The former “Claude Code Features Analysis” was a January 2026 reverse-engineering snapshot that counted strings in a vendor bundle and then marked similarly named OpenClaudia modules as implemented. The full audit shows that this produced false parity claims, especially for MCP, planning, sessions, hooks, permissions, subagents, memory, reasoning, and tool loops.
This file is no longer a feature checklist or implementation authority.
- Compatibility is defined at an explicit protocol/schema boundary, not by copying identity text, private prompt wording, paths, or UI behavior.
- A compatibility claim requires a versioned fixture and an end-to-end conformance test through every supported frontend.
- Unsupported fields and events are rejected or surfaced, not silently parsed and ignored.
- Provider authentication uses documented provider flows. Subscription-client impersonation and inherited vendor identity prompts are not supported target architecture.
- Project, hook, tool, memory, plugin, MCP, and fetched text keep their real provenance and do not become system authority merely because a vendor prompt once used a reminder tag.
The explicit browser feature contains free DuckDuckGo/Bing browser scraping
and has no search-key requirement. It is not part of the default build and
requires an operator-installed compatible browser; runtime executable download
is disabled. Backend availability is not evidence that browser egress,
cancellation, result provenance, or completion semantics are production-ready.
capabilities/registry.jsonrecords typed maturity, entrypoints, effects, limitations, and executable evidence links.capabilities/evaluation-corpus.jsonand its digest-bound quality review record define repeated final-environment graders.docs/binary-capability-matrix.mdis generated from the validated registry.docs/full-codebase-audit-2026-08-16.mdrecords the detailed findings.docs/production-remediation-design.mddefines the preserved outcomes and target contracts.- Future compatibility matrices must add released conformance scenarios to the same evidence boundary rather than reintroduce prose checklists.