github kenn-io/agentsview v0.48.0

30 minutes ago

New features

  • Sync ChatGPT history through Chrome from the Import dialog or with agentsview source import --type chatgpt --sync. Later runs fetch only chats whose update time changed. If you already use Chrome Sync, rerun agentsview source chrome setup and reload the extension.
  • Sync text from signed-in Gemini Apps chats through Chrome with agentsview source import --type gemini-apps --sync. Each chat imports as one session.
  • Sync Grok web chats through Chrome while signed in with agentsview source import --type grok-web --sync. Each import contains the text of the branch shown on grok.com.
  • Press t in a session to search typed prompts and jump to one, even in transcript history that hasn't loaded yet. The picker follows the transcript's block filters.
  • Find messages by words in any order on SQLite with FTS5 and on PostgreSQL and ClickHouse mirrors. By default, natural multi-word queries rank exact text first, then matches with every word, then partial matches. On SQLite, this word matching applies to non-CJK queries.
  • Find sessions by text inside tool-result images from the search palette or agentsview session search. Enable [image_text] with a loopback GLM-OCR endpoint and restart the daemon to index new and retained images.
  • Use the PR filter to find sessions whose prompts name a pull request. Local archives pick up retained prompt mentions when opened after upgrade. For hosted sessions, run agentsview replica pg raw-reparse hosted --run-id <id> --batch-size 64.
  • Browse Factory Droid sessions with subagent links when the transcript identifies the parent. Token usage is available when the settings file can be read, and subagent usage appears on the parent session.
  • Browse Muse Personal sessions with thinking, tool calls and results, and compaction summaries. Have Muse install the AgentsView CLI on its hosted machine, sync there, and push to a shared PostgreSQL database.
  • Report token usage for tools that call model APIs directly with the direct-api provider. AgentsView reads usage records without message bodies and counts them on the Usage page.
  • Install conversation search skills and register the MCP server for Claude Code and Codex with agentsview agent setup. Use --check to inspect setup or --mcp=false to work with skills while keeping a custom MCP entry.
  • Install AgentsView Memory from the AgentsView repository marketplace for Claude Code and choose the server each project uses. The plugin's search agent can search past conversations and read the source in its own context.
  • First-time users with session history get a Get started checklist on the Dashboard for opening sessions, searching history, viewing token use, and exploring insights. It goes away when completed or dismissed.
  • Choose Match device in Settings > Appearance > Theme to follow the light or dark setting reported by your browser or desktop app. AgentsView switches when the reported setting changes.
  • Turn on Always show year in Settings > Appearance to include the year in sidebar dates, message timestamps, the session header, and command palette results.
  • Apply a saved tool-result image policy with Restart now in Settings. Servers that can't restart themselves show manual steps, and remote access auth changes apply when saved.
  • Read saved tool results through MCP with include_results on get_tool_calls. Results mask detected secrets, and result_cursor reads the next chunk of an unchanged result.
  • Page through large MCP pattern searches, message reads, and tool-call results. Responses report omitted content and return a cursor where paging is supported. For truncated semantic or hybrid searches, retry with a smaller limit.
  • Check MCP call counts, error rates, and rates of server-marked truncation by tool with agentsview mcp calls. The report also shows empty-result rates for supported first-page searches.
  • Read compaction summaries with agentsview session compactions or MCP get_compaction_summaries. Look up one session or list sessions active since a timestamp. Daemon-backed reads require API 19 or later, and RooCode and legacy Kilo sessions are excluded.
