Remvu v0.2.0 — portable project memory

Remvu v0.2.0

Portable project memory for coding agents. Give a fresh agent the project code and one REMVU.md file with the intent, constraints, decisions, checks and unfinished work.

This release is a reference implementation and an open format proposal, remvu/1.

• Private memory and observable events stay in ignored .remvu/ storage. The active agent writes semantic updates; the runtime makes no model API requests.
• Claude Code and Codex adapters provide project instructions, lifecycle recording and bounded repair. Public sharing is explicit, with revision checks, reviewed-event watermarks, retained conflicts and audited public withdrawal.
• Old checks retain their code fingerprints. Failed or unsupported checks remain visible. The version 0.1 exporter and import bridge remain available.
• Schemas, conformance examples, a validator, adapter guidance and English/Chinese documentation are included in the source repository.

Verified results

Both real handoffs passed with fresh receivers that had code and REMVU.md, without the original chat or a Remvu installation: Claude Code → Codex passed 4/4 code checks; Codex → Claude Code passed 5/5. Both recalled 11/11 fixed context term groups. Code-only controls requested missing context and left the unfinished step open.

The local suite passed 137 tests. The release CI covers Python 3.10/3.12 on Linux, macOS and Windows, strict format examples and archive audits. Windows skips two POSIX-permission symlink fixtures. Native lifecycle and guided clean consolidation were observed on macOS with Claude Code 2.1.302 and Codex 0.163.0-alpha.10.2; native Windows automatic clean consolidation is not supported.

Release CI · Validation report · 30-second evidence replay

These two tasks establish functional examples, not general coding performance or a token-saving rate. Actual native usage and the initial blocked attempt are published in the report.

Downloads

• remvu-skill-v0.2.0.zip: standalone remvu skill, bundled runtime, references and license.
• remvu-plugin-v0.2.0.zip: portable plugin components with both Remvu and the legacy skill.
• SHA256SUMS: checksums of both archives.

Complete source, tests, reports and the replay are in GitHub's source archives. Keep extracted runtime files at their installation path and follow the README's initialization and client trust steps. Invoke /remvu in Claude Code or $remvu in Codex.

Agent Task Export v0.1.0

Local AI task evidence export with a compact handoff and on-demand tool records.

• Standard-library Python; no model API calls or automatic uploads.
• Codex rollout, a subset of exec JSON events, and a portable evidence format.
• Deduplicated full redacted outputs, call/result pairing, edit references and integrity checks.
• Observable decision summaries; hidden reasoning and system/developer messages are excluded.
• Chinese and English documentation, edited using Humanizer, plus synthetic examples.

Validation: 18 local tests passed, skill validation passed, and both synthetic and current-session exports passed integrity checks. Remote CI has not run; an Actions configuration example is provided in docs/github-actions-example.yml because the upload credential lacks the workflow scope.

The asset is a skills-only Agent Plugins 1.0 package. You can also copy skills/agent-task-export directly. Scripts need a local Python execution environment with file access. Redaction is best effort; review exported bundles before sharing.