

Pumpkin 2026.10.393

Pumpkin now removes old, saved git worktrees of the repos next to your workspace too. Agents make those by hand, and they filled the disk.
On Linux and WSL, Pumpkin deletes your own temp files left untouched for 3 days that no program is using. WSL never cleans them itself.
A web sync stopped halfway no longer leaves a big scratch copy behind.

Pumpkin 2026.10.392

Pumpkin no longer gets stuck at startup with "startup transaction busy or unsafe" after WSL restarts or a window dies mid-start.
It clears the leftover startup lock itself and connects. You no longer delete the lock folder by hand.

Pumpkin 2026.10.391

An older chat stays in the chat list when you send a new message, even while the list refreshes.
Right-click a link in chat and choose Copy Link to copy its URL or file path.

Pumpkin 2026.10.390

One frozen browser tab no longer makes every browser action in its chat time out. Only that tab reports an error, and it works again once it recovers.
Browser and canvas tools no longer fail with "Presentation authorization changed during lookup" when another chat opens or the phone is watching.

Pumpkin 2026.10.389

Remote can open very long chats again. Chats over about 12 MB used to show up empty or fail to open. The first page no longer loses its space to the list of older pages.

Pumpkin 2026.10.388

Remote on the phone and in a desktop browser now shows each turn's cost on one line right under that turn, as the IDE does, instead of a stack of cost lines at the bottom of the chat.
Remote names models the way the IDE does. A chat running on a Codex or Claude Code account shows the model name, never a long id.
Remote in a desktop browser keeps the model control in the message box footer, where the IDE has it.

Pumpkin 2026.10.387

Remote on the phone now has a Search chats box above the chat list. It matches chat titles at once, then chat text from your computer.
Remote on a desktop browser searches chat text too, not only titles. A computer on an older Pumpkin still matches titles and says it needs an update.

Pumpkin 2026.10.386

The chat beside the SQL editor on the Data page now sees the editor on its own: connection, access, SQL, selection, last result and last error reach it every turn, with no extra click.
That chat can write SQL straight into the editor and run it under its own access to the connection. The write is one undo step, and the status line says the chat wrote it.
The chat opens by default on a new Data tab. The quick helper's answer gains Continue in chat, which carries your question and its answer into the chat's message box.
Send to chat names its two files query and rows so the chips no longer look the same.

Pumpkin 2026.10.385

Memory’s Create doc button now asks for a name, creates the doc, and opens it. Existing docs open without being overwritten.
Chat tabs can wrap onto more rows. Turn on Wrap chat tabs in Settings.
Chat max width can keep wide chats easier to read. The default is 0, with no limit.
Search settings now label saved code folders and the folder path. Add folder is the single folder picker and keeps your other saved folders.
Vue 2.7 files now use their own compiler when indexed. A bad template no longer drops the file’s script, and source line links are more accurate.
Loops set to silence no-op runs now stay quiet when they only update their own memory files. Failed runs and other recorded file edits still post.

Pumpkin 2026.10.384

Media now sits beside Activity and Context. View chat images, videos, audio, PDFs, data files and canvases in a list or grid.
Click a Media item to open it. Sent text files open the copy the chat used.

Pumpkin 2026.10.383

A loop question that times out is no longer lost. It stays in Loop questions, also when Slack is off or not linked, and an answer you give there later reaches the loop's next run.
An answer sent from Loop questions now reaches the next run even when Slack could not deliver it.
Mark resolved with a note now passes that note to the loop's next run.
Questions kept only in Pumpkin no longer show an Open in Slack button.

Pumpkin 2026.10.382

A workspace whose VS Code profile changed without a handover is refused right away again, instead of after a 15 second wait.

Pumpkin 2026.10.381

A workspace last opened in another VS Code profile now opens in your default profile instead of failing until you restart several times.
A command-line run with no profile uses the profile whose window has the workspace.
A second Pumpkin with its own user data folder on the same computer user now stops at once with a clear message, instead of retrying and failing three times.

Pumpkin 2026.10.380

When Claude Code says you are out of usage credits, the turn now moves to your next login or another provider instead of stopping with an invalid-request error.

Pumpkin 2026.10.379

Auto picks a new model when a watcher, job or subagent result wakes a chat or loop and the last model's subscription ran out. Before, the woken turn kept the old model.

Pumpkin 2026.10.378

Adding a model to all accounts or copying models to a new account now uses each account's name. Existing saved model names stay unchanged.

Pumpkin 2026.10.377

Custom memory folders keep their own doc layout in prompts and new folder notes.
New Slack chats use your workspace's worktree setting. Existing chats keep their folder.

Pumpkin 2026.10.376

Extensions from the Open VSX gallery install again. Pumpkin no longer stops signed extensions with "cannot verify the extension signature".

Pumpkin 2026.10.375

The headless server now starts inside Docker and ECS containers. It used to refuse the container root filesystem (overlayfs) and exit before any run could begin.

Pumpkin 2026.10.374

A loop stopped by its time limit now says so on an open question, not "aborted by user". The question stays open for the next run.

Pumpkin 2026.10.373

Muse Wonder is now in the model list with a Meta key. It supports tools, images, and thinking for hard coding and long tasks. Costs use Meta's Standard prices.

Pumpkin 2026.10.372

Drop files into your message to add references at that spot, with selected lines kept. Send keeps your draft while files load.
Slack drafts keep bold text, links, lists, quotes and code.
Numbered lists continued after a label keep their numbers in chat and copied text.
Copying a table selection no longer adds a blank column at the next cell's edge.
Help > Release Notes opens the full version history on the web.

Pumpkin 2026.10.371

Gemini 4 Argon is now in the model list. It works with tools, images, and low, medium, or high thinking. Costs follow Google's dated prices.

Pumpkin 2026.10.370

Silent model calls now retry. Loops no longer lose whole runs while waiting for the first reply. Stop stays quick when a stream hangs.

Pumpkin 2026.10.369

Auto now judges a loop iteration by its instructions, not the standing frame, so loops get a model fit for the job.
Auto no longer loses the recent conversation to tool-only turns; a short follow-up is judged with the task it continues.
Auto tells its picker which subagent profile the parent chose, so review and expert work stop landing on low-skill models.
The built-in picker prompt rates work by a skill ladder with examples instead of a fixed "skill 8" rule, and moves up after a complaint or a repeated request.

Pumpkin 2026.10.368

Copying part of a chat table cell keeps just the selected text, including text in bold.
Repeated malformed tool replies can switch to another allowed model. Failed replies keep their reported costs, and no bad tool call runs.

Pumpkin 2026.10.367

Remote chats can use the connected desktop's browser tools, even when that chat has no desktop tab open.
Remote chat refreshes old history links once and keeps loaded text on screen when a page fails.
Phone model and run status are easier to read. Chat buttons stay together, and their icons draw correctly.

Pumpkin 2026.10.366

Models shows usage windows, reset times and past usage or limit hits. New accounts can copy models; new models can be added to all accounts of the same kind.
The headless server and bundled terminal tools start even when the host has no tools on PATH.

Pumpkin 2026.10.365

New headless server for Linux boxes and containers: no desktop, no keyring. Keys live in an encrypted file or the system secret service. The Linux release ships it as a separate server archive.
The terminal client can add Codex or Claude Code accounts, sign them in with a device code, pin a model to one account, and finish an MCP sign-in by pasting the redirect link.
Terminal runs get a typed streaming output with clear error codes, an effort setting, throwaway sessions, attached files and images, and flags that limit which folders a run may read or write.
On the server, runs can be capped and timed out, old sessions are swept, tools stay inside allowed folders, child programs lose secret variables, and shell commands run in a kernel sandbox.

Pumpkin 2026.10.364

Models now shows each account's usage windows, reset times and reading age. Usage history keeps past readings and limit hits, and tells reported caps apart from refused turns.
Adding an account can copy saved models from another account of the same kind. Adding a model can add it to all accounts of that kind, keeping existing choices.

Pumpkin 2026.10.363

Project memory now keeps short facts about the project. Old chat summaries are cleared on update. Feature and bug notes stay in feature memory.

Pumpkin 2026.10.362

Loops now fail before calling a model you turned off. A card names the model and shows how to fix the loop.

Pumpkin 2026.10.361

Sending and starting a turn avoid saving the same queue twice.
Jobs with secrets finish after their safe logs reach disk.

Pumpkin 2026.10.360

The built editor now includes the Mac code watcher CPU and RAM checks.

Pumpkin 2026.10.359

Sending a message no longer fails with "sources changed during refresh" while Pumpkin is busy. Pumpkin now waits up to 10 seconds for settings to settle instead of giving up after a split second.
Editing rules, skills or MCP files no longer makes every open folder reload its settings. Semantic Search and other lists stop refreshing every few seconds.

Pumpkin 2026.10.358

On Mac, a busy code scan can keep working through quiet steps, instead of being stopped after ten minutes with no log output.
Mac code watchers now get the same RAM checks as Linux watchers.
Code search health shows the real stall reason and file counts, including when a watcher stays stalled after one restart.

Pumpkin 2026.10.357

Loop run summaries in your Slack DM now come from the Pumpkin bot, so Slack notifies you. Before, they landed in your note-to-self DM with no notification.
Set the loop question time limit to 0 in Settings and a loop question waits until you answer it, up to the run's max runtime.
If a loop question cannot be posted to Slack, its card keeps waiting and Pumpkin retries the post every 20 seconds, instead of giving up at once.

Pumpkin 2026.10.356

Questions now wait 1 hour by default.
When a question's wait ends, the agent re-checks whether it still needs your answer. It no longer picks product or design choices for you; it finishes what it can and says what it needs.

Pumpkin 2026.10.355

Loops ignore Question wait time. A loop question waits for the loop's own question time limit, so Off never stalls an unattended loop.

Pumpkin 2026.10.354

Agent browser windows stay behind the IDE and keep the last window’s spot and size across chats and screens.
On Windows, opening browser windows again no longer makes them grow or drift across the screen.

Pumpkin 2026.10.353

Cost popups stay hidden across the chat list while you rename a chat. They return when you finish or cancel.
If a chat disappears during a name save, its late reply no longer clears another chat’s new name.

Pumpkin 2026.10.352

Update checks renew your pumpkin.club sign-in when it has run out, so they no longer fail after your computer sleeps or sits idle.
When pumpkin.club rejects an old sign-in, Pumpkin renews it at once instead of waiting for your next chat. The once-a-minute background check is gone.

Pumpkin 2026.10.351

Image previews and ZIP files now have a Download button. Pick a folder on your computer, then use Show in Folder to find the saved file.
Failed downloads keep your old files and folders safe.

Pumpkin 2026.10.350

The Data page connection form now shows the CA certificate and server name fields under TLS. Versions 346 to 349 carried the runtime side only.

Pumpkin 2026.10.349

If the background service files on disk get damaged, Pumpkin now sets them aside and reinstalls them when it starts.
Before, Pumpkin showed "Runtime inventory does not match actual files" at every start, and Restart did not help.

Pumpkin 2026.10.348

Custom GitHub SSH server lookups reuse your saved GitHub or gh sign-in. Missing or stale sign-in falls back safely, and draft builds stay out of the choices.

Pumpkin 2026.10.347

When the model-picking helper is slow or fails, Auto asks a second helper on another provider before falling back to its fixed rule.

Pumpkin 2026.10.346

Data connections can trust a private certificate authority. Paste or choose a CA certificate (PEM) under TLS, for example the Amazon RDS bundle, and keep certificate checks on.
A server name field covers PostgreSQL, SQL Server and Trino connections that go through a tunnel.
When a connection fails on the server certificate, the message now says what to do, and the connection form warns when TLS is off.

Pumpkin 2026.10.345

New workspace folders start with rules and context on. Folder switches keep your saved off choices.
A process with a missing skill shows a warning and lets chat continue. Settings and save messages show the warning too.

Pumpkin 2026.10.344

Chats no longer scan shared saved files or add their names to each step. Shared files stay available through the data tools.
Semantic search no longer skips folders inside a chosen project when they share a name with a project you left out, such as docs or video. They are searched again after indexing restarts.

Pumpkin 2026.10.343

Auto no longer asks the helper which model to use when a job, watcher or subagent result arrives; the model already doing the work keeps the turn.

Pumpkin 2026.10.342

Chats no longer scan shared saved files or add their names to each step. Shared files stay available through the data tools.

Pumpkin 2026.10.341

Add separate code folders in Semantic Search without choosing their common parent. Each folder keeps its own index, subfolder picks and exclusions. Edits from other windows stay safe.

Pumpkin 2026.10.340

Choose how long questions wait in Settings → Agent behavior: 15 or 30 minutes, 1, 2, 6, 12, 24, 48 or 72 hours, or Off. Changes apply to open questions. Off keeps them waiting through restarts.

Pumpkin 2026.10.339

The Subfolders list in Semantic Search keeps your place while it refreshes or saves a choice.
The thin strip above the title bar matches your workspace color, including when the window loses focus.

Pumpkin 2026.10.338

Long chats can switch to another ready model to sum up past work when the first runs out of credits or fails. The chat keeps its chosen model and saved history.

Pumpkin 2026.10.337

Pumpkin keeps you signed in after your laptop sleeps. Update checks and pumpkin.club settings no longer fail with an expired sign-in until you start a chat.

Pumpkin 2026.10.336

Stopping work or reaching a time limit no longer lets late errors crash Pumpkin.
Large code scans keep running while busy, even when they are quiet for more than ten minutes.
Settings load faster with many chats open, while connection state and live account details stay fresh.

Pumpkin 2026.10.335

The Context tab loads for a chat Pumpkin has closed instead of failing.
The long-turn reminder log no longer reads as if the final answer was hidden. Only a repeat of the answer is dropped.

Pumpkin 2026.10.334

Clicking a chat notification opens the chat in the window that already has its folder, even inside a multi-folder workspace. It no longer switches the folder of another window.

Pumpkin 2026.10.333

Shortened rules and tool descriptions are no longer thrown away over small wording changes. The check now looks for lost rules and exact names, paths and numbers only.
The merged rule block and shortened skill and tool descriptions now actually reach new chats.

Pumpkin 2026.10.332

Long chats keep your words. Saved answers keep their source.
Folds must make requests smaller. Bad summaries leave history intact.
Saved tool results support exact reads. Needed reads stay allowed.
Chat copies keep saved links. Queued answers survive a reload.
When room runs out, old replies go before the summary of earlier work.

Pumpkin 2026.10.331

Chat about SQL beside your query and results, with Pumpkin's usual models, tools and chat history.
Add exact SQL, saved schema context and results to your draft, then use a chat answer in the SQL editor after reviewing it. Nothing runs by itself.

Pumpkin 2026.10.329

A refused big re-read no longer stays refused after Auto hands the turn to another model, and hook messages no longer let an unchanged re-read through.

Pumpkin 2026.10.328

When context fills up mid-task, the agent is told which of its big reads were summarized away and to pass file paths instead of pasting file content into the chat.
An identical big re-read of unchanged content is refused every time, not just once. Edited files are read again as normal.
Loading a past chat's errors now also lists failed tool calls, and cache markers no longer look like summaries.

Pumpkin 2026.10.327

Auto now sends the widest-scope work, such as whole-company code audits, millions of lines or all history, to the highest-skill model you have instead of the cheapest model that could manage.
Each Auto turn now saves the models Auto weighed, up to eight, always including the one it chose, so a chat's choices can be checked later.

Pumpkin 2026.10.326

SQL help works with Claude Code and Codex when it uses saved database maps.
Duplicate a data connection, edit the copy, and keep its saved password without changing the source.

Pumpkin 2026.10.325

Semantic search starts its local Qdrant before the workspace watcher, even when a project also has its own index. A watcher no longer retries forever with "connection refused".

Pumpkin 2026.10.324

When WSL prints an error instead of starting the remote bridge, Pumpkin names that error and retries on its own. The window no longer needs a reload.

Pumpkin 2026.10.323

Data shares keep the full SQL and shown result in chat files. Open chat tabs get the draft before the side bar.
Type # in chat to find a data connection or saved query by name. Pick one to add its link text.
SQL help shows the model it uses. You can pick a model, and Write and Fix use your saved database map.
SQL help asks for missing facts and keeps your model choice when the host reconnects. Review its answer before you accept SQL.

Pumpkin 2026.10.322

Auto judges the full task before input size and cost. Scope changes keep the skill needed for work still in progress.
Auto treats room for chat input as a limit, rather than proof that a model can do the work.
Auto adds the current model and other provider choices when they meet the helper's skill bar, then compares route and cache costs.

Pumpkin 2026.10.321

Startup keeps the first frame error when closing a failed remote bridge. A fresh helper can connect while retries continue.
Old startup errors stay cleared when the agent host connects.

Pumpkin 2026.10.320

Finished child agents can take follow-up messages in the same chat and keep their history.
Later child work keeps its cost in the parent chat. Failed saves keep the costs until they can be saved.
Queued shell jobs keep their chat data paths and secrets.

Pumpkin 2026.10.319

Auto seeks cheap models that fit the task across providers. It includes the current model only when its skill fits the work.
Auto compares keeping the current model's cached input with sending the full chat to a new model.

Pumpkin 2026.10.318

Send to Pumpkin adds selected code to a chat draft with its file and lines. It keeps the copied text when the file changes.
Paste copied code or terminal text into chat to keep its source. Shift-paste keeps plain text. Review the draft before sending.
With code selected, Ctrl/Cmd+Shift+L adds it to the current chat. Ctrl/Cmd+L adds it to a new chat.

Pumpkin 2026.10.317

The Data page is now called Pumpkin Data. The sidebar has a Data tab with your connections and five most recently saved queries.
Every Data form has Back to query and Escape so you can return to SQL.
Stop frees the query controls at once and warns if Pumpkin cannot reach the server to cancel.
Ask AI can write SQL, fix the last error, or explain a query using your saved database map. Review and accept its SQL before using it.
Send to chat adds SQL, a short result table and the full grid as a CSV file to a new or chosen chat draft. You review the draft before sending.
New saved queries start with their connection's access. Projects with the same name show their ids.

