Six tools with discriminator-based oneOf schemas and action enums. All tools have descriptions (50-150 chars) and documented parameters. However, critical gaps: (1) No output schemas documented anywhere, LLMs cannot predict return structure or chain calls; (2) Parameter descriptions are minimal (5-20 chars) and lack constraints, ranges, or format guidance; (3) No error handling guidance, tools return CLI output but no recovery hints; (4) Tool names are generic (harness_*) and do not follow verb_noun pattern; (5) No indication of idempotency, retryability, or side effects; (6) Discriminator pattern is valid but complex, LLMs may struggle with oneOf branching. Per-tool scores average 52 due to present but incomplete schemas and sparse descriptions.

• Document output schema for each tool action. E.g., 'capabilities returns {capabilities: string[], version: string}'. Include field types and descriptions.
• Expand parameter descriptions to 50 - 100 chars with format/constraint guidance. E.g., 'dbPath: SQLite database file path (default: ~/.harness/db.sqlite). Must exist and be readable.'
• Rename tools to verb_noun pattern: 'inspect_harness_capabilities', 'orchestrate_campaign', 'manage_session', 'store_artifacts', 'administer_harness'.
• Add error handling guidance to tool descriptions. E.g., 'Returns error if dbPath does not exist (retryable: check file permissions). Returns error if campaign not found (user-fixable: verify campaign ID).'
• Add idempotentHint, destructiveHint, and readOnlyHint annotations to tool definitions. Mark WRITE tools with confirmation pattern or dry-run action.