New features

• Review daily problems in the Friction Log, including corrections, tool errors, and recurring patterns linked to exact messages. Browse dates, download Markdown, or read findings through the CLI and MCP. Enable it with [friction] enabled = true.
• Turn Friction Log patterns into Kata issues, link existing issues, or remove links from the page or CLI. Enable hourly automatic filing with [friction.kata] auto_file = true.
• Generate an optional model-written Friction Log summary with your configured insights provider. Summaries run only when requested and stay attached to the reviewed digest.
• Find the recorded session and prompt behind code with agentsview why file.go:120-140 or agentsview why --commit <sha>. List a session's recorded commits with agentsview session commits <id>.
• Turn on search by meaning from Settings or the search palette without editing configuration or restarting. Local installations use an available Ollama server or download the built-in embedding server with progress.
• Find sessions covering two to five concepts through MCP or the content search API. Results include a supporting passage for each concept.
• Browse saved Claude and Codex history from Docbank archives or backup snapshots alongside local sessions. Imported history remains searchable when the archive is disconnected.
• Find matching Docbank file versions beneath retained transcript images and attachments. The optional integration also exposes matches through the API and MCP.
• Import Open Interpreter sessions with their own label, usage totals, and resume commands. Command Code v3 sessions also import with thinking, tool calls, and token usage.
• Browse Kiro CLI subagent runs beneath their launching sessions, including nested subagents. The launching transcript shows each delegation's prompt and result.
• Choose message colors and a Color-blind safe chart palette in Settings > Appearance. Optional message markers, chart patterns, and dashed trend lines distinguish content beyond color alone.
• Diagnose an expensive session with agentsview session diagnose <id>. The report combines usage, context growth, compactions, large tool results, retry runs, and child-session links.
• Inspect per-session tokens and costs through MCP get_session_usage, including model totals and request breakdowns. Batch CLI lookups use agentsview session usage-batch <id>....
• Import xAI billing CSV files with agentsview usage grok <csv>. Reimports deduplicate billing records, but overlapping local sessions or Cursor billing imports may double count usage.
• Filter hosted sessions, searches, analytics, and usage by contributing source devices. Sessions supplied by several selected devices count once.
• Run PostgreSQL serving and watched local ingestion together with agentsview replica pg serve --push-watch. Add --embed to build and publish embeddings in the same process.
• Export visible dialogue with --text-mode visible, omitting structured tool and reasoning content. Reporting exports support hourly snapshots by originating machine with --schema-version 5 --by-machine.
• Discover commands from a task description with agentsview cli search. Inspect supported agents, session folders, and captured details with agentsview agents, without a daemon or archive.

Improvements

• Usage loads faster on large archives. In separate 30,000-session benchmarks, repeat visits drop from about 24 seconds to under a second on PostgreSQL and DuckDB.
• Activity, Signals, Trends, and session statistics load faster through reduced reads and reuse of unchanged results. ClickHouse session navigation also stays responsive while dashboard queries run.
• Session lists and long transcripts open and scroll with fewer stalls. Sidebar rows appear before their titles finish loading, and code highlighting blocks the page less.
• Sidebar and content searches return sooner, including substring, regex, and multi-term searches. The first local startup after upgrading builds a new search index before serving requests.
• Imports, archive rebuilds, and live Claude and Codex updates finish sooner. PostgreSQL pushes preserve unchanged transcript sections, DuckDB rebuilds write faster, and S3 sync reuses shared Codex parent downloads.
• Usage charts refresh automatically while visible and catch up when you return. Hidden tabs cancel pending Usage queries instead of holding browser connections.
• Usage charts omit zero-value tooltip rows and group positive series below 0.5% into Other. Token Attribution adds Linear, Sqrt, and Log choices for treemap label sizes.
• Activity timelines merge narrow bars into readable time intervals on smaller screens while preserving totals and selection accuracy.
• Read Claude hook-command output in collapsible cards controlled by the Hooks filter. Hook text is searchable and appears in HTML and Markdown exports.
• See recorded reasoning effort on Pi-family and Posit Assistant messages.
• Use grouped CLI commands under source, report, replica, and index. Previous spellings remain available for at least one release and print their replacements. Failures include specific recovery hints and structured JSON diagnostics where supported.
• Reduce background work while idle or on battery. Local sync pauses after ten idle minutes and resumes on archive access; scheduled embedding builds, Recall model passes, and archive audits wait for AC power. Set sync_interval to adjust watcher-triggered sync frequency.
• Keep using AgentsView while Claude.ai Sync runs by closing its dialog; only Stop cancels the run within the page. Closing or reloading the page still ends Sync. An organization allowlist also limits future imports after a restart.
• Recover useful search results when search by meaning is unavailable. Hybrid CLI and MCP searches return keyword matches with an explanation, and empty searches report matches hidden by session filters.
• Check stalled PostgreSQL and ClickHouse push watchers with status --max-stale. Status reports failed, overdue, or unfinished pushes with a nonzero exit code.