Pumpkin 2026.10.316

Claude Code account usage updates from chat turns when Claude sends a reading. Each account keeps its own counts, and old readings keep their age.
Finished chats release their Codex and Claude Code child processes. Active chats keep running, and only a few idle children stay open.
Semantic searches stop after 45 seconds if they cannot finish, so a stuck search does not hold up the chat.
Large first indexes using Ollama can pause when the measured work would take more than six hours. Choose Restart indexing to keep going.
Index progress shows chunks and clearer time estimates. Paused indexes stay paused until you choose to resume them.

Pumpkin 2026.10.315

Auto now picks the least capable model that can do the turn, so simple asks run on cheaper models. Among subscription models the cheaper one wins, since it uses less of your quota.

Pumpkin 2026.10.314

The Default model setting no longer has its own Min skill box. Auto's lowest skill is set once, in Auto Agent Selection, and the Default model row now links there.

Pumpkin 2026.10.313

A chat no longer stops with a "request too large" error when the agent's last step was very big, such as writing a large file.
Pumpkin now shortens that step in what it sends and keeps going. Your saved chat does not change.

Pumpkin 2026.10.312

Windows installers are back. Version 2026.10.311 shipped for Mac and Linux only; this version brings data connections to Windows too.

Pumpkin 2026.10.311

Connect to MySQL, PostgreSQL, Oracle, SQL Server and Trino from the new Data page, and check each connection before you save it.
Write SQL, run it, stop it, and see results in a grid you can sort, copy and export as CSV.
Save queries as files and open them again from the Data page.
Connections and saved queries are off for agents until you allow them for all chats or chosen projects, as read or read and write.
Map a database's tables and columns, add notes, and let allowed agents read the map and run queries.

Pumpkin 2026.10.310

Saved model choices now apply to your own agent files, including files that share a name with a built-in agent.
To give an agent file's own model line control again, remove that agent's entry from the pumpkin.subagents.overrides setting.

Pumpkin 2026.10.309

Start terminal chats and runs with Autopilot goals, guardrails and a round limit, or change them on a saved chat.
Choose whether Autopilot may answer questions. See spent rounds, its last choice and why it paused.
Runs wait for Autopilot to finish and can resume after their round limit is raised.

Pumpkin 2026.10.308

Sites that ask for a username and password now show a sign-in box in the agent browser; what you type stays with you and the agent never sees it.
Browser navigation results no longer include a stray internal response object.

Pumpkin 2026.10.307

Replies to questions reach Codex when its turn ends before it reads the answer.
Stop still stops the turn while a lost reply is being recovered.

Pumpkin 2026.10.306

Release notes stay readable after an update or restart, even while the agent starts.
Side chats keep their question links so replies reach the right ask.

Pumpkin 2026.10.305

MCP tools keep fresh host values after the shell PATH check.
Values changed during that check no longer get stuck on old values.

Pumpkin 2026.10.304

Feedback asks now require a clear Send choice. Typed replies stay distinct from skipped asks.
Reused answers now say that no new card was shown.

Pumpkin 2026.10.303

Codex and MCP tools on Linux and WSL now find Node installed through nvm, without a system-wide Node link.
The local proxy also uses your login-shell PATH on macOS.

Pumpkin 2026.10.302

Claude Code's new limit message now switches the chat to another ready route and explains why, while keeping your chosen model.

Pumpkin 2026.10.301

MCP sign-in works after a server URL changes, while keeping each URL's saved sign-in separate.
Failed MCP sign-in callbacks show a clear reason and a way to try again.

Pumpkin 2026.10.300

When an MCP tool returns a picture (for example a feedback screenshot), the agent now sees the picture instead of a text note about its size.
Editing and resending a message while other messages were queued no longer risks a "Failed to save queued input" notice.
A chat's cost total no longer fails with a metadata error when one of its sub-chats was archived a moment earlier.

Pumpkin 2026.10.299

Claude Code sign-in explains that the code you paste stays hidden, and offers Paste code into the terminal, Show the terminal, and How to paste buttons.
Paste code into the terminal sends your copied code to the Claude sign-in terminal even when another terminal is selected, and tells you if that terminal has closed.
How to paste walks through the Command Palette route with Terminal: Paste into Active Terminal and pressing Enter.

Pumpkin 2026.10.298

Chat update notices now show only builds offered by the app updater, so checking cannot leave a false update offer behind.

Pumpkin 2026.10.297

Claude Code and Codex chats keep their browser, canvas and other on-demand tools for the whole turn; using one no longer restarts the chat behind the scenes.
A restarted Claude Code or Codex chat no longer tries tool names it cannot call.
When OpenAI has run out of credits the error says so, and an Auto turn that gave up after four tries says that instead of naming a ten-minute budget.

Pumpkin 2026.10.296

The app menu shows Check for Updates just once when an update is ready to restart.
Fixed a settings load error in desktop builds.

Pumpkin 2026.10.295

Settings › Models has an "Edit as JSON" row that opens your skill ratings or descriptions in the settings file, so you can change many models at once.

Pumpkin 2026.10.294

Add more than one Codex or Claude Code account in Settings and sign in from Pumpkin; each account keeps its own login.
When one account hits its limit, the chat carries on with another ready account for the same model, keeping finished tool results.
Each account shows its own usage and reset times; old counts show their age and missing counts say "Usage unknown".
Sign out of or remove one account without touching the others; removing never deletes login files outside Pumpkin.
Account settings belong to the computer running the work, including WSL and SSH hosts.

Pumpkin 2026.10.293

Canvas tabs work again after Pumpkin restarts or updates, without reopening them.
Servers the agent started can be brought back after a restart: ask in chat, or click Restart.
A new Background jobs list in Activity shows running servers with Stop and Restart buttons.

Pumpkin 2026.10.292

A chat no longer stops with "requested bundle is unavailable" when Pumpkin's model helper restarts in the middle of a turn; the turn carries on with the current prompts.

Pumpkin 2026.10.291

Right-click a chat tab for Close, Close Others, Close to the Right, Close to the Left, Close All and Copy Chat Link.

Pumpkin 2026.10.290

Send Feedback works while Pumpkin is updating. Before, it could fail saying the chat was finishing on the previous Pumpkin runtime.

Pumpkin 2026.10.289

When one subscription is out of quota, Auto now picks a fitting model on another subscription with quota left instead of paying API rates.
Auto sees which subscription or API each model would run on and whether that subscription has quota.

Pumpkin 2026.10.288

Open an existing chat from outside Pumpkin: run pumpkin agent sessions open with the chat's id, or open a pumpkin link to it. Pumpkin comes to the front and shows that chat.

Pumpkin 2026.10.287

The Autopilot button is clearly on: it turns into an orange pill that says Autopilot on. Clicking it no longer leaves an orange outline that looks like on.
Clicking the Autopilot button while its goals panel is open now just closes the panel instead of turning Autopilot on or off.
Chats with Autopilot on show a small robot mark in the Chats list.

Pumpkin 2026.10.286

Loop run summaries in Slack now show bold, links, headings, lists and tables properly instead of stray asterisks and raw link text.

Pumpkin 2026.10.285

When a turn finishes, the agent's steps fold into one Worked for line above the answer. Click it to see the steps. Turn this off in Settings, Display.

Pumpkin 2026.10.284

Saved files with names like publisher>exchange>DSP.png are now kept on Mac and Linux instead of being skipped with a red error.

Pumpkin 2026.10.283

You can edit the same message again after editing it once, without reopening the chat.
Dragging chat tabs and dragging chats into projects now shows what you are moving and where it will land, and the drop shows up at once.

Pumpkin 2026.10.282

In Settings, the model box and the Auto Min skill box now sit side by side instead of stacked.
The switch that keeps the search index fresh moved from General to the Semantic search page.

Pumpkin 2026.10.281

Agent browser windows now remember their screen position and size across moves and reopens, and stay behind the main window when opening.

Pumpkin 2026.10.280

Clicking What's new on the update notice keeps the notice open, so Restart to Update is still there after you read the notes.
Check for Updates stays in the menu after an update has downloaded, so you can look for a newer one.

Pumpkin 2026.10.279

Clear in the Projects picker now removes the chat from its projects right away. The picker stays open so you can choose another.
Renaming a chat in the Chats list no longer shows the chat's tooltip over the name you are typing.

Pumpkin 2026.10.278

An agent started from Slack keeps reporting into its thread when the Slack relay restarts or drops the connection: your machine posts the progress and the answer itself instead of the thread failing.
Chats started from Slack are titled with the request you typed, not the "Started from Slack by …" header.

Pumpkin 2026.10.277

The Agents settings page is tidier: labelled Model and Description lines, a description box that shows the whole text, and tool lists shown as a count.
Your own agent files get a Model picker on the Agents page, including Auto with a minimum skill.

Pumpkin 2026.10.276

Canvases now offer everything Cursor canvases do: the same theme colours and helpers, CalloutToneIcon and CanvasChevron, and every prop type.
The agent can read typed definitions for every canvas part, and each canvas build is type-checked. Wrong props come back as line and column errors; the canvas still opens.
Canvas buttons can open a file at a line and column, or select a range.
Charts drawn with raw recharts and no colour now use the theme colours instead of black.
Canvas updates are about 3.7 times faster and swap in place without reloading the page, so the scroll position and saved values stay. A second canvas opens about 2.4 times faster.

Pumpkin 2026.10.275

Rules from all your rule files can now be merged into one shorter block. Lines the merge would lose are added back word for word.
Short tool and skill descriptions are now shortened too, one batch per tool server.
Tool servers with only a few tools but very long inputs are now listed instead of fully loaded, which saves room in every chat.

Pumpkin 2026.10.274

Auto now retries Codex outages that were shown as invalid requests, then switches to another provider.
Picked models can switch to another provider after retries fail. This is on by default; turn it off with requests.providerFallback. The choice uses skill and price, and the chat says why it switched.

Pumpkin 2026.10.273

On a Mac, the semantic-search watcher no longer opens a file handle for every file in your code. A big code root could fill the machine's file table and keep Pumpkin from starting at all.
On a Mac, the watcher now checks folders for changes every 30 seconds instead.
A short shortage of open files while Pumpkin starts no longer blocks the chat; the start waits a few seconds and tries again.

Pumpkin 2026.10.272

Web storage sync now carries your projects with your chats: name, color, description, memory notes and deletes.
A synced chat shows its project on your other computers. Project folders stay on each computer.

Pumpkin 2026.10.271

The bottom bar's loop question count drops as soon as you answer or resolve a question. Before, it stayed until something else changed.
Loop Questions pane shows each question cleanly: a title, the question in bold, its details, and the choices as buttons. Click a choice to fill in your answer.

Pumpkin 2026.10.270

A sparkle button in the Source Control message box writes the message for you. It describes your staged changes, or all changes when nothing is staged.

Pumpkin 2026.10.269

Chat tabs show the same status marks as the Chats list, such as Waiting when a chat needs your answer.
Drag chat tabs to reorder them. The order stays after a restart.
The scrollbar in the Chats list no longer covers the time on the right of each chat.

Pumpkin 2026.10.268

Loop questions accept more Slack reply styles, such as 1. a or a, b. A reply Pumpkin cannot read now gets a note in the thread saying what to send.
Mark resolved in the Loop Questions pane no longer throws away answers already in the Slack thread. It answers with them, or passes them to the loop.

Pumpkin 2026.10.267

MCP tools keep working when their server reconnects in the middle of a reply. Before, every later call in that reply failed with "MCP connection replaced".

Pumpkin 2026.10.266

On macOS, search indexing starts right after Pumpkin restarts. Before, indexing from the previous session that was still shutting down could block it for several minutes.
When indexing fails, the status now says Pumpkin retries within 5 minutes, which is how long it really waits.

Pumpkin 2026.10.265

Project settings: code folders picked in a window attached to WSL or SSH now save. Before, a Windows desktop flipped the slashes in the Linux path and the save was refused.
Agents can now read and edit project settings through the Pumpkin settings tool: name, color, description and code folders. Project memory and which chats belong stay untouched.

Pumpkin 2026.10.264

On Windows, semantic search setup now installs Ollama without stalling on a hidden winget prompt, and carries on when Ollama is already installed.
If the Windows installer still fails, the message shows winget's error code so you can look it up.

Pumpkin 2026.10.263

When an agent tries out a canvas, its button clicks no longer pull you to another chat or file. Files open in a background tab, and chats wait for you as a notification.
Canvases can ask for the same quiet open on any button.

Pumpkin 2026.10.262

Resizing the chat panel keeps your place in the conversation instead of jumping.
The message box is one even colour with its frame on every computer.
Attached files show as a single neat line: name, size, and a close button.
New Display section in Settings. Choose the message box font, text size, and how tall it grows before it scrolls. Chat list and tab options moved there too.

Pumpkin 2026.10.261

The local model no longer stays stuck on "Another Pumpkin window is using the local model" after an update when no other window is open.
A leftover process from the old version used to keep the local model reserved; Pumpkin now takes it back right away.

Pumpkin 2026.10.260

Grok no longer repeats its earlier thinking at the start of every step, so long Grok turns think less, cost less and fold their context far less often.

Pumpkin 2026.10.259

Auto can take a lowest skill level: pick Auto, then a number in the Min skill row. Also in Settings for the default model and subagents, and in the loop form.

Pumpkin 2026.10.258

Editing a long message opens the edit box at full size, so you can see the whole message without scrolling.

Pumpkin 2026.10.257

A chat that was mid-turn when Pumpkin updated no longer stops when it tries to start a subagent. It moves to the new version and carries on there by itself.

Pumpkin 2026.10.256

Settings pages keep showing your configuration when Pumpkin is busy. A slow refresh shows a short note with Retry instead of hiding everything.
Settings pages load faster on a busy machine.
The edit button on your own messages no longer covers the end of the text.

Pumpkin 2026.10.255

When the agent asks a question, the card now says you can type a longer answer in the message box below.

Pumpkin 2026.10.254

Agent browser tabs opened during UI testing no longer steal focus or switch your active tab away from the chat.

Pumpkin 2026.10.253

Canvases can be shared. Click Share beside Design, or ask the agent, and you get a pumpkin.club link anyone can open in a browser. Sharing again updates the same link.
A canvas keeps the editor theme after it reloads. Before, a reload could leave it with the wrong background.

Pumpkin 2026.10.252

The command line can now list a chat's open questions, with how to answer each. Scripts that start chats in the background can see one is waiting and reply.
Switching chats with a chat tab no longer leaves the old chat gray in the Chats list. Only the open chat is marked.
Chat tabs now sit at the top of the chat itself, not across the Chats list.

Pumpkin 2026.10.251

The model chip under the chat box now names the model Auto picked, such as Opus 5.5 · Claude Code, instead of a long internal id.

Pumpkin 2026.10.250

The Claude Code model picker no longer says Fable needs usage credits. That warning was wrong.

Pumpkin 2026.10.249

Loops keep reading a question's Slack thread for three days after the first answer. Later replies reach the next run.
Send a loop a message from the Loop Questions pane or from Message loop in the Loops tab. The next run gets it, even with Slack off.

Pumpkin 2026.10.248

Check for Updates now shows its progress (checking, downloading with a percentage) and always ends with a result or the reason it failed.
After an update has downloaded, Check for Updates stays in the menu and looks for an even newer version.

Pumpkin 2026.10.247

Codex chats no longer stop after the model starts a helper: it now uses Pumpkin's own subagents, which wake the chat when they finish.
Codex models can no longer edit files or sleep outside Pumpkin, so every change shows up in the chat.

Pumpkin 2026.10.246

Click "N loop questions" in the bottom bar to open a Loop Questions pane, like the old Cursor extension.
Each open question has an answer box, Send answer, Mark resolved, Open in Slack and Open chat.
The pane also lists answers you already sent and a history of resolved questions.

Pumpkin 2026.10.245

Project memory now saves long notes within its size limit. Empty or failed note updates keep the old note and no longer put a failure notice after your chat reply.

Pumpkin 2026.10.244

The model and mode dropdown works again. It went empty, and new chats showed "Invalid wire payload (at 1.routes.0.quota: unknown key)", when Claude Code reported a weekly limit for one model.

Pumpkin 2026.10.243

Topics now sit below Agents in the Activity column instead of at the top.
Chat topics no longer pick up build numbers or version numbers, which only added noise to note searches.

Pumpkin 2026.10.242

Usage logs now name the Codex or Claude Code model used, not a long saved key. Auto logs the route that ran.

Pumpkin 2026.10.241

Each chat now keeps a short list of topics, shown in the Activity column. You can remove a topic with its x or add your own.
Memory notes are now found by the chat's topics instead of the last message, and at most three are shown.

Pumpkin 2026.10.240

Auto no longer flips between Claude Code and Codex every turn. Switching threw away the cached conversation each time.

Pumpkin 2026.10.239

Short replies now find notes for the work at hand from recent chat and active tasks. Clear new topics use your latest words.

Pumpkin 2026.10.238

Projects now have a full settings page for descriptions and code folders, plus shared notes that give each chat a head start.

Pumpkin 2026.10.237

Remote folder picks now keep workspace rules on, Context views stay fresh, memory search finds local project notes, and stalled context rewrites recover without needless resets.

Pumpkin 2026.10.236

Clicking the Pumpkin icon after closing every window now reopens your last used workspace instead of an empty window.

Pumpkin 2026.10.235

On Windows, an update now installs its background service while it waits for a restart, so chat is ready seconds after the restart instead of about a minute later.

Pumpkin 2026.10.234

Pinned chats now sort by recent activity: the chat you or the agent used last moves to the top of the Pinned list.

Pumpkin 2026.10.233

Long tool-heavy turns on Codex and Grok no longer summarize their own work on every round and then stop with "re-ran the same reads after 3 context compactions".
Between tool rounds, the summary trigger now expects the next few rounds, not a whole new turn, so a fresh summary no longer re-opens the gate at once.
A reply can reserve at most a quarter of the model's context window (68K on a 272K Codex model, not 128K), leaving far more room for the conversation.
When a turn's own results get summarized away three times, the rest of that turn runs in the model's full window (long context) instead of folding again.
On Auto, a model without a wider window hands a read-compact-re-read loop to another model instead of stopping.
Re-running a small command such as git status after a summary no longer counts as a read loop.