  • Use recall list --updated-since or --created-since, or MCP list_recall, to list Recall entries by timestamp without a text query. Time-bounded lists include archived entries unless filtered by status. Without change tracking, narrow the bounds or filters if more than one page matches.
  • Read term counts through MCP get_term_trends on daemon, PostgreSQL, and hosted backends. The Trends page was removed, and old Trends links open Sessions.
  • Wait for startup ingestion or an archive rebuild with agentsview session usage <id> --wait through the daemon. If the wait expires after the session is found, the command shows saved totals and reports the incomplete wait. Archive-only waits work with API 15, while waits that refresh require API 17.
  • Wait for project identification before exporting with agentsview export sessions --wait. The command starts the local daemon when needed and shows progress.
  • Include per-session tool summaries in agentsview export conversations changes --tool-summaries. The export includes counts, errors, time by tool category, and invoked skills for sessions with stored tool calls. Archives last written by an older version need a writable open, such as a daemon restart, before listing summaries.
  • Follow session and message changes over GET /api/v1/changes/stream?after=<cursor>. The stream can resume from a saved cursor and carries token and cost usage and origin machine IDs when available.
  • Export Cursor account usage with agentsview export hour --schema-version 6 --by-machine <hour>. Only closed-hour events in the unfiltered poll that appear in exactly one numeric member query get account IDs. To count each account-hour once, use a reporting server that accepts schema 6 rows with polled_at and avoid mixing schemas 5 and 6.
  • Search hosted raw-sync uploads by meaning with agentsview session search --pg --semantic after the server embeds them. Set [vector] and [vector.embeddings] on the server, run agentsview replica pg hosted-provision provision, and grant the runtime role access to the embedding tables. PostgreSQL semantic search requires pgvector 0.8 or later.
  • Read connection secrets from private files with [kata] token_file, api_key_file for Docbank, insights, embeddings, or Recall extraction, and cursor_admin_api_key_file for Cursor. Configure one source per secret and restart the server after replacing a file.
  • Map a gateway's model name to a catalog model with price_as under [custom_model_pricing."<model>"]. The mapped name follows the catalog model's rates and pricing tiers unless an exact custom rate applies to the reported name.
  • Connect to Docbank over HTTP on a trusted private network with allow_http = true. HTTP sends the API key in plaintext.
  • Record script and background job failures outside an agent session in the daily Friction Log with agentsview diagnostics append on SQLite or PostgreSQL. Repeated failures with the same source and message appear as one Diagnostics entry with a count for the day.
  • With automatic Friction Log filing on, linked Kata issues closed as done reopen when a saved digest within backfill_days records a recurrence that happened after the close. The hourly job comments with the date it recurred.
  • Add an Agent-Session trailer to commits made from a detected agent session with agentsview git-hook install. If a commit-msg hook already exists, the installer prints the line to add.
  • Compare recorded request costs from supported agents with AgentsView's calculations in agentsview session usage. Positive recorded costs fill in when the catalog has no rate, and Claude Code's recorded session total appears separately as Agent total when available.
  • Turn on search by meaning and build its index on a fresh install with agentsview index embeddings build. The command asks for approval before setup and requires a writable daemon.
  • Capture a CPU profile with --pprof when serving the web UI from a PostgreSQL or ClickHouse replica.