Bug fixes

• Usage totals include blank Claude thinking responses, unfinished Copilot CLI sessions, and Hermes or Augure Desktop background model calls.
• Codex Fast-tier sessions use Fast rate estimates. ClickHouse and DuckDB use the latest Copilot-reported session cost.
• VS Code Copilot and Positron replies keep updating as they grow, preserve pins and notes through rebuilds and pushes, and render without duplicate-tool-call failures.
• DeepSeek Harness sessions import when logs contain job lifecycle events or compressed frames without checksums.
• Claude channel and peer messages remain visible and searchable. Slash commands containing < in their arguments no longer disappear or leave answers beneath the wrong prompt.
• Codex forks link to their source sessions. Pi subagent runs and OMP parent links also receive the correct relationships.
• Archive rebuilds complete despite unavailable Codex fork parents while preserving saved child history. The Sync button performs required rebuilds and reports failures.
• Remote and temporary worktree sessions reuse established repository project names when recorded repository evidence supports the match. Explicit assignments and mapping rules retain priority.
• All-history Usage opens during startup price refreshes without internal errors. Concurrent Usage requests in a new time zone avoid database-lock failures.
• Analytics panels remain usable when another panel fails. Failed panels show their own error and Retry control, and old filter results no longer appear current.
• Activity's Sessions table keeps its selected sort order and column widths after syncs.
• Desktop windows reconnect after daemon exits or endpoint changes, accept localhost and IPv6 loopback endpoints, and avoid reloads caused by archive contention. The Linux AppImage also starts on affected Wayland systems.
• Explicitly started daemons stay running until stopped. Daemon restart also completes in Linux containers whose exited children remain unreaped.
• Claude and Codex Cloud transcripts show generated images inline with full-size viewing. Codex code-mode wait calls show their returned output, and Codex Cloud tool calls no longer appear twice.
• Hosted search and session browsing remain consistent during incoming uploads and publication. Empty source files upload successfully, and permanent upload rejections stop retrying indefinitely.
• Background skill discovery avoids protected macOS folders by default. Live watches recover after macOS session folders are recreated, and large Linux archives retain watches beyond 8,192 folders.
• Transcript-only archives stop reporting retries based solely on repeated tool names. Session links using an unambiguous raw UUID also open correctly.

Improvements

• Antigravity CLI sessions show plaintext transcripts with model and token usage attached to the matching model replies.

Bug fixes

• Antigravity sessions parse again on read-only mounts, including read-only Docker volumes.
• Archived Antigravity CLI replies stay intact when the session database cannot be read.
• Repeated prompts remain separate turns in Antigravity CLI sessions. Existing archives are reprocessed to recover missing turns.
• Docker builds no longer fail because a required Claude.ai Chrome extension file is missing.

New features

• Sync Claude.ai chats through Chrome, including Google sign-in. Run agentsview chrome setup, load the printed extension folder at chrome://extensions, and restart AgentsView before using Import > Claude.ai.
• Import ChatGPT dot, Codex web, and Codex cloud task chats as searchable sessions with the opt-in codex-cloud source. Chats without token counts show OpenAI's plan usage figures. Configure this source on one machine only.
• Record Codex cloud token usage with optional live_tokens = true. Capture covers only turns heard live, and connecting can send queued messages. Sessions with incomplete capture show a "partly captured" note.
• Preserve a retired computer's Claude and Codex history with archive capture and archive import --seed, including stars, pins, names, and original files. Archive commands also extract originals and verify backups. Use a separate data directory for new local sessions.
• Sync remote Linux and macOS machines over SSH without manually installing AgentsView or running a remote daemon. Set transport = "ssh" in the remote host configuration. SSH authentication must work without a prompt.
• Browse and search Muse Code sessions, including tool output, child sessions, token usage, and costs.
• Browse and search fx schema 4 sessions with tool output and child sessions. Costs appear when an older fx version saved usage data.
• Find sessions by label, linked pull request, or launcher. Claude Code pull request links appear on session pages and in CLI, API, and MCP results. Linked workers contribute to their launcher's total cost.
• Group model variants in Usage with [model_aliases] in config.toml. Each variant retains its own pricing before totals combine. Restart AgentsView after editing the groups.
• Narrow Usage totals and charts by selecting an item in the cost breakdown panel. Open hermes-cron to compare individual jobs, with renamed jobs and continued runs grouped together.
• Analyze a session's tool effectiveness from its actions menu. Reports mark conclusions as helped, did not help, or unclear and link each conclusion to the calls it cites.
• Compare empty-result, repeat-call, and recovery rates in Tool Usage under project and date filters. Click a nonzero rate to inspect contributing sessions and calls.
• Ask connected assistants about file edits, tool calls, child sessions, Activity reports, tool usage, and cost comparisons through eight new read-only MCP tools. Built-in guide search works offline. list_sessions also accepts include_one_shot: true for single-prompt sessions.
• Remove an agent's archived history with agentsview prune --agent <name>. The filter combines with project and date filters and includes matching sessions in Trash. Prune also deletes matched source files, except shared app-owned stores.
• Sort Activity reports by Session or Model across pages. Cleanup, worktree mapping, and session health tables also support sorting by their data columns.
• Build and push embeddings in one process with pg push --embed. Containers accept deployment and embedding settings directly through environment variables, and the container image now requires a bearer token.
• Sign native HTTP reader and raw-sync requests with an independent HMAC key. A separate machine listener enforces signatures, existing authentication, and reader or contributor grants.
• Export repository-scoped activity and usage with reporting schema 5. Known repositories remain selectable when other sessions share their project name. Integrations adopting schema 5 must fetch new project keys and refresh saved digests. Schema 3 remains the default.

