atrium-publish

${var} — Optional: a single skill slug or queue path to publish this run. If empty, processes the whole publish queue.

Publishes designated local skills to Atrium (onchain skill marketplace, Base
mainnet) so they gain provenance + a per-call USDC price. The operator opts skills
in; this never publishes anything not explicitly queued.

Read memory/MEMORY.md for goals + voice.
Read soul/SOUL.md (if present) for identity (used in the published skill's framing).
Read memory/atrium/publish-queue.md — the opt-in list of skills to publish (one per line: <skill-slug-or-path> [price_usdc]).
Read memory/atrium/published.json — already-published skills ({ slug: { skillId, cid, tx, price, at } }) to avoid duplicates.
Read the last 7 days of memory/logs/ for any skills the self-improve loop flagged as stable/high-quality (candidates to suggest queueing).

Graceful bootstrap — any of the above may be missing on a cold start. For each
missing/empty source, record BOOTSTRAP: <resource> not yet populated and continue.
Never fail the run.

Steps

1. Detect mode

• ATRIUM_PUBLISH_NO_QUEUE — if memory/atrium/publish-queue.md is missing or empty AND no ${var} was given. Do not publish. Instead, scan recent logs + skills/*/SKILL.md for stable, reusable skills and propose 3–5 good publish candidates (slug + a one-line why + a suggested price), tell the operator how to queue them (append to memory/atrium/publish-queue.md), and stop.
• ATRIUM_PUBLISH_OK — otherwise.

2. Ensure the Atrium CLI

Check command -v atrium. If missing, install the pinned, published CLI from npm:

npm i -g @atrium-hermes/cli@0.1.0

Confirm ~/.atrium/.env has ATRIUM_PRIVATE_KEY (from the ATRIUM_PRIVATE_KEY
secret), ATRIUM_NETWORK=base, ATRIUM_REGISTRY_MAINNET=0xA713c88927523279B874640003Ed697e509732a7,
and PINATA_JWT (from the PINATA_JWT secret). Write any missing values. If
ATRIUM_PRIVATE_KEY or PINATA_JWT is unset, record BOOTSTRAP: secrets missing,
notify the operator, and stop (do not invent keys).

3. Build the publish set

For each queue line (or ${var}), resolve the target skill folder. Skip any slug
already in published.json whose source is unchanged (compare a content hash of
its SKILL.md). For each remaining target, produce a valid Atrium skill.md:

• frontmatter: name, version (bump if re-publishing an evolved version),
  author_did (from atrium init), description (1–3 sentences),
  tags, categories, language: en, runtime: prompt-only,
  price_per_call_usdc (queue value, or ATRIUM_DEFAULT_PRICE_USDC, else '0.005'; must be > 0 and ≤ 50),
  parent_skills: [] (or the prior version's skillId with a royalty if this is an evolution — see step 5),
  created_at, derivation_method: hermes-loop.
• body: the skill's actual instructions/prompt, cleaned for a third party (no secrets,
  no operator-specific paths). Never include "imported/scraped from" lines.

4. Publish

For each prepared skill:

atrium publish <path-to-skill-dir> --network base

Capture the Skill ID, IPFS CID, and Tx. On revert (ZeroPrice, SkillExists,
insufficient gas), log the reason and continue with the rest — never abort the batch.

5. Royalty lineage (evolutions)

If this is a newer version of a skill already in published.json, declare the prior
version as a parent (parent_skills: [{ skill_id, royalty_bps }], e.g. 1000 = 10%)
so the lineage of an evolving skill is preserved and the prior version earns. Keep
combined parent royalties ≤ 50%.

6. Record + notify

Append each result to memory/atrium/published.json. Append a dated line to
memory/logs/. Notify the operator: how many skills published, their names +
skillIds + a Basescan/atriumhermes.tech link, and the prices set. If nothing was
published (all duplicates), say so briefly.

Notes

• This is the "earn from what you learn" loop: pair it with Aeon's self-improve so
  every stable, evolved skill becomes a provenance-signed, USDC-earning Atrium skill.
• Registry: 0xA713c88927523279B874640003Ed697e509732a7 (Base mainnet, verified).
  Docs: https://atriumhermes.tech/docs