[SaladDay]

Claude Code can now execute in the V1 Docker Runtime beside the daemon and local workspace, while Core remains independently deployed. Public clients use the existing Session, Files and Artifacts interfaces. The pinned native SDK owns execution; preparation, cancellation, resource management and file operations reuse the existing contracts.

The adapter protects provider credentials and native history, resumes the same native Session after reconnect/restart, and refuses missing history instead of silently starting again. A small qualified engine profile separates public admission from advertised Runtime capabilities. The Docker nested-sandbox option supplies only the native procfs/init prerequisite; native sandbox fallback and weaker isolation remain disabled. The deployment guide documents the minimal registration path for additional engines.

Validation:

• Linux adapter suite: 95/95; focused Go tests and race checks passed.
• make check passed with a dedicated Agents API test database; make openapi regenerated the contract.
• Fixed official SDK 3.13.0/raw HTTP with real Kimi K3 through native Claude Code: standalone startup, execution, Files/Artifacts, tenant and credential isolation, enabled/disabled network, cancellation, Core/Runtime restart and crash recovery without replay.
• Final Runtime image: provider-secret isolation, loss of Core native identity, and missing-history refusal passed. Controlled reservation aging verifies the existing expiry path without waiting five minutes.
• Fresh GPT-6 Astra high blind review of the complete diff: no blocking findings; reviewer independently ran 13 JS tests. Reviewer Go execution was unavailable; server Go/check results are recorded separately.
• Candidate Core + accepted Codex Runtime real Kimi regression: all 13 hosted workflow checks passed, including Files, isolation, restart, reconnect, cancellation and network policy; cleanup passed.

Limits: this qualifies the Claude hosted text/native-tools profile, not full upstream compatibility. Hosted function/MCP combinations, multi-agent, structured output and user-managed enrollment remain gaps. Real acceptance used Kimi K3, not an Anthropic Claude model. Missing history fails closed but currently follows the shared preparation timeout; a pending pre-Turn cancellation conflicts. Faster error feedback is deferred.

Task/evidence: CLAUDE-V1-INTEGRATION-001 in the Feishu board; ~/.parsar/remediation/20260919/claude-v1/ on zju_a100_2. The protocol pin and Codex/native versions are unchanged. No product cutover, new loop, scheduler, framework or unrelated backlog work.