Improvements

  • Transcripts show tighter tool-call rows and use a consistent reading column across layouts. Choose Default, Wider, or Full in the new Reading width setting under Appearance.
  • Browser tab titles name the open screen, such as Activity or Usage. Open sessions get a short identifier in the title.
  • After you pick a command palette result, the next Cmd+K restores the query and filters if you haven't reloaded the page. It highlights your pick if the results had finished loading, and the clear button starts a fresh search.
  • HTML and Markdown session exports, published gists, and CLI exports with message text mask detected secrets by default. Use --reveal for CLI message exports or reveal=true for local, unproxied API requests.
  • AgentsView sends usage events only after opt-in through Kenn or an explicit TELEMETRY_ENABLED=1. Changes to the shared choice take effect the next time AgentsView starts.
  • The search index check at restart takes less time on large archives. On a 148 GB archive, the check took over 200 s before and now takes 0.07 s.
  • The Dashboard and its agent filter finish sooner after the first load on long histories. On a synthetic archive of 91,711 sessions and 2,897,301 messages, the Dashboard finished in 7.5 s instead of 13 s after a message write, measured to the last panel as the median of five paired runs.
  • Saved Tools, Velocity, and Signals panels appear sooner on repeat Dashboard visits on SQLite and PostgreSQL. SQLite and native PostgreSQL also reuse those panels after restart when the inputs match and the view has no active-since filter.
  • On SQLite, Dashboard Tools refreshes changed sessions after a small sync. On PostgreSQL, the first Dashboard visit can reuse frustration counts once the startup fill finishes.
  • On SQLite, unchanged Usage ranges reopen faster, and Usage and Activity load sooner when their cache starts empty. Upgrading from v0.47.0 or later to cache format 24 can reuse compatible facts from formats 21 to 23.
  • All-time Usage loads faster on PostgreSQL after a restart or sync once the priced cache is ready, on journals newer than version 1. On ClickHouse, unchanged, fully prepared Usage selections within cache limits reopen faster after a same-day restart. After a next-day restart, only the all-history range gets this benefit.
  • Activity queries run faster on SQLite and ClickHouse. PostgreSQL month and year views reopen sooner when saved report data is available after a graceful restart, including when sessions were pushed while the server was stopped.
  • Repeat global search pages reopen sooner, even after a restart, on SQLite, native PostgreSQL, and ClickHouse when the query, filters, sort, and paging match and searchable data is unchanged. PostgreSQL also refreshes saved matches after a sync, including after a restart, when the saved checkpoint is still usable.
  • On SQLite, session search stops scanning older sessions if recent exact matches fill the page. Filtered multi-word searches rank only matches in scope.
  • When nobody is watching the archive, the daemon syncs changed session files at most once per background_sync_interval, which defaults to ten minutes. A CLI or MCP read through the daemon keeps the live sync pace for two minutes, so repeated reads during that time skip a new catch-up.
  • agentsview sync is back in root help and runs without a deprecation warning. With a running daemon and up-to-date watcher coverage, manual sync applies the local changes the watchers collected. Use --rescan for a full pass.
  • When native file watching fails, fallback polling can sync just the changed source files for Claude, Codex, and supported single-file JSONL providers.
  • Remote sync can download just the appended bytes of a growing JSONL transcript if both ends support it and the saved copy's tail matches. Connecting to an offline HTTP remote now times out after 10 seconds instead of 30.
  • ClickHouse pushes reuse stored messages and usage when a session only grows, and reuse unchanged session content after rebuilds and resyncs. Upgrade ClickHouse serving with the pushers.
  • A file-triggered ClickHouse watch push skips writes when the archive has no changes for the replica. Daemon-backed PostgreSQL pushes can skip repeated schema checks in the same daemon run while the target schema still matches.
  • The help for agentsview serve stop and agentsview serve restart now names read-only ClickHouse servers.

