--- title: "2026.08.27" date: 2026-08-27 extra: version: "2026.08.27" build: 15 change_id: mlqwrkzuzutr highlights: - text: "Every conversation shows what it truly cost, with a full spend breakdown" - text: "The embedded engine now runs your local models on Android" - text: "Fetch tools reshape responses with jq, model and template filters" path: "/tools" - text: "Ask your other devices for a missing API key" path: "/settings/devices" - text: "Unsent drafts are kept, and follow you between devices" path: "/conversations" --- This release is about trust and reach: costs you can verify down to a single tool call, local models on Android through the embedded engine, fetch tools that reshape a response before your agent reads it, and API keys and unsent drafts that move between your devices. As always, BYLM is a constructor — you bring the provider account or the model file, the tools, and the prompts. ## Costs you can trust Every assistant reply, title, summary, tool call and sub-agent run now records its spend durably, so a conversation's cost no longer changes with scroll position or shrinks when a failed reply is retried. A one-time migration brings your existing conversations' recorded costs in too. The conversation's total now includes what its tools spent, itemised on its own row. A new per-conversation spend page breaks the total down by purpose, model (input/cache/output tokens and per-phase price), tool, and sub-conversation — and a Steps view shows each request with what answering it cost, endpoint fees and filter stages priced apart. Each tool tile shows its call's duration, tokens and dollars, ticking live while it runs. Money stopped leaking at the edges as well: a call that spends and then fails, parks for your input, or runs as a single-shot (non-streaming) completion now counts what it charged, and parallel OpenRouter calls no longer swap or double-count their costs. ## Local models on Android The embedded llama.cpp engine now works on Android: point it at a model file and generate on the phone's own CPU. On every platform the engine gained device-local tuning — context size, KV-cache quantization and flash attention, settable per provider and overridable per model — plus a thinking budget that stops a reasoning model from spending the whole turn thinking. Speculative decoding grew past Gemma-style MTP heads: DSpark and DFlash draft heads pair in the bundle form, speculation now works on tool-calling and structured-output turns too, and the response's performance panel shows the draft head's speed-up and acceptance histogram. The pinned llama.cpp moved forward, bringing upstream's newer model architectures and backend fixes, and the engine's per-token debug chatter no longer floods the log. ## Fetch tools that shape their own responses A fetch tool can now run its response through a chain of output filters before the agent reads it: - A **jq filter** reshapes a JSON response — `.items | map(.name)` instead of a paid summarisation step — and can read the response headers. - A **model filter** runs the response past any configured model, one question and no loop, with an input cap, a timeout, and an optional JSON schema to hold the answer's shape. Its prompts can quote the response headers. - **Template filters** can declare their output as text, Markdown or JSON, render the response headers and the request, and treat an empty render as an error. The form warns while you build a chain whose stages cannot read each other, instead of letting a paid call discover it. Error handling is configurable per endpoint: which HTTP statuses count as errors, and an optional error template that turns a raw 500 body into a clear failure message. A tool can pin the response content-type it expects, advertise it in the Accept header, and refuse anything else. Diagnosis got a real home: every fetch call keeps its raw response, each filter stage's output, timings, and the exact prompt a model filter sent — opened from a Response details row on the tile, syntax highlighted, with re-indent, line-wrap and render-as-Markdown switches. A running call says which stage it is on ("Template · step 2 of 4"), a waiting one names the failure and the attempt ("Rate-limited — retrying in 42s, attempt 2 of 4"), and a paid endpoint's own fee can be read from its response and counted with everything else. The tool test screen runs through the production retry loop and shows the same captures, attempts and spend. ## API keys and drafts across your devices A device missing a provider's or fetch tool's API key can now ask your other devices for it — from the key field, the detail page, or in batch from a banner on the provider and tool lists. The device holding the key approves or declines on its Devices page (several asks in one tap), and the key arrives right under the field you asked from, ready to accept. Missing keys are flagged on lists and detail pages before a request fails, a 401/403 points at the key rather than dumping a raw error, and a provider that needs no authentication can say so. Unsent messages stopped being disposable: leave a conversation or the whole app mid-sentence and the draft — attachments and caret included — is waiting when you come back, marked on