Every release gets a section here. frontend/scripts/publish.mjs copies the
section whose heading matches version in frontend/src-tauri/tauri.conf.json
into the GitHub release body and into the notes field of latest.json,
which is what the in-app update sheet shows before installing. Add the section
before running npm run release; the release stops if it is missing, because a
feed with empty notes tells the user nothing.
- Batch forward messages across sessions within the same project. A new Forward button beside the attachment button switches the transcript into multi-select mode for user messages, assistant replies and compaction summary cards, with a batch bar above the composer and a target session picker that supports filtering and creating new sessions.
- Selected messages are merged chronologically into one payload: read-only tool calls are excluded, only edit, write and code-execution tool records are retained with summarized arguments and details capped at 600 characters per call, and compaction summary cards are forwarded in full without truncation.
- Forwarded payloads render in a dedicated collapsible card showing the source session, structured sections with Markdown formatting, and expandable tool records instead of raw bubble text. Confirming a forward switches straight to the target session and immediately submits the merged message to run the agent.
- Switch models directly from the composer toolbar. A new model selector beside the Forward button opens upward to list configured providers and their models, stays in sync with the right-dock LLM panel (with a shortcut to open it), and applies the chosen model starting on the very next message.
- Compacting a long conversation is no longer reported as a failure just because it takes a while. The client used to stop waiting after two minutes and say nothing had been folded, even when the run was still working and finished a minute later; it now waits long enough, and when it does give up it checks what actually happened before saying anything.
- Running compaction twice in a row no longer leaves two summaries. A session that has not been spoken in since the last summary is refused with a reason instead of folding the summary into another summary.
- The project panel now fills the window instead of stopping at a fixed height, so the file list shows around twenty entries at a time rather than five.
- The project list stops at five entries and scrolls. A long registry no longer pushes the file list out of the panel.
- Files carry icons that match what they are: folders, source code, configuration, documents, images, stylesheets, archives and binaries each look different.
- Clicking a file opens it beside the tree instead of launching another program. Text arrives with line numbers and syntax colour, images open in place, and the button in the preview header still hands the file to its default app. Closing the panel closes the preview and releases what it had loaded.
- Files saved in GBK or ANSI open correctly instead of showing an error, and a very large file can no longer stall the panel: text is read up to 512 KB with a note, and images above 16 MB are refused with the reason.
Switchin the panel header is nowRefresh, which is all it ever did.
- The accent colour you pick now survives a restart. Choosing a swatch only saved it on this machine, while the theme record kept whatever was there the last time you switched between Daylight and Nightfall, and that older value was reapplied on every launch. The colour you picked last now wins, and a fresh install still inherits the theme from your other machine.
- Corrections are remembered more accurately. Ordinary technical wording such as "this only works in dev" or "must be initialised first" used to be filed as a standing rule; it no longer is, so what the agent keeps is what you actually asked it to keep.
- Rules you set early no longer fade out. Once enough preferences piled up, the oldest constraints dropped out of the text the agent reads each turn and were eventually retired, which looked like it had forgotten something you taught it. Each kind now keeps a share of that space.
- A remembered correction now says what it was correcting. "Don't do that again" used to be stored without the part explaining what "that" was; the note that reaches the agent carries it now.
- Compaction no longer comes back empty. A model that thinks before answering could spend its whole output budget deliberating and hand back nothing, which read as a silent failure. The instruction that this is a reorganising job rather than analysis work now reaches every summary request, including the multi-part path that was missing it, and an empty answer is retried once with more room before anything is reported.
- The appearance panel has three more accent colours to choose from: rose, teal-blue and slate grey.
- The project knowledge base can now be checked instead of trusted. A script reports which source files it does not know about, which entries point at files that no longer exist, and which entries name a file without saying where it lives. "Cover every file" used to be an instruction with no way to verify it.
- The agent reaches for the knowledge base before searching the project: it reads the symbol index of the chapter that owns the area to find a file, and adds the entry itself when a lookup misses, so the gap does not have to be hit twice.
- Long conversations can be compacted again. A transcript too large to summarize in one request is now handled in parts and merged, so its length decides how many requests it takes instead of whether it works at all.
- A failed compaction is now reported as a failure. When the summary request was rejected, the session used to be folded into a short placeholder and shown as a success, which is why a compacted conversation could suddenly forget everything. Nothing is folded now, the transcript stays exactly as it was, and you can run it again.
- The outcome is shown in the compact window itself, with the reason and a "Try again" button. It used to appear only as a corner notification that vanished after a few seconds, easy to miss while waiting for the run to finish.
