Reinstate Research and Compatibility Evidence

Evidence levels

Current evidence assets

Compatibility data

Machine-readable JSON, reviewed 2026-09-07

Fixture commit

cf6e430f3aa7

Acceptance methodology

Reinstate MacBook + Windows runbook

Architecture and security

Architecture and security model

Encrypted format specification

Manifest and snapshot format v1, including framing, validation, and pre-1.0 limits

Synthetic path explainer

macOS ↔ Windows path-mapping visualizerwith fixed non-private fixtures

Adapter and compatibility evidence

The compatibility matrix links each supported agent range to the exact adapter test source. Versions outside the inclusive stable ranges—including prereleases—are treated asUNTESTED and mutating operations fail closed. The matrix separately marks open physical platform gates instead of borrowing a source-test date as platform certification.

• Claude Code: 2.1.219–2.1.263; test source; source checks, deterministic synthetic fixtures, and dual-platform tagged-artifact acceptance.
• OpenAI Codex CLI: 0.133.0–0.149.0; test source; source checks, deterministic synthetic fixtures, and dual-platform tagged-artifact acceptance.
• Gemini CLI: 0.55.1–0.55.1; test source; read-only fixtures and a Windows physical path; no dual-platform T3 resume journeys.
• OpenCode: 1.18.21–1.18.29; test source; T3 resume and T4 destination device journeys recorded on macOS and native Windows (2026-08-22); T5 encrypted-sync journeys recorded on macOS and native Windows (2026-08-23); v0.6.0 tagged-artifact acceptance pending on native Windows and deferred on macOS.
• Grok Build: 1.0.5–1.0.13; test source; deterministic synthetic fixtures, a measured grok --version range, and T3 resume and T4 destination device journeys recorded on macOS and native Windows (2026-08-22 and 2026-08-23); GD8 collected on macOS only; native Windows range-widening read-only evidence (2026-09-08), completed-turn evidence deferred to the v0.6.0-rc.6 tagged run.
• Pi: 0.73.1–0.73.1; test source; dual-platform AGENT-PROBE-V1; F1 index source; no T3 device journey.
• Qwen Code: 0.21.12–0.23.0; test source; dual-platform AGENT-PROBE-V1; F1 index source; T2 reader; dual-platform T3 device journey; T4 destination journey; native Windows range-widening evidence.

Physical results and current limitation

The repository contains dated RC5 macOS and Windows observations. They are useful historical evidence, but they do not close the Reinstate acceptance matrix. The current Reinstate runbook requires both agents, both directions, checksum-verifying installers, wrong-passphrase refusal, ciphertext-only remote objects, conflict handling, restore backups, and redacted evidence.

Review theraw acceptance-result directory and the qualified public compatibility summary. Do not infer a stable support claim from an older passing row.

Why the fixtures are synthetic

Real coding-agent sessions can contain proprietary source code, prompts, internal paths, terminal output, or credentials. Committed fixtures are deterministic and synthetic, and the repository's secret checks protect against accidentally turning a compatibility test into a data leak. Public evidence uses versions, exit codes, counts, snapshot identifiers, and redacted paths rather than transcript text.

Benchmark publication contract

Before publishing a speed, success-rate, or session-size result, Reinstate must disclose:

• the question and success criteria;
• hardware, operating systems, agent versions, Reinstate version, and commit;
• the synthetic dataset or safely consented data source;
• exact commands, repetitions, warm-up and cache policy, and raw output;
• every failure, exclusion, limitation, and denominator; and
• a stable canonical page, methodology date, and downloadable evidence.