Pumpkin 2026.10.231

Auto now notes when a subscription has plenty of quota left before it resets, so spare plan quota gets used instead of wasted.
Auto knows whether it is picking for a main chat or a subagent, and favors larger context models for main chats.
The Auto model picker uses an updated, fuller set of instructions from pumpkin.club.

Pumpkin 2026.10.230

Large folders using local search models no longer restart from scratch if a batch takes too long. Slow batches split into smaller pieces and retry, and search indexing resumes from where it left off.

Pumpkin 2026.10.229

When a loop has questions waiting for you, the bottom bar shows "N loop questions". Click it, pick one, and its chat opens so you can answer right there. The count drops once you answer.

Pumpkin 2026.10.228

File edits on Windows no longer fail over and over with "Editor transaction cleanup is unacknowledged". The edit is written, and a real failure now shows its own message.

Pumpkin 2026.10.227

Pumpkin now keeps your machine awake while turns, subagents, or background jobs are running.
Subagents queue when the machine limit is reached instead of failing with an error.
Turns pause and wait for quota reset when all model routes are exhausted.
Failed voice recordings are kept so you can retry dictation.
Subagents accept an optional output schema to enforce structured responses.

Pumpkin 2026.10.226

Restart to Update no longer waits forever if a window never answers. After one minute Pumpkin installs the downloaded update anyway.

Pumpkin 2026.10.225

Codex and Claude Code models no longer vanish in the middle of a turn when another Pumpkin window has none saved.
When every subscription route hits its limit, Pumpkin now tries another provider before waiting for the limit to reset.

Pumpkin 2026.10.224

Static context condensing no longer cuts off on large sets of rules or permanently disables condensing after a cut-off.

Pumpkin 2026.10.223

Settings has a new Share chat titles switch. Turn it on and the report shows your chat names instead of session ids. With Remote on, the report also links each chat so you can open it in Remote.

Pumpkin 2026.10.222

Pumpkin now keeps your machine awake while turns, subagents, or background jobs are running.
Subagents queue when the machine limit is reached instead of failing with an error.
Turns pause and wait for quota reset when all model routes are exhausted.
Failed voice recordings are kept so you can retry dictation.
Subagents accept an optional output schema to enforce structured responses.

Pumpkin 2026.10.221

A secret the agent shows you or saves to settings is now the real value. Before, the agent could pass a placeholder, so Show and Copy gave [redacted].

Pumpkin 2026.10.220

Stop now also ends the chat's background commands, such as a local database or server. Before, a chat kept by a leftover command showed as running, and Stop spun forever.

Pumpkin 2026.10.219

Settings has a new Context management page, after Agent behavior. Window fill, Condense context, and Include rules and context from live there now.

Pumpkin 2026.10.218

The Models page now shows Claude Code plan use for the session, the week, and Fable, plus when each one resets.
Subagent cards now show which provider serves a model family, such as Codex or direct API routes.
Subagents without an explicit model pin fall back to a model with working sign-in or keys instead of failing on a missing API key.

Pumpkin 2026.10.217

Opening a browser tab to look at the UI no longer pulls you off the chat when you are previewing a file.
The Models page now shows Claude Code plan use for the session, the week, and Fable, plus when each one resets.
Unattended loop questions now require plain-language context, everyday terms, and clear decision choices with trade-offs.

Pumpkin 2026.10.216

Messages sent while a chat loads its agent now keep their side-thread routing.

Pumpkin 2026.10.215

Windows: long-running command logs are now capped at 64 MB like on Mac and Linux.
Windows: the fix that stops Pumpkin closing an unrelated program after a command finishes now actually ships.

Pumpkin 2026.10.214

Auto picks use the real input limit and price for normal and long modes across chats, child agents, side threads and model upgrades.
Model pins in agent files and child calls can add :long to request the larger window. The chosen mode stays saved with the chat.

Pumpkin 2026.10.213

Attaching a large text file (for example a CSV with thousands of rows) no longer fails with "Event exceeds the frame limit".
Windows: stopping a finished command can no longer close an unrelated program that reused its process number.

Pumpkin 2026.10.212

Background command logs stop at 64 MB: the start and the latest output are kept, the middle is dropped with a note. One log had reached 2.4 GB.
Every six hours Pumpkin removes git worktrees that are on origin, clean, and untouched for two days. Trees with edits, unpushed work, or saved chats stay.
A new worktree setting sets how many idle days count; zero turns the cleanup off.
New session worktrees share the main checkout's node_modules instead of needing a fresh install (Pumpkin › Worktree › Share Node Modules).
Auto now picks a high-skill model for questions that need investigating a machine or codebase, such as "what is filling up the disk?", instead of treating them as simple lookups.

Pumpkin 2026.10.211

File edits no longer fail the first time on a Mac when the system touches a file's details.
If a file changes just before an edit is saved, the agent now reads it and retries instead of stopping.

Pumpkin 2026.10.210

Signed-in tools keep working when a chat finishes on the previous Pumpkin version during an update.

Pumpkin 2026.10.209

Chat tabs: keep your open chats as tabs across the top of the chat panel so you can switch between active chats. Turn it on in Settings > General > Chat tabs.
Inline code in chat messages now has a rounded background pill that matches your theme.

Pumpkin 2026.10.208

Pumpkin finds tools from task words and shows the best matches first. Exact tool names come first, and tools your agent cannot use stay hidden.
Patched five packages used by MCP and network tools.

Pumpkin 2026.10.207

Loop questions accept Slack answers sent in separate thread replies. A single question also accepts its shown question id before the answer.

Pumpkin 2026.10.206

Windows: no more console windows flashing on every prompt. Pumpkin's git checkpoints now run hidden.

Pumpkin 2026.10.205

A second window opened while Pumpkin is already running in another window now starts its agent. Before, it could stop with "Operation store is not natively admitted for this client".
When the agent cannot start in a window, the chat shows why, with Show log and Reload Window, instead of "Waiting for the Pumpkin agent host to start" forever.

Pumpkin 2026.10.204

The agent writes fewer useless checks for your code. Each one must prove real behavior or bad input, never just that a setting or list entry exists.

Pumpkin 2026.10.203

Settings fit narrow windows. The right panel button sits beside the left panel button.
Plan updates save even when the editor is busy or closed.
Web sync and Slack no longer fail when two new runs make the machine id at the same time.

Pumpkin 2026.10.202

A chat whose background run was cancelled can be deleted, stopped and written to again. Before, it refused every action with "reserved by run … (cancelled)".

Pumpkin 2026.10.201

Settings > Agent behavior: choose where rules and context come from. All folders the agent works in, or only folders you tick from a list of your workspace folders, plus any you add.

Pumpkin 2026.10.200

Autopilot no longer calls a chat finished while a goal you set for it is still unmet, and says which goal it is judging.
The reason Autopilot gives for calling a chat finished is no longer cut off in the middle of a word.

Pumpkin 2026.10.199

Settings > Agent behavior has a Condense context choice (smartest, local, off), next to Communication style.
On Windows, a command that changes into another project folder now brings that project's rules and skills into the chat.

Pumpkin 2026.10.198

Agent persistence no longer has an End goal choice. How far the work goes now comes from what you ask and your project's own rules.

Pumpkin 2026.10.197

A chat no longer fails on every message with "tool_use.input: Input should be an object" after a model sends a broken tool call. Old chats stuck this way work again.

Pumpkin 2026.10.196

Codex and Claude Code chats no longer refuse to send with a false "too big" error after a busy round.

Pumpkin 2026.10.195

Switching chats no longer flashes a different part of the conversation for a frame while the transcript settles.

Pumpkin 2026.10.194

Agent browsers on slow pages no longer fail after 30 seconds with "Mutation acknowledgement unavailable". They wait for the page or give a clear timeout.

Pumpkin 2026.10.193

Rules, project notes, skill and tool descriptions are condensed once by a model and cached, so each chat spends fewer tokens on setup.
A new setting, Context: condense, lets you use the local model for this or turn it off.
The memory index now shows a short per-project catalog plus the notes relevant to your message, instead of the first 45 titles.
The workspace data list shows one summary line plus files relevant to your message, instead of the 20 newest.
Built-in tool descriptions and process steps are shorter, with the same rules.

Pumpkin 2026.10.192

Loop questions in Slack show clear, bold questions and spaced answer choices. Replies sent through Pumpkin can answer them.
Loops keep their owner when a Mac name changes. The Take over menu names the folder that owns a loop, and pause or resume does not undo a new claim.
Sign-in windows opened by agent browser pages now appear. Failed takeovers keep the page usable, and background popups do not steal typing focus.

Pumpkin 2026.10.191

When a chat's setup is too big for a model, the error now lists the largest parts, so you know what to turn off.
Auto no longer picks a model whose window cannot hold the chat's setup.

Pumpkin 2026.10.190

Release Notes no longer fails when the editor build number is used. Notes open for the installed Pumpkin version.

Pumpkin 2026.10.189

Reading a file in another project no longer loads that project's rules into the chat. Editing a file there, or running a command in it, still does.
Rules from another project now arrive once instead of twice. Very large rules, and rules already shown, are listed by file name instead of pasted in full.

Pumpkin 2026.10.188

On Windows, opening changes for a file no longer shows a false edit when only line endings differ from the saved copy.

Pumpkin 2026.10.187

A short file read hiccup on the Mac no longer stops a chat with a "source could not be read" error.
Clicking a file in the Activity column shows its changes, even for files in another folder or already committed.
Each project header and project filter offers a button to start a new chat inside that project.
A new chat no longer fails before its first reply because two size counts disagreed. A setup too big for the model now says what to trim.
The end-goal setting now gives way to your own instructions or the project's rules, and a plan can also be saved in the project.
On a WSL or SSH window, Retry on a tool server no longer opens a Windows app store prompt, and Settings shows the real connection error.

Pumpkin 2026.10.186

A canvas button can open another chat, or start a new chat with this canvas attached.

Pumpkin 2026.10.185

On Windows, Settings opens even when a saved tool server path differs only in drive-letter case. One refused server no longer blanks the whole page.
A local tool server starts in your project folder, so a relative command or argument in its setup finds the file it names.

Pumpkin 2026.10.184

When the agent writes the same plan more than once, the chat shows one plan card with Open and Build. Earlier writes show as updated. Renaming a plan keeps one card, and Open goes to the new file.

Pumpkin 2026.10.183

A chat that keeps running the same tool call over and over (eight times in one turn, whatever it returns) now stops instead of running for an hour. Waiting and polling tools are not counted.
The hidden repeat warning now shows the start of the earlier result, warns about each looping call, and no longer says the result is still on screen.
When a big file read was summarized away to save room, reading the whole file again is refused once with a note to read only the lines needed.

Pumpkin 2026.10.182

The agent can put a Slack draft in your composer for you to send. It does not post. A Slack setting can add a Sent from Pumpkin line. That setting stays off until you turn it on.

Pumpkin 2026.10.180

Tool results from connected servers now arrive as the text the server sent. JSON comes through as JSON, so it can be read directly instead of being unwrapped first.

Pumpkin 2026.10.179

Drag a chat onto a project in Chats to move it there. Drag it onto a date group to take it out of that project.

Pumpkin 2026.10.178

If a file changes just before Pumpkin saves an edit, the edit now fails cleanly. Nothing is written, and the agent can read the file again and retry.

Pumpkin 2026.10.177

Settings no longer lists your archived chats on the page. The Archive idle chats setting has a Show archived button that opens the chat picker to restore one.
A loop can use Auto as its agent, and a loop can name a model by its short name such as astra instead of a full provider id.

Pumpkin 2026.10.176

Canvases have a Design button. Turn it on, click the parts you want changed, write a note, and it goes to the chat with exactly which parts you picked.
Loop questions now reach you from the Pumpkin bot in Slack, so they can ping you. Reply in that thread to answer the loop without starting a new job.
Long questions keep their answer choices in Slack. If the bot cannot post, Pumpkin logs why and falls back to your own DM.

Pumpkin 2026.10.175

A loop no longer asks the same question twice in one turn. Its Slack question is sent from the window that owns the loop, not swallowed by another window.
Tools stay ready when a new folder adds rules during a Claude Code turn. A fresh Claude Code session also sees valid tool names in its chat history.
Settings has a switch for the line that marks Slack posts sent by Pumpkin. It is on by default. Turning it off sends your text alone.

Pumpkin 2026.10.174

Canvases get built-in parts: cards, notes, stats, tables, bar, line and pie charts, diffs, forms, to-do lists and diagrams.
Canvas parts follow your theme, and charts come labelled with a legend and values on hover.
A canvas button can send a prompt to its chat, and a canvas can open a file from your workspace.
Canvas settings such as the selected tab are saved with the canvas, so they come back after a restart.
A misspelled canvas part fails right away with the line number and the right name, instead of a blank page.

Pumpkin 2026.10.173

Browser actions on huge pages no longer flood the chat. The agent sees a short preview and the full page is saved to a file, so long browser work stops looping.

Pumpkin 2026.10.172

A chart canvas shows its chart on the first try. A chart sized "100% tall" used to come out empty; it now gets a sensible height, and the agent is told so.
Canvas screenshots wait for the canvas's data to load, so the agent sees the real chart instead of an empty one.
Canvas warnings from chart libraries now reach the agent, so it can fix a broken layout without guessing.

Pumpkin 2026.10.171

Scrollbars in Pumpkin settings and panels match a dark theme instead of showing a bright bar.
Settings buttons no longer have blue outlines, and the last blue accents in the default Dark 2026 and Light 2026 themes are orange.

Pumpkin 2026.10.170

On a Mac, What's new on a downloaded update opens the list of changes instead of an "Invalid Pumpkin version" error.

Pumpkin 2026.10.169

Archived chats no longer push the General settings out of reach. They sit at the bottom, folded into one line with the count and size.
Show opens the archived chats with a search box, newest first, ten at a time with Show more.

Pumpkin 2026.10.168

MCP sign-in failures now show the real reason, such as a missing client id when the server has no self-registration, instead of an opaque internal error.
A failed or cancelled MCP sign-in no longer breaks the next attempt for the same server; retrying works without restarting.
A Cursor-style OAuth client id on an MCP server entry is honored when Cursor compatibility is on, so the same config works in both editors.

Pumpkin 2026.10.167

A slow or timed-out Codex probe no longer blocks turns in your thread.

Pumpkin 2026.10.166

Pumpkin: Sign In now asks you to open a folder first when no folder is open, instead of showing that the command was not found.

Pumpkin 2026.10.165

Organizations can now limit which AI providers and models their members may use. Admins set the policy on pumpkin.club under Team; blocked models disappear from every picker and cannot be selected.

Pumpkin 2026.10.164

Chat titles no longer come out as sentences such as "X is the subject being discussed". Pumpkin keeps just the subject, or keeps the old title when the reply is a sentence.

Pumpkin 2026.10.163

Settings › Auto Agent Selection now has separate minimum and maximum skill ranges for main agents and subagents.

Pumpkin 2026.10.162

Help dropdown now has a single update section instead of duplicate items (F-37).
Resending or retrying an edited message after cancelling a starting turn now succeeds (F-40).

Pumpkin 2026.10.161

Loops stay visible while the previous Pumpkin version finishes their work after an update. The new version waits before it runs them.

Pumpkin 2026.10.160

WSL upgrades now replace old keepalive processes, so old runtime files can be freed.
A failed chat handoff now retries while keeping its saved work safe.

Pumpkin 2026.10.159

A question asked by a loop now shows the question alert, like any other chat. Loop finishes and failures stay quiet.
When a loop question is not sent to Slack, the log says why (for example, Slack not linked).

Pumpkin 2026.10.158

Sign in a computer from your signed-in phone or another desktop. Copy the sign-in link, check the code and account, then approve. Only the waiting computer gets signed in.

Pumpkin 2026.10.157

Adding a Codex model no longer fails with "Invalid wire payload (at rateLimits: unknown key)" for accounts signed in with ChatGPT (F-41).

Pumpkin 2026.10.156

Help > Show Release Notes now opens the release notes for the installed version when Pumpkin is up to date (F-38).

Pumpkin 2026.10.155

A model reached through Codex, Claude Code or Bedrock now shares the skill rating of that model by API key. One rating in Settings covers every way to reach it.
Settings: images now shrink to 1024 pixels on the long edge by default.

Pumpkin 2026.10.154

Shell commands now find tools installed in common user folders on Windows, Mac, and Linux, even when Pumpkin starts with a limited PATH.

Pumpkin 2026.10.153

Pinned chats stay out of auto archive, even when a prior archive stopped partway through.
Settings has a new Archive idle chats control. Choose how many days a chat can sit unused before it is archived, or turn auto archive off.
Your archived chats are listed in Settings with the date they were last used. Unarchive brings one back with its saved work.
New option: keep chats that are in a project out of auto archive. It is on by default, so filing a chat under a project also keeps it out of the archive.
Opening a chat now keeps it live while you read it. Restoring a chat refreshes its activity time, so it will not archive again straight away.
Restoring the same archived chat twice now shares one restore job and opens it once. Pumpkin shows progress while large chats restore.

Pumpkin 2026.10.152

Mac updates now close stray app processes while leaving the updater and running jobs alive. If an update fails to install, its log is copied into the main log when Pumpkin next starts.

Pumpkin 2026.10.151

Chats no longer scan the whole system temp folder for project rules. Projects inside it still load their own rules. Folders you add yourself keep their rules.

Pumpkin 2026.10.150

On Windows, installing a new Pumpkin version copies runtime files in small groups, so chat setup finishes sooner after an update.
A large repair of old chats now refreshes the chat list once, even when the first few finish before the rest.

Pumpkin 2026.10.149

A long chat now keeps the finished work and results from earlier in the conversation when older messages are folded into a summary, so answers no longer repeat work that was already done.
When Pumpkin notices you are repeating something, the notice now quotes the earlier tool result and points to where it is in the chat, so you can see what was already returned.

