fix(skills): resolve SKILLS_DIR at call time, not at module load

[Ahmett101]

Summary

tools/skills_tool._find_all_skills() was resolving [SKILLS_DIR](file:///home/pstk/.hermes/profiles/merlin/skills) once at module import via [SKILLS_DIR = HERMES_HOME / "skills"](file:///home/pstk/.hermes/profiles/merlin/skills). A long-running gateway process (systemd unit, multi-profile setup) reads HERMES_HOME once at boot — when the env drift's between launch and tool calls later, the function silently returns a stale skill index that diverges from a fresh Python subprocess (80 vs 88 for a profile with 88 on-disk skills, spanning 8 entire categories: meta, hermes-config, content-creation, scheduled, thinking, tools, transfer, video).

get_skill_commands() (which the report confirms returns all 88) reads HERMES_HOME / "skills" per call. Converge to the same pattern in _find_all_skills() so the count from the gateway matches the count from any fresh subprocess.

The fix preserves backward compatibility for tests / callers that patch the module-level SKILLS_DIR: when patched, that override is honoured. When not patched (production runtime), the function re-reads HERMES_HOME via get_hermes_home() and rescans the current home root.

Changes

• [tools/skills_tool.py](file:///data/data/com.termux/files/home/.hermes/hermes-agent/tools/skills_tool.py) — _find_all_skills() resolves SKILLS_DIR per call via get_hermes_home(); falls back to the module-level SKILLS_DIR only when tests have patched it.
• [tests/tools/test_skills_tool.py](file:///data/data/com.termux/files/home/.hermes/hermes-agent/tests/tools/test_skills_tool.py) — two regression pins:
  • test_find_all_skills_source_re_resolves_skills_dir pins the live-resolve behaviour by source inspection (a future refactor that goes back to module-level-only read silently regresses this fix without breaking any other test).
  • test_find_all_skills_results_use_call_time_skills_dir exercises the end-to-end behaviour: swapping the module-level SKILLS_DIR between two isolated home roots toggles which roots are scanned.

How to Test

venv/bin/python -m pytest tests/tools/test_skills_tool.py -q
# 90/91 passed (1 pre-existing inline-shell test fails on this host because
# /bin/bash isn't executable under our sandboxed runner; pre-existing on
# main and unrelated to this PR)

Risk & Impact

Low. The change only adds a per-call get_hermes_home() lookup, which is already the path get_skill_commands() and most other adapter entry-points take. The path is a single stat for a non-existent or shallow directory — negligible cost relative to the existing disk walk over iter_skill_index_files(...) for every skill subdirectory. No public API change.

The production scenarios where this matters are bounded:

• Single-profile single-launch CLI invocations: HERMES_HOME is constant from start to exit — no behavioural change.
• Multi-profile gateways with stable HERMES_HOME per profile across the process lifetime — same.
• Long-running gateway whose gateway manager reloads an associated profile mid-process (the reported case) — _find_all_skills() would now match the fresh-subprocess count on the next call.

Type: Bug fix
Closes: #58908

[teknium1]

Closing as duplicate: the same call-time skills-dir resolution fix was submitted earlier in #56689 by @jplew (Jul 1) and has now been merged via PR #60180 (#60180), which covers skills_list, skill_view, and skill_manage plus the TUI slash-worker subprocess env. Your diagnosis of the module-level SKILLS_DIR pin was spot on — thanks for the PR and the regression tests, sorry it collided with an earlier submission.