fix(tui): dispatch custom skill bundles as agent turns

[Adolanium]

What does this PR do?

Fixes custom skill-bundle slash commands hanging under hermes --tui.

The normal CLI recognizes bundles in cli.py:8988, expands them, prints the loading banner, and places the expanded agent message on _pending_input at cli.py:9005. The normal CLI process loop consumes that queue at cli.py:15170.

The TUI followed a different path. slash.exec intercepted individual skill commands and a fixed set of other _pending_input commands, but it did not recognize custom bundles. The bundle therefore reached the persistent slash worker. tui_gateway/slash_worker.py:101-133 calls HermesCLI.process_command() and returns only captured stdout. It never runs the normal CLI process loop, so the loading banner came back while the expanded agent message stayed on the worker's private queue.

The fix detects any installed custom bundle in slash.exec and routes it directly through command.dispatch. That handler now expands bundles through the shared agent.skill_bundles implementation and returns the existing structured send response with both the expanded message and a loading notice. The TUI already handles that response by displaying notice and submitting message as an agent turn.

Dispatch precedence stays aligned with the CLI and gateway paths: quick commands and plugins are checked first, bundles beat individual skills, and built-in commands continue through their existing routes. The shared builder receives the resolved TUI or desktop platform so global and platform-disabled skills keep using the existing filtering behavior.

Related Issue

Fixes #62863

Type of Change

• 🐛 Bug fix (non-breaking change that fixes an issue)
• ✨ New feature (non-breaking change that adds functionality)
• 🔒 Security fix
• 📝 Documentation update
• ✅ Tests (adding or improving test coverage)
• ♻️ Refactor (no behavior change)
• 🎯 New skill (bundled or hub)

Changes Made

• tui_gateway/server.py
  • Recognize generic custom bundles in slash.exec before the slash worker.
  • Expand bundles in command.dispatch through build_bundle_invocation_message.
  • Return {type: "send", message, notice} so the TUI starts a normal agent turn.
  • Preserve bundle-over-skill precedence and pass the resolved surface to disabled-skill filtering.
  • Include skipped missing skills in the loading notice.
• tests/tui_gateway/test_protocol.py
  • Prove slash.exec routes a generically named bundle away from the worker.
  • Prove command.dispatch preserves the user arguments, returns a sendable payload, and checks the bundle before the individual-skill builder.
  • Verify the platform is passed into the shared bundle builder.

How to Test

1. Focused gateway regression and existing individual-skill dispatch checks:

   scripts/run_tests.sh tests/tui_gateway/test_protocol.py -k 'slash_exec_rejects_skill_commands or command_dispatch_returns_skill_payload or custom_skill_bundle or custom_bundle_payload' -q
   

   Result: 4 passed, 0 failed.
2. Complete relevant protocol file:

   scripts/run_tests.sh tests/tui_gateway/test_protocol.py -q
   

   Result: 77 passed, 0 failed.
3. Shared bundle behavior:

   scripts/run_tests.sh tests/agent/test_skill_bundles.py -q
   

   Result: 35 passed, 0 failed.
4. Existing gateway bundle dispatch tests:

   python -m pytest tests/gateway/test_bundles_command.py -q
   

   Result: 4 passed, 0 failed. The canonical wrapper could not collect this file on native Windows because its clean environment converted HOME to a POSIX path before launching Windows Python, so this adjacent file used the project venv directly.
5. Direct integration check against an isolated real HERMES_HOME containing a generic integration-pack YAML and one real local skill. Both slash.exec and command.dispatch returned type=send, a bundle loading notice, the bundle header, and User instruction: compare self-hosted vector databases. The worker stub was configured to fail if called, and was not called.
6. Lint:

   ruff check tui_gateway/server.py tests/tui_gateway/test_protocol.py
   

   Result: all checks passed.

Checklist

Code

• I've read the Contributing Guide
• My commit messages follow Conventional Commits (fix(scope):, feat(scope):, etc.)
• I searched for existing PRs to make sure this isn't a duplicate
• My PR contains only changes related to this fix/feature (no unrelated commits)
• I've run pytest tests/ -q and all tests pass
• I've added tests for my changes (required for bug fixes, strongly encouraged for features)
• I've tested on my platform: Windows 11, Python 3.11.15

Documentation & Housekeeping

• I've updated relevant documentation (README, docs/, docstrings) or N/A
• I've updated cli-config.yaml.example if I added/changed config keys or N/A
• I've updated CONTRIBUTING.md or AGENTS.md if I changed architecture or workflows or N/A
• I've considered cross-platform impact (Windows, macOS) or N/A
• I've updated tool descriptions/schemas if I changed tool behavior or N/A

Screenshots / Logs

Before the fix, the real slash-worker seam returned only:

⚡ Loading bundle: tui-bundle-probe (2 skills)

The worker's private _pending_input queue contained one unconsumed expanded message.

After the fix, both RPC routes return a structured agent turn:

slash.exec: type=send, bundle_header=True, instruction=True
command.dispatch: type=send, bundle_header=True, instruction=True

Remaining limitation: the TUI reports an explicit bundle-load error when every referenced skill is missing or disabled. It does not start an empty agent turn.

[teknium1]

Thanks for tracing the TUI/worker seam; the reported hang is present on current main. tui_gateway/server.py:13090-13140 has no bundle interception before the slash worker, and tui_gateway/slash_worker.py:80-112 only captures HermesCLI.process_command() output while CLI bundle handling queues the actual prompt (cli.py:8980-8997).

Problems

• The new bundle lookup at tui_gateway/server.py:11876 precedes the built-in pending-input handlers, including /queue at tui_gateway/server.py:11937. Bundle parsing permits arbitrary slugs (agent/skill_bundles.py:136-165), so a custom bundle named queue changes /queue <prompt> into a bundle invocation. slash.exec routes /queue directly into this dispatcher (tui_gateway/server.py:11603-11616).

Suggested changes

• Apply the same resolve_command(name) is None built-in guard used by the new slash.exec route at tui_gateway/server.py:13148-13151 before resolving a bundle in command.dispatch.
• Add a collision regression for a queue bundle and retain the existing normal /queue send-payload contract (tests/tui_gateway/test_protocol.py:1404-1418).

Automated hermes-sweeper review.

[teknium1]

This runs before the built-in pending-input handlers below. Because bundle slugs are unrestricted, a custom /queue bundle changes /queue <prompt> into a bundle invocation after slash.exec routes it here. Gate this on resolve_command(name) is None, as the slash.exec bundle route does, and add a collision regression.

[Adolanium]

Addressed the requested precedence issue in follow-up commit 300c18087.

• command.dispatch now applies the same resolve_command(name) is None guard as slash.exec before resolving a bundle.
• A custom bundle named queue can no longer shadow the built-in /queue command.
• Added a regression test proving /queue still returns the existing {type: "send", message: ...} payload and the bundle builder is not called.

Verification:

• Focused TUI protocol tests: 6 passed.
• Complete tests/tui_gateway/test_protocol.py with the project venv: 78 passed.
• Ruff and git diff --check: passed.