Skip to content

compatibility-ledger.md

Latest commit

 

History

History
83 lines (62 loc) · 3.48 KB

compatibility-ledger.md

File metadata and controls

83 lines (62 loc) · 3.48 KB
Raw

Compatibility ledger

This document lists compatibility behavior intentionally retained by the current release. New code should use the canonical APIs even when older forms remain readable.

Configuration schemas

  • Schema v4 is canonical for runtime/profile/model-route/target configuration.
  • Schema v1-v3 remain readable and migratable during the compatibility window.
  • Legacy provider-only configuration is normalized into the canonical execution model rather than becoming a second source of truth.
  • The legacy opencode/providers.yaml endpoint format remains readable.

Agent/profile terminology

  • Scheduler identity is an agent profile/candidate.
  • Historical provider_id, --provider, and provider-named health/state fields remain readable where documented.
  • Model-provider names such as OpenAI, Anthropic, and Ollama remain correctly described as providers.
  • AgentAdapterCapabilities remains a compatibility alias for the canonical RuntimeCapabilities type.

Invocation and runtime records

Older invocation/runtime-session records are migrated or read conservatively. Missing newer dimensions do not authorize stronger execution. Opaque runtime session references remain disposable optimization state.

Native runtime

Native adapter command behavior and live-session compatibility remain stable behind the normalized runtime facade. OS process/PTY supervision is owned by execraft.process even when callers use higher-level orchestration APIs.

OpenClaw Gateway

  • The validated OpenClaw release is 2026.7.1-2 with Gateway protocol 4.
  • Managed and external Gateway modes remain distinct ownership models.
  • Explicit external configuration is never overwritten by Execraft.
  • Credential references remain references; secret values are not projected into browser/read-model payloads.
  • Public Gateway APIs are the compatibility boundary; private databases and transcripts are not authoritative inputs.

Model projection and targets

  • Canonical model routes and execution targets own provider/model/endpoint and placement data.
  • Legacy provider aliases may still resolve through the compatibility registry.
  • Supported target kinds are local and inference_endpoint.
  • remote_runtime configuration remains readable but is experimental-disabled in normal execution.

OpenClaw continuation and optimization

  • Cold reconstruction remains the fallback when exact session compatibility cannot be proven.
  • Session bindings include the context needed to invalidate stale reuse.
  • Managed workflow-skill projection is optional; external Gateways fall back to inline skill content when projection availability cannot be proven.
  • Context compaction/reuse is optimization only and must pass authoritative guards.

OpenClaw security

  • Managed children use isolated Execraft-owned runtime state and a restricted environment.
  • Hard sandbox/tool claims fail closed when enforcement cannot be provided.
  • Configuration-level security diagnostics do not fabricate live enforcement evidence.

Experimental-disabled definitions

OpenClaw specialist/sub-agent delegation and full remote-runtime placement remain readable for compatibility but are not supported normal execution paths. Release code must fail closed rather than activating them implicitly.

Runtime extensions

Third-party runtime kinds remain supported through the runtime registry seam. A runtime extension must use canonical execution requests/results and cannot create parallel scheduler/configuration authority.