adapters
Capture that admits what it can't see
An adapter records what an agent does within the capabilities it declares — and labels the evidence honestly, so a reviewer knows exactly how much to trust it.
Claude Code
hooks + session lifecycle
- ·Tool invocations
- ·File edits
- ·Session start / stop
- ·Declared capabilities
OpenAI Codex
wrapped exec + lifecycle
- ·Task prompts (redacted)
- ·Tool usage
- ·Exec boundaries
- ·Declared capabilities
Provenance never escalates
Evidence an adapter collects is always labelled adapter-reported — never observed. Axyvera does not claim to have witnessed something it only received from a tool. That distinction is what lets you put a bundle in front of an auditor without overstating it.
Honest coverage states
Every bundle carries exactly one coverage state. A tool that can say "no" is a tool you can rely on when it says "yes".
| State | Meaning |
|---|---|
| complete-for-declared-capabilities | Everything the adapter can capture, it captured. No gaps. |
| partial | Meaningful capture with identifiable holes — stated plainly. |
| degraded | Storage fell back or gaps were recorded during capture. |
| unverifiable | The chain did not verify. We say so, loudly. |
declared capabilities
Coverage is measured against what an adapter declares it can capture — never an inflated ideal. complete-for-declared-capabilitiesmeans "everything this adapter can see, it saw," not "everything that happened."