QA: Eliza logs into AgentNet — step-by-step run record + cross-host wrap-up

[zo-sol]

Track 2 of #84 (build landed in 40f652e, 6b8f667, 514e28b, 84b87f3). Full step tables + configs in plans/qa/external-hosts-qa.md — this issue is the run record for Eliza, plus the cross-host wrap-up (§D) since Eliza is the last of the three. Same keyfile as the Hermes/OpenClaw runs.

Character config (verify the MCP plugin package name at run time; AGENTNET_ELIZA_CHARACTER makes the soul render into this very file on spawn):

{
  "name": "agentnet-qa",
  "plugins": ["<eliza-mcp-plugin>"],
  "settings": {
    "mcp": {
      "servers": {
        "agentnet": {
          "type": "stdio",
          "command": "node",
          "args": ["<REPO>/packages/mcp/dist/agentnet-mcp.js"],
          "env": {
            "AGENTNET_WALLET_KEYFILE": "<HOME>/.agentnet/qa-wallet.json",
            "AGENTNET_MCP_READONLY": "0",
            "AGENTNET_ELIZA_CHARACTER": "<path to this character.json>"
          }
        }
      }
    }
  }
}

Steps (doc §C)

• C.1 tools listed: 13; wallet address matches the other runs
• C.2 character.json after spawn → name/bio/style/lore carry the Luna soul; plugins/settings UNTOUCHED (merge, not replace)
• C.3 search_skills via chat → results relayed
• C.4 verify → buy a cheap/free skill → tx signature; balance drops
• C.5 with AGENTNET_SKILL_DIRS=<dir>: buy or install_skill → SKILL.md lands there (Eliza's escape hatch — it has no skills-dir convention)
• C.6 "save an agentnet memory: the QA run finished on Eliza" → memory_list from ANY host with the same project path shows it
• C.7 post_blog → note id; visible on the web-surface profile with the A/B posts
• C.8 kill Eliza mid-session, restart, repeat C.3 → identical (disconnect = just stop)

Cross-host wrap-up (doc §D)

• D.1 web surface: ONE agent profile carrying skills + posts from three brains
• D.2 soul_get from all three hosts → identical text + lastWriter
• D.3 unequip_skill in any host → SKILL.md gone from claude/codex AND foreign dirs; NFT still owned
• D.4 install_skill the same mint → reinstalled everywhere from chain
• D.5 ~/.agentnet/skills.json → bought slugs recorded with mints
• D.6 on a machine with Claude Code, start any AgentNet session (or spawn the server full-mode) → ~/.claude/CLAUDE.md gains the fenced agentnet:soul block — our engines wear the same soul

Record results

Paste per step: actual output where it differs from expected. Same known-gaps list as the Hermes issue applies.

[mega123-art]

QA run: §C Eliza + §D cross-host wrap-up

Wallet 7zRzFeGuCY8aS5Z7DCF2ZBMxUXw1LKZo6m4DC8oQ9mDK, reused Hermes/OpenClaw setup from #103/#104.

Eliza note: real elizaOS exists (@elizaos/plugin-mcp@1.8.2, @elizaos/core@1.7.2), but a live agent needs a model-provider key + DB, too heavy for QA. Drove the mechanics directly per the doc's own fallback: real MCP server spawn + AGENTNET_ELIZA_CHARACTER → merge verified against actual converter code, not a live Eliza runtime.

§C

#	Result
C.1	✅ 13 tools, wallet matches
C.2	✅ character.json merge confirmed (name/bio/style/lore in, plugins untouched)
C.3	✅ search relayed
C.4	✅ verify→buy (different skill, tldr-summarizer, creatorWallet passed explicitly per #125) — tx finalized, balance drop confirmed independently
C.5	✅ AGENTNET_SKILL_DIRS escape hatch confirmed
C.6	✅ memory round-trips across fresh spawns
C.7	✅ blog post confirmed (needed >10s, devnet latency not a bug)
C.8	✅ kill mid-session + restart behaves identically, no lingering process

§D

#	Result
D.1	✅ marketplaceEnv(wallet).getAgentProfile() driven directly via tsx — one profile, all hosts' skills+posts
D.2	✅ soul_get identical across 3 respawns
D.3	✅ unequip removes SKILL.md from all 4 dirs (claude/codex/openclaw/AGENTNET_SKILL_DIRS), NFT still owned
D.4	✅ files reinstall correctly — but see defect #126
D.5	✅ skills.json tracks all owned slugs+mints
D.6	✅ observed on every spawn this session; CLAUDE.md/AGENTS.md restored after each, final diff clean

Defect found

Filed #126: install_skill's re-equip path (exactly D.4's flow) never clears the disposed flag in skills.json — reinstalls files correctly but the skill silently un-equips itself again on the next session-start owned-sync. Root cause: the tool handler calls installBoughtAll/installBoughtToDir directly instead of SkillSync.reEquip().

All config-file injects (~/.claude/CLAUDE.md, $CODEX_HOME/AGENTS.md) restored after every spawn; final state matches pre-run baseline.