Improvements

• Browse and search the existing archive while a full rebuild runs. Results carry a stale-data warning, and changes pause until the replacement is ready. Failed startup rebuilds retry automatically.
• Keep new local sessions arriving in the archive and UI while a PostgreSQL, ClickHouse, or DuckDB push runs.
• Resume interrupted full PostgreSQL and ClickHouse pushes and DuckDB rebuilds. Recent sessions transfer first, and summaries show resumed progress and why a full transfer runs. Passing --full discards unfinished progress and starts over.
• Load Usage and Activity faster on large archives. The daemon prepares daily usage totals after sync, and the Usage page starts its model comparison alongside totals.
• Keep quick page reads responsive while slow SQLite analytics run. Large session trees also load faster instead of remaining at "0 SESSIONS" or "LOADING".
• Keep semantic indexing up with long-running sessions by re-embedding only the newest portion of an assistant run. The first build after upgrading re-embeds oversized runs once.
• Transfer less data during HTTP remote sync by leaving unchanged transcripts out of downloads. Updated peers use Zstandard compression, and combined rebuilds prepare several remote hosts at once.
• Push ClickHouse embeddings in batches to reduce transfer time. Large archives also support a configurable push_batch_size for session transfers, with higher values using more memory.
• Sync only changed Hermes and Augure Desktop sessions instead of rewriting every session in their shared database. Unchanged remote imports also avoid repeated parsing after a one-time refresh.
• Show Antigravity IDE and local Qoder sessions by their app-saved titles, including existing sessions. Names set in AgentsView take precedence. Qoder renames appear at the next sync.
• Show the recorded starting branch for Codex, TraeX, and Augure Code sessions. Recorded repository identity also survives a missing checkout or changed remote.
• Use provider-reported billed costs for gptme messages when available, including zero. Estimate-only messages retain catalog pricing, and provider-routed model names resolve catalog prices.
• Return stored session usage without waiting for full startup ingestion. Targeted refresh failures retain those totals with a warning.
• Reuse MCP daemon connections to reduce repeated discovery delays.
• Refresh open session lists after imports, uploads, and restores.

Bug fixes

• Preserve Codex undo history and daily usage as linked continuations instead of letting rollout files replace one another. Older Desktop undo files also link back to their thread. The upgrade rebuilds existing archives and preserves supported saved annotations.
• Keep Codex rollouts under their own thread when a file contains another thread's metadata, preventing one transcript from replacing another.
• Read older Codex rollouts after Codex compresses them to .jsonl.zst. Decoding failures leave archived messages intact.
• Import current Mistral Vibe sessions from its unified store, including running turns, tool output, child sessions, and token usage.
• Import DeepSeek Harness v4 and newer TraeX transcripts with their messages, tools, and usage. OpenClaw databases with compressed events also sync without blocking healthy sessions.
• Preserve both Cursor conversations when S3 imports contain the same session ID in different projects.
• Keep archived Claude and iCodeMate CLI tool output after Claude Code deletes its external tool-result files.
• Preserve Hermes and Augure Desktop's shared app database when pruning sessions. Dry-run and deletion summaries report source files left in place.
• Keep message jumps in view while nearby rows finish rendering. Background tabs also release browser connections so new session tabs no longer hang at "Loading messages...".
• Limit SQLite message matches and snippets to conversation text after reparse. Use --in messages for dialogue and --in thinking for reasoning. Ordinary session search still includes the full transcript.
• Fill MCP content-search pages with eligible older matches instead of excluding active sessions after applying the limit. Restart searches begun before upgrading.
• Accept displayed machine names in filters across the web UI, CLI, API, and MCP. Unknown or ambiguous names return an explanatory error.
• Allow chart palette and zoom changes on read-only servers without editing configuration files by hand.
• Keep archive commands working when unrelated configuration sections are invalid. SSH remote sync also ignores server-only configuration errors.
• Stop repeated Antigravity and Gemini imports caused only by SQLite bookkeeping changes. Recovered file watchers also stop rescanning the entire Claude folder.
• Avoid rewriting unchanged PostgreSQL messages and usage events during metadata-only pushes. Shared PostgreSQL mirrors also stop reporting other machines' unrelated vector sessions as conflicts.
• Prepare ClickHouse usage totals on filesystems that cannot exchange tables, including NFS. Unsupported-exchange errors appear once per server start instead of on every push.
• Produce stable reporting digests when several checkouts or worktrees share a repository.
• Wait through concurrent daemon restarts and preserve host daemon discovery across process namespaces. Auto-starting commands report a held archive lock instead of launching a server that cannot open it.
• Keep Windows daemon commands from marking the entire data directory as changed for backup and sync tools. Stop the daemon before downgrading to 0.45.x or earlier.
• Honor DO_NOT_TRACK=1 and suppress telemetry in CI and non-release builds. Browser telemetry also waits for the selected server's response before sending further reports.
• Reject compromised eMudhra/emSign certificate chains in HTTPS connections and desktop updates. Linux and Windows desktop builds also update their bundled trust roots.

