v0.45.0

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 local sessions and reachable remotes syncing when another remote's hostname cannot resolve. Sync failures retain their actionable error details.
• Keep Linux servers discoverable after clock corrections, and let Windows startup progress update while commands read it.
• Detect same-size file replacements on Windows and sync Positron project changes when file timestamps stay the same.
• Open the desktop app against an authenticated local daemon without a readiness timeout. The app also recovers after an observed backend outage while hidden and avoids stale startup errors during an in-app update.
• Replace a served DuckDB mirror with duckdb push on Windows without stopping duckdb serve.
• Load current SQLite-backed OpenClaw sessions and discover Hermes Agent sessions in the default Windows location.
• Report incomplete Git and pull-request totals in agentsview stats, and wait beyond the former 30-second client timeout for slow results.
• Count repositories once across clones and worktrees, and count each commit once while retaining work from diverged branches.
• Clear the minimum-prompt filter by clicking its selected value again. Usage cards grow to contain wrapped text, and password-manager menus no longer cover typeahead search boxes.
• Refuse writable archives when the build cannot load their existing FTS5 index, with an error that explains how to fix the executable.