Skip to content

Agenturm 0.2.4

Choose a tag to compare

@abhishekash abhishekash released this 16 Nov 07:06
· 236 commits to main since this release
a56b6c1

Agenturm 0.2.4

Agenturm now makes recorded work usable outside the live cockpit. agenturm review provides the same scoped evidence as text, JSON, or a frozen Markdown/HTML receipt: initiating intent, recorded edits and checks, later changes, and explicit capture/ownership limits. Checks followed by edits remain reasons to inspect and rerun affected checks; agent-reported completion does not certify the worktree.

Findings retain each occurrence’s source ownership, with exact-source inspection and capture-policy fences through copy and export. Saved-task continuity preserves authored decisions and linked execution context. Typed investigations search recorded paths, commands and instructions with explicit coverage limits.

This release also improves how the receiver prepares retained history: conservative safety checks avoid repeated work only when the complete operation path cannot add information. Frozen comparisons cover redaction, source mutations, ownership, retry reminders and duplicate occurrences. Initial source indexing is bounded so discovery cannot inherit an unrelated history batch. Activity, Metrics, Sessions and raw MCP reads certify their own dependencies while other indexes catch up.

Upgrade

uv tool upgrade agenturm
agenturm

For a fresh isolated install: uv tool install agenturm==0.2.4. Restart Agenturm and any separately running receiver.

Evidence and limits

A retained workload reconstructed all 297 turns over 3,622 active seconds, but missed the existing eight-second final cockpit shutdown deadline. Both cockpit processes ultimately exited with code 0 and preserved the expected review receipts; the overall soak remains failed. Startup, idle and API resource targets are reported separately and are not implied by the functional checks.

See the current quality report and 38-item acceptance ledger for exact tested revisions, measured outcomes and open targets. Source-derived synthetic fixtures establish specific supported contracts; they do not prove universal runtime compatibility or human reconstruction outcomes. Agenturm continues to monitor agents rather than launch or control them.