Pumpkin 2026.10.148

Exporting a long chat as JSON from the command line now works. Before, any export larger than about one megabyte failed with an invalid payload error.
Very large files can now be compared with their original version from before the chat changed them.
On Windows, a new version copies its files in small groups instead of one by one, so the chat should be ready sooner after an update.

Pumpkin 2026.10.147

A chat you opened on another computer no longer starts running here, so it cannot interrupt work still going on that machine.
Project files are no longer picked up from a Windows user folder, which removes a slow folder scan at the start of every message.
A patch with the wrong line counts now says exactly what is wrong with it, instead of a vague safety message.

Pumpkin 2026.10.146

Auto now keeps the same good model between your messages instead of picking a new one every time, so the chat it already knows is not thrown away.
Auto moves to a stronger model only when the work is clearly hard or the last message failed, for example after repeated tool errors.

Pumpkin 2026.10.145

Edits that add a line above or below an existing line now work, including at the very top or bottom of a file. Before, any edit whose new text kept the search text was refused.
Running the same edit a second time is still refused, with a message that says the edit is already applied, so text is not added twice.

Pumpkin 2026.10.144

Windows: clicking Restart to update now installs the downloaded update right away instead of starting another download when a newer build is out.
Windows and Mac: an expired sign-in no longer throws away a downloaded update every five minutes, so Restart no longer closes Pumpkin without updating.

Pumpkin 2026.10.143

When an edit is refused for repeating its own search text, the message now shows that text and explains how to add a line next to it.
That refusal used to name no fix, and one reported case read it as "this file cannot be edited". The message now names the fix.

Pumpkin 2026.10.142

A running chat is saved once per round instead of on every event, and only the changed part is written. Thirty saves on an 8 MB chat went from 289 MB written to 1.6 MB.
Web-sync passes skip chats that have not changed instead of walking every file.
The remote status heartbeat reuses its 50 rows and lists nothing when nothing changed.
Archiving or restoring a chat now republishes the remote status.
The service log reports save, scan and publish counts once a minute.

Pumpkin 2026.10.141

Claude Code Opus and Sonnet now show their real 1M-token context instead of 200k, so big chats no longer fail with a false "setup is too large" error. Haiku stays at 200k.

Pumpkin 2026.10.140

Loops tab: the row hover now shows the last pre-check result and how long ago it ran.
Loops can choose to skip runs when their pre-check fails, and can start promptly when you answer an asked question on Slack.

Pumpkin 2026.10.139

Loops no longer pop desktop notifications, play sounds, or raise the attention badge when a run finishes, fails, or asks a question. Loops report through their run summary instead.

Pumpkin 2026.10.138

Restoring a checkpoint is faster. It uses a handful of git processes and reads only the files that changed, instead of one process and one read per file.
When git is not installed, the checkpoint shows a clear notice and a clear error instead of a confusing failure.
An upgrade no longer retries forever when the previous service is stuck shutting down. The stuck service is replaced within 30 seconds.
The startup status now says the new service is handing over from the previous one, not that it is stopping.
Old server downloads, stale payloads, leftover pid files, old CachedData, temp folders, and old runtimes are cleaned up instead of piling up on disk.
The service log rotates so it no longer grows without limit.
When a download fails because the disk is full, the error says how much space is free and how much is needed.

Pumpkin 2026.10.137

Settings has a new Agent behavior page. Thinking level, communication style, and question frequency moved there from General.
Pick how many code comments the agent writes and how much checking code it adds: low, medium, high, or your own words.

Pumpkin 2026.10.136

Opening three or more WSL or SSH windows at once after an update no longer shows a "busy or unsafe" error. Each window waits its turn.

Pumpkin 2026.10.135

When you send a message in a chat from yesterday, it moves up to Today right away. Other chats under Today stay where they are.

Pumpkin 2026.10.134

A chat no longer stops with an error when the desktop app restarts in the middle of a reply. It waits for the desktop to come back, then carries on.

Pumpkin 2026.10.133

A chat no longer shows a "could not be refreshed" error when the auto archive moves one of its helper chats away. The chat just stops counting that helper.

Pumpkin 2026.10.132

Clicking a chat now puts the cursor in the message box. Double-click still lets you rename the chat.
Show more works on question cards after you answer them.
Right-click a chat to mark it unread, or mark an unread chat as read.

Pumpkin 2026.10.131

Scrolling up while a reply is still being written now stops the chat from following it, even after a single small turn of the mouse wheel. Scroll back to the bottom to follow again.

Pumpkin 2026.10.130

New setting: Image size limit. Pick the largest size, in pixels on the long side, for images sent to a model. Leave it on Model default to keep each model's own limit.
Find it in Settings under General. Choose 768, 1024, 1568, 2048 or 2576 pixels, or type your own number from 100 to 8000.
Images keep their shape when shrunk, and your saved attachments never change.

Pumpkin 2026.10.129

Opening several WSL or SSH windows at once after an update no longer fails with "startup transaction busy". Extra windows wait for the first one to start the server, then connect.

Pumpkin 2026.10.128

Chats idle for five days are now zipped and moved out of the chat list, so old chats stop slowing the list down. Set the number of days in Settings, or set it to 0 to turn auto archive off.
Archived chats are not lost. Run Pumpkin: Show Archived Chats to see them and pick one to restore; it comes back exactly as it was, with its helper chats.
The archive sweep runs two minutes after Pumpkin starts and then every six hours. Pumpkin: Archive Idle Chats Now does it right away.
Pinned chats, loops, chats with pending work and chats with a running turn are never archived. A chat is only removed once its zip has been read back and every file matches.
Press Up in an empty chat box to walk back through your sent prompts. Down walks forward and clears the box past the newest prompt.

Pumpkin 2026.10.127

A stuck agent on Auto now moves to a stronger model instead of stopping. A pinned model still stops as before.
Reading a file again after its result was compacted away no longer counts as a repeat. The agent gets one note to read small windows and write findings down. Three such cycles in one turn is a loop.

Pumpkin 2026.10.126

A Claude Code or Codex model can now be the side model. Chat summaries, titles and other helper calls no longer fail with "the workspace root is required for CLI models".
A helper call on the same Claude Code or Codex model as the chat no longer restarts or blocks the chat's own model session.

Pumpkin 2026.10.125

Token counting for a model round no longer re-reads the whole transcript every time. Counts are remembered per item, so a round on a large chat costs milliseconds instead of blocking the service.
Searching the chat list no longer parses every saved transcript on the main thread per keystroke. Each file is parsed once in a worker thread and its searchable text is cached.
Chat-list requests arriving together now share one scan of the sessions folder instead of each walking it, so the list stops stacking up 10 to 50 second reads.
The service log reports once a minute how much token counting and transcript parsing it did.

Pumpkin 2026.10.124

Builds fetch their full source in one pass, so a second source fetch cannot block the fixes for search and host links.

Pumpkin 2026.10.123

Chat lists, search and chat delete now skip saved loop runs and their child chats before reading their files. Loop runs can still be opened from Loops.

Pumpkin 2026.10.122

Chats can now load large tool lists on demand when they would make a request too big. All tools stay in reach, and work already done stays done.

Pumpkin 2026.10.121

A paired host keeps its reply alive while server setup runs. Slow setup can report progress before it ends.

Pumpkin 2026.10.120

The host now ships with the current shared code for paired windows. Build checks cover the real link code before release.

Pumpkin 2026.10.119

You can now open a remote window on another computer that is paired with your account. It uses the same secure link your phone uses. Files and the terminal run on that computer.
A new command pairs this desktop with another computer by code, the same way a phone pairs.
Chat inside a paired remote window is not ready yet. The window says so. Chat still works in a local window.

Pumpkin 2026.10.117

Running loop rows now keep their age up to date. Idle or hidden rows stop their clocks.
The search helper now reports the code it runs, so its build checks match the shipped code.

Pumpkin 2026.10.116

Deleting a chat is instant: the row leaves the list on the click. If the delete is refused (a child chat still running, a worktree prompt cancelled), the row comes back where it was with the reason.
Deleting no longer reads the whole transcript of the chat twice just to check its worktree; it uses the saved summary.

Pumpkin 2026.10.115

A full disk no longer leaves Pumpkin stuck at "acquiring service" after space is freed. A half-written service ownership record is cleaned up or taken over instead of blocking every start.
A slow background service no longer makes the editor drop its connection and reconnect in a loop ("unsubscribe did not settle, 128 in flight"). Cleanup calls keep reserved slots and are retried.
Chat-list reads are paced so they cannot fill the request pool, and the log names which calls are pending and which replies were slow.

Pumpkin 2026.10.114

Loops tab: clicking a running loop opens its live chat right away, and its row shows how long it has been running. The arrow still opens the list of files and runs.

Pumpkin 2026.10.113

Search shows a full scan as done only once all queued projects pass. One project ending no longer marks the whole scan done.
Failed scans stay shown. Split log lines keep their full counts.

Pumpkin 2026.10.112

Chats synced from another computer no longer run here. Their scheduled tasks, waits and subagents stay with the computer that last used the chat.
Sending a message in such a chat takes it over. Subagents that ran on the other computer are marked as finished there, not waited on.

Pumpkin 2026.10.111

Semantic search: files the index now ignores (editor folders, build output, lock files) are dropped on the next walk, so clean projects stop re-walking on every start.

Pumpkin 2026.10.110

Settings › Models: the Stand-in model row no longer offers Nemotron 3.5 Lightning. GPT OSS 120B is the one recommended cheap choice, matching the setting text.

Pumpkin 2026.10.109

Loops tab: stopping a running iteration asks for confirmation first, and uses the solid square icon from chat.

Pumpkin 2026.10.108

Semantic search no longer stops its index worker at 6 GB on machines with lots of memory. It may use up to 30 % of RAM, so a large code folder can finish its first scan.

Pumpkin 2026.10.107

Chat tags are now called projects. The chat menu, the picker, and the sidebar filter and group buttons all say project.
The phone app shows projects in the chat menu and settings too.

Pumpkin 2026.10.106

A slash command in the middle of a message now opens the list and runs. It no longer has to be the first word. A path such as src/file still stays plain text.

Pumpkin 2026.10.105

Auto: when the local model is choosing Auto's model for a chat, the chat shows a note once a month saying it is not recommended for that and suggesting GPT OSS 120B as the stand-in.

Pumpkin 2026.10.104

Semantic search uses far less memory when it starts: it reads the index one page at a time and only once for the whole workspace, instead of loading all of it for every project.

Pumpkin 2026.10.103

Claude Code models no longer get stuck when the model asks for two tools in one step. Claude Code runs them one at a time, so Pumpkin now runs the first and comes back for the next.

Pumpkin 2026.10.102

Auto no longer ends a turn after repeated tool failures. With no more capable model it tries another model at least as capable, or carries on with the current one and says why.
Auto blames a model for a tool failure only when it is the model's doing: bad arguments, or the same tool failing the same way again. A first-time error does not count.
The trigger is now three blamed rounds in six, up from two in five.

Pumpkin 2026.10.101

Semantic search health says Unknown instead of Healthy when this window cannot check a workspace, and repairs name the folder to open instead of saying the workspace is missing.

Pumpkin 2026.10.100

Linux and WSL: shell commands that turn on stop-on-error now report their real exit code. A command that worked no longer comes back as exit 1.

Pumpkin 2026.10.99

The stand-in model setting no longer recommends Nemotron 3.5 Lightning. It timed out and gave unusable answers as the Auto helper; GPT OSS 120B is the recommended cheap choice.

Pumpkin 2026.10.98

Auto: a model description you wrote yourself now counts more than the catalog text, skill or price. Say a model is good for math and Auto sends math there.

Pumpkin 2026.10.97

Settings › Models: a model whose description or skill rating you changed now has a Reset to default button that puts both back to the catalog values.

Pumpkin 2026.10.96

Codex and Claude Code are now found when installed through pnpm, asdf, mise, their own installers, or a custom home folder, so Add Model no longer says not installed.

Pumpkin 2026.10.95

Auto no longer forces main chats onto skill-8 models. Simple requests can use a cheap or specialised model again, so your own model descriptions matter.

Pumpkin 2026.10.94

Code search keeps its RAM cap steady on retries, so a smaller worker can finish its scan. Stop logs show the known cause.

Pumpkin 2026.10.93

Linux: a new install now starts when the home folder is writable by your own private group (the normal Ubuntu and Fedora setup). Sign In no longer says command not found there.

Pumpkin 2026.10.92

Windows: a cancelled in-app update no longer leaves files behind that break the next update's background service start.

Pumpkin 2026.10.91

Remote: a computer no longer drops off pumpkin.club/remote for an hour after a burst of reconnects. It registers once per account, and any "come back later" from the site is capped at five minutes.

Pumpkin 2026.10.90

Chat names are the subject only, two to five words. Renames keep the chat's topic instead of naming the latest step.

Pumpkin 2026.10.89

Busy chats keep their link. Errors keep the real cause.
Search picks the right index. Shared indexes skip false errors.
Shell tasks keep true results.

Pumpkin 2026.10.88

During an update, a helper chat still finishing on the old service no longer shows a red "waiting for its owner" message. Real faults still show.

Pumpkin 2026.10.87

Mac and Windows: the update check right after launch no longer gives up while sign-in is still loading. It runs again on its own, so you need not click Check for Updates.

Pumpkin 2026.10.86

The model picker shows one row per model. Claude Code, Codex, Bedrock and direct API keys that reach the same model are routes behind it, tried in your chosen order.
Subscriptions are used first by default. When a Codex or Claude Code quota runs out mid-run, the turn carries on through the next route and the route comes back after its reset time.
Settings shows a Provider order list with up and down buttons, plus quota chips: Codex shows percent used and the reset time; Claude Code shows a quota hit and when it resets.
A Bedrock model that reads images no longer says text only next to its images chip.

Pumpkin 2026.10.84

Auto keeps main chats on strong models and moves up after repeated tool faults. Your chosen models and Stop still win.
On Windows, a failed local access check waits for its helper to close before cleanup.

Pumpkin 2026.10.83

Auto keeps main chats on strong models. Repeated tool faults cause a bounded upgrade, or a clear stop when no stronger model is available. Your chosen models and Stop still win.

Pumpkin 2026.10.82

Waiting rows show what the wait is doing. A job shows its latest output line, run time and output size, with the command on hover. Other waits show their last check.

Pumpkin 2026.10.81

Mac: the menu bar now shows the Pumpkin icon with its alert count. The count has its own switch in Settings, even when Mac alerts are off.

Pumpkin 2026.10.80

Chats copied from another PC now arrive with their full activity log. Before, a long copied chat could not open.
Other chats no longer show a red "could not be refreshed from its saved metadata" message because of one broken chat.
A chat with a missing activity log still opens and shows in the list.

Pumpkin 2026.10.79

The chat modes are renamed: Agent is now Build, Ask is now Discuss. Same tools, same behavior, clearer names. The message box says Build with Pumpkin or Discuss with Pumpkin.

Pumpkin 2026.10.78

Windows: in-app updates now work. Pumpkin finds a newer build and offers Restart to Update; the portable zip offers a download. Older Windows builds never checked, so install this one by hand once.

Pumpkin 2026.10.77

Long chats keep going. Old chat gets cut as needed. The three-reset stop is gone. Saved chat stays intact.

Pumpkin 2026.10.76

Remote web app: New chat now uses the normal message box, so photos, paste, Autopilot, secrets, dictation and slash commands work on the first message.

Pumpkin 2026.10.75

Remote web app on the phone: one-line status row, full-width message box with its buttons beneath, compact queued messages, no blank gaps, and your own messages stand out.

Pumpkin 2026.10.74

Remote web app: a just-sent message no longer disappears from a long chat; pasted image thumbnails show; session settings are a compact popover on desktop and a bottom sheet on phone.

Pumpkin 2026.10.73

A question nobody answers for 15 minutes goes back to the agent: it re-reads its rules, decides where allowed, and may ask once more if the answer is truly needed. Stop still stops.
A question card that timed out shows "unanswered"; only the Decline button shows "declined".

Pumpkin 2026.10.72

Remote web app: a big chat shows its recent messages right away and loads earlier ones behind; each chat row has a menu with Rename, Rename with AI, Auto-rename, Pin, Tags, Fork, Copy id, and Delete.

Pumpkin 2026.10.71

Remote web app: runs of searches and file reads fold into one row like the editor; pairing a new device asks you to name it, so the Devices list no longer shows every device as "Phone".

Pumpkin 2026.10.70

Remote web app: tool rows read like the editor, the message box has an Autopilot toggle and a Secret button, and delivered secrets show in a Show/Copy card.

Pumpkin 2026.10.69

Synced computers no longer show your pumpkin.club account number.

Pumpkin 2026.10.68

Semantic search now really skips unchanged projects when it starts: it reads the whole index instead of only the first thousand entries, so saved progress is trusted instead of thrown away.

Pumpkin 2026.10.67

Chats see all open folders, including folders added after a chat starts.
Alerts keep working when folders change, without repeating old alerts.
Windows on different hosts no longer mix up folders that share the same path.

Pumpkin 2026.10.66

While the desktop is reconnecting, a chat shows one notice instead of an error row that counts into the thousands.
On Windows with WSL, no terminal window opens when Pumpkin reconnects to the background service.

Pumpkin 2026.10.65

Bad settings show their error. Saved keys still work.
Fix settings to restore tools.
Windows starts without slow setup scripts.

Pumpkin 2026.10.64

Sign in on another computer with the same account and your chats sync there. No pairing code needed.
Web storage still has to be turned on for sync to work.
When sync cannot run, Pumpkin now says what went wrong instead of staying quiet.
Approving a request from your phone is still a separate step.

Pumpkin 2026.10.63

Semantic search no longer re-scans every project after a short connection drop to the desktop; the watcher waits up to two minutes for the connection to return.
Search skips work without changes. Its own files stay ignored.

Pumpkin 2026.10.62

Folder changes keep host access. Changed sources get fresh checks.
Saved tools recover after changes. Other windows keep their login.

