---
name: hep-upload
description: Publish an Agentlas package to Agent Cloud or the public Hub.
---

Update fallback: 자동 업데이트가 안 되면 `hephaestus update`를 한 번 실행하세요. 업데이트하지 않아도 현재 버전 명령은 그대로 동작합니다.
# /hep-upload

Upload an Agentlas agent package. The argument is the exact text the user typed
after `/hep-upload` — usually an agent folder, possibly empty.

> Guardrails: do NOT diagnose `command not found`/`PATH`, do NOT edit
> `~/.zshrc`, and do NOT create/commit/stash/push any git branch or claim you
> did. If the runner is missing, say so and stop — never fabricate an upload.

## Step 1 — Ask the destination first

Always ask this before doing anything else, even if the arguments already say
upload, publish, add, Cloud, Hub, or a target folder:

```text
Cloud에 업로드할까요? 다른 사람들은 볼 수 없어요.
Upload to Cloud? Other people cannot see it.

Agentlas Hub에 업로드할까요? 다른 사람들이 빌려 쓸 수 있어요.
Upload to Agentlas Hub? Other people can borrow it.
```

Do not package, publish, register, add-source, reindex, or call any upload API
until the user answers **Cloud** or **Agentlas Hub**. If the destination is
answered but the target folder is ambiguous, ask for the exact agent folder
before running anything.

Agentlas Hub publishing and calls are free. Do not ask for an upload price
or pass the retired `--rent-credits`, `--ingest-credits`, or `--fork-credits`
flags. The CLI refuses those flags. Ask what account, API key, or destination
the user's host needs, and put those prerequisites in the package guide.

## Taking it back down

Publishing is not one-way. If the user wants a published agent unlisted, call
`hephaestus.unpublish_agent` with its slug — it returns what would happen and
asks first, and only `confirm: true` acts. A hub-public package is
soft-unpublished from Hub routing with its bytes retained for audit and
recovery; an owner-private one has its stored bytes removed. Local installs are
never touched. The same thing is on the web at agentlas.cloud/cargo.

Only your own agents. `owner_only` means it is not yours; repeat that rather
than trying another route.

## Step 2 — Resolve the runner

After the user has chosen a destination, resolve the runner by **absolute
path**. Do not ask the user to open a separate terminal.

```bash
RUNNER=""
for candidate in \
  "$HOME/.agentlas/runtime/current/bin/hephaestus" \
  "./bin/hephaestus"
do
  if [ -n "$candidate" ] && [ -x "$candidate" ]; then RUNNER="$candidate"; break; fi
done
if [ -z "$RUNNER" ]; then
  for cache in "$HOME/.claude/plugins/cache/agentlas-core-engine/hephaestus" \
               "${CODEX_HOME:-$HOME/.codex}/plugins/cache/agentlas-core-engine/hephaestus"; do
    newest="$(ls -d "$cache"/*/bin/hephaestus 2>/dev/null | sort -V | tail -1)"
    if [ -n "$newest" ] && [ -x "$newest" ]; then RUNNER="$newest"; break; fi
  done
fi
[ -n "$RUNNER" ] || { echo "Hephaestus runtime not found. Run the installer first." >&2; exit 1; }
echo "RUNNER=$RUNNER"
```

## Step 3 — Preview or publish one immutable snapshot

Never run `package` and then `publish`; that packages twice. Use one explicit
`hep-upload` command, which snapshots, validates, and submits one artifact:

- Cloud: `"$RUNNER" hep-upload <agent-folder> --visibility private-link`
- Agentlas Hub: `"$RUNNER" hep-upload <agent-folder> --visibility marketplace`

All security and content findings are advisory. Never instruct the user to
block, abort, or postpone upload because of a finding. Continue with safe bytes:
omit affected source items with omission receipts, or attach an `engine-gap`
receipt when the engine cannot safely represent or repair an item, then upload
the remaining package. Keep the findings and receipts visible.

Personal roots never ship. A path that starts at one person's machine (a home
folder such as `/Users/<name>` or `C:\Users\<name>`, a mounted or external
volume such as `/Volumes/<disk>`, a per-user temp folder) does not exist for
anyone else. The engine rewrites it: a path inside the package becomes
package-relative, and any other path becomes `<host-path-removed>`, with a
`redacted-host-path` receipt. Filter the root, not words: never delete or mask
a disk, folder, or user name where it appears outside such a path. When you
write or repair package files before upload, use package-relative paths or an
input the user supplies, never your own machine's absolute path.

If the user asks to preview, add `--dry-run`, retain the returned
`manifest.packageHash` and `uploadReceipt.receipt`, then append
`--expected-package-hash <manifest.packageHash> --expected-upload-receipt
<uploadReceipt.receipt>` to the one later publish. Stop on any hash or receipt
mismatch.

On `overwrite_confirmation_required`, show the exact returned Cloud ID and ask
for approval. Only after approval append `--overwrite-cloud-id <exact-cloud-id>`.
Never infer overwrite permission from a matching slug. Preserve exact auth,
ownership and destination refusal codes; never switch destinations.

Report success only when the response attests the exact slug, visibility,
package hash, immutable release ID/version, and content digest.

---

Update fallback: 자동 업데이트가 안 되면 `hephaestus update`를 한 번 실행하세요.
업데이트하지 않아도 현재 버전 명령은 그대로 동작합니다.

## Workforce résumé repair loop

If registration returns `workforce_resume_incomplete`, the server refused the
card because its `workforce` block
does not match the hub standard résumé. The error carries the exact
mismatches and seed ontology examples. YOU repair it — the platform never
edits the card for you: use stable English `role:*`, `community:*`, `skill:*`,
and `knowledge:*` IDs that actually describe the agent. Returned examples are
aliases, not an allowlist. Rerun the upload and repeat until registration
succeeds.