New features

• Choose Español in Settings. Browsers set to Spanish use it on first load.
• Browse Junie CLI, OMO, and StepCode sessions with their usage and costs. Junie ingestion excludes tool calls and IDE-only conversations. If you previously used PI_DIR for OMO, remove that override to avoid indexing new sessions twice.
• Read Antigravity conversations from plaintext transcripts when no session database exists, without a decryption key. These transcripts include reasoning and tool calls but no token usage or cost.
• See commits, lines added and removed, and files changed on the dashboard for the selected dates. Load pull-request totals on demand after GitHub setup.
• Switch the Activity timeline between concurrency, user-message frequency, and assistant-message frequency.
• Compare user, assistant, and other message counts in the Analytics timeline with Absolute and Percent views. Other includes system entries and tool results.
• See all prompt tokens in the Usage page's Total Input card, including cache reads and writes. The previous Input Tokens count is labeled Uncached Input.
• Choose Smooth, Lines, or Bars for Cost Over Time. The chart remembers your choice, shows zero-spend days, and supports dragging to select a date range.
• Follow tool errors, retries, switches, and recovery in a session's tool sequences panel. Each step links to its transcript call and shows available timing and evidence. Opening the panel adds no token charges.
• Give Claude Code and Codex access to past conversations through native conversation-memory plugins. Session-start refresh and readiness diagnostics help identify missing setup or search coverage.
• Install proactive conversation recall with agentsview skills install. The package guides agents to search relevant history, distinguish proposals from accepted decisions, and cite transcript evidence.
• Approve or archive extracted entries in Recall, the collection of memories extracted from sessions. Review decisions survive later extraction changes.
• Activate the first Recall collection before the initial backlog finishes, once a session is complete and usable entries exist. Recent sessions extract first, and later passes add to the active collection.
• Extract Recall entries from several sessions at once by setting concurrency on an extraction server. The default remains 1.
• Choose the default CLI for generated insights with [insights] default_agent. Leaving it unset keeps Claude as the default.
• Replace selected ChatGPT or Claude.ai imports with agentsview import --replace <session-id>. The old transcript moves to Trash as a restorable copy, while the live session keeps its name and pins on matching messages.
• Identify refused ChatGPT and Claude.ai imports by session ID and reason. CLI summaries separate changed history, shorter exports, trashed sessions, and temporary failures.
• Use semantic and hybrid search on ClickHouse mirrors. Embeddings upload with clickhouse push when vector search is enabled; --no-vectors or push_vectors = false opts out.
• Browse hosted raw-sync uploads after the server parses them into PostgreSQL sessions. This opt-in mode removes the need for a local SQLite archive or a separate pg push.
• Upload existing sources in a finite, resumable run with raw-sync backfill --run-id ID --provider NAME. Start raw-sync watch after the run completes to upload later changes.
• Inspect hosted uploads and parsing with raw-sync server-status and the raw-sync health API. Run raw-sync clean-uploads to expire upload sessions and remove orphaned temporary files without starting the server.
• Serve AgentsView under a reverse-proxy subpath with serve --base-path. Keep --public-url origin-only.
• Search several literal terms within one user/assistant exchange over MCP on SQLite and PostgreSQL. Narrow results by session, branch, project, agent, or date.
• Read MCP search evidence from the same transcript revision, and continue reading messages beyond the previous 20,000-character limit. Revision checks report when the source changed.
• Build clients for multiple machines with host keys and session titles in search results, batch session lookup, and optional source size and archive-update metadata.