Pumpkin 2026.10.61

An agent can stop a helper agent immediately, without waiting for it to wind down, and it reports when the helper has settled.

Pumpkin 2026.10.60

Windows avoids slow folder checks. Host access keeps strict checks.
Folder groups load saved tools. Other windows keep their login.
Failed startup reads can recover. Unclear writes are never repeated.

Pumpkin 2026.10.59

Remote lists hide child chats. New chat has its button.
Enter sends on desktop screens. Shift+Enter adds a new line.
Chat settings keep their height. Long lists scroll inside them.
Ctrl+V now pastes your images.

Pumpkin 2026.10.58

The /feedback command now has the agent add details from the chat (errors, steps, model) to your report; you approve the full report before it is sent.

Pumpkin 2026.10.57

Folder groups load saved tools. Other windows keep their login.
Failed startup reads can recover. Folder changes cannot trap work.
File write errors show causes. Unclear writes are never repeated.

Pumpkin 2026.10.56

The desktop "Finished" alert now waits until the agent has truly stopped.
No alert when a turn ends while helpers, background jobs or wait timers will wake the chat again.

Pumpkin 2026.10.55

Long agent turns no longer stop with a "request exceeds its safe budget" error when the request grows too big a second time in one turn.
The agent now uses the newer summary and keeps working, up to three recoveries per turn.
When a request truly cannot fit, the error lists every large part, so the numbers add up.

Pumpkin 2026.10.54

Fixed a crash in long agent runs. After the agent repeated the same tool call three times, its next model request was rejected and the run died.
Background agents using the browser tools hit this most. The hidden repeat notice is now saved in the right place.
Claude requests also repair older chats that were saved with the bad ordering, so those runs can continue.
Typing a slash-feedback command followed by your words sends a bug report to the team, with optional diagnostics, chat export, or screenshots. No setup needed.
When the agent itself wants to send a report, an approval card shows the full message and destination first. Nothing is sent until Send is clicked.

Pumpkin 2026.10.53

Amazon Bedrock Claude models now accept image attachments (screenshots, pasted images, tool screenshots). Other Bedrock families stay text-only with a clear message.

Pumpkin 2026.10.52

Folder changes keep login intact. Other windows keep their keys.
Folder groups load saved settings. One broken folder stays isolated.

Pumpkin 2026.10.50

Synced chats show new titles right away. An open chat keeps downloaded text even when its list details cannot be saved.

Pumpkin 2026.10.49

Phone and web: the chat list now carries each chat's tags and pin, and a paired device can rename a chat, set its tags, or pin it.

Pumpkin 2026.10.48

Windows: Pumpkin now updates itself. Signed in, it downloads the new installer in the background and shows Restart to Update, like Mac. Running work and WSL carry on through the restart.
Windows: the portable zip install still shows Download, since a zip cannot update in place.

Pumpkin 2026.10.47

A file name mentioned in chat that exists in more than one workspace folder now links to the file the agent actually worked on, or to none, instead of always the first folder.

Pumpkin 2026.10.46

The chat menu item for recurring renaming is now always called Toggle auto-renaming. A checkmark means it is on; no checkmark means it is off.

Pumpkin 2026.10.45

The agent's browser stays behind the chat. A hidden browser tab no longer jumps to the front every time the agent clicks or types in it.
In window mode the agent's browser window opens behind the chat and does not take focus; the chat keeps the keyboard when the page tries to grab it.

Pumpkin 2026.10.44

Local stores now publish whole files. Reads cannot see half-written data. Failed writes keep old files safe.

Pumpkin 2026.10.43

Adding a folder to a remote workspace now keeps its settings intact and does not break sign-in reads in other windows.

Pumpkin 2026.10.42

Codex and Claude Code subscription models now accept images: pasted screenshots, attached pictures and tool screenshots reach the model instead of being refused.

Pumpkin 2026.10.41

Clicking a desktop alert now opens the chat in your Pumpkin editor tab whenever one is open, even if that tab was showing a different chat. The side bar opens only when you have no chat tab.

Pumpkin 2026.10.40

Mac: an update check no longer fails with "The server sent an invalid response". If your pumpkin.club sign-in has expired, the message says so and asks you to sign in again.

Pumpkin 2026.10.39

Long tool-heavy chats no longer pay for a background summary every minute or two. Summaries now start when the chat is actually near its limit and each one folds much more history.

Pumpkin 2026.10.38

Adding a second folder to a WSL or SSH window now keeps chat and settings working when the workspace file lives on your desktop.
Changing folders or reconnecting one window no longer clears sign-in reads in other windows.

Pumpkin 2026.10.37

Mac: the update line in the chat pane and in Settings now follows the in-app updater: Checking, Downloading, then Restart to Update once a build is staged. No reopen needed, and the button stays.

Pumpkin 2026.10.36

Auto model picks, chat titles and file descriptions from a Fireworks helper model now arrive instead of timing out. The helper is asked not to reason for these small requests.

Pumpkin 2026.10.35

Codex subscription models now edit files when asked. Codex had been telling the model it could only read, so it refused to make changes.

Pumpkin 2026.10.33

New Auto Agent Selection settings page, right after Models: switch off any model you never want Auto to pick, and set the lowest and highest skill Auto may use.
The page explains that Auto picks by the skill ratings and descriptions on the Models page, so editing a description steers what each model gets picked for.

Pumpkin 2026.10.32

Remote web: the working indicator is the same three pulsing dots as the desktop, and the orange matches pumpkin.club everywhere (links, buttons, badges, theme color).

Pumpkin 2026.10.31

Remote: with remote access on and a paired phone, the desktop keeps its link to the computer alive with every window closed, so the phone stays connected overnight.

Pumpkin 2026.10.30

Mac: when an update has finished downloading, a notice appears right away with a Restart to Update button. One click restarts into the new version; running work carries on.
Mac: the update line in the chat pane and in Settings now says Restart to Update when a version is staged, or Update to check now. Download appears only where the in-app updater is off.

Pumpkin 2026.10.29

Desktop alerts now name the chat and the workspace, such as "Finished: Fix login" from "Pumpkin · my-project", instead of a bare "Turn finished".
Clicking a desktop alert opens the chat where you keep it: an editor tab that shows the chat is brought forward, and the side bar is used only when no tab shows it.

Pumpkin 2026.10.28

A message you send while waiting on other agents now reaches the chat at once. The wait stops. Those agents keep running.

Pumpkin 2026.10.27

Phone remote: the pumpkin.club orange stripe now sits at the top of the remote app too.

Pumpkin 2026.10.25

Phone remote: the Activity list is grouped like the desktop column, newest first, eight rows per group with "Show N more". Saved data files show their size and description instead of a bare name.
A turn now stops after the agent runs the same tool call with the same result six times. Before, it was only warned and could loop for minutes.
Remote access connects in a few seconds on the first try from a new device: direct and relay paths are offered together instead of waiting twelve seconds on a direct path first.

Pumpkin 2026.10.24

A very long chat that still does not fit after one automatic summary now stops with one clear message naming what filled the space, instead of summarizing again every few seconds until it dies.
Chat summaries now record a failed command with its exact error, and no longer turn it into advice or suggest git steps that were never checked.

Pumpkin 2026.10.23

The chat list stays a side column in narrower panes when it fits beside the chat. The dropdown appears only when there is no room.

Pumpkin 2026.10.22

The Tags dialog has a Clear button. It unchecks every tag and stays open. OK then saves none.
Each tag shows its own color in the dot beside the name.

Pumpkin 2026.10.21

Chats now sync while saves happen. Failed reads keep chats safe.
Pair Computers now guides setup. You must compare both codes.
Sync shows trust and faults. Failed runs no longer look done.

Pumpkin 2026.10.20

Multi-root workspaces over WSL or SSH work again: adding a folder to a remote window no longer leaves the chat refusing every request until the connection drops.

Pumpkin 2026.10.19

Remote access: a computer whose clock runs a little ahead no longer shows as not connected on the phone; heartbeats and messages allow for clock differences.

Pumpkin 2026.10.18

Startup no longer logs an error when alerts for a workspace are checked before that workspace is ready.

Pumpkin 2026.10.17

A draft typed into a new chat now survives closing or reloading the window; it comes back as a draft in the chats list.

Pumpkin 2026.10.16

An empty chat no longer shows Set up Pumpkin after you are signed in and have a key. Semantic search and old Cursor or Claude files stay optional.

Pumpkin 2026.10.15

The agent can now search files in a folder outside the workspace, and a search with no hits says which folder it searched, so it stops repeating the same miss when the code lives next door.

Pumpkin 2026.10.14

A workspace the agent service has not adopted yet no longer logs a desktop error for its alerts while it waits.

Pumpkin 2026.10.13

The agent can now share a file in Slack as you: a file from disk or text it wrote, in a channel or a thread, with an optional comment. Reconnect Slack once to allow it.

Pumpkin 2026.10.12

A long attention note now stays inside its gray box and no longer runs over the waiting line.
Clicking a plan opens the preview when Open plans in Preview is on. Turning that setting off opens the file itself. A link to a line still opens the file.

Pumpkin 2026.10.11

Semantic search no longer scans every project at startup. Watching starts at once, unchanged projects are skipped, and catch-up work runs one project at a time in the background.

Pumpkin 2026.10.10

Windows and macOS show a desktop alert when an agent finishes, fails, or asks you something, even with every Pumpkin window closed. Clicking it opens that chat.
Hook permission alerts have Allow and Deny buttons, so you can answer from the alert.
The Dock and taskbar show how many chats need you, the taskbar flashes on new ones, and an optional sound plays when a turn finishes.
On macOS a menu-bar icon shows running and waiting chats and opens any of them.

Pumpkin 2026.10.9

Quitting or updating Pumpkin on Windows no longer stops agents working in WSL. WSL stays up while work runs, and the next Pumpkin takes those chats over instead of marking them interrupted.

Pumpkin 2026.10.8

Reading the signed-in account token no longer writes a record every time. The token is reused until it nears expiry, and a record is kept only when the token is actually refreshed.

Pumpkin 2026.10.7

The Chats column no longer waits for a chat that is in the middle of saving before it refreshes; it shows the last saved row at once and the fresh one when the save lands.
Opening a chat row's summary no longer waits behind other chats' saves.

Pumpkin 2026.10.6

Text typed into a new chat but not sent now shows in the chats list as a draft, so you can start another new chat and come back to it later.

Pumpkin 2026.10.5

Auto and the model picker no longer offer weak models; low-rated models stay reserved for small background jobs like titles.

Pumpkin 2026.10.4

The phone app can now change agent, thinking, long context and autopilot, edit or remove queued messages, reply in side threads, edit and resend, resume, and dismiss notices.

Pumpkin 2026.10.3

Opening or switching to a long chat paints the newest messages right away and fills in the rest a moment later, instead of pausing while every message is drawn at once.

Pumpkin 2026.10.2

The phone app now shows what the agent is doing: the plan with a Build button, the step list, file edits, subagent cards with an Open link, per-turn cost, and notices. Questions appear inline.

Pumpkin 2026.10.1

Search stays up to date when many files change at once.

Pumpkin 2026.9.430

Repairing a chat's saved summary (after a crash, an update, or a chat written by another computer) now runs on a separate thread, so the pane and sending never stall while a large chat is re-read.

Pumpkin 2026.9.429

Plan status badges are now soft rounded pills with room around the text, instead of a tight outline that looked like a focus box.

Pumpkin 2026.9.428

A flag can no longer ask you a question. Questions go to the question card, where you can answer. Got it on a flag only means you saw it.

Pumpkin 2026.9.427

A new chat's first message shows in the pane the instant you press Send, and the chat appears in the Chats column at once; the chat is created in the background.
Sending the first message no longer waits for the next-message thinking or long-context setting to save; a failed save shows an error without blocking the message.
When "worktree by default" is on, a new chat's worktree is created in the background after the chat opens.

Pumpkin 2026.9.426

Remote access sends its connection offer sooner: both the phone and the computer stop waiting a full second after the relay path is found.

Pumpkin 2026.9.425

Remote access keeps its relay credentials between page reloads, so reloading the phone page several times no longer runs out of credentials and falls back to a slow direct-only attempt.

Pumpkin 2026.9.423

Phone remote: while you are scrolled up reading, new messages no longer stutter or stop your scroll.
Phone remote: attach photos to a message. Tap 📷, pick from the library or camera; photos are resized on the phone and sent through to the agent.

Pumpkin 2026.9.422

Remote access connects in about five seconds instead of eighteen: when a computer was last reached over the relay, the phone starts with the relay instead of retrying a direct path first.

Pumpkin 2026.9.421

The remote window fixes from 2026.9.415 now ship inside the agent host as well, so they take effect once the editor updates.

Pumpkin 2026.9.420

Semantic search: a window folder opened outside your chosen code folder no longer gets its parent directory indexed, and Health no longer reports a watcher Pumpkin started as missing.

Pumpkin 2026.9.419

Phone remote: scrolling up loads earlier messages by itself; no more tapping Load older for each page (the button remains as a fallback).

Pumpkin 2026.9.418

Auto now actually hears from its helper model. Helper requests give reasoning models room to think before answering, so model picks, chat titles and file descriptions are no longer cut off empty.

Pumpkin 2026.9.417

Turning a model off in Settings now works; the switch no longer fails with an error.

Pumpkin 2026.9.416

Phone remote: your sent message no longer shows twice after the machine files it, and the row of repeated "Earlier conversation (summarized)" boxes is gone.

Pumpkin 2026.9.415

Opening a second window, adding a folder, or reloading a window on a remote machine no longer drops the other windows' agent connection or stalls their running turns.
A remote machine is treated as one machine across Pumpkin upgrades, so windows no longer fight over the connection after an update.
New remote windows get their connection settings right away instead of retrying for minutes while a folder is being added.

Pumpkin 2026.9.414

The "Caveman" communication style now demands the caveman voice on every reply and shows good and bad examples, so models actually follow it.

Pumpkin 2026.9.413

Phone remote: older messages you have loaded stay on screen when the machine's view moves ahead of yours; the app fetches the missing page itself instead of hiding what you loaded.

Pumpkin 2026.9.412

Phone remote: a chat whose newest turn is very long no longer opens with only its last message and a Load older button; the latest part of that turn is shown right away.

Pumpkin 2026.9.411

Adding a folder to a window or opening more windows no longer leaves WSL and SSH chats refused by the desktop: a failed grant handshake is now retried instead of abandoned.

Pumpkin 2026.9.410

Mac: in-app updates really switched on this time; 2026.9.391 through .409 shipped without the updater code, so install this one from the dmg and later versions arrive by Restart to Update.

Pumpkin 2026.9.409

Pumpkin connects to WSL and SSH hosts again after an update when the old runtime is still running there: the new runtime now replaces it instead of waiting forever at "acquiring service".

Pumpkin 2026.9.408

The chat list appears quickly after an update: saved chats are no longer re-read one by one when model prices change, and the list is scanned several chats at a time.

Pumpkin 2026.9.407

Semantic search starts much faster on large workspaces: files that have not changed since the last run are no longer looked up one by one in the index at startup.

Pumpkin 2026.9.406

Semantic search starts much faster on large workspaces: files that have not changed since the last run are no longer looked up one by one in the index at startup.

Pumpkin 2026.9.405

Remote-access enrollment now rejects a tampered request every time.

Pumpkin 2026.9.404

Chats whose finished background jobs still carried an old completion record are tidied once at startup and then left alone; a notice that never reached its chat is still delivered.

Pumpkin 2026.9.403

Startup no longer reopens chats whose background jobs finished long ago. A finished job's completion notice that was already delivered is no longer mistaken for unfinished work.

Pumpkin 2026.9.402

A message saying it is for the main thread now always goes to the main turn, never to a side thread.
A message that suggests a cause, a fix or an idea while the assistant works now steers the main turn instead of opening a side thread.

Pumpkin 2026.9.401

Remote access from a phone or browser connects again. After one rejected poll the computer stopped listening for connections until restarted while still showing online; it now resumes within seconds.
A remote session streaming a long turn no longer drops the connection after about a minute with "the device is not reading the channel".
The remote machine page says "Waiting for the computer to answer…" or "Connected · waiting for the session list…" instead of "No state from this machine yet".

Pumpkin 2026.9.400

A question asked in a side thread can now open and use the browser and canvas. Before, every browser call from a side thread failed with "Presentation requires complete local ancestry".

Pumpkin 2026.9.399

Subagents now report in while they run: a progress note every 15 minutes with rounds, cost, files edited and the current step.
A subagent silent for 10 minutes mid-turn raises an alert naming its last step, so you can inspect, steer or stop it instead of wondering if it hung.
A warning appears when a subagent keeps re-reading the same files after every summary fold without editing anything.
A message sent while the assistant works is now routed correctly when it reports a stuck session or tool error: the main turn is steered with it instead of opening a side thread.

Pumpkin 2026.9.398

A quick question asked while the assistant waits on subagents no longer breaks every later message in that chat with an invalid-request error; the chat heals on its own the next time it is opened.
A subagent rebuilding a large tool call after hitting the output limit now shows live progress in the Thinking row instead of sitting silent for minutes.

Pumpkin 2026.9.397

Mac: the 2026.9.396 release did not finish building for Intel Macs; this release carries the same fixes for every platform.

Pumpkin 2026.9.396

Opening a second window on the same remote no longer breaks the chats running in the first one, and the new window no longer sits on "Waiting for the Pumpkin agent host to start".
The desktop now adds the new window's folder to its existing permission instead of replacing it, so work in progress keeps running while a window opens.
A new window waits for the desktop to admit it (up to five minutes) instead of giving up after one failed attempt.
While the desktop is away (window reload, upgrade) a session no longer fills its log with hundreds of "desktop broker is unavailable" failures; it waits once and refreshes when the desktop is back.
One machine is one environment again: restarting a remote no longer registers it as a new machine.

Pumpkin 2026.9.395

While the assistant works, each message you send shows what happens next: answered right away, applied at the next step, or held until the turn ends.
You can change that choice from the waiting row, and the new choice takes effect at once.

