v0.10.0: Investigate, compare and share agent executions

English | 中文

Released on October 19, 2026.

v0.10.0 connects related sessions, saved-content search, Run comparison and selective offline sharing in the existing Dashboard. A new signed demo follows a real task from failed diagnostics to a successful repair.

Start with the new demo

From an extracted v0.10.0 archive:

./agentprov demo investigation-workflow

The command loads two signed captures. Choose Diagnostic repair: before as A and Diagnostic repair: after as B in Compare. Follow the delegated review, read the saved test output, inspect access decisions and compare the resulting files. The demo library now includes nine signed replays and eleven demo guides.

Follow the walkthrough.

Investigate without leaving the evidence

• Related sessions exposes parent/child and continuation relationships, alongside their supporting records.
• Search finds saved messages, tool inputs and results, configuration, file bodies and runtime records. Results link to the original content and graph.
• Compare shows configuration fields, file evidence, network destinations, outbound data surfaces and access-control records. One-sided observations retain their collection context.
• Export replay lets you select a Run or session, review the included content and download a self-contained, read-only ZIP. The recipient opens index.html without installing AgentProvenance. Selected exports are unsigned views; original source bundles retain their signatures.

The graph, risk list and investigation controls remain in place. Long-form reading, search and comparison open on demand.

Read AppArmor and seccomp evidence

Import or follow existing audit logs, and observe process security state through the native sensor. The Dashboard separates configuration, observed state, audit decisions and runtime results. Its detail reader connects attempts, mechanism decisions and actual outcomes where evidence supports the association.

Audit ingestion is read-only. The system reports ambiguous associations and collection gaps alongside the records. The daemon API, remote CLI, Python SDK and AI/MCP tools expose audit associations with English machine-readable output.

Access-control guide · API access

Choose context retention

launch and context import accept --capture basic|detailed; detailed is the default. Use the basic preset to retain less supplemental Agent context, or override tool-result retention with --capture-tool-results.

Kernel events, TLS bodies and workspace snapshots have independent controls. Saved capture settings appear in session details and follow the evidence into signed bundles and selected replays.

Capture settings

Reliability and upgrades

This release tightens process-lifetime and audit correlation, records syscall outcomes separately from attempts, and strengthens TLS/HTTP evidence handling. Offline readers preserve source references and selected export boundaries.

Seccomp rejections no longer borrow neighboring syscall results. Basic context retention preserves Codex fork relationships. Incomplete comparison scans keep unseen records unknown, and replay export enumerates all sessions independently of the overview's display limit.

Back up your data directory before upgrading. Keep an older binary paired with its original database backup. Historical demo bundles remain available unchanged.

Full changelog · Getting started