Summary

~/.openclaw/agents/<name>/sessions/sessions.json keeps cliSessionId, previousSessionIds, and transcriptPath as null indefinitely on the main agent record. As a result, when the runtime spawns a fresh claude -p for the same agent (e.g. after a cron callback or compaction-induced restart), it can't --resume the prior CLI session — every CLI process exit becomes a context wipe.

Versions

• openclaw 2026.4.21
• claude CLI 2.1.114
• macOS Darwin 25.3.0

Repro

1. Run main agent for an extended session via claude -p. Confirm transcript accumulates under ~/.claude/projects/<workspace-path>/<sessionId>.jsonl.
2. Trigger a fresh CLI process (cron callback with wakeMode: now, or let the underlying CLI session rotate from a config change / plugin reload).
3. Inspect ~/.openclaw/agents/<name>/sessions/sessions.json for the agent.

Observed:
```json
{
"sessionId": "41eda629-c62a-4e36-a98c-86653422682a",
"cliSessionId": null,
"previousSessionIds": null,
"transcriptPath": null,
"sessionFile": "/Users/.../agents/main/sessions/41eda629-...jsonl"
}
```

The runtime never populated the resume-chain fields, even though the actual CLI transcripts exist on disk under ~/.claude/projects/.

Expected

On each CLI process spawn, the runtime should:

• Capture the new claude -p session ID into cliSessionId
• Append the prior cliSessionId to previousSessionIds (or maintain a chain)
• Resolve and store the transcript path so resume can be reattempted

When spawning a fresh CLI for the same agent (e.g. cron callback), pass --resume <cliSessionId> so the new process inherits the prior session's history.

Real-world impact

Today's session went: long productive run ended at session 3b7fe27b... (982KB transcript, 18:22 EDT) → cron callback at 18:23 EDT spawned 6d0a999c... with no resume → user's next message in 336dcf95... at 18:26 EDT had zero context for the immediately-prior conversation. From the user's perspective: "are you sure the callbacks need to be that long?" → "what callbacks?"

Workaround in place

At the content layer, installed a Claude Code SessionStart hook in ~/.claude/settings.json that reads STATE.md and MEMORY.md from the workspace and emits them via hookSpecificOutput.additionalContext. Plus overrode agents.defaults.heartbeat.prompt in ~/.openclaw/openclaw.json because the default prompt (in dist/heartbeat-BARAswcJ.js) contains "Do not infer or repeat old tasks from prior chats" — which directly contradicted the hook-injected continuity context.

These workarounds restore summarized continuity. They can't substitute for actual --resume because STATE.md only captures the current state, not the conversation history (intermediate tool calls, sub-agent reasoning, etc.).

Suggested fix scope

• Where the runtime invokes claude -p, capture the resulting session ID (visible via Claude Code SDK output / transcript filename) and persist it onto the agent record before any subsequent spawn.
• On subsequent spawns, prefer --resume <previous> and only fall back to fresh-session when resume fails (e.g. transcript was rotated/cleaned up).

Happy to provide more transcripts or a minimal repro on request.