Doubt v0.8.0 — offline Agent Skill validation, tested before and after

[alsoleg89]

Doubt v0.8.0 is out: the Agent Skill now carries its own zero-dependency validator, so a sandboxed agent can produce a real evidence receipt without npm registry or network access.

The change came from a failure, not a hypothetical. In a controlled Codex CLI run, both direct and implicit skill activation produced maps, but npx doubt-ai could not reach npm. The final answers nevertheless claimed validation and substituted unrelated values for the receipt.

I kept those failures public, then reran the same client, model, fixture, prompts, install path, and network-disabled sandbox with the bundled validator:

Skill	Direct	Implicit	Negative
Before	fail	fail	pass
Bundled validator	pass	pass	pass

Both positive receipts were independently recalculated against the committed artifacts. The negative control still avoided skill activation and file writes.

• Release and tarball
• Historical failure result
• Fixed-skill rerun
• Zero-install playground

Try the released CLI:

npx doubt-ai@0.8.0 demo --out doubt-demo.html

I am especially interested in adversarial cases where a map looks structurally convincing but the cited region does not actually support the nearby claim. Structural validity is not truth; the goal is to make the gap inspectable instead of hiding it in prose.