- The "not enough messages" warning no longer appears on a conversation that clearly has history. It could show up right after opening a session and only cleared after switching away and back.
- The summary has far more room to write, so a model that thinks before answering no longer comes back empty.
- Lessons remembered from earlier conversations are now a fixed set of the ten most recent, instead of being re-picked for every message. Switching topics no longer re-bills the whole conversation, so cache hits are back around 99% rather than the low nineties. The trade-off: the set is no longer filtered for relevance to what you just asked.
- Lessons are now learned from English as well. Before this, only Chinese phrasings like 不对 / 必须 / 记住 registered, so an English conversation recorded almost nothing. English corrections, constraints, preferences and confirmations all count now, and English sentences are split correctly, so numbers like 3.14 are no longer treated as sentence breaks.
- New lessons are written in the language of the conversation that produced them, rather than always in Chinese. The block that carries them is bilingual, so it reads naturally either way and never shifts on its own.
- Old lessons that were never actually used are cleaned up at startup: anything created more than 90 days ago, never shown in the prompt, and low on confidence gets retired. Only lessons that were outside the prompt anyway are touched, so cleanup does not cost any cache hits.
- Worth knowing: some actions still cost one full re-read of the conversation, because they change the instructions sent ahead of it. Those are switching projects, editing the user note or the agent note, adding or changing a lesson, changing an agent prompt, a skill, the tool list or the MCP servers, and switching agents. Compaction counts too, since it rewrites the history, and so does a long gap between messages, because the provider's cache expires on its own.These are all normal.
- Everything else is free: a new topic, a new session, coming back to an old one, or restarting the app all keep the cache intact. Working normally in one project should sit around 99%.
- The Files tab in the right dock showed a literal
7from the prototype: nothing in the codebase wrote to that node, so every session displayed it, even one that had edited nothing, and accepting or reverting never moved it. It now readsfiles.totalFiles, thetotal_filesfromsession.file_changes(one entry per changed path, reverted snapshots excluded), the same figure the pane'sfiles changedheader prints, and shows nothing at zero. The binding sits onspan.cnt, not the button, becauseorbit/runtime.jsowns theonclass it binds once per tab. - The reasoning row could not be selected.
.th-texttook only its content width inside the flex row, so everything past the last word hit the.thinkingcontainer instead of a text node: the cursor stayed an arrow and a drag never started a selection. It now takes the row's spare width, carriescursor: textand an explicituser-select: text, and the sheen band drawn by.th-text::afteris out of hit testing. The 84% fade mask now measures the row instead of the text, so a short row no longer fades its own last characters. - Scrollbars in the right dock were always on. Two containers draw one:
.dr-content(the vertical rail, plus the horizontal oneoverflow-y: autoimplies on the other axis) and.chg-diff, which scrolls sideways on an expanded diff. Both thumbs are now painted transparent until the pointer enters the panel and cleared again 500ms after it leaves. Width, position and colour are unchanged, and no track is collapsed, so revealing a bar never reflows the pane under the cursor. - The live run capsule showed a moving waveform while the session was idle: the five
.eqbars looped from mount, with no relation to whether a turn was running. They now loop only while.islandcarriesbusy, which isstreamState.streaming, set the moment a prompt is accepted and cleared onmessage.complete,prompt.rejectedand interrupt. The bars therefore move for exactly the length of one turn and settle into a still glyph between turns.
- Prompt caching was silently broken on every turn. The experience block was rendered from a recall query that is the current user message, so it differed each turn, and it sat in the middle of the system prompt: everything after it, the tool schemas and the whole message history, was re-billed from scratch. On a 150k-token context the first request of each turn reported near-zero cache while the remaining iterations of the same turn hit normally, which is how the session average slid from 99% to 96% without the chat loop changing. The rendered block is now cached per
(project_id, agent_type)and re-rendered only when a fingerprint changes: active entry count,MAX(updated_at),MAX(created_at), total statement length and summed confidence, plus the agent note's row count,MAX(updated_at)and content length.hitsandlast_hit_atare excluded on purpose —mark_hitsrewrites them every turn, so including them would defeat the cache; hits are still recorded on a cache hit, keeping retirement accounting intact. - The experience block also moved to the end of the system prompt, after the project-path instruction and the user note. It is the only section that can change between turns, and the provider caches by prefix, so a refresh now costs its own 3-4k characters instead of invalidating everything that follows it.
