v0.9.0 — Follow an agent from task to result
English | 中文
v0.9.0 adds Agent session to the execution view. Read the task, conversation, tool inputs and results, inspect the configuration recorded at the time, then follow a tool into its process and file evidence. Export the run and those saved records travel with it, ready to read on another machine.
Explore the online demo or download a portable archive from Releases.
Try it in one command
From an extracted v0.9.0 archive:
./agentprov demo deepseek-context
Watch a real DeepSeek Harness session add daily totals to a Python report and pass seven tests. The replay works offline without Go, a model account or an API key. The demo library now includes seven signed replays and nine guides.
The online library has the same graph, session and saved-content views. The full project documentation sits alongside the nine demo walkthroughs, in English and Chinese. Online replay uses the bundled public captures; it needs no local daemon.
What is new
Conversations and tool results alongside the graph
The graph stays at the top; Agent session sits below it, collapsed by default. Expand it to read user and assistant messages, tool inputs, results and errors. Recorded links let you move between a tool entry and its runtime evidence. The existing timeline, process tree, network and risk views remain available.
Task, permission and configuration history
See the task, model, working directory, approval policy and sandbox settings recorded by the agent. MCP, skills and plugin details are included when the source provides them. Select two task, configuration or approval records to compare their values, within a run or across runs.
Configuration and actual approval decisions are separate records. If the log has no decision, the interface shows it as unrecorded.
Full saved content, including files
Long tool results and original source records open in a separate paged reader.
With --file-diff, recording also saves the final text of changed files. Open a
file from the graph and keep reading even with Agent session collapsed.
The saved text remains available after signed export and offline import.
A clear account of what was collected
Collection status reports session sources, missing fields and parsing errors, with runtime probe coverage and association gaps alongside them. Unknown counts stay unknown. The same reports are available in the CLI, HTTP API and evidence bundle, so another machine sees the original capture conditions.
DeepSeek Harness, with stronger existing adapters
| Agent | Supported context |
|---|---|
| Claude Code | Hooks, native messages and tools, permission-mode changes and session state |
| Codex | Native rollout messages and tools, resumed threads, settings and recorded child identities |
| DeepSeek Harness | v3 JSONL and native v4 Zstandard sessions, tools and configuration |
| Kimi / Grok | Existing supported transcript formats |
Real task and resume checks used Claude Code 2.1.293 and Codex 0.159.2. The signed DeepSeek development task uses Harness 0.1.7-rc.2. See the compatibility table for the checked formats. Session collection runs after the agent exits; custom directories, explicit session selection, resumed history and repeat imports are supported.
Record your own session
With the agent installed and authenticated:
./agentprov launch --file-diff -- claude
./agentprov launch --context-dir /path/to/sessions -- codex
./agentprov launch --file-diff -- dsh headless --json 'Inspect the project and run its tests.'
./agentprov context coverage --run RUN_ID
./agentprov context list --run RUN_ID --kind configuration --revisions
Agent session guide · Context API · DeepSeek walkthrough
Reliability and compatibility
This release also strengthens process identity, cgroup migration, late event attribution and durable capture. The amd64 precompiled probe now includes the child birth-cgroup fix, and live lifecycle checks run in CI.
Acceptance includes 485 browser checks, six original historical signed bundles, database migration from schema 17 to 19, and live x86 KVM/K3s and TLS checks. Remote CI and Linux/macOS archives for amd64/arm64 passed. Details and reproduction steps are in the acceptance report. The 24-hour soak remains deferred; this round did not repeat ARM live validation.
Before upgrading
- Back up your data directory. v0.9.0 migrates it to schema 19; an older binary should not open the upgraded database.
- The six historical signed demos keep their original signatures and content. Their missing session fields remain unrecorded.
- Linux native capture requires the appropriate kernel and permissions. macOS supports application recording and replay. Platform and TLS guide.
- Signing is enabled with
--sign-key; saved final changed-file text with--file-diff. Both are opt-in.