tg-recall 0.8.1

Fixed

• stats(query=…) also counts voice messages whose transcript matches, each message once, like search.
• transcribe uses a local speech-to-text executable configured by path (transcription.executable), not only one found on PATH. Checked live with Faster-Whisper-XXL on a GPU: two voice messages downloaded and transcribed in about 20 seconds.
• sync from an AI agent's shell (CLI, not MCP) runs for 90 seconds by default instead of an hour, so the agent's command is not cut off by its harness; it resumes on the next call. The owner's terminal keeps 3600 seconds.
• tg-recall stats no longer offers an unused --budget option.

Added

• scripts/agent_bench.py: runs your own questions through a real agent (headless Claude Code with only the tg-recall tools) and records turns, tool calls, tool-result tokens, total tokens, cost, time and keyword checks; --compare shows the difference between two runs. Questions and results stay in your profile because they name your chats; scripts/bench_questions.example.json shows the format.

tg-recall v0.8.0

tg-recall 0.8.0 gives AI agents the tools for big jobs over your Telegram archive: a year of a forum topic, trends over months, voice messages.

Highlights

• export(): a whole period or topic, full text, a date on every line, in one local file the agent reads in chunks. No more paging read() 100 messages at a time.
• stats(): counts per day/week/month, query hits per period, top senders, topics. A topic's year in a few hundred tokens.
• Background sync: an MCP call returns within seconds; the download continues with Telegram's message count, progress and ETA while the agent keeps working. sync() reports status.
• Search syntax: a | b alternatives (other languages, synonyms), "exact phrase", -word.
• transcribe(refs) (opt-in, ai_access.allow_transcribe): voice/audio/video by citation.
• tg-recall usage: what agents called, how many tokens it cost, which searches came back empty, where they repeated themselves. Arguments only, never message text.

Upgrade notes

• ai_access.sync_max_seconds now means how long an MCP sync call waits (default 20). Profiles saved by 0.7 keep 50; tg-recall config set ai_access.sync_max_seconds 20 to switch.
• New keys: sync_background_minutes (30), max_export_messages (50000; 0 disables export), allow_transcribe (false).

Install

uv tool install --force https://github.com/Rerowros/tg-recall/releases/download/v0.8.0/tg_recall-0.8.0-py3-none-any.whl

Full notes: CHANGELOG · Русский

tg-recall v0.7.0

tg-recall is now a small core: a local Telegram archive plus search, read, chats and sync for you and your AI agents, over MCP or the CLI. Breaking release: back up first (tg-recall backup create --mode essential --output <zip>).

Highlights

• Compact search / read / chats output (one line per message, token budget, tg:// citations), the same for MCP and CLI.
• Forum topics: select them by t.me link or chat/topic; a topic is fetched alone, not the whole group.
• Fast tg-recall sync [TARGET...]: one connection for many chats, first sync goes 30 days back, --since for older history, interruption-safe, FloodWait respected, session lock (busy).
• MCP sync and auto-refresh before search/read when ai_access.allow_sync=true.
• ai_access.allow_all_chats to give agents every archived chat.
• The archive shrinks on upgrade: derived semantic/embedding tables and sync scopes are dropped and the file is compacted (a real 159 MB archive became 83 MB).

Removed

ask/retrieve and in-tool LLM answers, semantic/hybrid search and embeddings, research sessions and the knowledge catalog, wiki memory, export packs, integrate / agent guide / model routing, update, legacy tg-ecosystem aliases, sync scopes. Old configs still load; removed keys are ignored.

Install

uv tool install --force https://github.com/Rerowros/tg-recall/releases/download/v0.7.0/tg_recall-0.7.0-py3-none-any.whl

Full notes: CHANGELOG · Русский

tg-recall v0.6.0

tg-recall v0.6.0

Local-first Telegram archive for CLI and AI agents. Archives, media, sessions,
credentials, wiki, exports, and AI packs remain private by default.

Highlights

• Added a permanent versionless setup prompt for Codex, Claude Code, Cursor,
  and generic AI harnesses. It resolves only a verified stable GitHub Release
  wheel and keeps installation or updates as explicit human commands.
• Allowed data-blind integrate list, preview, and status diagnostics for
  automation without reading profiles, Telegram data, sessions, or SQLite and
  without writing harness files.
• Added profile-local typed transcription settings and a shell-free
  Faster-Whisper-XXL adapter with explicit local model checks, CUDA/device,
  compute type, language, VAD, and bounded timeout controls.
• Preserved sidecar as the transcription default and Telegram-first auto;
  agents still need one permitted explicit tg:// citation and cannot change
  configuration.
• Added complete Russian versions of the main public documentation and
  synchronized English/Russian local-transcription guidance.
• Corrected documented placement of the global --json flag and expanded
  distribution privacy checks.

Install

uv tool install https://github.com/Rerowros/tg-recall/releases/download/v0.6.0/tg_recall-0.6.0-py3-none-any.whl
tg-recall --json doctor

Privacy

No archive-wide automatic download, transcription, export, or provider upload
is introduced. Telegram sessions, credentials, SQLite archives, media, wiki,
exports, and generated agent integration state are excluded from release
artifacts.

tg-recall v0.5.0

tg-recall v0.5.0

Local-first Telegram archive for CLI and AI agents, with all archive, media, sessions, wiki, exports, and AI packs kept private by default.

Highlights

• Hardened agent and MCP scope enforcement across chat, date, media type, and result limits.
• Safer incremental sync and backfill: monotonic watermarks and resumable FLOOD_WAIT handling.
• Schema v6 maintenance: transactional migrations, doctor diagnostics, filtered jobs, explicit retry, and repair previews.
• Budgeted agent context routing with Codex Spark/Luna guidance, cited evidence sets, knowledge catalog, and research sessions.
• Profile-local private wiki memory with immutable cited snapshots and raw-source expansion.
• Optional fully local hybrid retrieval with explicit embedding setup, vector freshness, deterministic evidence, and keyword fallback.
• Bounded private AI export packs with verifiable manifests and scope/citation provenance.
• Optional OpenAI Responses synthesis behind explicit policy and scope checks, with cited local fallback.
• Updated CLI and read-only MCP metadata for release version 0.5.0.

Privacy

No archive-wide automatic export or provider upload is introduced. Telegram sessions, credentials, SQLite data, media, wiki, and exports remain local and excluded from release artifacts.

tg-recall v0.2.0

Profile-aware local Telegram archive release. Includes portable platform storage, safe legacy migration, consistent backups, media policies, incremental sync/backfill, local Whisper support, and AI-oriented CLI retrieval. MCP remains read-only.

tg-ecosystem v0.1.0

Initial public early-alpha release. Includes the local Telegram archive CLI, read-only MCP server, privacy-safe package contents, and CI. See README for current limitations and security guidance.