hermes-agent/hermes_cli
Teknium dad10a78d0 feat(computer-use): cua-driver backend, universal any-model schema
Background macOS desktop control via cua-driver MCP — does NOT steal the
user's cursor or keyboard focus, works with any tool-capable model.

Replaces the Anthropic-native `computer_20251124` approach from the
abandoned #4562 with a generic OpenAI function-calling schema plus SOM
(set-of-mark) captures so Claude, GPT, Gemini, and open models can all
drive the desktop via numbered element indices.

- `tools/computer_use/` package — swappable ComputerUseBackend ABC +
  CuaDriverBackend (stdio MCP client to trycua/cua's cua-driver binary).
- Universal `computer_use` tool with one schema for all providers.
  Actions: capture (som/vision/ax), click, double_click, right_click,
  middle_click, drag, scroll, type, key, wait, list_apps, focus_app.
- Multimodal tool-result envelope (`_multimodal=True`, OpenAI-style
  `content: [text, image_url]` parts) that flows through
  handle_function_call into the tool message. Anthropic adapter converts
  into native `tool_result` image blocks; OpenAI-compatible providers
  get the parts list directly.
- Image eviction in convert_messages_to_anthropic: only the 3 most
  recent screenshots carry real image data; older ones become text
  placeholders to cap per-turn token cost.
- Context compressor image pruning: old multimodal tool results have
  their image parts stripped instead of being skipped.
- Image-aware token estimation: each image counts as a flat 1500 tokens
  instead of its base64 char length (~1MB would have registered as
  ~250K tokens before).
- COMPUTER_USE_GUIDANCE system-prompt block — injected when the toolset
  is active.
- Session DB persistence strips base64 from multimodal tool messages.
- Trajectory saver normalises multimodal messages to text-only.
- `hermes tools` post-setup installs cua-driver via the upstream script
  and prints permission-grant instructions.
- CLI approval callback wired so destructive computer_use actions go
  through the same prompt_toolkit approval dialog as terminal commands.
- Hard safety guards at the tool level: blocked type patterns
  (curl|bash, sudo rm -rf, fork bomb), blocked key combos (empty trash,
  force delete, lock screen, log out).
- Skill `apple/macos-computer-use/SKILL.md` — universal (model-agnostic)
  workflow guide.
- Docs: `user-guide/features/computer-use.md` plus reference catalog
  entries.

44 new tests in tests/tools/test_computer_use.py covering schema
shape (universal, not Anthropic-native), dispatch routing, safety
guards, multimodal envelope, Anthropic adapter conversion, screenshot
eviction, context compressor pruning, image-aware token estimation,
run_agent helpers, and universality guarantees.

469/469 pass across tests/tools/test_computer_use.py + the affected
agent/ test suites.

- `model_tools.py` provider-gating: the tool is available to every
  provider. Providers without multi-part tool message support will see
  text-only tool results (graceful degradation via `text_summary`).
- Anthropic server-side `clear_tool_uses_20250919` — deferred;
  client-side eviction + compressor pruning cover the same cost ceiling
  without a beta header.

- macOS only. cua-driver uses private SkyLight SPIs
  (SLEventPostToPid, SLPSPostEventRecordTo,
  _AXObserverAddNotificationAndCheckRemote) that can break on any macOS
  update. Pin with HERMES_CUA_DRIVER_VERSION.
- Requires Accessibility + Screen Recording permissions — the post-setup
  prints the Settings path.

