Every agent harness tells its story differently. Codex logs one way, Claude Code another, Hermes and OpenClaw their own ways. If you’re trying to govern agents across harnesses, and everyone eventually is, you’re comparing differently-shaped assertions and pretending it’s a comparison.
The provider-neutral harness evidence intake fixes the shape problem. One intake layer. One strict contract. Every harness’s claims become comparable evidence.
What it does
safe2 evidence harness normalizes evidence from Codex, Claude Code, Hermes, OpenClaw, coding agents, wrappers, and hooks into a single contract chain (harness-source-v1 to harness-evidence-v1).
What makes it trustworthy is what it refuses to do:
- Seven-domain coverage inventory: task lifecycle, tool calls, artifacts, tests, usage, environment, and more. You see what’s covered and what isn’t.
- Observed vs. declared vs. provider-reported vs. estimated, kept in separate lanes. An agent’s claim about itself never sits in the same bucket as an observed fact.
- Partial, missing, and unknown are first-class results. A harness that can’t show you something says so explicitly instead of silently omitting it.
- Conservative defaults:
completion_verified: false,billing_verified: false,conformance_claim: false. Nothing is assumed true until evidence says so. - SHA-256 binding to source bytes, so the evidence is pinned to exactly what was observed.
The question changes from “what did the agent say?” to “what evidence exists, where did it come from, and what’s missing?” That’s the entire governance shift in one sentence.
One honest caveat
This release ships the intake layer, not native harness adapters. The adapters that plug each harness in natively come separately. What you get now is the contract and the normalization seam, which is the hard part. Adapters are plumbing; the contract is the standard.
Why this matters
- Compare harnesses on evidence, not marketing. Same contract, same coverage inventory, the gaps become visible.
- No provider becomes the source of truth. The intake is provider-neutral by design; no harness grades its own homework.
- Missing evidence is data. When a harness can’t produce evidence for a domain, that’s a finding, not a footnote.
- It compounds. Every CLI release after this one (system identity, failure diagnosis, operational truth) binds to evidence that entered through this intake.
Takeaways
- One contract normalizes agent evidence across all major harnesses.
- Observed, declared, provider-reported, and estimated are never mixed.
- Conservative defaults mean silence is never mistaken for success.
Bottom line: If you’re running more than one agent harness and governing them all with vibes, this is the release that gives you a common ruler. Adopt the contract, then demand every harness measure up to it.
Release notes: Provider-Neutral Harness Evidence Intake on GitHub, tag 2026-09-14_CLI_0.3.0_Provider_Neutral_Harness_Evidence_Intake.
