Pan Agent — WAL recovery, failure correction and Trace replay

Pan Agent — database recovery demonstration

Watch the 129.8-second TUI demonstration: investigate an unreadable SQLite WAL, correct a real checksum-endianness error, recover all 11 records, and inspect the same run through Trace replay. Key evidence is enlarged and annotated; speed changes and archived detail views are labelled.

This is a media/documentation release. The executable is unchanged: download the v0.1.0 runtime package and follow the README.

One real Kimi Code task produced 11 model exchanges and 11 tool results, including two natural errors. Known token usage: 49,050 input + 3,806 output = 52,856; unknown usage records: 0. Original live capture lasted 371 seconds including interaction and replay; idle/navigation time is shortened in this edit. The independent review verified the recovery scripts against original inputs, 7/7 path-adapted upstream tests, a negative test, and replay with zero provider/tool/network calls.

Based on Terminal-Bench 2.0's db-wal-recovery task, authored by Xiangning Lin. This fresh macOS demonstration adapts the original /app path and is not a new official benchmark score. Both model errors are retained in the original trace. The video comes from actual PTY output, not a fabricated UI. Raw records and test answers are not public assets.

Task provenance, adaptations and edit map · Independent accepted Verdict

Pan-authored code is MIT; the upstream Terminal-Bench task retains its Apache-2.0 license. Human specified the task framing and visual emphasis; Kimi generated the recovery implementation; Builder recorded/edited and Regulator independently checked evidence. No runtime reliability or benchmark improvement is inferred from this demonstration.

Assets: pan-agent-wal-demo.mp4 and SHA256SUMS. Previous v0.1.0 assets remain unchanged.