…search#58)

The backend CLI session bridge (PR NousResearch#56) was complete but the frontend
never connected to it:

1. css class never applied -- el.className never included 'cli-session'
   so the gold border and 'cli' badge CSS was dead code. Fixed: append
   ' cli-session' when s.is_cli_session is true.

2. import never triggered -- click handler always called loadSession()
   directly, never POST /api/session/import_cli. Fixed: for CLI sessions,
   call import_cli first (idempotent -- safe to call on every click),
   then fall through to loadSession() which now finds the imported copy.

3. profile filter silently hid CLI sessions -- filter required
   s.profile === S.activeProfile, but CLI sessions may have profile=null
   if the SQLite DB has no profile column. Fixed: CLI sessions always
   pass the filter (s.is_cli_session || s.profile === S.activeProfile).

Co-authored-by: Nathan Esquenazi <nesquena@gmail.com>

…k variant

Parity port from linxule/hermes-kimi-plugin@93476f5 — same four fixes plus
the NousResearch#60 fold-in:

1. connect# off-by-one: counter + ts now bumped pre-dispatch in
   _group_subscribe_once so the gap log inside _on_group_event reads the
   correct connect# (matches the subscribe-stream-live log emitted moments
   later). Subscribe-live log + backoff clamp stay post-dispatch as the
   operator-visible recovery announcement.

2. Monotonic clock: arrival-time tracking + since-reconnect correlation
   now use time.monotonic() (immune to NTP/leap-second/VM-suspend).

3. _BoundedLRU shared cap: _last_arrival_time_per_room gets its own
   _ARRIVAL_TIME_CACHE_DEFAULT_MAX (10000), so eviction can't blind
   Phase 0 gap detection under cardinality pressure.

4. Anchor-only-on-first-page in list_group_messages: pageToken and
   start_message_id/end_message_id are now mutually-exclusive cursoring
   modes.

Fold-in (NousResearch#60): _fetch_group_message uses max_pages=2 for tight-range
pagination resilience.

Tests: 200 passing on fork (194 unit + 6 plugin-integration). Same +10/-1
delta as standalone. Existing handler-exception test updated to reflect
new pre-dispatch state semantic.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

…gs, 31/31 tests pass

New architecture (4 layers):
- store.py — persist full result to ~/.autolycus/rtk-cache/<uuid>.txt
- compressor.py — type-aware dispatcher (terminal/read_file/search_files)
- strategies/* — per-tool compression (tail-first for terminal, offset-aware for read_file, compact for search_files)
- monitor.py — per-tool aggregated metrics w/ JSON export
- __init__.py — transform_tool_result hook + rtk_recover tool

Benchmark on production-like data:
- systemctl: 57.6% savings (525 tokens/call)
- MOEX scanner: 57.0% savings (513 tokens/call)
- search_files: 96.1% savings (1,319 tokens/call)
- read_file: 91.5% savings (4,180 tokens/call)
- Global: 84.4% savings, ALL data recoverable

Closes: NousResearch#58

…a dirty live tree

The hourly ~/.hermes auto-pull silently DEBUG-skipped on a dirty working
tree. The live tree is almost always dirty (continuous skill-edit churn),
so incoming merges piled up undelivered on disk with zero alerting (PRs
NousResearch#54-NousResearch#58, ~5 merges, only noticed when a cron broke).

- _pull_once now counts commits-behind BEFORE the clean check so it can
  distinguish the WEDGED state (on main, behind>0, dirty) from up-to-date.
  Wedged invokes an optional notify callback; clean+behind still
  fast-forwards; up-to-date and feature-branch stay silent DEBUG.
- HermesHomePuller takes a notifier + throttles wedged alerts (default 6h,
  tracked via _last_alert_ts) so the operator isn't paged hourly. The
  throttle clock only advances once the notifier actually accepts the
  message, so a failed send retries next tick.
- GatewayRunner gains _send_operator_alert (reuses the home-channel
  adapter.send surface used by shutdown/startup notifications) and
  make_operator_alert_callback (thread-safe sink via safe_schedule_threadsafe).
  start_gateway threads the callback into start_hermes_home_puller.
- Tests: wedged fires once then suppressed in-window then again after the
  window elapses; clean+behind fast-forwards no alert; up-to-date no alert
  no pull; feature branch no alert; notifier-exception non-fatal + no
  throttle advance. 30 -> 39 tests.

…mes#62)

Expõe action=edit na tool send_message, espelhando o caminho send mas
enxuto (texto-only, Telegram-only). _handle_edit valida message_ref
({channel,chat_id,message_id}) e despacha para _edit_telegram, que usa
bot.edit_message_text com HTML auto-detect, fallback de parse para plain
e trata 'message is not modified' como sucesso. Outras plataformas
retornam erro claro.

Desdobrado da NousResearch#58 (critério NousResearch#2: editar mensagem anterior in-place).
TDD: TestEditTelegram + TestHandleEditDispatch (mock só em telegram.Bot).

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

…a dirty live tree (#45)

The hourly ~/.hermes auto-pull silently DEBUG-skipped on a dirty working
tree. The live tree is almost always dirty (continuous skill-edit churn),
so incoming merges piled up undelivered on disk with zero alerting (PRs
NousResearch#54-NousResearch#58, ~5 merges, only noticed when a cron broke).

- _pull_once now counts commits-behind BEFORE the clean check so it can
  distinguish the WEDGED state (on main, behind>0, dirty) from up-to-date.
  Wedged invokes an optional notify callback; clean+behind still
  fast-forwards; up-to-date and feature-branch stay silent DEBUG.
- HermesHomePuller takes a notifier + throttles wedged alerts (default 6h,
  tracked via _last_alert_ts) so the operator isn't paged hourly. The
  throttle clock only advances once the notifier actually accepts the
  message, so a failed send retries next tick.
- GatewayRunner gains _send_operator_alert (reuses the home-channel
  adapter.send surface used by shutdown/startup notifications) and
  make_operator_alert_callback (thread-safe sink via safe_schedule_threadsafe).
  start_gateway threads the callback into start_hermes_home_puller.
- Tests: wedged fires once then suppressed in-window then again after the
  window elapses; clean+behind fast-forwards no alert; up-to-date no alert
  no pull; feature branch no alert; notifier-exception non-fatal + no
  throttle advance. 30 -> 39 tests.

Co-authored-by: Sahil (AI) <266772320+sahilm-ai@users.noreply.github.com>