[stevepridemore]

Requested by Steve

Summary

Two opt-in uses of TypeSafe's Jev, which returns typed judgments rather than text ($0.042 per million input tokens, output free). Both are off by default. With the flags unset, behaviour is unchanged.

1. Verb selection (JEV_VERB_SELECT=on)

When graph_relate (single or batch) is about to store an edge as RELATED_TO, because the writer asked for it or asked for a verb outside the vocabulary, Jev picks from the documented verbs. There is an explicit "keep RELATED_TO" option. The verb is used only when:

• it's in RELATIONSHIP_TYPES, and isn't ALIAS_OF, CONTRADICTS or SUPERSEDES
• its Choice confidence is ≥ JEV_VERB_MIN_CONFIDENCE (0.55), and the separate "evidence names a specific relationship" Noul is ≥ 0.5
• it's defined for the source node's type

Otherwise the edge stays RELATED_TO, including on a timeout (8 s), an HTTP error or a spent cap ($0.05 a day per server process). A relabelled edge records verb_selected_by, verb_selection_confidence and verb_requested. The tool result reports each decision. The work happens in the graph_relate handler, so neo4j-client.ts is untouched.

2. Duplicate check (JEV_DUPLICATE_CHECK=on)

After a completed nightly API-runner dream, the runner judges graph_merge_suggestions candidates. It reads descriptions, aliases and relationships through graph_query and appends each verdict to logs/jev-duplicates.jsonl. Log-only: nothing is merged or aliased. Failures are logged and never change the run's outcome. Capped at $0.02 per run. It also runs alone with --job duplicates.

Shared

• The key is read at run time from JEV_API_KEY_FILE. Compose mounts JEV_API_KEY_HOST_FILE read-only into graph-memory-mcp and graph-dream. The key is never logged.
• Names or notes that look like personal identifiers (email, SSN, date of birth, card number) are never sent.
• New files: src/shared/jev.ts (client, budget, settings), src/shared/verb-select.ts, src/dream-runner/duplicates.ts.
• Docs: docs/DREAM_PROCESS.md ("Jev judgments"), docs/MCP_SERVER.md, .env.example.

Testing

• npm run build passes.
• 27 new unit tests (jev.test.ts, verb-select.test.ts, duplicates.test.ts) pass. They use fictional fixtures and a fake fetch.
• Full vitest run: 156 passed, 89 skipped, 2 failed. The same 2 tests fail on main without this change (the symlink-escape tests, on Windows). neo4j-client.test.ts needs the test Neo4j from npm run test:local and wasn't run against one.
• A live call with a real key on fictional edges gave DEPLOYED_TO (0.99) and PART_OF (0.72), and kept RELATED_TO for two vague edges. About 1,900 input tokens per verb judgment.

🤖 Generated with Claude Code