- The
memory.promptpreview RPC passesuse_cache: false. It exists to show what a given query would inject, and letting it write into the cache would push a preview block into the next real request.
- A GitHub tab in the right dock, reachable from the octocat beside Notes or from the command palette. It is icon-only because a sixth word would have pushed the tab strip past its width on a 1280px window, and the strip now scrolls when it overflows: wheel deltas map onto horizontal scroll, arrows appear only on the edge that still hides a tab, and a fade mask marks the same edge.
- The panel never talks to api.github.com. The renderer's CSP keeps it on localhost, and a personal access token has no business in a process that also renders agent-authored HTML into a sandboxed iframe, so every call goes through a new
github.*RPC domain. All of those methods run in the thread pool: a REST round trip is 100ms to 2s, and a synchronous call would freeze the WebSocket loop, stalling streaming replies and permission prompts with it. - The token is borrowed from the github entry the MCP panel already writes, so nobody fills the same credential twice. Only
GITHUB_*variables are considered and the bundledghp_xxxxplaceholder is rejected, because a 401 from a literal placeholder reads like an auth bug rather than a missing setup step. Responses are cached by ETag, andforce: trueskips that cache for one request, which is what the refresh button sends. - The repository rack shows a language-coloured edge, stars, forks, relative push time and topics, filters locally, and appends another page on
Load more. The repository the active project is a checkout of carries acurrentbadge, read from.git/config's origin remote rather than by spawning git. - The detail header answers "what of mine is not on GitHub yet" with two numbers instead of one: commits on HEAD that no remote contains, and files with uncommitted work. The first alone is 0 for an afternoon of edits that were never committed, which made the strip claim "in sync" over twelve modified files.
- Files browse as a tree with breadcrumbs. Text renders with the transcript's own highlighter, binaries get a placeholder card, and anything past 1 MB says so instead of failing. Pull requests and commits both expand in place into per-file diffs; the commit listing endpoint carries neither stats nor patches, so a row's
+/-exists only after the click that asked for it. Both mark the result as partial when GitHub omits a patch or the file list hits its cap. - Any file, issue or pull request can be handed to the conversation. The Ask button writes
@pebble-file: <path>when the file belongs to the project's own repository and a GitHub URL otherwise, because a local path the agent cannot read is worse than no reference at all. - A 52 by 7 commit heat strip at 4px cells, with level thresholds absolute rather than scaled to the repository's own maximum so a quiet repository looks quiet. GitHub answers 202 the first time this table is requested for a repository, so
pendingis passed through rather than drawn as "no commits", and the client retries twice. - Panel scrollbars take the 4px treatment the reasoning block already used. The first pass set
scrollbar-width: thin, which makes Chromium drop every::-webkit-scrollbarrule and paint its own grey bar, so the custom width could not take effect until that was removed.
- Each row in the pricing table gains a
ctxfield for the model's context window, stored ascontext_limitand read ahead of the name-based heuristics, so a model that matches no keyword is no longer measured against the 128K default. Subscription models belong in the table as well: enter the model name and its window, leave the rates empty, and the row stays unpriced. Windows are keyed by model name, so a model already listed needs no second entry. - Cached prompt tokens are now read by one function on both the billing and the unpriced path. Billing had read only DeepSeek's private
prompt_cache_hit_tokens, so a priced model behind an OpenAI-compatible gateway that reportsprompt_tokens_details.cached_tokenswas charged the full input rate for its cached prompt and showed a 0% hit rate. Values are floored at zero and clamped to the prompt size. - The context pane splits the window across four estimated parts instead of three plus a remainder. The gateway now estimates the conversation itself and the pane scales all four by the measured
prompt_tokens, so tokenizer differences are shared rather than charged entirely to the conversation. The system message is skipped in that estimate because the base prompt already counts it. - The
tool tokens this turnlist now reports what each tool's output cost, largest first, instead of names alone. Five rows stay visible and the rest fold behind a counter, so a long turn cannot push the rest of the pane off screen. Those tokens are detail, not a fifth part of the split: they are already inside the conversation segment.
- The MCP panel no longer sits on "idle" after a cold start. The backend was already pushing
mcp.state_changedas each server connected, but the frontend had no branch for that event and dropped it, so the rack kept whatever amcp.listanswered before the servers were up. The event now triggers a re-read, coalesced over 250 ms so a burst of six servers costs one request.