Improvements

• Load Activity, Usage, and Analytics reports faster on ClickHouse with lower memory use. Opened past-day reports survive restarts, and new pushes refresh only changed sessions where possible. Upgraded deployments need a push to complete new report data and access to system.parts for serving.
• Load Tool Usage and Top Skills analytics faster on PostgreSQL, with less memory use on large archives.
• Load dashboard metadata with less database work on large SQLite archives. The first writable startup builds an index before serving, so that startup takes longer.
• Open Usage after a price update without waiting for a full cache rebuild. Only sessions using changed models rebuild, and rebuilding starts in the background. The first startup after upgrading rebuilds the usage cache once.
• Sync and rebuild large archives with less retained transcript data and less repeated processing. Fresh archives defer the conversation export index until the first export, and full resync removes the redundant stored copy of message text.
• Update Cursor IDE chats with less parsing and memory use. A database change reparses only chats whose saved documents changed.
• Keep file watching responsive on large Antigravity installs by skipping generated directory trees. Windows watchers also use smaller change buffers, reducing memory use.
• Build hourly export day and export digest reports faster, especially for long sessions.
• See the current sync stage, subagent-repair counts, and notices when a sync waits for another operation. Refresh controls show elapsed query time and a live request timeline.
• Load more Activity sessions as you scroll. New pages append to the table, and failed loads offer a retry.
• See agent-provided session titles consistently, with user-chosen names taking precedence. First-message previews retain the user's actual prompt.
• Read fuller Codebuff and Freebuff transcripts, including attachment markers, errors, and answers to agent questions. Nested agents appear as linked subagent sessions, and costs include every prompt's reported credits.
• See OpenCode tool durations from dispatch to result, including permission waits. Calls without a dispatch timestamp retain an unknown duration.
• Extract fewer Recall procedure entries from work an agent only proposed. Procedure extraction requires tool evidence in the relevant action window. Upgrading re-extracts existing sessions.
• Keep Recall vector and hybrid queries available while extraction adds entries. The index may trail by 256 changes by default; hybrid search still finds newer entries by keyword.
• Configure embedding endpoints with HTTPS for public IP addresses and without query strings. Builds leave documents pending when configuration, credentials, or server failures prevent embedding.
• Sync remote hosts over HTTP. SSH sync is removed; configure each remote with a daemon URL and token. Existing archived sessions remain available.
• Measure app opens, screen visits, core actions, and visible visit duration through the existing telemetry opt-out. Daily daemon activity counts at most once per installation per UTC day.
• Diagnose desktop startup and database delays with logs that show backend port selection, readiness results, stalled database checkpoints, and connection or write-lock waits.

Bug fixes

• Build from a fresh Go module cache without a missing-dependency error.
• Re-import ChatGPT exports to append new messages, fill previously empty tool results, and restore truncated message text while preserving existing message IDs and pins.
• Keep saved Claude.ai turns when a stale export contains fewer messages. Shorter exports are refused unless you explicitly request replacement.
• View, search, and copy raw Codex and Grok tool inputs. Codex patch edits also appear in Recent Edits.
• Keep both transcripts when Gemini CLI or Cursor files share a session ID, instead of letting one overwrite the other.
• Read Claude subagent continuations saved under another parent session. Pi subagents in nested directories also appear and link to their parent when the recorded layout identifies it.
• Keep Claude messages from other sessions out of human prompt counts and session-title selection. The transcript labels them "Message from another session".
• Count workers launched by another agent as Subagents on Activity, rather than inflating the Interactive peak.
• Price Codex cache writes at the cache-write rate and estimate auto-review costs using GPT-5.6 Luna rates. Offline reports and exports also include pricing for GPT-6 Astra, GPT-6 Sol, and GPT-6.1 Sol.
• Retain reported costs when model filters contain spaces between model names or are blank on SQLite and PostgreSQL.
• Count every message in term trends for time zones where daylight saving time skips local midnight.
• Remove permanently deleted sessions from semantic search on the next embedding build. Sessions removed by sync and later imported again receive embeddings again.
• Keep Cursor and Antigravity idle instead of repeatedly syncing unchanged SQLite databases.
• Show Cursor IDE sessions on the date of their earliest timestamped message.
• Update active Codex Desktop sessions on macOS within about 30 seconds, even without Codex history hints.
• Start background serving on macOS without hanging on permission prompts for external volumes. Enable scan_protected_paths = true to include Git details from those locations.
• Keep macOS file identities stable across reboots, avoiding repeated archive reparses. Upgrading performs one initial reparse and raw-source capture.
• Apply session-provider changes made in Settings without restarting the daemon. Separate push-watch processes still require a restart.
• Keep ...

New features

• Browse and correct project assignments in an opt-in project mapping workspace. Review folder suggestions and session previews, filter by date, and apply individual or bulk corrections.
• Filter searches by project and date directly in the command palette. Switch to All Projects without changing the sidebar, or choose relative, calendar, or custom dates.
• Open a session by ID or UUID with Ctrl/Cmd+G.
• Push your archive to ClickHouse with agentsview clickhouse push, then serve the dashboard with agentsview clickhouse serve. Session edits remain in SQLite. ClickHouse does not support vector search or hosted raw sync.
• Browse Cline CLI sessions with transcripts, token usage, costs, and remote sync. Continuation snapshots appear as one session. Cline VS Code extension sessions are not included.
• Browse Charm Crush and Tencent CodeBuddy CN sessions with their recorded conversations and usage. Crush reports session totals without per-message token or cache breakdowns.
• Track Augure Code separately from Codex and resume sessions through the Augure CLI. Proprietary Augure models remain unpriced.
• Browse, search, and export Augure Desktop 3 beta sessions with usage and cost reporting. Remote sync is not supported for this agent.
• Read DeepSeek Harness session formats through version 3, including compressed logs. AgentsView selects the newest supported generation and excludes inherited parent transcripts from child sessions.
• Resume Pi sessions from AgentsView or copy a complete pi --session command.
• Export new, edited, and deleted user and assistant messages with export conversations. Save progress to fetch subsequent changes, including messages from imported archives or sessions whose source files are gone. Only the latest message version is retained.
• Find the earliest recorded activity and latest complete reporting hour with export range.
• Generate and inspect Activity Insights from scripts with agentsview insight generate, insight list, and insight get.
• Query archived usage without refreshing source files with session usage --no-sync.
• Look up sessions by UUID through MCP search_sessions, including active sessions. MCP search_content also accepts include_one_shot and include_automated to include single-prompt runs and automated sessions.
• Include transcript paths in session-list JSON with session list --include-source.
• Inspect hosted raw sync through GET /api/v1/raw-sync/status, which reports source versions, parse-job counts, active devices, and upload backlog.
• Choose Azerbaijani in language settings.
• Hide AgentsView from the macOS Dock and Cmd-Tab when its window is closed using a new tray option. The setting is off by default and persists across relaunches. The app also adds a standard Window menu.