Supersedes PR #4562 (pyautogui/Quartz foreground backend, Anthropic-
native schema). Credit @0xbyt4 for the original #3816 groundwork whose
context/eviction/token design is preserved here in generic form.
2026-04-28 01:46:36 -07:00
..
__init__.py chore: release v0.11.0 (2026.4.23) (#14791) 2026-04-23 15:31:59 -07:00
auth.py feat(providers): add GMI Cloud as a first-class API-key provider (#11955) 2026-04-27 11:17:59 -07:00
auth_commands.py Add native Spotify tools with PKCE auth 2026-04-24 05:20:38 -07:00
azure_detect.py feat(azure-foundry): auto-detect transport, models, context length 2026-04-25 18:48:43 -07:00
backup.py feat(backup): exclude SQLite WAL/SHM/journal sidecars (#16576) 2026-04-27 06:43:52 -07:00
banner.py feat(banner): hyperlink startup banner title to latest GitHub release (#14945) 2026-04-23 23:28:34 -07:00
callbacks.py fix: ESC cancels secret/sudo prompts, clearer skip messaging (#9902) 2026-04-14 16:11:37 -07:00
claw.py Normalize claw workspace paths for Windows 2026-04-22 18:15:27 -07:00
cli_output.py refactor: remove dead code — 1,784 lines across 77 files (#9180) 2026-04-13 16:32:04 -07:00
clipboard.py feat: fix img pasting in new ink plus newline after tools 2026-04-11 13:14:32 -05:00
codex_models.py feat(codex): add gpt-5.5 and wire live model discovery into picker (#14720) 2026-04-23 13:32:43 -07:00
colors.py feat: respect NO_COLOR env var and TERM=dumb (#4079) 2026-03-30 17:07:21 -07:00
commands.py fix(cli): eliminate ghost status-bar + DSR input leaks from terminal drift 2026-04-27 05:31:47 -07:00
completion.py fix: preserve profile name completion in dynamic shell completion 2026-04-14 10:45:42 -07:00
config.py feat(plugins): add bundled observability/langfuse plugin 2026-04-28 01:40:59 -07:00
copilot_auth.py fix(copilot): exchange raw GitHub token for Copilot API JWT 2026-04-24 05:09:08 -07:00
cron.py feat(cron): per-job workdir for project-aware cron runs (#15110) 2026-04-24 05:07:01 -07:00
curses_ui.py feat: ungate Tool Gateway — subscription-based access with per-tool opt-in 2026-04-16 12:36:49 -07:00
debug.py fix(debug): sweep expired paste.rs uploads on a real timer (#16431) 2026-04-27 00:36:33 -07:00
default_soul.py fix: reset default SOUL.md to baseline identity text (#3159) 2026-03-26 01:34:27 -07:00
dingtalk_auth.py test(dingtalk): cover QR device-flow auth + OpenClaw branding disclosure 2026-04-17 05:08:07 -07:00
doctor.py feat(providers): add GMI Cloud as a first-class API-key provider (#11955) 2026-04-27 11:17:59 -07:00
dump.py fix(gemini): fail fast on missing API key + surface it in hermes dump (#15133) 2026-04-24 05:35:17 -07:00
env_loader.py fix(cli): ensure project .env is sanitized before loading 2026-04-22 05:51:44 -07:00
fallback_cmd.py feat(cli): add 'hermes fallback' command to manage fallback providers (#16052) 2026-04-26 06:19:04 -07:00
gateway.py yuanbao platform (#16298) 2026-04-26 18:50:49 -07:00
hooks.py feat(hooks): add duration_ms to post_tool_call + transform_tool_result (#15429) 2026-04-25 22:13:12 -07:00
logs.py feat: component-separated logging with session context and filtering (#7991) 2026-04-11 17:23:36 -07:00
main.py fix(mcp): move discovery out of model_tools import side effect (#16856) (#16899) 2026-04-28 01:17:58 -07:00
mcp_config.py fix(mcp): consolidate OAuth handling, pick up external token refreshes (#11383) 2026-04-16 21:57:10 -07:00
memory_setup.py fix(memory): discover user-installed memory providers from $HERMES_HOME/plugins/ (#10529) 2026-04-15 14:25:40 -07:00
model_catalog.py feat(models): remote model catalog manifest for OpenRouter + Nous Portal (#16033) 2026-04-26 05:46:43 -07:00
model_normalize.py fix(model-normalize): pass DeepSeek V-series IDs through instead of folding to deepseek-chat 2026-04-24 05:24:54 -07:00
model_switch.py fix(context): honor custom_providers context_length on /model switch + bump probe tier to 256K (#15844) 2026-04-25 18:47:53 -07:00
models.py fix(copilot): require successful exchange when walking credential_pool catalog tokens 2026-04-28 01:18:09 -07:00
nous_subscription.py fix(cli): coerce use_gateway config flags in tool routing 2026-04-26 19:02:55 -07:00
oneshot.py feat(oneshot): add --model / --provider / HERMES_INFERENCE_MODEL (#15704) 2026-04-25 08:55:36 -07:00
pairing.py fix(pairing): handle null user_name in pairing list display 2026-04-23 02:34:11 -07:00
platforms.py yuanbao platform (#16298) 2026-04-26 18:50:49 -07:00
plugins.py feat(plugins): add pre_approval_request / post_approval_response hooks (#16776) 2026-04-27 20:08:33 -07:00
plugins_cmd.py feat(plugins): make all plugins opt-in by default 2026-04-20 04:46:45 -07:00
profiles.py fix(profiles): stage profile imports to prevent directory clobbering 2026-04-23 03:02:34 -07:00
providers.py feat(providers): add GMI Cloud as a first-class API-key provider (#11955) 2026-04-27 11:17:59 -07:00
pty_bridge.py fix: mobile chat in new layout 2026-04-24 12:07:46 -04:00
runtime_provider.py fix(opencode): re-derive api_mode per target model on /model switch 2026-04-28 01:14:35 -07:00
setup.py yuanbao platform (#16298) 2026-04-26 18:50:49 -07:00
skills_config.py refactor: remove dead code — 1,784 lines across 77 files (#9180) 2026-04-13 16:32:04 -07:00
skills_hub.py feat(skills): install skills from a direct HTTP(S) URL (#16323) 2026-04-26 20:57:10 -07:00
skin_engine.py fix(skins): don't inherit status_bar_* into light-mode skins 2026-04-22 13:20:02 -07:00
slack_cli.py feat(slack): register every gateway command as a native slash (Discord/Telegram parity) (#16164) 2026-04-26 11:38:32 -07:00
status.py yuanbao platform (#16298) 2026-04-26 18:50:49 -07:00
timeouts.py refactor(timeouts): drop redundant ImportError in except clause 2026-04-26 20:48:20 -07:00
tips.py feat(busy): add 'steer' as a third display.busy_input_mode option (#16279) 2026-04-26 18:21:29 -07:00
tools_config.py feat(computer-use): cua-driver backend, universal any-model schema 2026-04-28 01:46:36 -07:00
uninstall.py feat(uninstall): offer to remove named profiles when uninstalling from default 2026-04-18 19:18:13 -07:00
voice.py fix(tui): ignore SIGPIPE so stderr back-pressure can't kill the gateway 2026-04-23 16:18:15 -07:00
web_server.py fix(tui): run built TUI with production React by default 2026-04-26 21:34:31 -05:00
webhook.py feat(webhook): direct delivery mode for zero-LLM push notifications (#12473) 2026-04-19 05:18:19 -07:00