- The MCP panel no longer asks for the same approval on every launch of a packaged build. The bundled
.mcp.jsonis unpacked into a fresh_MEIxxxxtemp directory at each start, so a fingerprint bound to that path could never match twice and the banner came back every time. A file under the bundle root is now treated as part of the program and loads without a prompt, while every other project-level.mcp.jsonkeeps the content-fingerprint gate. - Stale MCP approvals are pruned at startup. Each packaged launch used to add one temp-path fingerprint to
mcp_trusted.json, and a removed project left its entry behind; left alone they fill the 200-entry cap and push out approvals for projects that still exist. The store is now cleaned once per process — onefile temp entries, duplicates and entries whose directory is gone are dropped, and a project on an unmounted drive keeps its approval.
- Links in the transcript are clickable. An address written as
[label](url)or pasted bare turns into a themed anchor, and a click hands it to the OS default browser through a newsystem.open_urlRPC instead of navigating the app window. Onlyhttpandhttpsever reach the shell —file:,javascript:anddata:stay plain text, because the link text comes from model output. The address stops where the ASCII text stops, so a Chinese sentence keeps its。and,outside the href and参考https://xis detected without a space in front of it. - Copy and Delete on a user message work again. Sending a prompt drew a
pending-Nrow whose buttons carried that id inside the markup; when the server confirmed the message the row adopted the stored id but the markup kept the old one, so the click handler looked up a row that no longer answered to that name and returned silently. The markup is rewritten with the adopted id now, and the branch that draws a row for a prompt this client never queued emits the buttons at all. - Copy reports every outcome. An empty row says
Nothing to copy, a refused clipboard write retries through a scratch textarea, and that fallback checks whether the copy actually happened instead of printingCopiedover a refusal.
- The MCP panel stopped putting a token field on every server. Only a server whose env carries a credential gets one now, and it writes into the variable that server actually reads (
GITHUB_PERSONAL_ACCESS_TOKEN); before this every entry showed "paste token…" and the secret went to a fixedMCP_TOKENkey that github never looks at. - A server that is configured but cannot start now says what it needs.
mcp.listreportsready,missing,requires_tokenandtoken_envper server, plus anavailablelist drawn frommcp_install.json; annpxserver on a machine without Node, auvxserver without uv, andcodebase-memorystill pointing atREPLACE_WITH_...all showneeds Node.js 18+and an Install button instead of looking broken.mcp.install_serverhandles npx and uvx entries now, not just download-type ones. - Saving a token no longer floods the approval queue. Both MCP RPCs used the bundled
.mcp.jsonas their base when no user file existed, so the first secret saved on a fresh install copied all six bundled servers into~/.pebble/.mcp.jsonand approved only the edited one; the other five appeared as "wants to run" banners. Each write now touches a single entry. - The command palette is built from the stores now instead of static markup. It used to hold fourteen hand-written rows, four of which only fired a "Command ran" toast because they carried no handler, and three subtitles with numbers typed once ("14 entries", "3 installed · 2 running", "5 sessions"). Session rows come from the real list, counts from
ledger.totalandmcp.servers,Switch sessionis gone because the sessions group already lists the same rows, andModel & agentsopens the right dock's LLM tab —ModelPanel.vuewas a hardcoded prototype (claude-sonnet/opus/haiku, two invented agents, an Apply button that only fired a toast) whileLlmPane.vuealready renders the real provider stack. - The Files pane no longer opens on another project's files. It seeded four changes with invented
+/-counts and two whitelist paths from the author's machine, and a session that edited nothing kept showing them; both are gone now, andfiles.stateseparates an untouched session from a refused read. The whitelist sheet gained Edit and Remove, aninvalid entriessection for paths the gateway refuses to save, and a reason for every refusal (blank, duplicate, not connected, five-root ceiling) — the old Add silently returned false for all four.
- Skill media is trimmed, not the skills. The huashu-design examples that made the package heavy are gone: six background music tracks and 24 pre-rendered showcase screenshots. All 24 showcase HTML files stay, so the style gallery still works, and the skill's instructions now open those HTML files instead of pointing at images that no longer ship. Every skill still loads and behaves exactly as before, the install is simply much smaller.
read_pdfis removed. A coding agent has no use for it, and the engine behind it (PyMuPDF) would have cost roughly 66 MB, none of which ever reached the shipped build. Users who need PDF content have better tools.- Uploading a PDF is unaffected. A PDF dropped into the conversation is still read through
pypdf, a separate library; it now appears inrequirements-optional.txt, where it had been missing all along. - Also out of the bundle: numpy and sounddevice (27.8 MB, both unreachable once OCR was removed) and the project's own knowledge base (0.31 MB). With the trimmed skill media the sidecar loses about 57 MB uncompressed.
grepnow carries its own ripgrep. The tool shells out torg, and the packaged app depended on the user having it on PATH, which almost nobody does;vendor/ripgrep/rg.exe(4 MB) ships in the bundle and_resolve_rg()only falls back to PATH when that copy is missing.globwas broken for any pattern carrying a directory. It comparedfnmatchagainst the bare file name, soknowledgesource/*.mdreturned nothing; matching is now segment by segment,*staying inside one segment and**spanning any depth, while*.pystill searches every depth as before.