Pumpkin 2026.9.394

Mac: the in-app updater announced in 2026.9.385 is now actually switched on; it was dropped by mistake before 2026.9.391 shipped.

Pumpkin 2026.9.393

Phone remote: the app keeps a local copy of your machine list and viewed chats, so they open instantly and update once the machine answers; loaded older pages are never fetched twice.

Pumpkin 2026.9.392

After an update that adds a new kind of API key, chats still running on the previous Pumpkin service no longer all stop with "Invalid wire payload (at keys…slot …)".

Pumpkin 2026.9.391

Phone remote: this release carries the fixes listed under 2026.9.386 to 2026.9.390, whose releases did not publish.

Pumpkin 2026.9.390

Phone remote: a machine no longer shows offline the moment its desktop window disconnects; it keeps reporting in with its last sign-in token for up to an hour until the desktop reconnects.

Pumpkin 2026.9.389

Phone remote: "Load older" now fetches a few turns at a time (about 96 KB per page) instead of file-sized chunks, so paging back through a long chat is quick and incremental.

Pumpkin 2026.9.388

Phone remote: internal notices posted into a chat (watcher results, subagent reports, thread messages) now show as a small collapsed note you can expand, not as a raw message bubble.

Pumpkin 2026.9.387

Settings: the Models page is easier to scan. Each provider is its own card, with the provider name, model count and key status on its heading bar. Providers with no key start folded.
Model rows show one description line and short facts (context, output, images, tools, reasoning) instead of a sentence of facts.
The Description and Skill rating fields for a model sit behind a Customize button instead of being open on every row.
Phone remote: the "A new version is ready. Reload" notice no longer appears on every open; it shows only when a genuinely newer build is waiting.

Pumpkin 2026.9.386

Phone remote: opening a chat now shows the recent messages right away instead of an empty screen with only "Load older".
Phone remote: older messages you have loaded stay on screen when new messages arrive; the view no longer collapses back to the newest message.

Pumpkin 2026.9.385

Mac: Pumpkin now updates itself. The new build downloads in the background and **Restart to Update** installs it; no more dmg drag-install. Running jobs continue through the restart.

Pumpkin 2026.9.384

The Claude Code model list now includes Fable (Fable 5), with a note that it needs usage credits on your Claude plan.

Pumpkin 2026.9.383

Settings → Models is easier to scan: each provider is a card with its name, model count and key status; providers with no key start folded.
Model rows show one description line and short fact chips (context size, output size, images, tools, reasoning) instead of a sentence of facts.
The Description and Skill rating fields sit behind a Customize toggle instead of being open on every row.

Pumpkin 2026.9.382

When a browser click gives up, the error now says what is wrong with that element: hidden by a closed panel, zero size, disabled, scrolled out of view, or covered by a loading overlay.
Waiting for an element that appears in several places on the page (every row of a table, say) now succeeds as soon as the first one shows, instead of failing because the match was not unique.
A click on a selector that matches several elements now says how many matched and how to narrow it, instead of listing every one of them.

Pumpkin 2026.9.380

Google Gemini is a new model provider. **Pumpkin: Set API Key** offers "Google Gemini"; with a key set, Gemini 3.8 Flash, 3.7 Flash, 3.1 Pro Preview and 3.5 Flash-Lite appear in the model list.
Gemini models honour the Low/Medium/High thinking levels. Gemini 3 cannot turn thinking off, so "Off" uses the lowest level the model accepts and the chat says so.
Switching between a Claude model and a Gemini model in the same chat no longer risks a failed turn from the other provider's reasoning signatures.

Pumpkin 2026.9.379

MCP tool calls send each argument in the type the server expects. A number or true/false written as text (a width of "300") is converted before the call instead of being rejected as a string.

Pumpkin 2026.9.377

Closing a canvas tab and choosing Open again now shows the canvas, instead of doing nothing.

Pumpkin 2026.9.376

After an update, an old background service that no successor takes over now retires itself within ten minutes instead of blocking the workspace until it is killed by hand.
Reconnecting a remote workspace no longer fails a request in flight with "Native workspace source was removed"; it waits for the registration to finish.

Pumpkin 2026.9.375

The "Caveman" communication style now says to speak like a caveman: very short, very simple, blunt sentences, with file names and commands kept exact.
Settings: "Communication style" has a "Custom…" choice where you write your own style instructions; picking a preset clears a leftover custom style, and an Edit button reopens it.

Pumpkin 2026.9.374

After an update, chats still running on the previous Pumpkin service now show the running spinner in the chat list (hover: "running on the previous Pumpkin version") instead of looking parked.
A chat no longer stops with "Final context reduction changed frozen history" when Pumpkin was closed for an update while the chat trimmed an oversized request. The trim is redone when Pumpkin is back.
After an update, a chat no longer shows "Invalid wire payload (request getSessionSummary … got string(0))".

Pumpkin 2026.9.373

When Windows' WSL service briefly refuses new sessions, a new WSL window no longer stalls about 90 seconds and then fails: the connect notices at once and retries.

Pumpkin 2026.9.372

Opening a second Pumpkin window on the same remote machine no longer stops every running chat with "Current broker grant does not authorize this operation".
Chats now wait about 15 seconds (was 3.5) for the desktop to finish re-issuing access, then re-run the interrupted request once.

Pumpkin 2026.9.371

Adding a folder to an open workspace no longer leaves Pumpkin stuck on "Registered workspace is unavailable" until the background service is restarted.

Pumpkin 2026.9.370

Hover the context line under a message sent on Auto to see how the model was chosen: picked by the helper, or which fallback rule ran and what the helper actually replied.
The small request Auto makes to choose a model now counts toward that turn's cost.

Pumpkin 2026.9.369

Settings opened with no folder open now says to open a folder first, instead of a configuration error.
A server turned off in your Cursor setup is no longer loaded, so it no longer stops the agent from starting.

Pumpkin 2026.9.368

A slash command file can name a tool to run directly. The tool runs with no model call, and the local model fixes bad arguments once before the command falls back to a normal prompt.

Pumpkin 2026.9.367

A chat no longer stops with "Broker authorization changed during operation" when another Pumpkin window opens mid-request. The request is re-run once, with a notice in the chat.

Pumpkin 2026.9.366

New "Caveman" communication style: very short, very simple, blunt sentences, with file names and commands kept exact. Pick it under Communication style in Settings.

Pumpkin 2026.9.365

After an update, the chat list no longer fails with "Too many pending requests" while every chat's saved summary is refreshed in the background.
A chat still finishing its turn on the previous Pumpkin version no longer shows a "foreign metadata is missing or stale" error while the new version lists it.

Pumpkin 2026.9.364

Opening the same project as a plain folder, inside a multi-root workspace, or after adding a folder now works; Pumpkin no longer refuses it as belonging to another owner.

Pumpkin 2026.9.363

The plan's Build button stays disabled while the chat is still working after its turn ended (a subagent, job or watcher is pending), and re-enables on its own once the chat is idle.

Pumpkin 2026.9.362

Ticking plan steps or adding notes to the plan while it is being built no longer flips its badge back to "ready"; it stays "building".

Pumpkin 2026.9.361

A plan no longer shows "built" while its steps are still unticked or a subagent is still working on it; it stays "building" until the work is actually done.

Pumpkin 2026.9.360

Remote hosts: the "Session store is owned by profile … on broker …" error now lists exactly which registration fields differ, so the conflict can be resolved instead of guessed at.

Pumpkin 2026.9.359

Convert Existing AI Setup now opens the migration chat right away instead of failing after 90 seconds with "Pumpkin has not confirmed this request".

Pumpkin 2026.9.358

Mac: quitting Pumpkin now also ends any leftover Pumpkin processes still running from the app bundle, so dragging a new version into Applications is not blocked by them.

Pumpkin 2026.9.357

Adding a Codex or Claude Code subscription model is one step: Pumpkin finds the program and your login itself and shows a model list (Sonnet, Opus, Haiku; Codex lists its models). No paths to type.
Pumpkin now finds codex and claude installed through nvm, npm, Homebrew or ~/.local/bin, so a signed-in program no longer shows as "not found".
Not installed or not signed in is one plain message with Check again and Choose the program file buttons; a second account is an optional folder pick.

Pumpkin 2026.9.356

Mac: after you quit Pumpkin, nothing keeps the app open any more, so dragging a new version into Applications works.
A hidden Pumpkin process started by a terminal command or the loops autostart timer now exits shortly after that command finishes. The background service keeps running.

Pumpkin 2026.9.355

The agent can send feedback itself. Ask it to report a bug and it sends your message, optional diagnostics, the chat, and up to five screenshots to the team, then tells you the report id.

Pumpkin 2026.9.354

Remote on your phone: opening a chat sends only its recent part; earlier messages load when you scroll up, so a long chat opens in seconds instead of downloading everything.
Remote on your phone: the page no longer rebuilds itself on every streamed update, so your scroll position stays put, you can reach the bottom, and typing and Send respond at once.
Remote on your phone: the "A new version is ready" notice appears only when a newer version is really waiting; a leftover copy of the same version no longer triggers it.

Pumpkin 2026.9.353

The first snapshot in a repository only starts the cleanup clock; the scheduled cleanup runs from the next one on.

Pumpkin 2026.9.352

Account token reads keep their safety record again. Records still expire after a day, so the store no longer grows without limit.

Pumpkin 2026.9.351

Semantic search status changes no longer make every open chat re-read the settings inventory each time; refreshes are grouped and time out after 15 seconds instead of hanging.
The settings inventory is reused until a setting, account, or connection actually changes, so many open chats no longer pile up repeated reads at startup.

Pumpkin 2026.9.350

Startup no longer slows down over time. Account token reads leave no record behind, old records are removed after a day, and finished background jobs from past chats are not re-checked on every start.
Per-turn file snapshots are cleaned up on a schedule so a workspace's git folder stops growing with every turn.
Chat messages keep a reserved lane so bursts of settings reads cannot delay sending.

Pumpkin 2026.9.349

A rejected or expired Amazon Bedrock key now says only that, instead of also claiming no supported models exist in the chosen region.

Pumpkin 2026.9.348

Adding an Amazon Bedrock model now offers its global profile before regional profiles and labels it "Global routing".

Pumpkin 2026.9.347

Old remote-server downloads no longer pile up on WSL and SSH: connecting now removes earlier ones automatically, keeping the current one and any still in use.

Pumpkin 2026.9.346

Dragging a long chat's scrollbar to the very top now lands on the first message instead of somewhere in the middle.

Pumpkin 2026.9.345

Use your ChatGPT (Codex) or Claude (Claude Code) subscription as a model: new Settings sections let you add entries by model, with an optional separate login folder per account.
Chats on a subscription model run normally: streamed text, Pumpkin's own tools, multi-step tool loops and Stop. The cost line says "subscription" instead of a made-up price.
Missing, signed-out or rate-limited subscription programs show the reason in Settings and the model picker; Auto model and subagents skip them until they are ready.

Pumpkin 2026.9.344

The semantic search panel no longer says "0 files" for a full index whose store cannot count files; it says the file count is not reported, and after a scan it shows the real total.

Pumpkin 2026.9.343

A browser a subagent opened now closes itself once the subagent finishes, instead of staying open until you delete its chat.

Pumpkin 2026.9.342

Quitting Pumpkin on a Mac now really exits it while the background service keeps working; dragging a new Pumpkin into Applications no longer fails with "Pumpkin is in use".

Pumpkin 2026.9.341

After a Pumpkin update, a chat that was waiting on a job or watcher no longer loses its shell and wait tools with "job registry disposed" until you restart.

Pumpkin 2026.9.340

Pumpkin now starts on Macs where the background service failed with "listen EINVAL": its private socket file name was too long for macOS, so most Mac usernames hit the limit.

Pumpkin 2026.9.339

Releases are no longer held up on Windows by a cleanup step that ran before the previous chat had fully let go of its files.

Pumpkin 2026.9.338

Claude no longer remarks on the background context Pumpkin refreshes each step (data lists, memory index, todo list); replies stay on your request.

Pumpkin 2026.9.337

Waiting on a GitHub Actions run now works when the GitHub CLI is only on your shell PATH; a previous fix for this did not make it into a release.

Pumpkin 2026.9.336

A long chat no longer repeats its status line before every tool call, and when a tool result is cut for size it is told how much was cut and to read a smaller window.

Pumpkin 2026.9.335

Waiting on a GitHub Actions run no longer fails with "gh is not installed" when the GitHub CLI is only on your shell PATH.

Pumpkin 2026.9.334

A long attention note from an agent now wraps in its own box beside the Got it button, instead of spilling over the rows around it and under the button.

Pumpkin 2026.9.333

A chat that keeps saying the same status in different words now stops after five rounds, instead of running until you stop it.

Pumpkin 2026.9.332

MCP servers set up with a pasted token now offer Sign In and Sign Out. An expired pasted token is no longer sent, so you are asked to sign in instead of seeing Connected with broken tools.

Pumpkin 2026.9.331

A chat that was still working when Pumpkin updated no longer breaks afterwards with a red "cannot call" error. Messages sent to it meanwhile now wait and run once it moves over.

Pumpkin 2026.9.330

A running subagent on Auto now shows the model it picked, in its row and in its own header, instead of Auto twice. Before, the model only appeared once the subagent finished.

Pumpkin 2026.9.329

Auto's model picker now also knows which kind of turn it is picking for (first message or a chat in progress, main thread or subagent) and which price mostly pays for it.

Pumpkin 2026.9.328

Auto's model picker now sees each model's full price list, including what it costs to read and write cached conversation, and knows that reading is most of a chat's cost.

Pumpkin 2026.9.327

Semantic search indexing is several times faster on Macs and other machines without a dedicated GPU: many code chunks are now embedded per request instead of one at a time.
Large semantic search indexes no longer keep one processor core busy looking up files one by one; per-file lookups are indexed and writes are grouped.

Pumpkin 2026.9.326

Messages you sent while a turn was running are no longer dropped when that turn fails or when Pumpkin shuts down. They stay queued, and the chat reports whether they can be run.
A message that could not be confirmed is never sent twice. If the connection changes mid-send, Pumpkin reports the unknown outcome instead of quietly resubmitting.
A send whose deadline passes before it reaches the background service is not sent later.
Autopilot controls refuse cleanly when the connected runtime is too old, including after a reconnect mid-request, instead of sending a request the service cannot understand.
If saving the message queue fails, the chat shows a notice instead of only writing to the log.

Pumpkin 2026.9.325

Auto now sees both prices (input and output per million tokens) for each candidate model, and picks the cheaper one when models fit the task equally.
Semantic search: a code folder moved to Qdrant now finishes the move. The Current index panel shows the real file count instead of 0 files, and the old index file is removed once the scan completes.
Semantic search: Check now on the Health row reports the folder index result instead of flipping to Not checked yet when there is no workspace registry.

Pumpkin 2026.9.324

Amazon Bedrock: Claude turns with thinking no longer fail with "malformed tool input cannot be replayed" when the model calls a tool that takes no arguments.

Pumpkin 2026.9.323

Typing / in the message box now offers /search: a one-shot semantic code search. It runs a search turn that answers with file and line evidence.

Pumpkin 2026.9.322

The agent can hand you a secret (a generated password, token or key) with the new deliver_secret tool: a masked card with Show and Copy. The value never enters the chat transcript or the next request.

Pumpkin 2026.9.321

New Source control section in settings: Pull behavior says whether the agent brings the branch up to date before it starts. Default: always, so the agent works from the latest state.
Conflict management moved from Agent persistence into the Source control section; a saved choice carries over.

Pumpkin 2026.9.320

A chat that lost its desktop connection while idle (screen lock, sleep, restart) keeps reconnecting and picks up where it left off, instead of stopping after a few tries or greying out the input box.
A turn that starts while the desktop connection is coming back waits for it (with one notice) rather than failing straight away; nothing already sent is repeated.
The background service's startup log is no longer wiped by a restart or upgrade: earlier records stay in place, and WSL/SSH remotes keep the last three startup logs per installed version.
On Windows, the private ownership files the service creates get the right per-user security from the start, so a second window or a service restart no longer fails with an access error.
Browser file uploads only send files the desktop browser can actually read: local files, or files in a WSL workspace (handed over as their Windows path). Other remotes get a clear error instead.
A loop can turn Slack off for itself, and that choice survives edits made in the loop form. A loop's question stays in its chat until answered even when Slack is not linked.

Pumpkin 2026.9.319

Amazon Bedrock: Claude turns with thinking no longer fail with "malformed tool input cannot be replayed" when the model calls a tool that takes no arguments.

Pumpkin 2026.9.318

The "Waiting for N subagents" footer under a chat now clears the moment each subagent reports back, together with the Agents list, instead of lingering while the finished subagent shuts down.

Pumpkin 2026.9.317

A first message sent while the new chat was still being created, then switching chats or pressing Escape, now reaches the created chat instead of being dropped.

Pumpkin 2026.9.316

Updates no longer stop chats that are still running: the previous background service finishes its turns while the new one takes new work, and a chat moves over when its turn ends.
The previous background service is only forced out during an update if it does not answer.
While the desktop reconnects a workspace to the new background service, a chat shows one plain notice and keeps checking until it is back, instead of filling with red Retry cards or giving up.
Requests that reach the new background service before the desktop has reconnected the workspace wait briefly instead of failing at once.

Pumpkin 2026.9.315

In WSL, browser_file_upload now sends the browser the Windows path of the file it checked, so /tmp, /mnt/c and C:\ paths upload the right file.

Pumpkin 2026.9.314

On Auto, a model that is stuck or keeps hitting errors can escalate: Auto picks a more capable model, says so in the chat, and the turn carries on. The next message remembers it.

Pumpkin 2026.9.313

Remote (phone) question cards get the desktop's Queue instead toggle: a typed message answers the pending question by default, or queues for the agent when the toggle is on.
Remote composer no longer loses text typed while the previous message is still sending.

