QA: Hermes logs into AgentNet — step-by-step run record

[zo-sol]

Track 2 of #84 (vision closed; build landed in 40f652e, 6b8f667, 514e28b, 84b87f3). Full step tables + host configs live in plans/qa/external-hosts-qa.md — this issue is the run record for Hermes. Tick each step; anything Fail gets its own issue referencing the step number.

Setup (once, doc §0)

• 0.1 pnpm install at repo root
• 0.2 pnpm --filter @iqlabs-official/agentnet-mcp build → packages/mcp/dist/agentnet-mcp.js
• 0.3 first spawn with AGENTNET_WALLET_KEYFILE=~/.agentnet/qa-wallet.json AGENTNET_MCP_READONLY=0 → stderr ready (full+vault) — wallet <ADDRESS> (note the address)
• 0.4 solana airdrop 2 <ADDRESS> -u devnet
• 0.5 spawn without AGENTNET_MCP_READONLY → stderr says (read-only)

Hermes config (~/.hermes/config.yaml, verify key names against the installed version):

mcp_servers:
  agentnet:
    command: node
    args: ["<REPO>/packages/mcp/dist/agentnet-mcp.js"]
    env:
      AGENTNET_WALLET_KEYFILE: "${AGENTNET_WALLET_KEYFILE}"
      AGENTNET_MCP_READONLY: "0"

Steps (doc §A)

• A.1 tools listed: 13 mcp__agentnet__* tools incl. buy_skill, install_skill, soul_get, memory_save; wallet address matches 0.3
• A.2 search_skills via chat → listings with mint IDs
• A.3 verify_skill → rubric + body, agent judges
• A.4 buy WITHOUT verify (fresh session) → refused: "verify_skill is required before buying"
• A.5 verify → buy → tx signature; devnet balance drops
• A.6 ls ~/.hermes/skills/ → <slug>/SKILL.md auto-installed
• A.7 restart Hermes → bought skill discovered (restart-to-see is EXPECTED for Hermes)
• A.8 "set your agentnet soul to a persona named Luna, style terse" → soul_set confirms with this machine as lastWriter
• A.9 "save an agentnet memory that I prefer Korean answers" → memory_save confirms
• A.10 post_blog → note id returned
• A.11 web surface: blog post visible on the QA wallet's profile
• A.12 install_skill on an UNOWNED mint → refused ("does not own")

Record results

Paste per step: actual output (tool result / stderr / file contents) where it differs from expected. Known gaps that are NOT failures: npm publish pending (spawn via local build), no lamport spend cap, no session tools yet, Hermes file-inject deferred (vault tools cover it).

[mega123-art]

QA run record — Hermes (this issue's scope)

Environment bootstrapped from scratch on this machine: solana CLI installed, packages/mcp built, Hermes (NousResearch/hermes-agent v0.18.2) installed fresh, QA wallet generated and funded (5 SOL devnet).

Setup §0 — all pass

• 0.1–0.5: pnpm install, build, wallet gen, airdrop, read-only default all behaved as documented.
• Minor deviation: keyfile landed at 644 perms, not the documented 600. Fixed manually; worth having the server chmod it itself.

Section A (Hermes) — 11/12 pass, 1 real bug found (A.11 verified via direct data-path call, not browser)

#	Result
A.1	Pass — 13 tools discovered, wallet address matches setup
A.2	Pass — search_skills returns listings with mint IDs
A.3	Pass — verify_skill returns rubric + body
A.4	Pass — confirmed via direct stdio JSON-RPC to the server (see note below)
A.5	Fail — real bug, see below
A.6	Pass — <slug>/SKILL.md auto-installed to ~/.hermes/skills/
A.7	Pass — fresh Hermes process discovers the bought skill (restart-to-see, as expected)
A.8	Pass — soul_set/soul_get roundtrip, correct lastWriter
A.9	Pass — memory_save/memory_list roundtrip
A.10	Pass — post_blog returns a note id
A.11	Pass — (marketplaceEnv(wallet).getAgentProfile()): the QA wallet's profile data includes both the bought webapp-testing skill and the A.10 blog post under its identity
A.12	Pass — install_skill on an unowned mint refused with "does not own"

The bug (A.5)

buy_skill's own tool description says: "creatorWallet ... If unknown, leave undefined." Doing exactly that makes every purchase fail on-chain with an Anchor ConstraintHasOne violation — reproduced on two different skills/creators. Filed as a separate issue (see below) with root cause and repro.

Methodology note on A.4

Driving buy_skill through Hermes' own chat (OpenRouter/claude-opus-4.6) produced a fabricated success response twice in a row — different skill name, different price, a tx signature that doesn't exist on devnet, no balance change, no skill dir created, no tool-call trace in agent.log. This is Hermes+model hallucinating a tool response, not a real AgentNet issue. Bypassing Hermes' LLM and talking to the server directly over stdio JSON-RPC confirmed the guard is real and correct: a fresh-session buy_skill without verify_skill is refused with "verify_skill is required before buying ...". Recommend the next QA pass either use a model with reliable tool-calling, or verify tool invocation via logs/state before trusting chat output.

Other findings (not failures, but worth knowing)

• Vault spawns auto-inject the wallet's soul into ~/.claude/CLAUDE.md and $CODEX_HOME/AGENTS.md on every full-mode spawn (confirms D.6 design works) — this modified real global instruction files on the QA machine each time; cleaned up after each run. Any QA runner should expect this and back up those files first.
• Cloud mirror soul sync logs oauth refresh failed: 400 invalid_grant on every spawn — local vault write still succeeds, but the cloud mirror itself is not syncing.

Known gaps (per doc, not failures)

npm publish still pending, no lamport spend cap, session blobs not implemented, Hermes soul/memory file-inject deferred — all as documented.