Validate Agent configuration with official error fields

[SaladDay]

Core-owned validation of Agent configuration (saved Agent create/update and the inline agent on Session create) now rejects what the official service rejects, with the official error fields. Native harness admission limits are unchanged. The pinned baseline is unchanged (SDK 3.13.0 / d7c41ef / agents=v1).

Behavior

• Structural errors. Missing required parameter, unknown parameter, wrong JSON type and invalid enum value, anywhere in tools[], text.format, reasoning, service_tier, multi_agent and so on, return 400 invalid_request_error:
  • the param is the JSON path, prefixed with agent. on Session create;
  • the messages use the observed official forms, e.g. "Missing required parameter: ''.", "Unknown parameter: ''.", "Invalid type for '': expected …, but got … instead." and "Invalid value: ''. Supported values are: …";
  • supported-value lists come from the pinned enums.
• Duplicates and member names.
  • Duplicate function names, and more than one web_search or tool_search tool, return 400 with the observed messages.
  • A repeated member, or a member name that differs from a known one only by case, returns 400 with its path. This local message was not observed officially. It stops Go's merge of repeated keys from bypassing validation.
• Schema roots. Function parameters and text.format json_schema whose root is not type: object return the observed 400 messages, for every harness, before harness admission.
• Session create order. An invalid inline agent is reported before the none input requirement, as officially observed.
• Unchanged.
  • Values the official service accepts: arbitrary function names, PTC enabled on a saved Agent, reasoning.effort: max, service_tier: flex.
  • Harness admission limits, which keep unsupported_or_invalid_configuration.
  • Saved web_search modes (TV-05, pending a separate decision).
• Echo bound. Echoed names and values are limited to 256 printable bytes.

Evidence

Campaign scan 3 findings TV-01..07 (owned official requests, all resources deleted). All 44 official TV records were replayed against this branch: 42 match on status, code, param and message, and the other 2 are the intentional K3 rows. Recorded in official-semantics-alignment.md and operation-evidence.md (entry M; rows 1, 3, 6).

Validation

• Live check that valid configurations still execute, through real Core, the daemon and native harnesses with Kimi K3:
  • A Codex object-root function was called and completed.
  • A Claude SDK object-root json_schema answer parsed correctly.
  • These passed on both main and the candidate.
  • Invalid roots now return the new 400 on both harnesses. On main, Codex had admitted a string root.
  • 2 Turns per phase. Cleanup and secret scans passed.
• Tests.
  • Go handler tests cover every row on Agent create, update and Session create.
  • A real-PostgreSQL replay covers two tenants and owned, foreign, missing and malformed IDs, with no-write digests.
  • The pinned-SDK scripts and the full official_client.py harness pass.
• Server gate on the rebased code. All make check targets plus the Web typecheck, core-doctor, unit tests and build pass. The Playwright browser cases were not run on the server (no Google Chrome; skip approved by the user). make openapi is byte-identical.
• Review. A fresh full-diff blind review by a Claude Code subagent (the user-approved replacement for GPT-6 Astra) found no K-boundary regressions. Its in-scope finding, the repeated-member bypass, is fixed in the last two commits and verified by tests and the gate.

Documented differences

• Same-key retries of inline Sessions created before this change fail if their original configuration is now invalid; saved-Agent retries still recover.
• An invalid inline override is reported before a missing saved Agent.

No full protocol compatibility is claimed.

---

^{Need help on this PR? Tag @codesmith with what you need. Autofix is disabled.}