Improvements

• Session Vitals, the session timing view, separates measured tool execution from time that cannot be assigned to a phase. Activity rows link to the corresponding transcript entries.
• Export large session archives and multi-day digests with fewer repeated database reads.
• Startup sync and usage-cache backfills use less background CPU.
• Database upgrades show opening, schema, index, and migration progress before session syncing begins.
• Session list, detail, sync, and search JSON include browser links when connected to a server. MCP session results preserve those links.
• Session links give the provider and UUID separate URL segments, so replacing the UUID is easier. Existing encoded links still open.
• JSON output from session list --json and usage daily --json --breakdown includes machine labels. The Activity machine filter keeps friendly labels visible and adds a short ID only when labels collide.
• Failed pricing refreshes retry after about 1.5 hours, with increasing delays after further failures, instead of waiting a full day.
• Profile export sessions, export hour, export day, and export digest with --cpuprofile, --memprofile, and --trace.

Bug fixes

• Devin usage totals no longer discard messages from different sessions that share the same message number.
• StepFun step-5-preview sessions receive cost estimates, including in existing archives without a resync.
• Codex prompts no longer mix user text with injected context. Existing archives refresh affected text, previews, and message counts during startup resync when source files remain available.
• Session previews skip complete leading injected-context blocks. WorkBuddy sessions also use their generated titles when present.
• Every codex exec session is classified as automated. Roborev-tagged sessions are identified as automated code reviews.
• Grok child sessions appear beneath their spawning parent and count as subagent activity. Native Pi branches regain parent links when the parent session can be resolved.
• Older Piebald and Kilo databases import despite missing columns that previously caused discovery or parsing failures.
• Cursor IDE imports accept structured tool results, including object-valued todo_write results.
• VS Code Copilot assistant messages retain the model used for each request, including model changes within a session.
• Healthy archived sessions continue copying to PostgreSQL and DuckDB when daemon-managed incremental pushes encounter a local ingestion failure.
• Sync avoids repeatedly parsing unchanged malformed or missing source files where failure caching applies. Unrelated file events no longer bypass retry delays, and temporary lock-file changes no longer trigger session syncing.
• Configuration loading rejects unknown keys under [vector], including misspelled or misplaced settings. Correct or remove these keys after upgrading, even if vector search is disabled.
• An explicit nonzero --port fails when occupied instead of silently choosing another port. Updates preserve the explicit port choice.
• PostgreSQL and DuckDB servers mounted under --base-path are discoverable, and API requests retain the path prefix.
• Version-mismatch errors identify whether to restart the daemon or upgrade and restart a stale command or background service.
• Clicking the macOS Dock icon or activating AgentsView through Cmd-Tab restores a hidden main window.
• Copying a session's directory path works after the directory has been deleted.
• Quality-page excerpts preserve complete Unicode characters when shortened.
• Activity concurrency tooltips clear when the date changes or the report refreshes.
• The Analysis panel header stays on one line.

New features

• Browse and search Tau, Evener v2, and Open Code Review sessions, including recorded reasoning, tool results, and token usage. Open Code Review also includes structured review comments and resume links.
• Import OpenCode 2.0 beta conversations while preserving history across upgrades. Tool results retain embedded images and files. PDFs remain available in raw results without a preview.
• View Cursor CLI subagent transcripts under the session that launched them, including transcripts imported from S3. Run agentsview sync to import existing children if you normally run only usage commands.
• Open desktop sessions through agentsview://sessions/<session-id> links. Add ?msg=<ordinal|last> to jump to a message.
• Copy a remote session's resume command with "Copy command", then paste it into a shell on the machine that owns the transcript.
• Choose how much content to archive with archive_content in config.toml. full remains the default, transcripts omits tool inputs and outputs, and usage keeps accounting without conversation text. Apply stricter retention to existing data by restarting the daemon and running agentsview sync --full. Recovering removed content requires the source files.
• Choose Keep, Drop, or Offload for tool-result images in Settings. Offload stores supported images as separate assets while preserving image access. Restart the daemon after changing the setting.
• Preview and remove archived inline images in Settings, with project and date filters. Use agentsview db migrate --images to move images into separate assets instead of removing them.
• Reclaim unused archive space with agentsview db compact. Preview savings with --dry-run, choose a staging directory, or retain the original with --keep-backup.
• Register additional Claude, Codex, and Pi home directories in Settings or configuration. Directory settings now use [agents.<id>] tables, and startup converts older keys automatically. Upgrade both HTTP sync peers together.
• Discover Pi sessions through PI_CODING_AGENT_DIR and PI_CODING_AGENT_SESSION_DIR, locally and over SSH. PI_DIR retains priority.
• Search Chinese, Japanese, and Korean messages with optional SQLite CJK full-text search, installed through make install-cjk-fts. Chinese queries use word segmentation, while Japanese kana and Korean queries preserve character order and adjacency.
• Export usage grouped jointly by project, model, and agent with opt-in reporting schema 4. Select projects with --project-key and choose time intervals with --bucket. Exports include repository identity information. Schema 3 remains the default; integrations using schemas 1 or 2 must update and refresh saved digests.
• Filter MCP search_sessions results by inclusive activity dates using date_from and date_to.
• Discover running HTTP MCP listeners with agentsview mcp status --json, including their bound ports, backend targets, and token-file paths.