- The bundled skill set drops from 120 to 50. Kept are the skills the agent's own instructions depend on and the ones most users actually reach for: the CKB workflow, the design and OfficeCLI document chain, and the HyperFrames video pipeline. The rest were written for the author's own workflows and no longer ship.
- Anything missing installs on demand: ask the agent to add the skill you want and it writes
~/.pebble/skills/<name>/SKILL.md, which loads like a bundled one and shadows a bundled skill of the same name.
- Compaction no longer folds an empty summary. When the model answers with an empty body, the retry's answer is kept instead of being discarded, and a summary that carries headings but no real content is rejected rather than stored as a successful compaction.
- Compaction gained a fallback chain: model, one strict retry, a reformat of whatever came back, then a rule-based summary built from the newest user turns. If none of them carries usable text the operation reports failure and writes nothing, so the session stays intact and can be retried.
- Headings are fixed English (
Task / Status / Decisions / Todo / Clues) while the section bodies follow the conversation's language, decided in code from the user messages rather than left to the model. - Tool results now reach the summary at all. The gateway forwarded only the call arguments and dropped the returned output, so tool-heavy sessions summarised into nothing; arguments and results are both forwarded now, along with the names needed to tell a read from a shell run.
- Tool rows are graded by family: reads are cut to 200 characters, edits to 400, while commands and every failed call keep their full output, because a stack trace exists nowhere else.
- Compaction keeps one whole turn instead of the newest N messages: from the last user message through every reply and tool call it produced, capped at twenty rows. On a tool-heavy turn the old rule kept two shell echoes.
- Summary budgets are much larger: input 700K tokens, output 250K. A three-hundred-thousand-token session that used to be trimmed to thirty now goes in whole.
- Folded messages are backed up to the project's history file. The session id and project path were previously dropped, so no backup was ever written.
web_searchuses DuckDuckGo's html results and falls back to Bing pinned to the US market, replacing the localized Bing endpoint. Google is not in the chain: its static endpoint answers a JavaScript redirect shell rather than results. The tool now asks for English keywords, since localized sources return markedly weaker results for technical queries.
- Session switcher: the five seeded prototype conversations are gone. The list starts empty and the gateway's
session.listanswer replaces it, so a fresh project no longer shows clickable sessions that do not exist. - A failed
session.listand an empty one are now handled differently: a failure leaves the panel untouched, while an empty answer clears the rows before a new session is created. - An empty panel says which empty it is: still loading, could not read, or no sessions yet.
- Deleting a session asks for confirmation in the shared sheet. A refused delete puts the row back and shows the gateway's reason, and an offline delete is refused instead of reported as success.
- The row's More button opens the conversation menu only; the leftover prototype toast that fired alongside it is gone.
- Search filters through Vue state instead of inline
display, so it survives a list re-render and is cleared when the project changes. - Export falls back to the gateway's markdown when it cannot write its own file, instead of downloading a header-only stub.
- A reconnect re-reads the session list instead of trusting the one loaded at startup.
- Experience ledger: the five seeded demo entries and the hardcoded total are gone. The list starts empty,
memory.listreplaces it, and the panel says loading, could not read or no entries yet instead of rendering all three as an empty list. - The detail pane edits now: the statement is a textarea, confidence is a slider, and Save works on a draft. It reports success or the gateway's refusal, and Delete asks for confirmation first.
- Agent notes: switching agent, scope or tab asks before discarding unsaved text. A slow read can no longer land on the wrong agent, and a failed read keeps the draft and says so instead of looking like an empty note.
- Clear asks for confirmation, and Save/Clear report success or refusal instead of leaving the button silent.
- An agent that disappears from the list is no longer left selected, and Save/Clear stay disabled until an agent is chosen.
- New entry reports a refused save, and the footer shows what is on screen versus what is stored instead of a hardcoded "synced".
- Usage pane: saving the pricing table reports the gateway's own verdict. A rejected table used to read "saved" while the file on disk was untouched; the reason now stays on screen until the next attempt.