Pumpkin 2026.9.312

Remote (phone) session screen: the model and the running indicator now sit on their own row under the title, and an Auto chat shows which model Auto picked.
Remote question cards have Decline and Submit like the desktop; unanswered questions are reported, one-tap Allow/Deny stays, and answered cards remain visible with the choice.
Remote replies render markdown (headings, lists, code, links), tool steps show their description, and a status line says what the agent is doing.

Pumpkin 2026.9.311

On Auto, a chat with images no longer stops when the chosen model can't read them. Auto now skips those models, and if one is reached anyway it switches to a model that reads images and says so.

Pumpkin 2026.9.310

Semantic search Health no longer says "nothing is indexing this folder" for a folder the code-root index already covers; it now points at the root watcher that keeps it fresh.

Pumpkin 2026.9.309

Semantic search keeps working in large code roots. Linux allows about 128 file-watch slots per user and the watcher used one per repository, so roots with over a hundred repositories kept dying.
One repository that cannot be watched no longer stops indexing for the others; it is indexed on each start and the log says why live updates are off for it.
When the semantic search watcher does stop, its real error is shown in the Health panel and the doctor instead of a generic "command failed".

Pumpkin 2026.9.308

Starting Pumpkin no longer re-checks every file of the agent runtime on each launch; on Windows that took about 11 seconds, and connecting to WSL another 6.
The first start after an upgrade sets up the runtime in a fraction of the time (about 10 seconds on Windows instead of up to 45).
The startup log now records how long each start step and the runtime setup took.

Pumpkin 2026.9.307

Windows: the runtime install after an upgrade completes again; 2026.9.306 could fail to start with a "directory not empty" error on Windows.

Pumpkin 2026.9.306

Connecting to WSL after an upgrade is faster. The server files are unpacked straight from the Windows install, and the runtime is set up in parallel with fewer repeated checks.
Old runtime versions are cleaned up after each upgrade, keeping only the current and previous one instead of piling up gigabytes.

Pumpkin 2026.9.305

After installing a new version the window opens right away instead of waiting about 45 seconds for setup to finish behind the scenes.
Chat is ready a few seconds after the window opens. The previous version hands over at once, and startup retries every second instead of waiting up to a minute.

Pumpkin 2026.9.304

A chat that was summarized partway through a turn no longer stalls when everything recent is already kept; the older part is summarized as before.

Pumpkin 2026.9.303

After installing a new version the window opens right away instead of waiting about 45 seconds for setup to finish behind the scenes.
Chat is ready a few seconds after the window opens. The previous version hands over at once, and startup retries every second instead of waiting up to a minute.

Pumpkin 2026.9.302

A model that spent its whole reply on hidden thinking and showed nothing is retried once, then the chat continues on another model with more room instead of stopping with an error.

Pumpkin 2026.9.301

The chats list no longer fails with "Invalid wire payload" when many chats with subagent cost figures are loaded: Pumpkin sends the list in smaller parts and keeps every chat you had loaded.

Pumpkin 2026.9.300

When a long turn is summarized partway through, the newest several rounds now stay word for word instead of one, so the agent keeps the files it was working from.
Every summary keeps the project, goal, direction, what was learned and the current tasks, and ends with the next concrete action.
On Grok and OpenAI models, summaries start before a request can run out of room, not only as an emergency after the limit is hit.

Pumpkin 2026.9.299

Settings › Semantic search opens with one plain sentence about what Pumpkin is doing (preparing, indexing with time left, moving to Qdrant, up to date, or paused and why).
A single button appears only when you are truly needed, such as trusting the folder.
Health for a folder index now works without a workspace and reads healthy only while the index is current or being updated; the "Not installed" message is gone.
Buttons renamed: Restart indexing, Rebuild index from scratch, Check now. Words like watcher and health check no longer appear.
The background indexing setting moved to an advanced section; changing it no longer restarts anything (Pumpkin checks every minute anyway).

Pumpkin 2026.9.298

Semantic search now keeps itself running: every minute Pumpkin re-checks each folder's index and starts a stopped indexer or resumes an interrupted move to Qdrant.
A start that was refused earlier (for example before the folder was trusted) is retried on its own. No button needs pressing after an install or update.

Pumpkin 2026.9.297

Semantic search: after a workspace watcher starts, an indexed folder is no longer briefly treated as not indexed, so Pumpkin does not skip or redo its setup.

Pumpkin 2026.9.296

When a chat action is refused with "Peer role … cannot call …" during a Pumpkin update, the runtime log now records which side refused it and which call, so the failure can be traced.

Pumpkin 2026.9.295

Amazon Bedrock: global inference profiles (names starting with "global.") can now be added; before, Add failed saying the destination was not a foundation model.
The Amazon Bedrock setting description now lists the supported models, including Fable 5.1, instead of only Sonnet 4.6 and Haiku 4.5.

Pumpkin 2026.9.294

Settings has a new Agent persistence section: choose issue ownership, effort, end goal (saved locally, a pull request, or deployed) and how the agent handles conflicts.
Each Agent persistence choice has three presets or your own text, and adds one line to the agent's instructions.

Pumpkin 2026.9.293

Semantic search: a folder's index now lives in a small local database (Qdrant) that Pumpkin downloads and runs, instead of one big file loaded into memory on every status check and search.
An existing folder index is moved over once at the next start; the old file is removed only after the new index is complete, and Settings shows the progress.
If Qdrant cannot start on a machine, the folder keeps its old index and Settings says why; a new semantic-search setting ("backend: gob") keeps the old file for good.
Search results during the move say the index is still being rebuilt.

Pumpkin 2026.9.292

Semantic search: a workspace project the indexer cannot write no longer stops the rest; Pumpkin repairs it if it can, else indexes the others and Health names the fix.

Pumpkin 2026.9.291

Semantic search: a search workspace with a project you cannot write no longer restarts forever; Settings names the project, its owner and how to fix it.

Pumpkin 2026.9.290

Browser tools accept a setting sent in the wrong form, such as true or false written as text or a file list written as text, instead of failing; the result says what was converted.
On Windows, deleting a chat that had a background command running no longer fails with a directory-not-empty error.
Semantic search: when Pumpkin stops a folder watcher, Settings now says why, such as a rebuild from scratch, new default ignore patterns, or the runtime shutting down.

Pumpkin 2026.9.289

The "Thinking…" row no longer stays on screen after a chat has already finished and is waiting on you; the chat settles as soon as the runtime reports it idle.
Sending the first message of a new chat keeps that message on screen instead of the pane going blank for a moment and the message popping back in.

Pumpkin 2026.9.288

Usage logs now say which part of a turn spent the money: your chat, each subagent by name, or a background job such as a title or summary, alongside the model that actually ran.

Pumpkin 2026.9.287

Long chats are summarized a little earlier, and much earlier once a chat has sat idle long enough for its prompt cache to expire, so the next message re-reads a smaller history.
The summary is written knowing which recent messages the assistant still sees, so it no longer repeats them and keeps conclusions rather than raw command output.

Pumpkin 2026.9.286

Semantic search: a Pumpkin update no longer stops the folder watcher and re-scans every file; the new runtime picks up the running watcher where it left off.
The Settings page no longer waits on slow index counts, so large indexes stop showing "the agent host did not answer in 15 s".

Pumpkin 2026.9.285

Built-in subagents now run on Auto: the helper picks the cheapest capable model for each of their turns instead of a fixed top-tier model.
Custom agents can set their model to auto; Settings shows "Auto — chosen per turn" for subagents on that choice.

Pumpkin 2026.9.284

Seven Bedrock models now use compatible output limits, prompts and response modes; models with unreliable tool use run as text-only chat.

Pumpkin 2026.9.283

Adding a Bedrock model with an expired or invalid key now says to set a new key instead of asking for extra permissions.

Pumpkin 2026.9.282

Retired Fireworks models no longer appear in model choices or get picked by Auto, preventing requests to models Fireworks no longer serves.

Pumpkin 2026.9.281

Semantic search: changing which folders are indexed restarts the watcher in place instead of wiping the index and re-embedding everything; the folder you picked stays watched.

Pumpkin 2026.9.280

Semantic search: unselected subfolders and linked worktrees now really stay out of the index, even when they carry their own ignore file that re-included everything.

Pumpkin 2026.9.279

Remote (WSL/SSH) start helper that fails now really exits in installed builds; the previous release only fixed it in development builds.

Pumpkin 2026.9.278

After WSL or the computer restarts, Pumpkin no longer waits on a leftover record from its previous run when that process number now belongs to something else; it starts right away.
A remote start helper that fails now exits instead of lingering until the next retry.

Pumpkin 2026.9.277

Remote: the session list on your phone now appears in seconds. Each update carried hundreds of kilobytes of cost details the phone never showed; it now sends only what the list displays.

Pumpkin 2026.9.276

Semantic search Subfolders picker: unchecked folders are excluded the plain way, so other exclusions keep working inside kept folders. Before, a kept folder quietly re-included excluded files.
A folder created after you saved your choice is excluded at the next watcher start until you check it.

Pumpkin 2026.9.275

Questions from the agent can now explain why they are being asked and what each answer means, shown a few lines at a time with Show more.
Question cards, including requests for a secret, have a Decline button that tells the agent to decide for itself.
Answering only some questions on a card now works instead of failing.

Pumpkin 2026.9.274

Questions left waiting by a helper agent in a closed chat show up on the right parent chat again after a restart.

Pumpkin 2026.9.273

Semantic search no longer indexes the worktree checkouts Pumpkin creates beside your repositories; the Subfolders picker no longer lists them either.
Data dumps such as CSV, TSV, JSON Lines, Parquet, SQLite and HAR files are skipped by default, so indexing spends its time on code.

Pumpkin 2026.9.272

New Loops settings control how loops reach you: Slack summaries, channel, whether they may ask questions, and how long they wait. Each loop can still override these.

Pumpkin 2026.9.271

Scheduled tasks, watchers and other saved follow-ups now fire even when their chat was never reopened: Pumpkin keeps a small list of chats with pending work and checks it on start.
Startup no longer reads every chat's summary file; with hundreds of chats the first check finishes in milliseconds instead of seconds.
The first start after this update does one last full pass to build that list.

Pumpkin 2026.9.270

Agents can now hand a web page a local file to upload, with the new browser_file_upload tool, instead of being stuck at upload forms.
The path must exist on disk; the tool checks and tells you exactly which one is missing.

Pumpkin 2026.9.269

Loops can ask you questions without Slack: the question appears in the loop's chat and the run waits for your answer. With Slack linked it is sent there too.

Pumpkin 2026.9.268

Loops now message you on Slack by default: once Slack is linked, every loop DMs you a summary after each run and can ask you questions, with no per-loop setup. Turn it off per loop if you prefer.

Pumpkin 2026.9.267

When an agent hands work to a helper it can say which development-process steps the helper owns; the helper sees only those steps and is told the parent handles the rest.
Built-in coding helpers default to the build, verification and code-check steps; reviewer and research helpers default to their own step.
Agents are now advised to delegate as a recommended pattern for parallel or isolated work, not as a rule, and to brief helpers so they do not redo research.

Pumpkin 2026.9.266

Open Anthropic chats keep their cached context warm during pauses while refreshes cost less than rebuilding it. Closing the chat or sending a message stops refreshes.
Cache refresh spending appears in the turn cost without adding extra rounds. You can turn automatic refreshes off in Settings.

Pumpkin 2026.9.265

Semantic search no longer re-indexes every file each time its watcher starts; only changed files are embedded, so indexing settles instead of churning all day.

Pumpkin 2026.9.264

When a browser an agent opened disappears (for example its tab was closed), the agent is now told when and why instead of "unknown browser", and the list shows it as closed for ten minutes.

Pumpkin 2026.9.263

After an upgrade on WSL, the previous background service gets time to finish and exit on its own instead of being force-killed.

Pumpkin 2026.9.262

While a WSL or SSH folder connects, the status bar shows what Pumpkin is doing and for how long, such as verifying the payload or acquiring the service, so a slow start no longer looks frozen.
If it fails, the status bar shows the reason, the attempt number and a countdown to the next retry; click it to retry now or open the log. After three failures in a row a warning appears once.
Starting the local background service shows the same live status, phase by phase.

Pumpkin 2026.9.261

After the computer wakes from sleep with no Pumpkin window open, the background service gets a fresh five minutes to reconnect before Pumpkin quits, instead of quitting right at wake.
A background service that failed to connect no longer counts as running; Pumpkin quits after the usual five minutes rather than staying in the background.

Pumpkin 2026.9.260

Closing the last window while the local background service is briefly reconnecting or restarting no longer quits Pumpkin under it; the service gets five minutes to come back first.
A background service that has just been started always gets its full five minutes to connect, even when Pumpkin was already counting down for something else.

Pumpkin 2026.9.259

When a Windows upgrade closes Pumpkin's windows itself, the previous WSL background service is now handed over the same way as when Pumpkin quits for the installer on its own.
If quitting for a Windows upgrade fails, Pumpkin retries after ten seconds instead of never trying again.

Pumpkin 2026.9.258

If the local background service exits or crashes after the last window closed, Pumpkin now quits about five minutes later instead of staying in the background with nothing to do.
On macOS, Quit now exits Pumpkin unless a local background service or a connected WSL/SSH host still needs it. This started in 2026.9.256; that note only mentioned Windows and Linux.

Pumpkin 2026.9.257

If a Windows upgrade is cancelled because a window has unsaved work, WSL/SSH folders stay connected. Before, that connection ended when the installer asked Pumpkin to close and did not come back.
Clearer than the 2026.9.255 note: a Windows upgrade leaves the previous WSL background service running and the new version takes it over; only the old Pumpkin process's own connection to it ends.

Pumpkin 2026.9.256

On Windows and Linux, closing the last window quits Pumpkin unless a local background service or a connected WSL/SSH host still needs it. A WSL-only setup no longer runs in the background forever.
A connected WSL/SSH host that drops after the last window closed gets five minutes to reconnect; then Pumpkin quits instead of retrying forever.

Pumpkin 2026.9.255

WSL folders connect again. 2026.9.254 started the WSL bridge without a console, so every connection failed with "Bridge bootstrap timed out" and a stray wsl.exe terminal window appeared.
Upgrades on WSL work as in 2026.9.253: the new version takes over the previous background service instead of keeping it alive across the installer.

Pumpkin 2026.9.254

Upgrading on WSL keeps the previous background service running until the new version takes it over, so Pumpkin comes back without restarting WSL.

Pumpkin 2026.9.253

WSL upgrades stop the previous background service instead of waiting for running chats to finish. Failed connections retry quickly without restarting WSL.

Pumpkin 2026.9.252

Use 75 verified Bedrock models, including 61 with tools, with reasoning controls matched to each model and clear reasons for unavailable choices.
Bedrock tool turns now work with models that cannot stream tool calls, including Llama 4 Maverick, Mistral Large and Palmyra X5.

Pumpkin 2026.9.251

On WSL, an update no longer sits on "host not available" forever when the previous runtime is frozen; the wait is bounded and the message names the stuck process.
The update helper's own failure text is now recorded in the desktop log, and the runtime log is rotated once it reaches 64 MB.

Pumpkin 2026.9.250

Add Bedrock models by choosing a region and a model discovered with your saved AWS key. AWS controls model access.
Supported Bedrock models reuse matching catalog ratings and descriptions. Prices remain Unpriced.

Pumpkin 2026.9.249

Recent chats load ahead without leaving background observers running after the chat closes.
Credential access stops when permissions are revoked, without retrying the operation under replacement permissions.

Pumpkin 2026.9.248

When a plan edit does not match the plan text, the agent is told to re-read the plan and retry instead of drifting to a separate copy of the plan elsewhere.

Pumpkin 2026.9.247

Cache and request diagnostics identify the chat and turn, including when a long chat uses a summary.

Pumpkin 2026.9.246

Long chats recover from oversized requests by compacting or trimming history, keeping current instructions and avoiding repeated actions.

Pumpkin 2026.9.245

Chats no longer stop with "Broker grant changed during operation" when a window reconnects mid-turn; the turn keeps going.

Pumpkin 2026.9.244

Pumpkin loads your twelve most recent chats ahead of time on startup, so switching back to one of them is instant.

Pumpkin 2026.9.243

Verified automatic oversized-request recovery alongside the compact tool, preserving current instructions and avoiding repeated actions.

Pumpkin 2026.9.242

Cache and request diagnostics identify the chat and turn, including when a long chat uses a summary.

Pumpkin 2026.9.241

The agent can ask to compact the conversation early instead of waiting for it to fill up.

Pumpkin 2026.9.240

Long chats automatically compact or trim oversized requests and continue without repeating completed actions. Explicit pins and current instructions stay protected.

Pumpkin 2026.9.239

Cache and request diagnostics identify the chat and turn, including when a long chat uses a summary.

Pumpkin 2026.9.238

The message history list under the pinned message now shows a thin line between messages so they're easier to tell apart.

Pumpkin 2026.9.237

Cache and request diagnostics identify the chat and turn, so activity from concurrent chats can be traced accurately.

Pumpkin 2026.9.236

Semantic search: a Subfolders button beside the code folder lets you pick which folders get indexed, with search and select all, even before indexing.
Semantic search: exclusion patterns saved in Settings now take effect, including file-type patterns such as *.json.

Pumpkin 2026.9.235

Requests are checked before sending and shed oldest unprotected history when needed. Tools, pinned content and latest reasoning stay protected.
Startup recovers when the previous runtime finishes shutting down while its replacement is starting.

Pumpkin 2026.9.234

Requests are checked before sending and shed oldest unprotected history when needed. Tools, pinned content and latest reasoning stay protected.
The deferred tool groups setting now appears in Settings; choose an empty list to show tool declarations directly.

Pumpkin 2026.9.233

Requests get a final size check before sending. Oversized chats shed oldest unprotected history first, keeping tools, pinned content and latest reasoning; protected-only requests stop safely.

Pumpkin 2026.9.232

Auto now picks a more skilled model for demanding work such as debugging, multi-file changes and design, and uses price only to break ties; simple chat and small edits still go to cheaper models.

Pumpkin 2026.9.231

When you switch to a chat that isn't already open, the chat you were on now fades out with a spinner until the new chat appears, so you can tell the switch is still loading.

