Mark Murphy

Hermes merged 60 PRs on August 30 The ones worth knowing: • Credential vault lives in Settings, with agent deep link prefill • Real profile browsing: Brave Origin detection, macOS signed in, headless (no focus steal), toggle under Capabilities to Tools to Browser • Desktop MCP OAuth sign in works against remote backends • /loop runs its first iteration immediately • /plan is a built in command that survives platform menu caps • Telegram: every command and skill searchable via @botname inline picker • CLI and TUI status show cache hit %, latency, and t/s with field toggles • System prompt always rebuilds at compaction commit, so updates reach long lived sessions • Skill activation drops repeated absolute paths (down 31% load on reference heavy skills) • Copied session cookies are no longer world readable • Unattended webhook/API sessions no longer stall 300s on dangerous command approval • Lean compaction is one auxiliary request again (7 to 11 min digest stalls gone) Here is the full list: DESKTOP IMPROVEMENTS (13 PRs) • feat(desktop): manage the credential vault from Settings, with agent deep link prefill • fix(computer use): stop launching retired browser grant runtimes • fix(tui): render nested todo subtasks via the parent field • fix(desktop): remote bot replies no longer die at the 30s relay deadline (#93911, salvage #93965) • feat(browser): Brave Origin no longer misdetected, real profile browsing + CDP connect (salvage #41505) • fix(todo): Tasks panel stays live across merges, reconnects, and hidden tool progress (salvage #97815, #97811) • fix(browser): real profile browsing on macOS no longer opens signed out or hangs (salvage #96763) • feat(mcp oauth): Desktop MCP OAuth sign in now works against remote backends • fix(cron): Desktop's advertised natural schedules now parse (#51975, salvage #51598 #52504) • feat(desktop): the real profile browsing toggle users couldn't find, Capabilities to Tools to Browser • feat(tui): status rule shows cache hit %, latency, and t/s with shared field toggles (extends #98250) • fix(browser): real profile browsing runs headless, no focus stealing window • fix(todo): unversioned tool.start merges again after desktop resume (#98304, salvage #98306) DEVELOPER TOOLS & PLUGINS (22 PRs) • fix(providers): register Alibaba China + Token Plan providers, 'Unknown provider' despite catalog entry (#73265) • fix(memory): keep Mem0 OSS OpenAI requests direct • feat(loop): /loop runs its first iteration immediately (salvage #97958) • fix(skills): skill_view directory file_path + skill_manage categorized name resolution • fix(prompt): sync DEFAULT_SOUL_MD with the #95681 identity rewrite • docs: sync stale /background references with the /bg + /btw split • fix(tools): restore setup_mcp's never hand edit instruction • fix(delegation): subagents on an explicit base_url keep their provider's request_overrides (#65035, salvage #65054) • feat(skills hub): impeccable joins the optional skills catalog, content pulled live from upstream • feat(cli): status bar gains cache hit %, latency, t/s, and per field toggles (salvage #92696, #39760, #97970) • hermes model shows curated lists for all six Alibaba tiers (salvage #77848) • Background review lands skill patches again, read tools whitelisted, cache parity intact (#61521, #39996) • docs(computer use): browser route and grant references match the shipped tool (salvage #96156) • Qwen Cloud picker gains qwen3.8 max + 4 more missing models (salvage #87808) • feat(delegation): delegation.request_overrides config key sends per child request settings on every resolution branch (salvage #90953) • feat: /plan survives the platform command menu caps as a built in command (closes #67264, #36821, salvage #67292) • Background review accepts profile configured extra tools via opt in whitelist (closes #44672, salvage #82146) • fix(skills hub): skill installs survive dangling refs, sibling links, glob prose, and unpinned fetches (salvages #66828 #82236 #90105 #96336) • feat(compaction): system prompt always rebuilds at the commit boundary, updates finally reach long lived sessions • fix: skill activation no longer repeats the absolute path on every supporting file line (down 31% load cost on reference heavy skills) • refactor(skills): move merge reconciler to optional catalog as agent merge conflict arbiter • refactor(skills): shipped set slim, 15 to optional, github 6 way merge, pdf absorbs OCR, channel gated teams pipeline (index down 26%) SECURITY & SECRETS (4 PRs) • Anthropic OAuth: shared credential refreshes stop burning single use tokens; dashboard PKCE removed (salvage #87891) • fix(browser): copied session cookies are no longer world readable (#96729) • fix(security): env vars with mid name API/KEY/TOKEN no longer flag as exfiltration (salvage #63994) • fix(approval): unattended webhook/API sessions no longer stall 300s on dangerous command approval (#37284, salvage #37317) INTEGRATIONS & CHAT (5 PRs) • fix(telegram): recover exhausted request pool • fix(runtime): provider switch, fallback, and recovery keep request settings in sync (#75091, salvage #75139 #60062) • fix(gateway): custom provider extra_body survives gateway turns and /model switches (#54922, salvage #39429 #52432 #53765) • fix(telegram): pinned skill commands survive the menu cap (salvage #82516) • feat(telegram): every command and skill is searchable via @botname, inline picker bypasses the menu cap STABILITY & FIXES (16 PRs) • fix(cli): slow /handoff transfers no longer misreported as "gateway not running" • 3x fmt(js): npm run fix auto fix • fix(cron): 'every hour' and 'every day' schedules no longer error (salvage #97783) • fix(cron): scheduled jobs no longer lose custom provider request settings (salvage #56876) • fix(cron): bare '30m' schedules recur as documented; 'in 30m' is the one shot (salvage #78928) • fix(cron): repeat='forever' no longer breaks job updates (completes the string repeat class) • fix(compression): stale todo snapshots no longer accumulate or resurrect finished plans (salvage #97604, #89910) • fix(compression): native Codex compaction no longer fires at the stale 200K default on 900K windows (#88695, salvage #88722) • fix(compression): compacted native Codex sessions no longer false trigger 600s local compression mid turn (#96995, #97602, salvage #97006) • fix(native compaction): image only user messages survive checkpoint pruning (salvage #91557, folds #98345 test) • fix(codex): compaction only turns no longer die after 3 continuation attempts (salvage #98008) • fix(compaction): native capability survives same provider /model switches and gateway resume (salvage #94036 + #97292) • fix(compression): native compaction settings hot apply to open sessions (salvage #96740) • fix(compression): lean compaction is one auxiliary request again, 7 to 11 min digest stalls eliminated (#96603) xitter.cf/iamlukethedev/status/2…

Hermes hidden gem: /btw Your agent is deep into a task Then you suddenly need something completely different: Read another file Pull some documentation Review recent commits Run a quick security check You do not want to interrupt the current work or babysit another window Just type: /btw summarize the last 20 commits and flag anything unsafe Hermes launches a separate background session and immediately gives your main chat back. The background agent inherits your: • Model • Provider • Toolsets • Reasoning settings • Fallback model It works independently, then sends the result back automatically when it finishes. In the CLI, the answer appears as a result panel. On Telegram, Discord, Slack, and other messaging platforms, it returns to the same chat. You can use any of these: /background /bg /btw The difference is simple: /queue waits until the current task finishes. /steer changes the task already running. /btw starts a completely separate task while you keep working It is basically telling Hermes: “By the way, handle this too.” Small command Huge multitasking unlock