fix(cli): preload skills in oneshot mode

[Vicenteut]

Summary

• forward -s / --skills through every one-shot CLI dispatch path
• load requested skills into the one-shot agent before its only model call
• preserve existing ephemeral prompt content and match chat behavior for missing skills

Scope

• no gateway, provider, model, or container changes
• no migrations
• no deployment changes

Verification

• pytest -q -p no:cacheprovider tests/hermes_cli/test_oneshot_skills.py tests/hermes_cli/test_tui_resume_flow.py tests/hermes_cli/test_argparse_flag_propagation.py tests/cli/test_cli_preloaded_skills.py — 19 passed
• ruff check --no-cache hermes_cli/main.py hermes_cli/oneshot.py tests/hermes_cli/test_oneshot_skills.py — passed
• offline integration against a real nested skill: loaded theclamai-content-intelligence, injected 7,721 prompt characters, no network or model call

[Enough1122]

AI code review — automated review for reference, author can ignore or act on any point.

fix(cli): preload skills in oneshot mode

The plumbing through _run_and_exit_oneshot and the _apply_preloaded_skills helper are straightforward, and the tests cover the forwarding and unknown-skill paths. Observations:

1. hermes_cli/oneshot.py::_apply_preloaded_skills: agent.preloaded_skills is only set when skills_prompt is truthy. If build_preloaded_skills_prompt ever returns a non-empty loaded_skills list with an empty prompt, the agent records no loaded skills even though the load happened. Keying the bookkeeping off loaded_skills (and only the prompt append off skills_prompt) would keep the two consistent.

2. The behavior split — warn-and-continue when some skills are unknown, raise ValueError when all are unknown — is reasonable, but it means a typo in a single-skill one-shot invocation fails hard while the same typo next to a valid skill is silently forgiven. For a scripted/one-shot context, a consistent warn-and-continue (matching cli.py's tolerance) may be less surprising.

3. _normalize_toolsets is reused to parse the skills value. If it applies toolset-specific normalization or validation, unknown-skill input may surface toolset-flavored messages before reaching the clearer build_preloaded_skills_prompt messaging — worth verifying the two vocabularies don't collide.

No blocking issues.

[teknium1]

Fixed on main via PR #90389 (salvage of #31591, the earliest submission for this bug, May 24). Thanks for the contribution!