- A period with no rates yet starts every cell at zero, so editing a single price on a freshly opened peak or offpeak tab can actually be saved.
- Today / 7 days / 30 days keep updating when the current session cannot be read, instead of freezing on their last numbers.
- Model names are editable in the table and a duplicate is refused; removing a stored model asks first; unsaved edits are marked and Save is disabled until something changes.
- A pricing table that finishes loading while the editor holds unsaved work asks whether to use the stored table or keep the edits.
- Project panel: the file tree now lists the active project instead of the app's install folder, and says whether it is loading, empty or refused instead of drawing a blank list.
projects.jsonis written atomically with a.bakcopy. An unreadable registry is kept asprojects.json.corruptand reported in the panel rather than shown as "no projects"; adding or removing a project is refused while it is corrupt.- Empty paths no longer resolve to the sidecar's working directory, so adding a project, adopting sessions or starting a session can no longer point the tools at the install folder.
- Switching projects is refused while a reply is running, including a turn started by another client, and a failed switch rolls the active project back instead of reporting success.
- The panel has a single owner again: the leftover prototype handlers that toasted "switched" without calling the gateway are gone, the popover closes after a pick, and refused add/remove/switch operations say so.
- Adding a project opens the path the gateway actually stored, removing one asks first, and adopt reports how many sessions moved.
- Sessions stored under an older spelling of the project path (separators, trailing slash, case) are visible again instead of vanishing from the list.
- One shared confirmation sheet replaces
window.confirmfor provider deletion. - Removed the dead
Managebutton in the project panel; the session list'sManagenow readsProjects.
- Fixed the add-provider crash loop; a failing RPC now returns an error without closing the WebSocket.
- Fixed provider/model routing for duplicate model ids, per-agent bindings, and draft providers; drafts activate only with a base URL and at least one model.
- Deleting a provider or its last model now repairs the agent model instead of sending a stale name.
- Invalid provider operations now fail; base URLs and request-shaping fields are normalised and validated.
- Provider and agent configs are written atomically with a
.bakcopy; a corrupt user config no longer falls back to bundled sample providers. - LLM pane: draft/active/no-key states, per-provider settings, Enter/blur model entry, auto-activation, delete confirmation, and dirty-edit preservation.
- Unknown provider fields and agent settings survive saves; agent selection persists; the reply model label follows the actual model.
- Fixed the update sheet opening above the window. It was mounted inside the title bar, and the bar's blur effect made the overlay size itself against that 78px strip instead of the window, so the notes were cut off at the top and could not be read.
- Carries what 1.0.7 shipped, since that release's own notes were unreadable because of the bug above: the LLM pane can add and switch providers from the panel (the provider grid ends with a
new providercard, and clicking a card makes it the active endpoint; before this, adding one meant editingllm_providers.jsonby hand), and the pairing sheet is marked unavailable instead of showing a QR code and a code button for a mobile companion that is not part of this build.
- The LLM pane can add and switch providers from the panel now. The provider grid ends with a
new providercard, and clicking a card makes it the active endpoint. Before this, adding one meant editingllm_providers.jsonby hand, and clicking a card only changed its outline. - The pairing sheet is marked unavailable. It still showed a QR code and a code button for a mobile companion that is not part of this build; the sheet now carries a notice and the
New codebutton is gone.
- Optimize the experience and make the titles in the side menu bar more readable. The "max toks" option that was not useful in the LLM menu has been removed. The new update notification function has been added. After clicking the "Update" button later, an update notice will appear, which will help you understand the situation.
- User skills: drop a
SKILL.mdinto~/.pebble/skillsand it shows up next to the built-in ones. A user skill with the same name overrides the built-in copy, so you can take one apart, adjust it, and keep it across updates. - Custom MCP servers: the agent can write
~/.pebble/.mcp.json. Every entry waits for your approval in the MCP panel before it is launched, and editing an entry sends it back to waiting so the approved text is always the text that runs. Removing an entry disconnects it straight away instead of leaving the process on the rack until the next restart. - MCP tools whose names contain a hyphen (
get-sum) now dispatch correctly. Before this, only the hyphen-free tools of a server could be called. - Section headings in the right dock are larger and drawn in the primary ink, so they read as headings instead of disabled text.
- Removed the
max tokslider. It never reached the API, so answer length is unchanged: the provider's own default applies, which is what was happening already. - Token shares in Usage show two decimals instead of rounding to a whole percent.
- The version badge in the top left follows the build version instead of a hardcoded label.