Improvements

• In-session search reveals the selected text inside collapsed blocks and nested scroll panes. Counts and navigation follow the active filters, and a result list and overview show each match's location. The underlined ab toggle enables whole-word matching.
• Activity separates interactive conversations, subagents, and automated sessions across counts, minutes, costs, and badges. Its concurrency chart stacks their simultaneous activity in one bar per time interval.
• View recorded reasoning effort beside the model in Claude Code and Codex session headers. Upgrading resyncs existing sessions to populate it, which takes time on large archives.
• Adjust one Zoom preference through Appearance, desktop shortcuts, or the desktop status bar. Set zoom_level in config.toml to supply a default while preserving each client's local choice.
• Hide continuation, resume, interruption, and task-notification cards independently with the "System boundaries" block filter.
• Preserve tags, whitespace, and literal lines inside unknown XML blocks by enabling "Render unknown XML blocks as preformatted text" in Appearance.
• Read assistant reasoning and recorded timestamps in local Cursor CLI sessions when matching chat databases are available.
• Track exported sessions across database rebuilds with archive_id. Session exports also include transcript_revision and local_modified_at to distinguish corrections from session activity.
• Exclude specific sessions from content search with repeatable --exclude-session options. Installed history-search skills also retain configured remote-server commands across reinstalls.
• See preparation phases and elapsed time while agentsview usage daily builds its cache. Daily reports use saved archive data; run agentsview sync first to include new source changes.
• Load Activity reports and tool and skill analytics with less work on large archives. Archives also avoid storing duplicate copies of single tool results, reducing storage and cold-read costs.
• Sync growing Codex and Claude sessions with less repeated parsing. Large imports retain less data in memory, and shared OpenCode databases allow more sessions to parse concurrently.
• Refresh the semantic search index after sync without rereading unrelated message content. Linux daemons using glibc also retain less memory during repeated queries.
• Recover disk space from inactive usage caches after upgrades without removing caches still in use.

Bug fixes

• Session transcripts load again under duckdb serve, fixing blank transcripts in 0.42.0.
• Usage reports complete while live sessions continue writing. Summary, comparison, and top-session requests also allow longer cache preparation without the normal server timeout.
• Windows daily usage and activity reports use the correct local timezone instead of failing or applying the wrong calendar dates.
• Copilot usage includes observed per-call tokens before shutdown and known output tokens when shutdown summaries are absent. Overlapping store and transcript records no longer count the same output twice.
• Cost estimates recognize Codex gpt-reserve as GPT-5.6 Luna, apply Bedrock rates to Bedrock model names, and resolve Ollama Cloud tags to base-model prices. Existing custom overrides retain priority.
• Antigravity CLI usage keeps the recorded effort-qualified model for the observed experimental serving variant, preventing one model's accounting from splitting into separate entries.
• The Usage model picker and charts agree on visibility. Clicking a model hides it, and checking it restores it. Older inclusion-only model links no longer restrict the view.
• Focused mode preserves Codex and TraeX answers followed by tool calls. Claude task notifications and stop-hook feedback no longer split an exchange or turn progress updates into final answers.
• Filtered code blocks remain as expandable placeholders. Code-only messages stay visible, and each block retains rendering and raw copy when expanded.
• Retained tool-result images render inside formatted output instead of appearing as enormous text strings that overflow the pane.
• The command palette stays open when text-selection drags end outside it. Two-character Chinese, Japanese, and Korean queries search the archive instead of only recent sessions.
• Generate starts reports and session analyses over HTTP outside localhost. Recall's Generated insights also accepts session-agent filters without rejecting the request.
• Selective hybrid Recall searches return results instead of failing when project or working-directory filters reach the vector search limit.
• Ollama embedding recovery reloads a failing Metal runner before falling back to CPU.
• Cursor imports preserve legacy tool-result text and use recorded user-turn timestamps when available. Empty Cursor IDE database rows no longer abort the entire sync.
• Claude sessions use the latest generated title when no /rename is present. Explicit renames, including an empty clear, retain priority.
• Codex guardian review sessions appear under their parent thread. Forks of readable parents with no prompts no longer retry indefinitely.
• Hermes session activity times advance as messages arrive, and Pi skill loads appear in Top Skills analytics.
• DeepSeek Harness v0 sessions accept valid compressed source-event ranges.
• Remote imports for OpenHands, OpenCode, Kilo, MiMoCode, Command Code, and Kiro derive project names from recorded paths instead of unrelated local repositories.
• Local history stays together across updates, restarts, and hostname changes. Saved hostname filters and URLs continue to resolve. Remove old machine values from local session_sources entries.
• Startup recovers automatically from stale state left by a crashed startup process. Missing sessions show a retryable not-found message instead of a blank pane.
• SQLite-backed sessions stop resyncing in a loop because of AgentsView's own database reads.
• Local incremental sync waits with visible status when the daemon is busy. Redirected sync output remains readable without terminal control sequences.