Bug fixes

  • AgentsView starts on x86-64-v2 CPUs. The DuckDB mirror and agentsview replica duckdb were removed. Use a PostgreSQL or ClickHouse replica for a read-only copy.
  • Open session tabs can share live updates over HTTP/1.1 without blocking new pages. Selected-session live updates require API 28, and reloading shows new messages on API 27.
  • Use an unambiguous plain UUID in MCP session tools, and keep it in the address bar when opening a session by that URL. Older remote servers before API 29 still require full stored IDs for required MCP lookups.
  • MCP content search accepts concepts with mode: "hybrid" and lists a session's messages in order when given session_id without pattern or concepts.
  • MCP errors for unknown arguments list the names the tool accepts. If a session search page is empty because matches were active in the last 10 minutes, it shows their count and suggests include_active: true.
  • Failed page and transcript loads show an error and Retry. The command palette also retries a transient full-text search error once and offers Retry if it still fails.
  • A live transcript refresh no longer jumps back to the current find match when new messages arrive.
  • Pages served from PostgreSQL, including hosted PostgreSQL, wait for a connection when the server's pool is busy, and the pool stays within the login role's connection limit. Broad PostgreSQL session searches from MCP server mode and the command palette can complete past the server's write timeout.
  • ClickHouse global search uses less query memory so large searches that hit the memory limit can return results.
  • CockroachDB 25.4.6 can finish schema setup and accept session pushes.
  • Hosted raw sync accepts uploads when the published container runs as a non-root UID with no passwd entry. A non-container install with such a UID still needs DOCBANK_LOCK_DIR set.
  • Daemon-backed PostgreSQL and ClickHouse push watches follow committed archive changes on API 27 or later. Different daemon and client source roots no longer cause those change-triggered pushes to fail.
  • Remote syncs requested through the daemon continue when the requesting client disconnects. On Windows, refused or unreachable remotes are skipped and reported as offline.
  • On macOS, live sessions keep updating after a recursive folder watch drops events. Watches on folders one level deep also avoid a failure that could hang the daemon.
  • Daemon restarts on Windows wait for the old process's archive write lock and can reopen the archive while an old shared-memory mapping lingers.
  • AgentsView Desktop on Windows can unpack compressed in-app updates. If you're updating from an affected Windows release starting with 0.43.0, run the setup .exe once by hand.
  • The desktop loading page shows database preparation and initial sync progress during background startup.
  • agentsview serve holds its chosen port through startup when you use --port 0 or when the default port is occupied.
  • Claude.ai Chrome Sync skips chats whose list entry is missing the current leaf field instead of reporting an error for each one.
  • agentsview source cursor sync-usage retries pages when Cursor rate-limits a request or returns a server error. It prints each wait so you can see why the sync paused.
  • Cursor IDE chats in the _v: 18 format import again. Mistral Vibe 3.1 sessions also sync from their unified stores.
  • Compacted Grok Build sessions show earlier turns and a divider with each compaction summary when the source history is available. Tool calls also appear in the transcript.
  • Codex transcripts show readable reasoning summaries and compaction dividers. Calls with identical token counts can still count as separate requests.
  • Antigravity sessions use recorded workspace information for project and working directory, and nest subagents when a parent is recorded. Antigravity CLI summary imports keep the typed prompt when a skill expands in the same step.
  • DeepSeek Harness messages keep their native IDs, so existing rows, pins, and exported message IDs survive live rescans as new frames arrive.
  • Sessions from <repo>.<branch> sibling worktrees group under their repository when the checkout is missing from disk and the suffix matches the recorded non-default branch. Existing sessions pick up the change when they are reparsed.
  • OpenCode-family usage counts reasoning tokens at the output rate when the usage record shows that output excludes them. Existing sources are reparsed in place.
  • Usage totals include recorded tokens from replies with no visible text or tool call for affected agents, including Codex, OpenCode, and Gemini CLI. Claude responses with a message ID are counted once when split by tool results, even if requestId is missing.
  • Gemini 4 display names in Antigravity and bare gemini-4-* model IDs match Google's catalog rates. Older Antigravity display names such as Gemini 3.5 Flash (High) also match the pricing catalog. Exact custom rates still take precedence.
  • Activity reports pick up changes when pricing is refreshed.
  • Search index builds on the built-in local embedding server retry timed-out batches in smaller pieces. On SQLite, Message and Recall embedding builds can also publish when they overlap.
  • PostgreSQL pushes stop if the local embedding index contains an all-zero vector. If a push reports one, run agentsview index embeddings build --repair-invalid.
  • With require_auth and managed Caddy, failed stream requests no longer leave bearer tokens from request URLs in the daemon log.
  • Growing Codex sessions no longer add deletion records for unchanged messages when sync can update the transcript in place. The daemon prunes expired records without native message IDs for reuse, and db compact can reclaim the space on disk.
  • Commit links are recovered from older Codex shell calls that recorded Git commands as argument lists. In PostgreSQL and ClickHouse replicas, a retained link pushed by another archive no longer replaces a live session's prompts for the same commit.
  • Prune removes matching parents and child sessions in one pass.
  • Full transcript parses count dropped oversized JSONL records in the session's malformed-lines badge and sync summary. Undecodable data also counts for affected OpenCode, Shelley, and Forge records.
  • capture report returns finalization_timeout when the wait expires before the first session lookup completes, so it can suggest retrying or allowing more time.

Dependency updates

  • Update Kit to v0.58.1 and go.kenn.io/kata to v0.23.0.
  • Update Huma to v2.41.0 and golang.org/x/crypto to v0.59.0.
  • Update the SQLite driver to v1.14.53, which bundles SQLite 3.54.1.

Don't miss a new agentsview release

NewReleases is sending notifications on new releases.