Pass official execution options through to each Harness

[SaladDay]

Summary

Core passes the official execution options straight to each Harness's native settings and validates them against the Harness declaration. Core sends a value only when the Agent sets it or the pinned API defines a default; an omitted web_search or programmatic_tool_calling keeps the Harness's native behavior.

• Codex: web_search disabled/cached/live with context_size, allowed_domains and location; reasoning.effort and reasoning.summary; service_tier default/flex/priority/fast (fast is Codex's native alias of priority); programmatic_tool_calling enabled: true turns on native code_mode, and is rejected together with multi_agent because Subagent observation turns code_mode off. An omitted programmatic tool calling setting stays native (off), which differs from the official default and is recorded in the compatibility notes.
• Claude Code: reasoning.effort maps to the SDK effort. Search stays disabled or absent for now.
• MiniMax Code: absent or disabled controls only.
• Each declaration now carries its accepted value sets in Declaration.Execution; a value outside them is rejected at Session admission with a message naming the field. Settings of a disabled search configure nothing and stay admitted on every Harness, so saved disabled tools keep working. A declaration must include text verbosity medium, the official default Core always sends.
• The Core–Runtime protocol version moves to 0.17.0.
• reasoning.effort is the only home for effort: the Codex model_reasoning_effort and Claude effort harness_config keys are removed, and migration 000102 moves stored Agent and Session values into reasoning.effort before dropping the keys.
• Deleted: the Codex verbosity catalog probe (Codex ignores unsupported verbosity itself), DISABLE_TELEMETRY, the per-control capability bits replaced by execution_controls, and the old Core gates.
• Each Harness declaration is now one file: harnessconfig/admission.go folds into harness.go, and Claude's restrictions and native parameters fold into claudesdk/declaration.go.

Operator notes

• Deployment default model configurations that set model_reasoning_effort (Codex) or effort (Claude) in harness_config lose that key on upgrade; set reasoning effort on the Agent instead.

Checks

• go build ./..., go vet, focused go test on harnessconfig, daemon agent and CLI, Core execution and API.
• make check-harness-catalog, make openapi (no tracked changes), make check-names, website translation tests.
• Local PostgreSQL: the official-client suite (official_client.py and its Go tests) and the three services/core/tests/integration shards.
• Not run: a live Codex run (TestNativeToolPolicyPublicExecution needs real model options).