openclaw - 💡(How to fix) Fix [Bug]: OpenClaw session transcript stops flushing mid-session on claude-cli runtime (Claude CLI side keeps updating) [1 comments, 2 participants]

Description

The OpenClaw-side session transcript at ~/.openclaw/agents/<agent>/sessions/<sessionId>.jsonl stops growing mid-conversation, while the Claude CLI-side transcript at ~/.claude/projects/<wd>/<sessionId>.jsonl continues updating normally for the same session. Gap widens turn over turn and does not recover.

This is a silent data loss for any tooling that reads the OpenClaw-side transcript (e.g. sessions_history, the session-logs skill, replays).

Environment

• OpenClaw 2026.5.4 (325df3e)
• Model: anthropic/claude-opus-4-7
• Runtime: Claude CLI
• Auth: oauth (anthropic:claude-cli) (fallback path because primary token auth was unavailable for this model)
• Surface: Discord channel session (agent:main:discord:channel:...)

Observed

Concrete numbers from a live session:

Over the ~24 minutes after the OpenClaw side stopped, the session continued normally end-to-end (replies delivered, tool calls executed and persisted, commits made on disk), but the OpenClaw transcript file's mtime and line count never advanced. Claude side advanced through ~7 additional turns in the same window.

Earlier in the same session, OpenClaw transcript was tracking the conversation correctly (lines were growing turn-over-turn), so the writer started fine and then silently stopped at some point after roughly the 193rd entry.

Expected

OpenClaw session transcript should continue to flush turn entries for the lifetime of the session, matching the Claude CLI side at its own granularity.

Impact

• sessions_history / replays / session-logs skill miss recent turns silently
• Long sessions on claude-cli runtime quietly lose transcript fidelity
• No user-visible error; only detectable by comparing mtimes/line counts against the Claude CLI side

Notes

• Distinct from #78194 (status display) — different code path: transcript writer vs. context-window formatter.
• Related but distinct from #77974 (Telegram DM context loss from a missing-transcript race before flush) — this is not a startup race; the writer flushes successfully early in the session and then stops mid-session. Surface here is Discord, not Telegram.
• Did not investigate root cause; flagging the symptom.