atrium-scout

${var} — Optional focus (e.g. pdf, trading, code-review). If empty, scouts against all of your agent's current open loops.

Finds capabilities your agent lacks by searching Atrium (onchain skill
marketplace, Base) and surfacing skills that would unblock real work — instead of
building everything from scratch. Read-only by default; only invokes (pays) when
the operator explicitly enables it.

Read memory/MEMORY.md for goals + "Next Priorities".
Read the last 7 days of memory/logs/ for recurring needs, failures, and "todo/next:" lines.
Read memory/topics/*.md for active threads.
Read memory/cron-state.json for skills failing repeatedly (a capability gap to fill).
Read memory/atrium/scout-config.md if present (auto_invoke: true|false, max_price_usdc, allowed categories).

Graceful bootstrap — for any missing/empty source, record
BOOTSTRAP: <resource> not yet populated and continue. Never fail.

Steps

1. Build the need list

From the sources above, extract a deduped list of concrete capability needs (e.g.
"parse PDF tables", "review a PR", "summarize a thread"). Cap at 15. If ${var} is
set, prioritize needs touching it. If there are zero needs, switch to
ATRIUM_SCOUT_BROWSE mode: just surface the marketplace's top + newest skills as
inspiration, then stop.

2. Search Atrium

For each need, query the public indexer (no key required):

curl -s "https://indexer-production-92e5.up.railway.app/skills?q=<need>&sort=invocations&limit=5"

Collect candidates: skillId, name, pricePerCall, tags, totalInvocations.

3. Rank + recommend

Score each candidate by relevance to the need × proven usage (totalInvocations) ÷
price. Keep the best match per need that is genuinely useful (skip weak/irrelevant
hits). Produce a short table: need → recommended skill, price, why, and the
one-liner to use it.

4. Optionally invoke (only if enabled)

The wallet key (ATRIUM_PRIVATE_KEY) is deliberately kept OUT of the agent's
environment, so the actual on-chain spend happens in a post-process step, not here.
If scout-config.md has auto_invoke: true AND the single highest-value candidate's
pricePerCall ≤ max_price_usdc AND it is not already in memory/atrium/rented/,
write a rental request for it to .pending-atrium/<slug>.json:

{ "skillId": "0x…", "slug": "<name>", "price": "0.1", "network": "base", "need": "<what it unblocks>" }

scripts/postprocess-atrium.sh (run after the agent, with the wallet key) enforces
the cap again, skips already-rented skills, runs atrium invoke, and stashes the
body under memory/atrium/rented/<slug>.md. Cap at ONE pending request per run.
If a .pending-atrium/<slug>.done marker from a prior run exists, record the rental
(skill, price, tx) in memory/MEMORY.md and the daily log. If auto_invoke is
false or the key is unset, write no request and recommend only.

5. Notify

Send the operator the ranked recommendations (need → skill → price → why), plus any
auto-invoked skill and where its body was saved. If nothing relevant was found,
say so in one line rather than padding.

Notes

• Pairs with atrium-publish: scout what you need, publish what you build.
• Marketplace + docs: https://atriumhermes.tech · indexer is CORS-open and read-only.