[OutThisLife]

What

Skills mounted from skills.external_dirs are externally authored, but every provenance surface flattened them into agent — the usage report, the dashboard's /api/skills set-based twin, the desktop catalog/detail UI, and the learning journey graph (where a used external mount rendered as a learned skill milestone).

This adds an external provenance tier at every site, without touching mutability.

Addresses #108032

Changes

• tools/skill_usage.py — provenance() gains an external tier: an external-only skill (present under skills.external_dirs, absent from the local store — the local copy keeps winning on name clashes, mirroring is_agent_created's existing external exclusion) classifies as external. New _external_skill_names() scans external dirs once for the router's set-based twin.
• hermes_cli/web_routers/skills.py — /api/skills classifies hub > bundled > external > agent.
• agent/learning_graph.py — skills whose path resolves under an external dir are tagged source='external' (this is the shape the graph actually sees: the common symlinked SKILL.md into ~/.hermes/skills/, which the profile-root scan yields) and _has_learning_signal excludes them: a mount is configured, not learned, even when used.
• Desktop — SkillInfo.provenance gains 'external'; the catalog labels mounts external (not local), and the detail pane shows an external skill as externally mounted while keeping it editable in place. The two provenance decisions are extracted into pure helpers (skill-provenance.ts) so they're testable directly.
• hermes_cli/curator.py — curator usage reports and filters the external origin.

What this deliberately does NOT do

The issue also asks for skills.external_dirs to be read-only by default for foreground skill management. That half is not included because maintainers rejected it: commit 8c8fc6c1ec (PR #17512, which settled the then-open #4759 and #4381) removed exactly that default read-only gate — the refusal made agents create silent local duplicates, and the standing regression test at tests/tools/test_skill_manager_tool.py:784+ pins in-place external mutation succeeding. Re-imposing default-deny would resurrect a closed direction; the opt-in variant (external_dirs_readonly) is a separate live design question (see open PR #51412). This PR is the origin-labeling half only: the label says where a skill came from, and edit rights stay keyed to the standing April decision.

Prior art

Two earlier PRs addressed this issue and are both gone — #108040 (deleted) and #108767 (closed 2026-09-12 as its duplicate) — so this is a fresh implementation, not a salvage; no code was reused from either.

Testing

• tests/tools/test_skill_provenance_external.py (new, 8 tests): tier classification (external-only → external, local stays agent, local copy wins, is_agent_created exclusion, names helper), the router twin through a real TestClient against the mounted app, and the journey graph (a used external skill produces no learning node via the symlinked-into-profile shape; a used local skill still does). 4 of 8 fail on origin/main sources, all pass with the change.
• apps/desktop/src/app/capabilities/skills/skill-provenance.test.ts (new, 6 vitest cases): both helpers, including that bundled/hub stay non-editable and absent provenance keeps its legacy meaning.
• Sibling suites green: test_skill_provenance.py, test_skill_usage.py, test_learning_graph.py, test_web_server_skills_profiles.py (38 tests).

[kiramakes]

🤖 Hermes Agent automated review: PR reviewed. Diff analyzed (424 lines changed). Check CI status and manual review recommended.

[whyyagswhy]

Independent verification on the PR head (2785b50): desktop provenance suite 6/6 green. Python provenance suite 7/8: test_used_local_skill_still_produces_learning_node fails under scripts/run_tests.sh but passes under direct pytest. Root cause is test isolation, not product code: the runner execs pytest under env -i, and the test only sees its fixture skill when ambient environment leaks in. If CI runs through run_tests.sh this test fails there too. Worth making the test hermetic before merge.

[arkheioncorp]

Verdict: APPROVE

Clean implementation of the 'external' provenance tier. Key points:

• Priority chain in provenance() is correct: hub > bundled > external > agent — local copy wins on name clashes via is_external_only()
• External skills correctly excluded from learning graph (_has_learning_signal), while keeping in-place edit rights (isEditableProvenance)
• Defensive import in learning_graph.py (try/except with fallback) is appropriate
• Set-based twin in web router matches the per-skill provenance() semantics
• Tests cover classification, learning graph exclusion, router twin, and the local-copies-external regression

[github-actions]

૮ >ﻌ< ა ci review

_{ran on 76dea6e — fix(desktop): fix lint in skill provenance files (hermes-age}

debug info

CI timings

CI timings · View report · View job

Wall time 30m14s (no baseline yet).

[OutThisLife]

@whyyagswhy investigated on both the head you reviewed (2785b50) and the current one — reproduction says the test is not env-dependent, and the suite is green under the scrubbed runner on both:

• scripts/run_tests.sh -j 1 tests/tools/test_skill_provenance_external.py → 8/8 green on 2785b50 and on the current head. The test was already hermetic: conftest's autouse _hermetic_environment fixture sandboxes HERMES_HOME (and every behavioral HERMES_* var) per test, and the fixture additionally sets its own HERMES_HOME + clears the external-dirs cache.
• What you hit was the 'external' source loop-variable leak in build_skill_nodes, fixed in 71485f5 on this branch: at the reviewed head source = "external" overwrote the loop variable, so any skill yielded after the mount in one rglob scan inherited "external" and was dropped from the graph. Scan order depends on the tmpdir's on-disk layout, and env -i changes TMPDIR — so the failure correlated with the runner without being caused by the env.

To keep that order-dependence from ever hiding again, de01606 pins the hostile order directly: the new test_external_source_does_not_leak_to_later_skills_in_a_root patches Path.rglob so the external mount always scans before the local skill and asserts the local skill still classifies 'agent' — red at 2785b50 (my-local dropped entirely), green on the current head. Also removed the fixture's no-op monkeypatch.setattr("hermes_constants._hermes_home_cache", ...) — that attribute doesn't exist, so it reset nothing.

scripts/run_tests.sh -j 1 on the file: 9/9 green. New head: de01606.