Here’s a scenario that should worry you: the agent says it finished. The test passed. A receipt exists. And the three of them are describing different tasks, or different revisions, or different environments.
Nobody lied. The agent completed something, the test tested something, the receipt recorded something. But without correlation, you have three true statements that add up to a false picture. That’s the gap 0.7.0 closes.
What’s new
safe2 evidence truth: correlates multi-harness evidence into one versioned operational-truth contract: revision-bound task receipts, coverage, usage, and completion claims, all pinned to the same task, revision, and environment. When sources conflict, the result isn’t a quiet win. It’s a visible hold. Conflicting completion claims from different harnesses become something a human reviews instead of something nobody notices.
safe2 evidence changes: a one-shot inventory of skills and agent configuration across SKILL.md packages, MCP servers, OpenClaw, and .codex / .claude / .hermes / .openclaw configs, diffed baseline-vs.-current. This is the silent attack surface: a skill updates, a config drifts, an MCP server gains a capability, and nothing in your pipeline notices. The change monitor notices, in CI, every run.
Both feed a human decision card generated from the same record the machine used. One source of truth, two audiences.
“Agent said it finished, test passed, receipt exists. Do they describe the same task, revision, and environment?” 0.7.0 is the first release that can actually answer yes or no.
Why upgrade
- Conflicts become holds, not hidden wins. When harnesses disagree about completion, you see it.
- Skill and config drift is caught in CI. The attack surface nobody watches is now watched automatically.
- One record, two readers. Machines get the versioned contract; humans get the decision card. No translation loss.
- It completes the evidence chain. Receipts (0.3.0) to intake (harness) to identity (0.4.0) to diagnosis (0.5.0) to attribution (0.6.0) to truth (0.7.0). Each release compounds.
Takeaways
- 0.7.0 correlates all agent evidence into one versioned operational-truth contract.
- Conflicting claims surface as reviewable holds.
- Change monitor covers SKILL.md, MCP, OpenClaw, and major agent configs.
- 805 tests passing, 7 skipped; Python 3.11 to 3.14.
Bottom line: Three true statements can add up to a false picture. 0.7.0 makes sure they don’t.
Release notes: AI SAFE² CLI 0.7.0 on GitHub, tag 2026_09_20_CLI_0.7.0_Reviewable_Operational_Truth.
