Include cache reads and writes in the complete input denominator for OpenCode usage. Use the same calculation for stored records, summary totals, and renderer fallbacks so existing history is corrected without migration.
Add regressions for full cache hits, cache writes, mixed Runtime aggregation, and persisted conversation/system usage. Focused validation passed 41 tests; repository-wide checks were blocked by concurrent work-mode changes.
Release note: 修复 OpenCode 缓存命中率偏高或完全命中时显示为空的问题，已有用量记录在重新查询时按完整输入量计算。
Require a root folder in the local project creation dialog and IPC while preserving updates to historical projects with blank paths. Show remote symlinks whose targets are directories and resolve their canonical paths when browsing.
Validation: focused UI, contract, IPC and SSH tests passed; typecheck and lint passed. Verified directory links on the shared Linux Host. Full test run timed out after a Magic Notes storage timing failure.
Release note: 新建本地项目必须选择目录；远程项目选择目录时可查看并进入指向目录的符号链接。
HTTP OCR failures in PPTX parsing were incorrectly labeled as local OCR failures. Report the selected provider and slide/image location, and retain bounded structured server diagnostics with configured-key redaction and log IDs. Preserve existing fallback behavior.
Validation: 38 focused tests, typecheck and lint passed. Full test run timed out after 200 seconds. Synthetic service probes reproduced HTTP 500; the server-side failure remains unresolved.
Release note: PPTX parsing warnings now identify the actual OCR provider and failed image location, with server diagnostics for troubleshooting.
The renderer no longer holds the whole history. Startup sends queued
changes, lets Main reconcile leftover pending/running records against the
stored tasks (same localized interruptedOnRestart line), then loads the
first page (200). The Activity page loads a full visible batch (500, as
before paging); "load more" shows the next batch and loads the next page
behind it; a filter switch refetches that filter's first page; leaving the
page trims to one batch. Counts, project/conversation groups, titles,
conversation status and the clear count come from activityHistory.summary
(refreshed on entry, by the page timer, the refresh button and after load
more / filter changes).
Records that are not loaded are still updated incrementally: the update
contract gains content-matched changes (upsert-call, update-request,
resolve-approval, settle-request, remove-request, remove-call) that Main
applies in SQL to shown rows; the store applies the same semantics to its
pages through the shared reference. update returns Main's stored call
records so a call whose earlier record was not loaded keeps one id. Task
refresh reconciles via Main SQL. Legacy localStorage records migrate as
incremental changes; the renderer no longer calls replace.
Tests: randomized sequences of every action, page loads, trims, flushes and
injected failures against the former full-list code on a real database;
first page <= 200 after startup; flush payload depends only on changed
records (including unloaded ones); ActivityPanel output with pages + summary
equals the full-list rendering for randomized datasets with duplicates;
restart interrupts leftover records with the same detail; clear empties the
database, pages and summary; SQL change semantics equal the reference.
Schema 57 rebuilds activity_history_records with order_seq (newer = larger;
a move to the front assigns a new maximum) and stored generated columns
(record_id, occurrence, status, status_group, kind, request_id,
conversation_id, project_id, created_at) with indexes. The order is migrated
from record_order_json exactly; legacy duplicates keep their occurrence and
only the first is shown. record_order_json is no longer written.
ActivityHistoryRepository owns the SQL: page({before, limit <= 500, filter,
projectId, conversationId}) newest first by index, summary() with the counts
and conversation titles/status the Activity page derived from the full list,
reconcile() joined with tasks with a renderer-localized detail line, update()
applied directly in SQL without reading the history, replace() for the
legacy migration, clear(). New activity-history:page/summary/reconcile IPC
with zod validation; page and summary run on the readonly worker.
perf:main measures 24,000 records: startup load 106 ms on Main (14 MB) ->
first page + summary 13.5 ms on the worker (0.6 ms loop block, 114 KB).
The renderer no longer sends the whole history on every change. Domain
actions on an external activity store update the list and queue the matching
contract changes; activity-sync.ts flushes them every 250 ms via
activityHistory.update (<= 20,000 per call), keeps failed changes queued in
order with backoff retry, and flushes on quit and unmount. replace remains
only for the legacy localStorage migration; clearing uses the new
activityHistory.clear IPC. App's activity state is gone; ActivityPanel reads
a selector. update is now required in the desktop API.
Add a supervisor text-model selector using existing connections and a per-conversation story graph switch above knowledge settings. Hide the switch when Supervisor is disabled and enforce conversation preferences at local and remote tool execution boundaries.
Release note: Choose a dedicated supervisor model and control whether each conversation can query existing stories and experiences. Enabling graph access does not start a review or automatically insert context.