Pumpkin 2026.9.230

Lower model costs: browser, canvas, Slack and config-authoring tools appear in a short catalog instead of full declarations on every request (about 8,000 fewer tokens); the agent loads them on demand.
Shorter descriptions for the shell, update_plan, ask_user, use_worktree and wait_for tools; the guidance moved into the system prompt once.
The evidence-analyst, code-reviewer and evidence-runner subagents no longer receive file-edit, spawn or browser tools they never use.

Pumpkin 2026.9.229

Long Grok turns can discard superseded reasoning to make room for useful context while preserving the latest reasoning, pinned content and available tools.

Pumpkin 2026.9.228

Lower model costs: coming back to a Claude chat after more than five minutes no longer re-sends the whole conversation; only the previous turn is re-sent.

Pumpkin 2026.9.227

Subagents inherit their parent's browser, canvas and diagnostics tools without opening a child chat. Unavailable optional tools no longer block startup.

Pumpkin 2026.9.226

The message history list under the pinned message now scrolls with the app's own scrollbar instead of a plain default one.

Pumpkin 2026.9.225

Amazon Bedrock: adding a model now picks the region and model from a list instead of typing them, so a mistyped region no longer fails with an unclear error.

Pumpkin 2026.9.224

Remote: what you send from your phone stays on screen and the streamed reply grows in place instead of blanking and coming back during a turn.
Remote: pick the model when starting a new chat from your phone, and switch a chat's model from its header when the computer offers a list.

Pumpkin 2026.9.223

Chats recover from a brief failure while reading sign-in or connection settings without repeating completed work, and persistent failures give clearer retry guidance.

Pumpkin 2026.9.222

A chat that was closed while waiting on a tool or question no longer gets stuck refusing every message; chats already stuck this way continue on your next message.

Pumpkin 2026.9.221

Answers and discussions can finish without being mistaken for unfinished work or a request for approval.
Plans wait for the Build button instead of asking for approval in a separate question card.
Required human reviews still pause work without adding redundant approval cards.
Fewer unnecessary refusals when a request can be completed safely.

Pumpkin 2026.9.220

Agents can search your other chats in the same folder and read them, including just their errors, so you can ask about another chat by its id or topic.

Pumpkin 2026.9.219

Long Grok turns no longer get refused for exceeding the context limit when the conversation itself was not too large; extra reasoning from earlier steps is dropped first.

Pumpkin 2026.9.218

Mentioning a Windows file path now loads that project's rules in chats, side threads, and worktrees, including drive-letter and network paths.

Pumpkin 2026.9.217

Large semantic-search indexes are no longer restarted for being quiet while finishing symbol indexing or saving results after the initial scan.

Pumpkin 2026.9.216

Lower model costs: a chat that starts or stops touching another project (such as a temporary git worktree) no longer rewrites its cached conversation; that project's rules arrive as a hidden update.

Pumpkin 2026.9.215

Lower model costs: the hourly prompt-bundle refresh no longer rewrites every open chat's cached conversation; a chat picks up a new bundle at its next natural rewrite.

Pumpkin 2026.9.214

Large tool results, errors and diffs in the chat now preview a bounded amount and need a click to show the rest.

Pumpkin 2026.9.213

Lower model costs: editing a rule file no longer rewrites the cached conversation of every open chat; the changed rule is delivered to each chat as a hidden update message instead.

Pumpkin 2026.9.212

Lower model costs: switching the Auto model, a tool-limit fallback, or an MCP server reconnecting no longer rewrites a chat's cached conversation on the next turn.

Pumpkin 2026.9.211

The Plans list no longer shows plans that were renamed away or whose file no longer exists.

Pumpkin 2026.9.210

Fix Windows packaging so secure semantic search is included reliably in new installations and updates.

Pumpkin 2026.9.209

Long context defaults Off. Grok and GPT chats compact before higher-priced prompt tiers; switch it On to use the full window.
Prompt prefixes stay locked between compactions. Unexpected cache changes now show a notice.
Anthropic prompt caching now defaults to five minutes; one-hour caching remains available for longer pauses.
Tool output caps and spill files keep large results from filling the conversation.
Cache-write costs now use the five-minute or one-hour rate reported by Anthropic.

Pumpkin 2026.9.208

A message waiting for a connection cannot be sent after its waiting time has expired.
If Pumpkin loses confirmation after sending, it says the outcome is unknown instead of resending or reporting an unrelated chat error.

Pumpkin 2026.9.207

Improve installation checks so the new semantic-search security protections ship reliably on Windows and Linux.

Pumpkin 2026.9.206

Reconnecting a window no longer interrupts chats just because unchanged permissions were refreshed.
An older failed connection check can no longer remove access from a newer working connection.

Pumpkin 2026.9.205

Protect OpenAI keys during semantic search and indexing, and stop background indexing when workspace trust is revoked.
Keep local indexing working when switching away from OpenAI, including searches from nested folders.

Pumpkin 2026.9.204

Agent browsers now have short names like b1 and b2, shown in the tab title, so agents no longer lose track of which browser they opened.
Agents can see browsers opened by other chats, subagents can use their parent's browser directly, and a chat can take over another chat's browser.
A new Use in Chat button on a browser tab hands that browser to the active chat and types its name into the message box.

Pumpkin 2026.9.203

Remote pairing requests stay visible even when the device list or QR code cannot load, and refresh while Remote settings is open.
Phone reconnects reuse relay credentials instead of requesting new ones each time. Connection failures are logged without repeating the same message constantly.

Pumpkin 2026.9.202

Agents can cancel their own obsolete question cards. Canceled questions stay canceled after restart and cannot resume an old task when answered later.

Pumpkin 2026.9.201

Long turns now summarize completed tool rounds before dropping earlier context, so useful findings can carry forward within one turn.
Cached input is counted once, and the context meter shows the limit Pumpkin uses for the selected model.
If context must be removed to fit, the chat explains what was hidden. Requests are checked again after formatting and secret masking.

Pumpkin 2026.9.200

Plan statuses now follow what happens: any Agent turn builds the plan, it turns built when the turn finishes, and stuck plans are repaired on open.
Plan badges are coloured by status: ready blue, building orange, built green, done dimmed.
Build a plan from the Plans row or Context tab too. Every Build asks in place whether to build in this chat or in a new one.
Switching between Ask, Plan, Debug and Agent now tells the model about the change, so earlier mode rules no longer apply.

Pumpkin 2026.9.199

Models now has collapsible provider headings, clearer missing-key warnings, and secure key setup directly in each provider section.
Add Amazon Bedrock models with your key, AWS region and model target. Claude Sonnet 4.6 and Haiku 4.5 support streaming, tools and thinking.

Pumpkin 2026.9.198

Right-click the Autopilot icon to give it goals and guardrails for that chat. Goals steer every decision it makes; guardrails are hard limits it will never cross.
A small dot on the Autopilot icon shows that goals or guardrails are set; hovering says so too.

Pumpkin 2026.9.197

A loop iteration that waits for a background job, a watcher or a subagent now stays open until that work finishes; its cost and outcome count, and the next iteration never starts on top of it.
While a loop iteration is parked waiting, the stall timeout does not count that time; the maximum runtime still applies.

Pumpkin 2026.9.196

Models now has collapsible provider headings, clearer missing-key warnings, and secure key setup directly in each provider section.

Pumpkin 2026.9.195

Settings › Semantic search: a new "Add exclusions using AI" button starts a chat that surveys your code root and adds the right exclusions for you.
"Fix with AI" now also appears when the semantic search health check reports a degraded state, not only a failure.

Pumpkin 2026.9.194

The chats search box now shows a spinner while older chats are still being searched, so the list is not mistaken for final before the search finishes.

Pumpkin 2026.9.193

Loops can now set a maximum run time and a stall timeout; an iteration that runs too long or stops making progress is stopped and recorded as an error, and the next run is scheduled as usual.
A loop can run a short pre-check command before each scheduled iteration; when it reports nothing to do, the iteration is skipped without starting a chat or spending tokens.
Loops can post a summary of every iteration to a Slack channel or your DMs, relay their questions to Slack and wait for your reply, and pick up late replies and feedback for the next iteration.
Loop memory is tidier: a missing changelog entry is added for you, old entries are archived when the file grows large, and shared machine-wide notes are given to every loop.
The Loops tab shows spend over the last day and week, how many iterations were skipped, and which runs timed out or stalled; the loop form has a new Advanced section for these options.

Pumpkin 2026.9.192

When Autopilot decides a chat is finished, it now says so in the chat with its reason instead of staying silent.
When the agent stops to ask for a decision, a spec, or an approval, Autopilot now supplies it instead of treating the work as done.

Pumpkin 2026.9.191

Semantic search now starts and looks after its own search database on your computer, and restarts it if it stops; you no longer have to start it yourself.
Settings › Semantic search shows what Pumpkin is doing with that database (downloading, starting, waiting, up) instead of asking you to fix it.

Pumpkin 2026.9.190

After an update on Windows with WSL (or over SSH), the remote connection no longer gives up after a minute: it waits for the remote helper to start and keeps retrying, so WSL need not be restarted.
The remote helper logs how long each start-up step took (payload check, taking over the old service, starting the new one), so a slow update can be traced.

Pumpkin 2026.9.189

The pinned message at the top of a chat now has a list icon: click it to see every message you sent in that chat, newest first, and jump to any of them.

Pumpkin 2026.9.188

Two unreliable internal checks that could block a release were removed; nothing changes in how Pumpkin works.

Pumpkin 2026.9.187

When the chats sidebar groups chats by tag, you can now put the tag sections in any order: drag a tag heading onto another, or use Move Up / Move Down in its menu. The order is remembered.

Pumpkin 2026.9.186

When the background service refuses a request because of who sent it, the error now names the sender and the action and is logged, so a chat that will not open or an answer that fails can be traced.

Pumpkin 2026.9.185

After an update on WSL or SSH, a chat that was mid-turn keeps going: the new Pumpkin passes settings and credentials to the old service until that turn finishes, instead of it failing.
When the background service inside WSL (or on any machine) is still running but stops answering, Pumpkin now replaces it on its own after two failed starts. Before, only shutting down WSL fixed it.

Pumpkin 2026.9.183

When Pumpkin ends a turn because the model kept repeating itself, the chat now says so in a line that stays after a reload, instead of stopping silently.
Grok chats are compacted from about 150k tokens of context, well before Grok's 500k window, because Grok degrades around 200k.

Pumpkin 2026.9.182

Autopilot is now an icon at the bottom of the chat, like Attach and Secret. It lights up in the Pumpkin orange while Autopilot is on, and hovering tells you whether it is on or off.

Pumpkin 2026.9.181

An agent that re-runs the same tool call or repeats the same sentence round after round is told to stop; if it keeps going, the turn ends instead of running until you press Stop.

Pumpkin 2026.9.180

Semantic search no longer indexes Pumpkin's own saved chat state, which had been slowing indexing to a crawl on some machines.
Semantic search exclusions for logs, minified files, media and archives now take effect; before, they were listed but ignored.

Pumpkin 2026.9.179

Settings › Models: each model's Description and Skill rating are now laid out as tidy labeled rows instead of stacked, unspaced controls.
The chat model picker lists model names only; a model's description now shows when you hover its row.

Pumpkin 2026.9.178

Auto: the picker now weighs the task and recent conversation against each model's description and capabilities, at the lowest cost that fits, and shows its reason on the Auto line.

Pumpkin 2026.9.177

New setting Auto-rename new chats: when on, new chats start with Auto-rename every 5 messages turned on. You can still turn it off per chat.

Pumpkin 2026.9.176

Chats no longer show "foreign metadata is missing or stale" errors after an upgrade, and a chat open in another Pumpkin window no longer blocks other chats from refreshing.
Code search skips generated content by default: vendored dependency trees, lockfiles and protobuf output. These exclusions now reach every project in a workspace.

Pumpkin 2026.9.175

Models settings lets you edit descriptions and personal skill ratings for all providers, including Fireworks. Choose Catalog to reset a rating.
Model capabilities and exact token limits are visible alongside your preferences. Auto uses both when choosing a model.
Auto has a Make default action on the Models page. Rapid edits, refreshes and failed saves preserve your other model preferences.

Pumpkin 2026.9.174

Models settings now lets you edit descriptions and personal skill ratings for every model, including Fireworks, and reset ratings to the catalog defaults.
Models settings shows image, tool and reasoning support plus exact context and output limits. Auto uses these facts and your personal ratings when choosing a model.
Auto now has a Make default action on the Models page.
Rapid model edits and settings refreshes no longer overwrite another edit or discard a description you are typing. Failed saves show an error and can be retried.

Pumpkin 2026.9.173

Lower model costs: the conversation cache is no longer rewritten when the current time, saved data, a memory doc or the process step changes; that text now follows the cached history.
Grok and Meta requests use the same request layout as OpenAI, so their caches also survive those changes.

Pumpkin 2026.9.172

Auto: model skill ratings were recalibrated on a 1–10 scale (open-weight Fireworks models such as GLM 5.3 and DeepSeek were rated too high), so Auto picks models that match the task.

Pumpkin 2026.9.171

Remote (phone): the pairing and QR pages no longer fail in Safari with "Can only call Window.fetch on instances of Window".
Remote (phone): "Device vault is unavailable; pair again" right after signing in is fixed. Signed-out visitors are sent to sign in once and come back to the page.

Pumpkin 2026.9.170

Remote (phone): the Machines list shows each computer by its name before pairing, not "Computer" plus eight random characters. Computers on an older Pumpkin keep the old label until they update.
Remote (phone): a brief connection drop no longer ends pairing with a raw browser error such as "Fetch is aborted"; it shows "Reconnecting…" and carries on.
Remote (phone): if pairing does stop, the message says "Pairing was interrupted. Tap Pair to try again." and tapping Pair works.

Pumpkin 2026.9.169

Slack tool: each message line from read_thread and read_channel now shows the exact Slack ts next to the time, so the agent can reply to, react to, or track a message without guessing it.

Pumpkin 2026.9.168

Loops tab: when Pumpkin is connected to a remote such as WSL, a loop's instructions, definition and memory files now open from the remote instead of showing "file was not found".
A turn no longer ends silently when the model returns an empty reply: the agent is asked once for a closing statement, and if that is also empty a visible notice says so.

Pumpkin 2026.9.167

Autopilot: a checkbox at the bottom of the chat. Turn it on and your highest-rated model takes your place in that chat.
With Autopilot on, the agent's questions get answered, permission prompts are allowed or denied, a failed turn is retried or continued, and the next message is sent until the request is done.
Messages Autopilot sends show an Autopilot badge, and question cards it answered say Answered by Autopilot.
When the agent asks for a secret, Autopilot picks one you already stored by name; the value stays masked. If none matches, the question is marked unanswered and the agent carries on.
Autopilot pauses after a set number of actions in a row (25 by default, changeable in settings) and resumes when you send a message. Stop or unticking the box cancels it.

Pumpkin 2026.9.166

Models settings links to pricing and capabilities on pumpkin.club. Text-only models no longer receive image inputs.
Token counting works in the packaged desktop, and supported images are resized without changing saved originals.

Pumpkin 2026.9.165

The model picker is grouped by provider, then by name. Auto stays at the top.

Pumpkin 2026.9.163

Settings Models links to provider pricing and model capabilities on pumpkin.club. The editor uses that shared catalog too, including dated prices and image limits.
Images are checked against the selected model before sending. A text-only model gives a clear error instead of sending an image it cannot read.
Supported images are resized to what the model accepts; the saved originals are not changed.

Pumpkin 2026.9.162

The agent no longer ends a long reply by asking you to say the word. If it needs a decision, it asks with the question card. If the next step is obvious, it just does it.

Pumpkin 2026.9.161

You can attach any file, including zips and other binary files. The agent gets the saved copy's path and can unzip or inspect it with its shell tools.
Settings → Semantic search shows a workspace index's size on disk and its current chunk count, and counts projects rather than files.
Indexing runs at a lower priority than the editor on macOS and Linux, so a long index no longer slows down the rest of the computer.

Pumpkin 2026.9.159

The local model no longer stays stuck on "owned by another host" after an update; a second window waits for the first to close instead of saying the model is CPU only.

Pumpkin 2026.9.158

Sending a message or answering a question while Pumpkin is updating itself no longer fails with "Mutation acknowledgement unavailable".
Pumpkin waits for its background service to come back, then says plainly whether your message went through, was never received (try again), or could not be confirmed (check the chat first).

Pumpkin 2026.9.156

Reissues the subagent-question update for all desktop platforms: questions appear in the parent chat, and clicking the question icon jumps to the answer controls.

Pumpkin 2026.9.155

Questions from subagents appear in the parent chat, labelled with the subagent that asked them. Answering in either chat updates both.
Clicking a chat's question icon takes you straight to the question and its answer choices, even when it came from a subagent.
Subagent questions count once in the question badge, and deleting their chat no longer leaves a question waiting in the parent.

Pumpkin 2026.9.154

Get a key opens the page where you create an Anthropic, OpenAI, or xAI key, from Settings, first-run setup, and a missing-key error.
Settings can select Claude Opus 5.5, GPT-6 Sol, and GPT-6 Luna.

Pumpkin 2026.9.153

Long chats no longer repeat the final answer twice or loop for hundreds of rounds re-checking something blocked; the hidden reminder behind both now fires at most once per turn.
A model that starts repeating itself mid-reply (seen on Grok and GLM) is cut off and retried once instead of streaming filler until you press Stop.
Attached text files have no size limit; a large file is read by the agent in windows instead of pasted whole into the request. Binary files are still refused.

Pumpkin 2026.9.152

What's new lists every change since your version, in plain English.
You can add a Meta AI key and use Meta AI models.
A hook permission card shows the command it is asking about and offers only Allow and Deny. Typing into it no longer produces an error you cannot retry; a typed message queues for the agent instead.
The repeated "could not be refreshed from its saved metadata … run initial session listing" errors when another Pumpkin window or the command line saves a chat in the same workspace are gone.