the conversation list. With sync on, a draft started on one device is offered on the next, named after the device it came from, with "Use here" and "Discard". Multi-device conversations got honest, too: a conversation being answered on another device says so (and for how long), the list marks rows still generating, opening a running turn mid-flight replays what has already been produced, and a second device asks before re-running a turn another device already owns — so you are never billed twice for one reply. ## Reading, selecting, and quoting Text selection in conversations now works the way you expect: drag past the edge of the screen and the list scrolls, one selection spans several messages, tables and code blocks select with the text around them, and a finished reply's last paragraph joins the selection. Copies keep their Markdown — bold markers, table pipes, fenced code — and a partial table selection pastes as a table under the headings it came from. Select any passage and Quote it into the composer as a Markdown quote. Tool results render by content type — Markdown, formatted JSON, inline image and file previews — with syntax highlighting, and the assistant's thinking renders as Markdown like the reply it precedes. Links in messages and tool results open behind a confirmation naming the real destination. The conversation header's cost figure, the centred jump-to-bottom button, and a progress line that names the running tools round out the reading experience. ## Context under your control Compacting a conversation now asks first: the dialog spells out the cut, lets you keep the last few turns in context, and can tell the summarizer what matters in the messages being dropped. A running compaction is visible and cannot be started twice. And a compaction can be taken back — from the snackbar right after, or from the boundary divider long after — releasing the cut and cancelling the summary it started. When compaction is configured to summarize, the manual action now actually summarizes; the bare cut is named "Clear context" for what it does. ## Tools you can tell apart, everywhere they appear Every tool can wear an icon — a glyph, an emoji, or artwork fetched from a URL, which the app can find from the endpoint itself — shown in the lists, the picker, an agent's tools, and on each call in a conversation. The tools list gained search and kind filters, agent tool-picking became a searchable full-screen picker, and deleting a tool first names the agents and counts the conversations still using it. A tool synced from a newer app version no longer vanishes (taking its whole kind with it): it is listed under "Not available on this device" with the reason, and a conversation names the tools it would run without, above the composer. The tools that ship with the app are labelled Built-in rather than Global. ## Providers and agents Star the models you actually use on a provider and they sort to the top of every model list, synced across devices. Copy a provider or a fetch tool into a pre-filled create form — asking first whether the copy carries its API key. Agents gained an optional short description, shown on the list and detail pages and to calling agents; the agents list gained search. Sub-agent calls became first-class: the spawned conversation streams to whoever opens it and its Stop button works, the open link appears the moment the run starts, the tile reports each step, and the run's tokens and cost land on the call that ordered it. Agent-built tools got sturdier rails: a mistyped config field is rejected by name instead of silently ignored or crashing opaquely, values outside a field's range are refused at save time, and an agent can reuse a secret one tool already holds — after you approve — without ever seeing its value. ## The app tells you what changed Settings → What's new lists every shipped release's highlights and full notes, readable offline, opening on the release you are running — and highlights navigate straight to the screen they describe. After an update, the conversation list greets you once with what the new version changed. Settings → Open source credits every component this build actually contains, with its license in full; the same inventory is published on the website. ## And the rest - The app opens to the conversation list about a second faster on cold start. - On wide screens the navigation drawer becomes a permanent, collapsible sidebar, and the menu button badges the jobs, unread conversations and key offers waiting behind it. - Back on a top-level page opens the navigation menu; leaving the app takes a second, deliberate press. - Settings that cascade (global → workspace → agent) now merge field by field, show the value they inherit in the field itself, and say what a reset would restore. The app-wide layer lives on one Global settings page beside the workspaces. - Conversation titles and summaries stopped breaking: a truncated or fenced enrichment reply no longer becomes the title, thinking models get room to think, and the reply cap fits the summary length you allow. - Prices in a reply no longer render the words between them as a math formula. - An image attached to an unsent draft survives the orphan-blob cleanup.