Create a response

Context management directives. Forwarded verbatim to providers that support server-side compaction (OpenAI, Azure OpenAI); for others the field is ignored at the adapter layer. See OpenAI's compaction guide for the schema; the canonical entry is {"type": "compaction", "compact_threshold": <tokens>}.

Items to include in response.

Hadrian Extension: also accepts "usage.incremental", which emits cumulative response.usage.updated SSE events at server-tool turn boundaries (stripped before forwarding upstream).

Hadrian Extension: List of models for multi-model routing (alternative to single model)

Hadrian Extension: Plugins to enable for this request

Hadrian Extension: Skills to mount into the shell-tool session. Hadrian treats skills as a top-level request field rather than threading them through the OpenAI shell-tool environment block, so they're accepted on every responses request regardless of whether the upstream provider has its own skills surface.

Mirrors OpenAI's typed shape — each entry is either a reference to a stored skill or an inline bundle:

[
  {"type": "skill_reference", "skill_id": "skill_…", "version": "latest"},
  {"type": "inline", "name": "extract-csv", "description": "...",
   "source": {"type": "base64", "media_type": "text/markdown", "data": "..."}}
]

For skill_reference, skill_id is a prefixed id (skill_…), a bare UUID, or the skill's name slug, owned by the caller's org. version is optional: omit for the skill's default version, "latest" for the newest, or a positive integer for that exact version (default and latest can differ).

For inline, the decoded source.data is mounted as an ephemeral skill bundle: text/markdown is treated as the SKILL.md content; other media types are rejected today. The name must be a lowercase slug — it becomes the mount directory.

Each resolved skill's SKILL.md is prepended to instructions so the model knows the skill is available; all skill files are materialized under /skills/<name>-<version>/ (stored skills) or /skills/<name>/ (inline skills) inside the sandbox. The path mirrors OpenAI's <name>-<version> container layout under Hadrian's /skills root.