Main module evaluation drops from ~420 ms to ~330 ms. Types stay as import type; the DeepSeek Harness chunks stay static because the protocol chunk shares code with Main startup paths.
Runtime creation, MCP launches, Obsidian, native clients and the tool-environment IPC await the initialization promise instead of startup awaiting it before the window path. A failure still fails startup with the same diagnostic.
Pause scroll writes while the jump-to-bottom smooth scroll runs; keep rows with pending questions, approvals, recent interactions and selection ends mounted; cache Mermaid results; copy text spanning unmounted rows; anchor the sidebar on id changes; disable native overflow-anchor for windowed lists; write terminal output on the leading edge and before clear; forward remote events every 16 ms; flush streaming text before copy/export; add performance principle P10 (no UX regressions).
conversations:list-summaries and conversations:get run on the readonly query worker. Activity history replace validates and serializes only changed records; a new activity-history:update IPC writes one record. Adds npm run perf:main.
Remote semantic events were written with one or two synchronous SQLite transactions each before being forwarded; recoverable conversations also rewrote the whole assistant message per event. Events now commit together at each remote checkpoint (before the Agent can acknowledge), and before non-remote, question and terminal events, on a size cap or a 50 ms timer. Only newly stored events are forwarded, in order; a failed batch rolls back and blocks later commits. Applies to interactive, background and recovery runs.
Leaf batches retry at half and a quarter of the text budget, navigation merges retry once; provider errors are not retried. Adds the long-session acceptance script and records the results.
Search matches query terms (CJK pairs) instead of only the whole phrase, previews show readable fields first, and the tool description says when to call it and when to stop.
Configuration apply previously required a separate native confirmation even in Agent Runtime Execute mode, and failed in clients without an approval callback. Use existing work-mode and tool authorization instead, preserving Ask read-only access and request-scoped plan validation. Update regression tests, tool descriptions and documentation.
Validation: 267 focused tests passed; typecheck and changed-file lint passed. Full tests timed out with failures in other modules; repository lint reported errors in unchanged files.
Release note: Execute 模式下通过 Agent Runtime 修改 GoodBuddy 配置不再额外弹出确认框，Ask 模式仍保持只读。
Also includes pending workspace changes: unsaved-change guard, keep-alive cache and live message store, knowledge workspace dialogs and perf driver updates.
Each source version is extracted once for every review scope; events carry spans and projects from their sources, and attention density is counted from messages. After publication, events are assigned to project features and sub-threads (cross-project stories only when enabled), with user rename, merge, remove, move and undo. The story graph gains a lazily loaded Three.js time-spiral view. Includes the 3D concept demo, real-data DeepSeek check scripts and docs.
Isolate Obsidian tests from the default browser capability, check the additional image-tool card, and center native overlay click targets below the frosted header. Retain existing permissions, identity, hit-testing and keyboard assertions; preserve unrelated uncommitted work.
Expose save_image across direct model, MCP gateway, DeepSeek Harness and remote Agents (written on the remote host via chunked reads), and show the image MCP tool list in settings.
Scroll settings controls into the unobscured area below the measured sticky header, retaining native interaction and hit-testing assertions. Give the full legacy outbox migration fixture a scoped 15-second timeout. Product code and bilingual release notes are unchanged; validation remains in CI.
- Segment streaming Markdown so completed blocks are reused and only the growing tail is re-parsed; fall back to whole-document parsing for document-wide constructs
- Memoize conversation sidebar rows and cache per-conversation derived activity, ordering and pending approvals
- Activity page loads on entry, refreshes every 15 s while visible, and has a single header refresh button
- Include pending settings, model download source, device sharing and knowledge gateway changes in the working tree
- Add Electron streaming input benchmark and regression tests
Check detail navigation focus before the auto-focused creation input takes over. Preserve return-to-list and reading-position assertions, and include the reviewed sharing first-read test and independent validation record. Validation runs in CI per user request.
Add an opt-in persistent frosted glass setting for the top bar and runtime checklist, with chat scrolling behind the header while preserving scroll anchors and window controls.
Retry transient local MCP initialization failures once after cleanup, include redacted connection diagnostics in feedback, and index lifecycle events to reduce execution statistics scans.
Release note: Add optional frosted glass appearance with chat visible behind the title bar; improve local tool connection recovery and diagnostic feedback.
Keep the notes list beside the open note, compact the detail view with a full-width sticky toolbar, and make Back to overview return to a full-width all-notes grid. Add manual note tags with AND filtering, a tag manager for rename, merge and delete, schema 51 tag tables, and tag support in the note Agent tools.
Use gbd-/gbc- prefixes with a 6-char conversation hash instead of 20-char per-request hashes, lease per-conversation slots so parallel requests never share a name, and expose gateway custom tools as <server hash>_<name> within the 64-char provider limit.