• Add hosted raw session sync with device authentication, resumable uploads, durable checkpointing, and continuous watching.

• Add one-shot capture with exact per-run usage reporting and durable raw artifact storage.

• Ingest Cursor IDE, IcodeMate CLI, and Posit Assistant usage-event sessions.

• Extend S3 discovery to all supported session providers.

• Add Japanese UI localization.

• Expand pricing coverage with OpenRouter and calculate historical costs using rates active at usage time.

• Allow API clients to opt in to session source paths.

• Add an opt-in recall setting for candidate findings.

• Serve usage reports from an exact daily rollup cache for faster results at scale.

• Skip unchanged OpenCode sessions inside changed containers.

• Keep embedding builds progressing through rate limits and cap batches by token budget.

• Make the analysis panel resizable.

• Require explicit consent for shortened session uploads.

• Improve remote sync target selection and skip offline HTTP hosts.

• Harden container builds with distributed Debian mirror selection and automatic failover.

• Align serialized JSON with Go 1.27 JSON v2 semantics.

• Keep previously synced sessions browsable when their source files disappear.

• Keep reconciliation moving past deferred sources and report full-resync finalization progress.

• Discover OpenCode channel databases and recover legacy project working directories.

• Recover Antigravity working directories and effort metadata, and group sessions across Git worktrees.

• Support current Kiro CLI and Qoder CN layouts while handling incomplete Gemini and unsupported Trae sessions safely.

• Show Devin session information in Usage.

• Correct Claude one-hour cache-write pricing and Posit Assistant cache-write accounting.

• Preserve legacy PostgreSQL timestamps with offsets.

• Restore Grok agent-minute reporting and non-interactive session classification.

• Keep daemon readiness waiting active throughout startup progress.

• Keep virtual session rows aligned after sorting and route arrow keys to the active pane.

• Stabilize Usage project filters and prevent password-manager prompts in the project picker.

• Adds Windows system tray support with proper lifecycle behavior.

• Keeps oversized VS Code Copilot snapshots searchable and indexable.

• Makes daily usage snapshots reliable regardless of database query planning.

• Ignores invalid self-referential subagent links.

• Add session support for DeepSeek Harness, TraeX, Codebuff/Freebuff, Prime Agent, and Goose.

• Import Gemini Apps activity from Google Takeout exports.

• Discover Qoder IDE sessions from SharedClientCache.

• Support OpenAI-compatible endpoints for generated insights.

• Show both formatted and raw tool output in session messages.

• Organize generated insights, expose extraction progress, and manage extraction generations.

• Report per-inference model and token usage for Amp.

• Include complete delegated-session costs in usage reports.

• Add JSON output to the usage statusline.

• Scale large activity reports with streaming aggregation and fewer repeated scans.

• Speed up daily usage reporting and duplicate-prompt analysis.

• Sync only changed sessions over HTTP.

• Reduce idle CPU usage and allow unused providers to be disabled.

• Persist Claude and Codex freshness tracking across restarts to avoid unnecessary parsing.

• Bound startup reconciliation and expose daemon identity and sync health.

• Improve Ollama embedding recovery on CPU systems.

• Populate working directories for Kimi sessions.

• Support legacy Zed thread schemas and restore VS Code Copilot response items.

• Preserve Codex tool outputs during incremental parsing and correctly associate forked turns with parents.

• Exclude replayed Codex subagent usage and avoid repeatedly parsing titled sessions without an index.

• Remove replayed prefixes from backgrounded Claude sessions and hide IDE-context envelopes from previews.

• Remove stored Claude sessions that no longer appear after a complete parse.

• Honor explicitly empty agent-directory settings.

• Avoid probing macOS-protected folders during agent discovery.

• Prevent OpenCode session containers from consuming recursive watch capacity.

• Keep root-session navigation unfiltered.

• Distinguish failed frontend actions from successful ones.

• Start reliably when IPv4 and IPv6 listeners encounter port collisions.

• Report stalled synchronization as unhealthy.

• Fixes Kimi Work tool-step usage tracking and K2.6 pricing.