Trace Skill Provenance

Use this when a user asks questions like:

• “这个 skill 是哪来的？”
• “是不是 builtin？”
• “什么时候创建的？为什么我没印象？”
• “这是导入的、提升的，还是你自己后来建的？”

Goal

Determine, with evidence:

1. whether a skill is

   builtin

   or

   local

2. its on-disk path
3. approximate creation time
4. whether it came from promotion/import (

   imported-codex

   / manifest) or was created later via

   skill_manage

5. the session that created or modified it, if available

Procedure

1. Confirm how Hermes classifies it

Run:

hermes skills list | grep -n '<skill-name>'

Interpretation:

• builtin | builtin

  => official built-in skill

• local | local

  => local skill in the runtime environment

Do not rely on memory alone.

2. Find the actual filesystem path

Check likely skill roots directly:

• ~/.hermes/skills

• ~/.codex/skills

  (older/shared runtime copies may still exist)

Example shell approach:

for d in ~/.hermes/skills ~/.codex/skills; do
  [ -d "$d" ] || continue
  find "$d" -type d -name '<skill-name>'
done

If needed, inspect

SKILL.md

under the found path.

3. Get file timestamps

For the skill directory and

SKILL.md

, inspect:

• birth time (

  st_birthtime

  on macOS when available)
• mtime
• ctime

Example Python snippet:

from pathlib import Path
from datetime import datetime
p = Path('~/.hermes/skills/<category>/<skill-name>/SKILL.md').expanduser()
st = p.stat()
print(datetime.fromtimestamp(getattr(st, 'st_birthtime', st.st_ctime)))
print(datetime.fromtimestamp(st.st_mtime))
print(datetime.fromtimestamp(st.st_ctime))

Use this only as approximate creation evidence; later edits can change mtime/ctime.

4. Check promotion/import manifests

Look for promotion manifests such as:

~/.hermes/skill-promotions/<timestamp>/promotion-manifest.json

Read them and see whether the skill appears in

entries

.

Interpretation:

• Present in manifest => likely promoted from

  imported-codex

  or another source
• Absent from manifest => likely not part of that promotion batch

Important: verify against the specific manifest timestamp instead of assuming all local skills came from promotion.

5. Search session history for creation evidence

Search Hermes session files and/or use session search for:

• the skill name

• skill_manage

• action":"create"

• action":"patch"

Useful targets:

• ~/.hermes/sessions/*.json

• session_search(query='"<skill-name>"')

High-signal evidence is a session snippet containing:

• skill_manage

• action: create

• name: <skill-name>

If found, report:

• session id
• session start time
• exact lines or file reference

6. Distinguish creation modes clearly

Use the evidence to classify one of these outcomes:

• Built-in: listed as builtin; no local creation needed
• Promoted/imported: appears in promotion manifest and local path matches promoted category
• Locally created later: not in manifest, but session logs show

  skill_manage(action='create', ...)

• Locally patched later: original source may differ, but session logs show subsequent

  patch

  /

  edit

Recommended evidence order in final answer

1. classification from

   hermes skills list

2. on-disk path
3. file timestamp
4. promotion-manifest result
5. session-log evidence of

   skill_manage

This makes the conclusion auditable and easy to trust.

Pitfalls

• hermes skills inspect <name>

  may fail even when

  skill_view(name)

  works; do not stop there.
• Searching only

  ~/.hermes/skills

  can miss older copies in

  ~/.codex/skills

  .
• mtime alone is not “creation time”; prefer birth time when available and corroborate with session logs.
• A skill can be

  local

  without being imported; do not equate

  local

  with

  promoted

  .
• Promotion manifests prove inclusion in a batch, but absence there does not by itself prove manual creation; verify with session evidence.

Done criteria

You should be able to answer, with citations/evidence:

• what kind of skill it is (

  builtin

  vs

  local

  )
• where it lives
• when it was likely created
• whether it was promoted/imported or created by a later agent action
• why the user may not remember it (for example, an agent auto-saved it after a complex task)