Every 'Open in terminal' launch for OpenCode copied the bundled plugin dependency tree (about 3,600 files, 52 MB) into a temporary directory, adding 2.2-2.4 s on Windows before the TUI started. The tree is now copied once into terminal/opencode-config under user data and reused until its bundled version marker changes. Skills, Ask-mode plugins and model configuration are still generated per launch and cleaned up on close. OpenCode's own TUI startup (about 3 s) is unchanged; the remote Agent path is unaffected.
Release note: 在终端中打开 OpenCode 时不再每次复制插件依赖，第二次及以后的打开速度明显加快。
Heartbeat no longer writes its own report; it triggers the shared incremental review and, per plan intervention, derives evidence-backed suggestions from the review's graph changes. Manual review is incremental by default with explicit reanalysis. The Supervisor tab is renamed Smart heartbeat with heartbeat plans, and the redundant inner plan frame is removed.
- Build runtime history per turn: a cancelled or failed reply keeps its partial
text plus an interruption marker (or a marker-only placeholder), so its user
message is never left unanswered and merged into the next request.
- Use the same history for send, manual compaction and Main compact validation.
- OpenCode: keep a cancelled prompt in the reused native session and prefix the
next turn with a one-time synthetic interruption notice; failed prompts are
still removed.
- Add an opt-in DeepSeek live test covering direct model and OpenCode.
- Store runtime selection as { provider?, model? } layers and resolve them per
field (conversation -> project -> global); drop the legacy auto provider.
- Migrate stored selections (database v49) and clear deleted profile ids.
- Re-resolve selections in Main per request and at queue dispatch.
- Add ProjectRuntimeSelector and a two-column RuntimeModelPicker; manage
link opens model connections; restore per-runtime default model tabs.
- Stop duplicating trusted instructions and history across OpenCode,
Continue and DeepSeek Harness runtimes.
- Document the layered design, global provider removal plan and planned
purpose-specific models.
Add a System tasks tab to run history that splits background usage (supervisor, auto-supervision, magic notes, knowledge base, delegation) from conversation usage, with a manual refresh button and automatic refresh when usage is recorded. Record previously dropped usage from supervisor reviews, image regeneration, knowledge graph extraction, and remote embeddings.
Local direct-model text conversations could not use manual context compaction,
and disabling automatic compression prevented saved summaries from being reused.
Restore the composer action regardless of the automatic setting or threshold,
use the selected summary connection, and persist summaries for subsequent
requests while retaining recent turns and full conversation history.
Image connections, remote channel conversations, and SSH projects remain outside
the local direct-model compaction workflow. Add focused regressions, a local
Electron end-to-end test, and an opt-in two-request live probe.
Validation: staged snapshot passed typecheck, changed-file ESLint, 13 focused
tests, and the Electron end-to-end test with two local HTTP fixture requests.
No additional real-model calls were made during commit verification.
Release note: 修复本地直连文本模型无法手动压缩上下文的问题。现在关闭自动压缩或尚未达到触发阈值时，也可点击“压缩上下文”；摘要会随会话保存并用于后续对话，同时保留最近轮次和完整历史。
Remove the fixed gateway capability lifetime so active requests retain knowledge, notes, browser, configuration, and custom MCP access. Preserve token validation and request cancellation or cleanup revocation, and replace the misleading knowledge expiry error with a tool authorization error.
Remove native-client clock overrides and cover long-running scoped and HTTP MCP calls. Validation: 34 focused tests passed; changed-file lint passed. Repository typecheck and lint remain blocked by unrelated manual-compaction and sidebar probe files.
Release note: Long-running Agent requests no longer lose tool access after ten minutes, including when the knowledge-base switch is off. Request cancellation and completion still revoke access.
Avoid hidden activity renders, no-op message saves and attachment scans when pinning conversations. Reuse note body inspection and skip unchanged attachment reference writes while preserving recovery and cleanup behavior.
Skip unnecessary Markdown normalization and memoize subagent lookups without delaying live rich text. Add storage operation counts and real Electron rendering equivalence benchmarks; full-render speedup remains unproven.
Validation: 5397 tests passed, 84 skipped; typecheck passed. Lint passed excluding two existing untracked sidebar probe files.
Release note: Reduce repeated rendering and disk work during chat, conversation pinning and note startup while preserving existing content and interactions.
Keep cancellation and failure details in message status while preserving existing replies. Render cancellation neutrally, restore terminal metadata from history, and offer Edit again only when input can be restored.
Release note: Cancelled and failed requests show a single notice; cancellation no longer appears as an error, and Edit again accurately describes the recovery action.
mesalogo